Merge branch 'master' into wii-network

Conflicts:
	Source/Core/Core/CMakeLists.txt

17 files changed, 162 insertions, 120 deletions

diff --git a/Source/Core/VideoCommon/Src/BPMemory.h b/Source/Core/VideoCommon/Src/BPMemory.h
index 1a29439b36..f0d0a502ed 100644
--- a/Source/Core/VideoCommon/Src/BPMemory.h
+++ b/Source/Core/VideoCommon/Src/BPMemory.h
@@ -621,6 +621,17 @@ union X10Y10
 
 // Framebuffer/pixel stuff (incl fog)
 
+#define GX_BL_ZERO         0
+#define GX_BL_ONE          1
+#define GX_BL_SRCCLR       2 // for dst factor
+#define GX_BL_INVSRCCLR    3 // for dst factor
+#define GX_BL_SRCALPHA     4
+#define GX_BL_INVSRCALPHA  5
+#define GX_BL_DSTALPHA     6
+#define GX_BL_INVDSTALPHA  7
+#define GX_BL_DSTCLR       GX_BL_SRCCLR // for src factor
+#define GX_BL_INVDSTCLR    GX_BL_INVSRCCLR // for src factor
+
 union BlendMode
 {
 	struct 
diff --git a/Source/Core/VideoCommon/Src/CommandProcessor.cpp b/Source/Core/VideoCommon/Src/CommandProcessor.cpp
index dacec67823..733c99df69 100644
--- a/Source/Core/VideoCommon/Src/CommandProcessor.cpp
+++ b/Source/Core/VideoCommon/Src/CommandProcessor.cpp
@@ -80,9 +80,9 @@ void UpdateInterrupts_Wrapper(u64 userdata, int cyclesLate)
 
 void DoState(PointerWrap &p)
 {
-	p.Do(m_CPStatusReg);
-	p.Do(m_CPCtrlReg);
-	p.Do(m_CPClearReg);
+	p.DoPOD(m_CPStatusReg);
+	p.DoPOD(m_CPCtrlReg);
+	p.DoPOD(m_CPClearReg);
 	p.Do(m_bboxleft);
 	p.Do(m_bboxtop);
 	p.Do(m_bboxright);
@@ -185,7 +185,7 @@ void Read16(u16& _rReturnValue, const u32 _Address)
 				_rReturnValue = ReadLow (fifo.CPEnd - fifo.SafeCPReadPointer + fifo.CPWritePointer - fifo.CPBase + 32);
 		else
 			_rReturnValue = ReadLow (fifo.CPReadWriteDistance);
-		DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_RW_DISTANCE_LO : %04x", _rReturnValue);
+		DEBUG_LOG(COMMANDPROCESSOR, "Read FIFO_RW_DISTANCE_LO : %04x", _rReturnValue);
 		return;
 	case FIFO_RW_DISTANCE_HI:
 		if (IsOnThread())
@@ -195,29 +195,29 @@ void Read16(u16& _rReturnValue, const u32 _Address)
 				_rReturnValue = ReadHigh (fifo.CPEnd - fifo.SafeCPReadPointer + fifo.CPWritePointer - fifo.CPBase + 32);
 		else
 			_rReturnValue = ReadHigh(fifo.CPReadWriteDistance);
-		DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_RW_DISTANCE_HI : %04x", _rReturnValue);
+		DEBUG_LOG(COMMANDPROCESSOR, "Read FIFO_RW_DISTANCE_HI : %04x", _rReturnValue);
 		return;
 	case FIFO_WRITE_POINTER_LO:
 		_rReturnValue = ReadLow (fifo.CPWritePointer);
-		DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_WRITE_POINTER_LO : %04x", _rReturnValue);
+		DEBUG_LOG(COMMANDPROCESSOR, "Read FIFO_WRITE_POINTER_LO : %04x", _rReturnValue);
 		return;
 	case FIFO_WRITE_POINTER_HI:
 		_rReturnValue = ReadHigh(fifo.CPWritePointer);
-		DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_WRITE_POINTER_HI : %04x", _rReturnValue);
+		DEBUG_LOG(COMMANDPROCESSOR, "Read FIFO_WRITE_POINTER_HI : %04x", _rReturnValue);
 		return;
 	case FIFO_READ_POINTER_LO:
 		if (IsOnThread())
 			_rReturnValue = ReadLow (fifo.SafeCPReadPointer);
 		else
 			_rReturnValue = ReadLow (fifo.CPReadPointer);
-		DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_READ_POINTER_LO : %04x", _rReturnValue);
+		DEBUG_LOG(COMMANDPROCESSOR, "Read FIFO_READ_POINTER_LO : %04x", _rReturnValue);
 		return;
 	case FIFO_READ_POINTER_HI:
 		if (IsOnThread())
 			_rReturnValue = ReadHigh (fifo.SafeCPReadPointer);
 		else
 			_rReturnValue = ReadHigh (fifo.CPReadPointer);
-		DEBUG_LOG(COMMANDPROCESSOR, "read FIFO_READ_POINTER_HI : %04x", _rReturnValue);
+		DEBUG_LOG(COMMANDPROCESSOR, "Read FIFO_READ_POINTER_HI : %04x", _rReturnValue);
 		return;
 
 	case FIFO_BP_LO: _rReturnValue = ReadLow (fifo.CPBreakpoint); return;
@@ -326,83 +326,83 @@ void Write16(const u16 _Value, const u32 _Address)
 		{
 			UCPClearReg tmpCtrl(_Value);
 			m_CPClearReg.Hex = tmpCtrl.Hex;
-			DEBUG_LOG(COMMANDPROCESSOR,"\t write to CLEAR_REGISTER : %04x", _Value);
+			DEBUG_LOG(COMMANDPROCESSOR,"\t Write to CLEAR_REGISTER : %04x", _Value);
 			SetCpClearRegister();
 		}
 		break;
 
 	case PERF_SELECT:
 		// Seems to select which set of perf registers should be exposed.
-		DEBUG_LOG(COMMANDPROCESSOR, "write to PERF_SELECT: %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR, "Write to PERF_SELECT: %04x", _Value);
 		break;
 
 	// Fifo Registers
 	case FIFO_TOKEN_REGISTER:
 		m_tokenReg = _Value;
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_TOKEN_REGISTER : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_TOKEN_REGISTER : %04x", _Value);
 		break;
 	case FIFO_BASE_LO:
 		WriteLow ((u32 &)fifo.CPBase, _Value & 0xFFE0);
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_BASE_LO : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_BASE_LO : %04x", _Value);
 		break;
 	case FIFO_BASE_HI:
 		WriteHigh((u32 &)fifo.CPBase, _Value);
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_BASE_HI : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_BASE_HI : %04x", _Value);
 		break;
 
 	case FIFO_END_LO:
 		WriteLow ((u32 &)fifo.CPEnd,  _Value & 0xFFE0);
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_END_LO : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_END_LO : %04x", _Value);
 		break;
 	case FIFO_END_HI:
 		WriteHigh((u32 &)fifo.CPEnd,  _Value);
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_END_HI : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_END_HI : %04x", _Value);
 		break;
 
 	case FIFO_WRITE_POINTER_LO:
 		WriteLow ((u32 &)fifo.CPWritePointer, _Value & 0xFFE0);
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_WRITE_POINTER_LO : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_WRITE_POINTER_LO : %04x", _Value);
 		break;
 	case FIFO_WRITE_POINTER_HI:
 		WriteHigh((u32 &)fifo.CPWritePointer, _Value);
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_WRITE_POINTER_HI : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_WRITE_POINTER_HI : %04x", _Value);
 		break;
 
 	case FIFO_READ_POINTER_LO:
 		WriteLow ((u32 &)fifo.CPReadPointer, _Value & 0xFFE0);
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_READ_POINTER_LO : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_READ_POINTER_LO : %04x", _Value);
 		break;
 	case FIFO_READ_POINTER_HI:
 		WriteHigh((u32 &)fifo.CPReadPointer, _Value);
 		fifo.SafeCPReadPointer = fifo.CPReadPointer;
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_READ_POINTER_HI : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_READ_POINTER_HI : %04x", _Value);
 		break;
 
 	case FIFO_HI_WATERMARK_LO:
 		WriteLow ((u32 &)fifo.CPHiWatermark, _Value);
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_HI_WATERMARK_LO : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_HI_WATERMARK_LO : %04x", _Value);
 		break;
 	case FIFO_HI_WATERMARK_HI:
 		WriteHigh((u32 &)fifo.CPHiWatermark, _Value);
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_HI_WATERMARK_HI : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_HI_WATERMARK_HI : %04x", _Value);
 		break;
 
 	case FIFO_LO_WATERMARK_LO:
 		WriteLow ((u32 &)fifo.CPLoWatermark, _Value);
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_LO_WATERMARK_LO : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_LO_WATERMARK_LO : %04x", _Value);
 		break;
 	case FIFO_LO_WATERMARK_HI:
 		WriteHigh((u32 &)fifo.CPLoWatermark, _Value);
-		DEBUG_LOG(COMMANDPROCESSOR,"\t write to FIFO_LO_WATERMARK_HI : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"\t Write to FIFO_LO_WATERMARK_HI : %04x", _Value);
 		break;
 
 	case FIFO_BP_LO:
 		WriteLow ((u32 &)fifo.CPBreakpoint, _Value & 0xFFE0);
-		DEBUG_LOG(COMMANDPROCESSOR,"write to FIFO_BP_LO : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"Write to FIFO_BP_LO : %04x", _Value);
 		break;
 	case FIFO_BP_HI:
 		WriteHigh((u32 &)fifo.CPBreakpoint, _Value);
-		DEBUG_LOG(COMMANDPROCESSOR,"write to FIFO_BP_HI : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"Write to FIFO_BP_HI : %04x", _Value);
 		break;
 
 	case FIFO_RW_DISTANCE_HI:
@@ -417,11 +417,11 @@ void Write16(const u16 _Value, const u32 _Address)
 			ResetVideoBuffer();
 		}
 		IncrementCheckContextId();
-		DEBUG_LOG(COMMANDPROCESSOR,"try to write to FIFO_RW_DISTANCE_HI : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"Try to write to FIFO_RW_DISTANCE_HI : %04x", _Value);
 		break;
 	case FIFO_RW_DISTANCE_LO:
 		WriteLow((u32 &)fifo.CPReadWriteDistance, _Value & 0xFFE0);
-		DEBUG_LOG(COMMANDPROCESSOR,"try to write to FIFO_RW_DISTANCE_LO : %04x", _Value);
+		DEBUG_LOG(COMMANDPROCESSOR,"Try to write to FIFO_RW_DISTANCE_LO : %04x", _Value);
 		break;
 
 	default:
@@ -453,7 +453,7 @@ void STACKALIGN GatherPipeBursted()
 			RunGpu();
 		else
 		{
-			// In multibuffer mode is not allowed write in the same fifo attached to the GPU.
+			// In multibuffer mode is not allowed write in the same FIFO attached to the GPU.
 			// Fix Pokemon XD in DC mode.
 			if((ProcessorInterface::Fifo_CPUEnd == fifo.CPEnd) && (ProcessorInterface::Fifo_CPUBase == fifo.CPBase)
 				 && fifo.CPReadWriteDistance > 0)
@@ -469,7 +469,7 @@ void STACKALIGN GatherPipeBursted()
 	if (IsOnThread())
 		SetCpStatus(true);
 
-	// update the fifo-pointer
+	// update the fifo pointer
 	if (fifo.CPWritePointer >= fifo.CPEnd)
 		fifo.CPWritePointer = fifo.CPBase;
 	else
@@ -481,7 +481,7 @@ void STACKALIGN GatherPipeBursted()
 		RunGpu();
 
 	_assert_msg_(COMMANDPROCESSOR, fifo.CPReadWriteDistance	<= fifo.CPEnd - fifo.CPBase,
-	"FIFO is overflown by GatherPipe !\nCPU thread is too fast!");
+	"FIFO is overflowed by GatherPipe !\nCPU thread is too fast!");
 
 	// check if we are in sync
 	_assert_msg_(COMMANDPROCESSOR, fifo.CPWritePointer	== ProcessorInterface::Fifo_CPUWritePointer, "FIFOs linked but out of sync");
diff --git a/Source/Core/VideoCommon/Src/EmuWindow.cpp b/Source/Core/VideoCommon/Src/EmuWindow.cpp
index bf9da4190b..8952432789 100644
--- a/Source/Core/VideoCommon/Src/EmuWindow.cpp
+++ b/Source/Core/VideoCommon/Src/EmuWindow.cpp
@@ -247,7 +247,7 @@ void OSDMenu(WPARAM wParam)
 		OSDChoice = 1;
 		// Toggle native resolution
 		g_Config.iEFBScale = g_Config.iEFBScale + 1;
-		if (g_Config.iEFBScale > 7) g_Config.iEFBScale = 0;
+		if (g_Config.iEFBScale > SCALE_4X) g_Config.iEFBScale = SCALE_AUTO;
 		break;
 	case '4':
 		OSDChoice = 2;
diff --git a/Source/Core/VideoCommon/Src/HiresTextures.cpp b/Source/Core/VideoCommon/Src/HiresTextures.cpp
index 5a3fb051c3..0207bc021c 100644
--- a/Source/Core/VideoCommon/Src/HiresTextures.cpp
+++ b/Source/Core/VideoCommon/Src/HiresTextures.cpp
@@ -148,7 +148,7 @@ PC_TexFormat GetHiresTex(const char *fileName, unsigned int *pWidth, unsigned in
 		break;
 	}
 
-	INFO_LOG(VIDEO, "loading custom texture from %s", textureMap[key].c_str());
+	INFO_LOG(VIDEO, "Loading custom texture from %s", textureMap[key].c_str());
 cleanup:
 	SOIL_free_image_data(temp);
 	return returnTex;
diff --git a/Source/Core/VideoCommon/Src/PixelEngine.cpp b/Source/Core/VideoCommon/Src/PixelEngine.cpp
index b014148d8d..348b704f5b 100644
--- a/Source/Core/VideoCommon/Src/PixelEngine.cpp
+++ b/Source/Core/VideoCommon/Src/PixelEngine.cpp
@@ -137,7 +137,7 @@ void DoState(PointerWrap &p)
 	p.Do(m_DstAlphaConf);
 	p.Do(m_AlphaModeConf);
 	p.Do(m_AlphaRead);
-	p.Do(m_Control);
+	p.DoPOD(m_Control);
 
 	p.Do(g_bSignalTokenInterrupt);
 	p.Do(g_bSignalFinishInterrupt);
diff --git a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
index d863a6e0a9..93f328fb61 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
+++ b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
@@ -390,7 +390,7 @@ static const char *tevCInputTable[] = // CC
 	"float3(0.5f, 0.5f, 0.5f)",              // HALF
 	"(konsttemp.rgb)", //"konsttemp.rgb",    // KONST
 	"float3(0.0f, 0.0f, 0.0f)",              // ZERO
-	///aded extra values to map clamped values
+	///added extra values to map clamped values
 	"(cprev.rgb)",        // CPREV,
 	"(cprev.aaa)",        // APREV,
 	"(cc0.rgb)",          // C0,
@@ -598,9 +598,9 @@ const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType
 
 	if (ApiType == API_OPENGL)
 	{
-		WRITE(p, "out float4 ocol0;\n");
+		WRITE(p, "COLOROUT(ocol0)\n");
 		if (dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
-			WRITE(p, "out float4 ocol1;\n");
+			WRITE(p, "COLOROUT(ocol1)\n");
 		
 		if (per_pixel_depth)
 			WRITE(p, "#define depth gl_FragDepth\n");
@@ -668,17 +668,20 @@ const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType
 				per_pixel_depth ? "\n  out float depth : SV_Depth," : "");
 		}
 
-		WRITE(p, "  in float4 colors_0 : COLOR0,\n");
-		WRITE(p, "  in float4 colors_1 : COLOR1");
+		// "centroid" attribute is only supported by D3D11
+		const char* optCentroid = (ApiType == API_D3D11 ? "centroid" : "");
+
+		WRITE(p, "  in %s float4 colors_0 : COLOR0,\n", optCentroid);
+		WRITE(p, "  in %s float4 colors_1 : COLOR1", optCentroid);
 
 		// compute window position if needed because binding semantic WPOS is not widely supported
 		if (numTexgen < 7)
 		{
 			for (int i = 0; i < numTexgen; ++i)
-				WRITE(p, ",\n  in float3 uv%d : TEXCOORD%d", i, i);
-			WRITE(p, ",\n  in float4 clipPos : TEXCOORD%d", numTexgen);
+				WRITE(p, ",\n  in %s float3 uv%d : TEXCOORD%d", optCentroid, i, i);
+			WRITE(p, ",\n  in %s float4 clipPos : TEXCOORD%d", optCentroid, numTexgen);
 			if(g_ActiveConfig.bEnablePixelLighting && g_ActiveConfig.backend_info.bSupportsPixelLighting)
-				WRITE(p, ",\n  in float4 Normal : TEXCOORD%d", numTexgen + 1);
+				WRITE(p, ",\n  in %s float4 Normal : TEXCOORD%d", optCentroid, numTexgen + 1);
 			WRITE(p, "        ) {\n");
 		}
 		else
@@ -844,12 +847,21 @@ const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType
 		WRITE(p, "\tocol0 = prev;\n");
 	}
 
-	// On D3D11, use dual-source color blending to perform dst alpha in a
+	// Use dual-source color blending to perform dst alpha in a
 	// single pass
 	if (dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
 	{
-		// Colors will be blended against the alpha from ocol1...
-		WRITE(p, "\tocol1 = prev;\n");
+		if(ApiType & API_D3D9)
+		{
+			// alpha component must be 0 or the shader will not compile (Direct3D 9Ex restriction)
+			// Colors will be blended against the color from ocol1 in D3D 9...			
+			WRITE(p, "\tocol1 = float4(prev.a, prev.a, prev.a, 0.0f);\n");			
+		}
+		else
+		{
+			// Colors will be blended against the alpha from ocol1...
+			WRITE(p, "\tocol1 = prev;\n");
+		}
 		// ...and the alpha from ocol0 will be written to the framebuffer.
 		WRITE(p, "\tocol0.a = " I_ALPHA"[0].a;\n");	
 	}
@@ -1324,7 +1336,7 @@ static void WriteFog(char *&p)
 
 	// x_adjust = sqrt((x-center)^2 + k^2)/k
 	// ze *= x_adjust
-	//this is complitly teorical as the real hard seems to use a table intead of calculate the values.
+	//this is completely theoretical as the real hardware seems to use a table instead of calculating the values.
 	if (bpmem.fogRange.Base.Enabled)
 	{
 		WRITE (p, "\tfloat x_adjust = (2.0f * (clipPos.x / " I_FOG"[2].y)) - 1.0f - " I_FOG"[2].x;\n");
diff --git a/Source/Core/VideoCommon/Src/PixelShaderManager.cpp b/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
index ad84f6876e..265bebf991 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
+++ b/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
@@ -245,14 +245,14 @@ void PixelShaderManager::SetConstants()
 		{
 			//bpmem.fogRange.Base.Center : center of the viewport in x axis. observation: bpmem.fogRange.Base.Center = realcenter + 342;
 			int center = ((u32)bpmem.fogRange.Base.Center) - 342;
-			// normalice center to make calculations easy
+			// normalize center to make calculations easy
 			float ScreenSpaceCenter = center / (2.0f * xfregs.viewport.wd);
 			ScreenSpaceCenter = (ScreenSpaceCenter * 2.0f) - 1.0f;
-			//bpmem.fogRange.K seems to be  a table of precalculated coeficients for the adjust factor
-			//observations: bpmem.fogRange.K[0].LO apears to be the lowest value and bpmem.fogRange.K[4].HI the largest
-			// they always seems to be larger than 256 so my teory is :
-			// they are the coeficients from the center to th e border of the screen
-			// so to simplify i use the hi coeficient as K in the shader taking 256 as the scale
+			//bpmem.fogRange.K seems to be  a table of precalculated coefficients for the adjust factor
+			//observations: bpmem.fogRange.K[0].LO appears to be the lowest value and bpmem.fogRange.K[4].HI the largest
+			// they always seems to be larger than 256 so my theory is :
+			// they are the coefficients from the center to the border of the screen
+			// so to simplify I use the hi coefficient as K in the shader taking 256 as the scale
 			SetPSConstant4f(C_FOG + 2, ScreenSpaceCenter, (float)Renderer::EFBToScaledX((int)(2.0f * xfregs.viewport.wd)), bpmem.fogRange.K[4].HI / 256.0f,0.0f);
 		}
 		else
diff --git a/Source/Core/VideoCommon/Src/RenderBase.cpp b/Source/Core/VideoCommon/Src/RenderBase.cpp
index 90a09aec43..652a613b2a 100644
--- a/Source/Core/VideoCommon/Src/RenderBase.cpp
+++ b/Source/Core/VideoCommon/Src/RenderBase.cpp
@@ -144,7 +144,7 @@ int Renderer::EFBToScaledX(int x)
 {
 	switch (g_ActiveConfig.iEFBScale)
 	{
-		case 0: // fractional
+	case SCALE_AUTO: // fractional
 			return (int)ssaa_multiplier * FramebufferManagerBase::ScaleToVirtualXfbWidth(x, s_backbuffer_width);
 
 		default:
@@ -156,7 +156,7 @@ int Renderer::EFBToScaledY(int y)
 {
 	switch (g_ActiveConfig.iEFBScale)
 	{
-		case 0: // fractional
+		case SCALE_AUTO: // fractional
 			return (int)ssaa_multiplier * FramebufferManagerBase::ScaleToVirtualXfbHeight(y, s_backbuffer_height);
 
 		default:
@@ -166,7 +166,7 @@ int Renderer::EFBToScaledY(int y)
 
 void Renderer::CalculateTargetScale(int x, int y, int &scaledX, int &scaledY)
 {
-	if (g_ActiveConfig.iEFBScale == 0 || g_ActiveConfig.iEFBScale == 1)
+	if (g_ActiveConfig.iEFBScale == SCALE_AUTO || g_ActiveConfig.iEFBScale == SCALE_AUTO_INTEGRAL)
 	{
 		scaledX = x;
 		scaledY = y;
@@ -283,28 +283,28 @@ void Renderer::DrawDebugText()
 	const char* res_text = "";
 	switch (g_ActiveConfig.iEFBScale)
 	{
-	case 0:
+	case SCALE_AUTO:
 		res_text = "Auto (fractional)";
 		break;
-	case 1:
+	case SCALE_AUTO_INTEGRAL:
 		res_text = "Auto (integral)";
 		break;
-	case 2:
+	case SCALE_1X:
 		res_text = "Native";
 		break;
-	case 3:
+	case SCALE_1_5X:
 		res_text = "1.5x";
 		break;
-	case 4:
+	case SCALE_2X:
 		res_text = "2x";
 		break;
-	case 5:
+	case SCALE_2_5X:
 		res_text = "2.5x";
 		break;
-	case 6:
+	case SCALE_3X:
 		res_text = "3x";
 		break;
-	case 7:
+	case SCALE_4X:
 		res_text = "4x";
 		break;
 	}
diff --git a/Source/Core/VideoCommon/Src/TextureConversionShader.cpp b/Source/Core/VideoCommon/Src/TextureConversionShader.cpp
index 5a81c9e7d5..314be94355 100644
--- a/Source/Core/VideoCommon/Src/TextureConversionShader.cpp
+++ b/Source/Core/VideoCommon/Src/TextureConversionShader.cpp
@@ -105,8 +105,8 @@ void WriteSwizzler(char*& p, u32 format, API_TYPE ApiType)
 
 	if (ApiType == API_OPENGL)
 	{
-		WRITE(p, "  out float4 ocol0;\n");
-		WRITE(p, "  in float2 uv0;\n");
+		WRITE(p, "  COLOROUT(ocol0)\n");
+		WRITE(p, "  VARYIN float2 uv0;\n");
 		WRITE(p, "void main()\n");
 	}
 	else
@@ -187,7 +187,7 @@ void Write32BitSwizzler(char*& p, u32 format, API_TYPE ApiType)
 	if (ApiType == API_OPENGL)
 	{
 		WRITE(p, "  out float4 ocol0;\n");
-		WRITE(p, "  in float2 uv0;\n");
+		WRITE(p, "  VARYIN float2 uv0;\n");
 		WRITE(p, "void main()\n");
 	}
 	else
diff --git a/Source/Core/VideoCommon/Src/VertexLoader.cpp b/Source/Core/VideoCommon/Src/VertexLoader.cpp
index 5df495a44c..d1e8b14e8a 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader.cpp
@@ -227,7 +227,7 @@ void VertexLoader::CompileVertexTranslator()
 
 #ifdef USE_JIT
 	if (m_compiledCode)
-		PanicAlert("trying to recompile a vtx translator");
+		PanicAlert("Trying to recompile a vtx translator");
 
 	m_compiledCode = GetCodePtr();
 	ABI_EmitPrologue(4);
diff --git a/Source/Core/VideoCommon/Src/VertexLoaderManager.cpp b/Source/Core/VideoCommon/Src/VertexLoaderManager.cpp
index 153e69a8c3..01db9f647c 100644
--- a/Source/Core/VideoCommon/Src/VertexLoaderManager.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoaderManager.cpp
@@ -16,17 +16,7 @@
 // http://code.google.com/p/dolphin-emu/
 
 #include <algorithm>
-#ifdef _MSC_VER
-#include <hash_map>
-using stdext::hash_map;
-#elif defined __APPLE__
-#include <ext/hash_map>
-using __gnu_cxx::hash_map;
-#else
 #include <unordered_map>
-using std::unordered_map;
-#endif
-#include <map>
 #include <vector>
 
 #include "VideoCommon.h"
@@ -41,32 +31,21 @@ static int s_attr_dirty;  // bitfield
 
 static VertexLoader *g_VertexLoaders[8];
 
-#ifdef _MSC_VER
-namespace stdext {
-	inline size_t hash_value(const VertexLoaderUID& uid) {
-		return uid.GetHash();
-	}
-}
-#else
-#ifdef __APPLE__
-namespace __gnu_cxx
-#else
 namespace std
-#endif
 {
-	template<> struct hash<VertexLoaderUID> {
-		size_t operator()(const VertexLoaderUID& uid) const {
-			return uid.GetHash();
-		}
-	};
+
+template <>
+struct hash<VertexLoaderUID>
+{
+	size_t operator()(const VertexLoaderUID& uid) const
+	{
+		return uid.GetHash();
+	}
+};
+
 }
-#endif
 
-#if defined _MSC_VER || defined __APPLE__
-typedef hash_map<VertexLoaderUID, VertexLoader*> VertexLoaderMap;
-#else
-typedef unordered_map<VertexLoaderUID, VertexLoader*> VertexLoaderMap;
-#endif
+typedef std::unordered_map<VertexLoaderUID, VertexLoader*> VertexLoaderMap;
 
 namespace VertexLoaderManager
 {
diff --git a/Source/Core/VideoCommon/Src/VertexManagerBase.cpp b/Source/Core/VideoCommon/Src/VertexManagerBase.cpp
index 87856b5aeb..3e2e88c55f 100644
--- a/Source/Core/VideoCommon/Src/VertexManagerBase.cpp
+++ b/Source/Core/VideoCommon/Src/VertexManagerBase.cpp
@@ -57,13 +57,13 @@ void VertexManager::PrepareForAdditionalData(int primitive, u32 count, u32 strid
 		Flush();
 		
 		if(count > IndexGenerator::GetRemainingIndices())
-			ERROR_LOG(VIDEO, "Too less index values. Use 32bit or reset them on flush.");
+			ERROR_LOG(VIDEO, "Too little remaining index values. Use 32-bit or reset them on flush.");
 		if (count > GetRemainingIndices(primitive))
 			ERROR_LOG(VIDEO, "VertexManager: Buffer not large enough for all indices! "
-				"Increase MAXIBUFFERSIZE or we need primitive breaking afterall.");
+				"Increase MAXIBUFFERSIZE or we need primitive breaking after all.");
 		if (needed_vertex_bytes > GetRemainingSize())
 			ERROR_LOG(VIDEO, "VertexManager: Buffer not large enough for all vertices! "
-				"Increase MAXVBUFFERSIZE or we need primitive breaking afterall.");
+				"Increase MAXVBUFFERSIZE or we need primitive breaking after all.");
 	}
 }
 
@@ -190,7 +190,7 @@ void VertexManager::Flush()
 				PixelShaderManager::SetTexDims(i, tentry->nativeW, tentry->nativeH, 0, 0);
 			}
 			else
-				ERROR_LOG(VIDEO, "error loading texture");
+				ERROR_LOG(VIDEO, "Error loading texture");
 		}
 	}
 
diff --git a/Source/Core/VideoCommon/Src/VertexShaderGen.cpp b/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
index 2c0f5676ed..e4a04102ff 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
+++ b/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
@@ -133,30 +133,34 @@ static char text[16384];
 
 char* GenerateVSOutputStruct(char* p, u32 components, API_TYPE ApiType)
 {
+
+	// "centroid" attribute is only supported by D3D11
+	const char* optCentroid = (ApiType == API_D3D11 ? "centroid" : "");
+
 	// GLSL makes this ugly
 	// TODO: Make pretty
 	WRITE(p, "struct VS_OUTPUT {\n");
-	WRITE(p, "  float4 pos %s POSITION;\n", ApiType == API_OPENGL ? ";//" : ":");
-	WRITE(p, "  float4 colors_0 %s COLOR0;\n", ApiType == API_OPENGL ? ";//" : ":");
-	WRITE(p, "  float4 colors_1 %s COLOR1;\n", ApiType == API_OPENGL ? ";//" : ":");
+	WRITE(p, "  %s float4 pos %s POSITION;\n", optCentroid, ApiType == API_OPENGL ? ";//" : ":");
+	WRITE(p, "  %s float4 colors_0 %s COLOR0;\n", optCentroid, ApiType == API_OPENGL ? ";//" : ":");
+	WRITE(p, "  %s float4 colors_1 %s COLOR1;\n", optCentroid, ApiType == API_OPENGL ? ";//" : ":");
 
 	if (xfregs.numTexGen.numTexGens < 7) {
 		for (unsigned int i = 0; i < xfregs.numTexGen.numTexGens; ++i)
-			WRITE(p, "  float3 tex%d %s TEXCOORD%d;\n", i, ApiType == API_OPENGL ? ";//" : ":", i);
-		WRITE(p, "  float4 clipPos %s TEXCOORD%d;\n", ApiType == API_OPENGL ? ";//" : ":", xfregs.numTexGen.numTexGens);
+			WRITE(p, "  %s float3 tex%d %s TEXCOORD%d;\n", optCentroid, i, ApiType == API_OPENGL ? ";//" : ":", i);
+		WRITE(p, "  %s float4 clipPos %s TEXCOORD%d;\n", optCentroid, ApiType == API_OPENGL ? ";//" : ":", xfregs.numTexGen.numTexGens);
 		if(g_ActiveConfig.bEnablePixelLighting && g_ActiveConfig.backend_info.bSupportsPixelLighting)
-			WRITE(p, "  float4 Normal %s TEXCOORD%d;\n", ApiType == API_OPENGL ? ";//" : ":", xfregs.numTexGen.numTexGens + 1);
+			WRITE(p, "  %s float4 Normal %s TEXCOORD%d;\n", optCentroid, ApiType == API_OPENGL ? ";//" : ":", xfregs.numTexGen.numTexGens + 1);
 	} else {
 		// clip position is in w of first 4 texcoords
 		if(g_ActiveConfig.bEnablePixelLighting && g_ActiveConfig.backend_info.bSupportsPixelLighting)
 		{
 			for (int i = 0; i < 8; ++i)
-				WRITE(p, "  float4 tex%d %s TEXCOORD%d;\n", i, ApiType == API_OPENGL? ";//" : ":", i);
+				WRITE(p, "  %s float4 tex%d %s TEXCOORD%d;\n", optCentroid, i, ApiType == API_OPENGL? ";//" : ":", i);
 		}
 		else
 		{
 			for (unsigned int i = 0; i < xfregs.numTexGen.numTexGens; ++i)
-				WRITE(p, "  float%d tex%d %s TEXCOORD%d;\n", i < 4 ? 4 : 3 , i, ApiType == API_OPENGL ? ";//" : ":", i);
+				WRITE(p, "  %s float%d tex%d %s TEXCOORD%d;\n", optCentroid, i < 4 ? 4 : 3 , i, ApiType == API_OPENGL ? ";//" : ":", i);
 		}
 	}
 	WRITE(p, "};\n");
@@ -413,7 +417,7 @@ const char *GenerateVertexShaderCode(u32 components, API_TYPE ApiType)
 
 				if (components & (VB_HAS_NRM1|VB_HAS_NRM2)) {
 					// transform the light dir into tangent space
-					WRITE(p, "ldir = normalize(" I_LIGHTS"[%d + 3].xyz - pos.xyz);\n", texinfo.embosslightshift);
+					WRITE(p, "ldir = normalize(" I_LIGHTS"[5*%d + 3].xyz - pos.xyz);\n", texinfo.embosslightshift);
 					WRITE(p, "o.tex%d.xyz = o.tex%d.xyz + float3(dot(ldir, _norm1), dot(ldir, _norm2), 0.0f);\n", i, texinfo.embosssourceshift);
 				}
 				else
diff --git a/Source/Core/VideoCommon/Src/VertexShaderManager.cpp b/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
index a5e84ee69b..b8c9b7d5a3 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
+++ b/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
@@ -122,7 +122,7 @@ void UpdateProjectionHack(int iPhackvalue[], std::string sPhackvalue[])
 
 	if (iPhackvalue[0] == 1)
 	{
-		NOTICE_LOG(VIDEO, "\t\t--- Ortographic Projection Hack ON ---");
+		NOTICE_LOG(VIDEO, "\t\t--- Orthographic Projection Hack ON ---");
 
 		fhacksign1 *= (iPhackvalue[1] == 1) ? -1.0f : fhacksign1;
 		sTemp[0] = (iPhackvalue[1] == 1) ? " * (-1)" : "";
@@ -191,7 +191,7 @@ void VertexShaderManager::Dirty()
 }
 
 // Syncs the shader constant buffers with xfmem
-// TODO: A cleaner way to control the matricies without making a mess in the parameters field
+// TODO: A cleaner way to control the matrices without making a mess in the parameters field
 void VertexShaderManager::SetConstants()
 {
 	if (g_ActiveConfig.backend_info.APIType == API_OPENGL && !g_ActiveConfig.backend_info.bSupportsGLSLUBO)
@@ -455,7 +455,7 @@ void VertexShaderManager::SetConstants()
 			break;
 
 		default:
-			ERROR_LOG(VIDEO, "unknown projection type: %d", xfregs.projection.type);
+			ERROR_LOG(VIDEO, "Unknown projection type: %d", xfregs.projection.type);
 		}
 
 		PRIM_LOG("Projection: %f %f %f %f %f %f\n", rawProjection[0], rawProjection[1], rawProjection[2], rawProjection[3], rawProjection[4], rawProjection[5]);
diff --git a/Source/Core/VideoCommon/Src/VideoConfig.cpp b/Source/Core/VideoCommon/Src/VideoConfig.cpp
index 60677ceeeb..ad040b9e24 100644
--- a/Source/Core/VideoCommon/Src/VideoConfig.cpp
+++ b/Source/Core/VideoCommon/Src/VideoConfig.cpp
@@ -49,7 +49,6 @@ VideoConfig::VideoConfig()
 
 void VideoConfig::Load(const char *ini_file)
 {
-	std::string temp;
 	IniFile iniFile;
 	iniFile.Load(ini_file);
 
@@ -80,7 +79,7 @@ void VideoConfig::Load(const char *ini_file)
 	iniFile.Get("Settings", "HackedBufferUpload", &bHackedBufferUpload, 0);
 
 	iniFile.Get("Settings", "MSAA", &iMultisampleMode, 0);
-	iniFile.Get("Settings", "EFBScale", &iEFBScale, 2); // native
+	iniFile.Get("Settings", "EFBScale", &iEFBScale, (int) SCALE_1X); // native
 
 	iniFile.Get("Settings", "DstAlphaPass", &bDstAlphaPass, false);
 
@@ -138,7 +137,32 @@ void VideoConfig::GameIniLoad(const char *ini_file)
 	iniFile.GetIfExists("Video_Settings", "EnablePixelLighting", &bEnablePixelLighting);
 	iniFile.GetIfExists("Video_Settings", "HackedBufferUpload", &bHackedBufferUpload);
 	iniFile.GetIfExists("Video_Settings", "MSAA", &iMultisampleMode);
-	iniFile.GetIfExists("Video_Settings", "EFBScale", &iEFBScale); // integral
+	int tmp = -9000;
+	iniFile.GetIfExists("Video_Settings", "EFBScale", &tmp); // integral
+	if (tmp != -9000)
+	{
+		if (tmp != SCALE_FORCE_INTEGRAL)
+			iEFBScale = tmp;
+		// Round down to multiple of native IR
+		else
+		{
+			switch (iEFBScale)
+			{
+			case SCALE_AUTO:
+				iEFBScale = SCALE_AUTO_INTEGRAL;
+				break;
+			case SCALE_1_5X:
+				iEFBScale = SCALE_1X;
+				break;
+			case SCALE_2_5X:
+				iEFBScale = SCALE_2X;
+				break;
+			default:
+				break;
+			}
+		}
+	}
+
 	iniFile.GetIfExists("Video_Settings", "DstAlphaPass", &bDstAlphaPass);
 	iniFile.GetIfExists("Video_Settings", "DisableFog", &bDisableFog);
 	iniFile.GetIfExists("Video_Settings", "EnableOpenCL", &bEnableOpenCL);
diff --git a/Source/Core/VideoCommon/Src/VideoConfig.h b/Source/Core/VideoCommon/Src/VideoConfig.h
index 442e95f9bf..1670e47222 100644
--- a/Source/Core/VideoCommon/Src/VideoConfig.h
+++ b/Source/Core/VideoCommon/Src/VideoConfig.h
@@ -44,6 +44,18 @@ enum AspectMode {
 	ASPECT_STRETCH = 3,
 };
 
+enum EFBScale {
+	SCALE_FORCE_INTEGRAL = -1,
+	SCALE_AUTO,
+	SCALE_AUTO_INTEGRAL,
+	SCALE_1X,
+	SCALE_1_5X,
+	SCALE_2X,
+	SCALE_2_5X,
+	SCALE_3X,
+	SCALE_4X,
+};
+
 class IniFile;
 
 // NEVER inherit from this class.
@@ -153,7 +165,7 @@ struct VideoConfig
 		bool bSupportsDualSourceBlend; // only supported by D3D11 and OpenGL
 		bool bSupportsFormatReinterpretation;
 		bool bSupportsPixelLighting;
-
+		bool bSupportsSeparateAlphaFunction;
 		bool bSupportsGLSLUBO; // needed by pixelShaderGen, so must stay in videoCommon
 	} backend_info;
 
diff --git a/Source/Core/VideoCommon/Src/XFStructs.cpp b/Source/Core/VideoCommon/Src/XFStructs.cpp
index df9624b6f0..efbe09e4b3 100644
--- a/Source/Core/VideoCommon/Src/XFStructs.cpp
+++ b/Source/Core/VideoCommon/Src/XFStructs.cpp
@@ -212,7 +212,7 @@ void LoadXFReg(u32 transferSize, u32 baseAddress, u32 *pData)
 	// do not allow writes past registers
 	if (baseAddress + transferSize > 0x1058)
 	{
-		INFO_LOG(VIDEO, "xf load exceeds address space: %x %d bytes\n", baseAddress, transferSize);
+		INFO_LOG(VIDEO, "XF load exceeds address space: %x %d bytes\n", baseAddress, transferSize);
 
 		if (baseAddress >= 0x1058)
 			transferSize = 0;