Merge branch 'master' into zcomploc-support

9 files changed, 40 insertions, 67 deletions

diff --git a/Source/Core/VideoCommon/Src/BPStructs.cpp b/Source/Core/VideoCommon/Src/BPStructs.cpp
index 6cb102555b..01b09e351b 100644
--- a/Source/Core/VideoCommon/Src/BPStructs.cpp
+++ b/Source/Core/VideoCommon/Src/BPStructs.cpp
@@ -476,10 +476,15 @@ void BPWritten(const BPCmd& bp)
 		// if this is different from 0, manual TMEM management is used.
 		if (bp.newvalue != 0)
 		{
-			// NOTE(neobrain): Apparently tmemodd doesn't affect hardware behavior at all (libogc uses it just as a buffer and switches its contents with tmemeven whenever this is called)
 			BPS_TmemConfig& tmem_cfg = bpmem.tmem_config;
 			u8* ram_ptr = Memory::GetPointer(tmem_cfg.preload_addr << 5);
-			u32 tmem_addr = tmem_cfg.preload_tmem_even * TMEM_LINE_SIZE;
+			u32 tmem_addr = 0;
+
+			if (bp.newvalue >> 16)
+				tmem_addr = tmem_cfg.preload_tmem_odd * TMEM_LINE_SIZE;
+			else
+				tmem_addr = tmem_cfg.preload_tmem_even * TMEM_LINE_SIZE;
+
 			u32 size = tmem_cfg.preload_tile_info.count * 32;
 			memcpy(texMem + tmem_addr, ram_ptr, size);
 		}
diff --git a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
index 79afc45974..89032c67ea 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
+++ b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
@@ -560,10 +560,6 @@ const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType
 	WRITE(p, "uniform float4 "I_KCOLORS"[4] : register(c%d);\n", C_KCOLORS);
 	WRITE(p, "uniform float4 "I_ALPHA"[1] : register(c%d);\n", C_ALPHA);
 	WRITE(p, "uniform float4 "I_TEXDIMS"[8] : register(c%d);\n", C_TEXDIMS);
-	if (ApiType & API_D3D9)
-	{
-		WRITE(p, "uniform float4 "I_VTEXSCALE"[4] : register(c%d);\n", C_VTEXSCALE);
-	}
 	WRITE(p, "uniform float4 "I_ZBIAS"[2] : register(c%d);\n", C_ZBIAS);
 	WRITE(p, "uniform float4 "I_INDTEXSCALE"[2] : register(c%d);\n", C_INDTEXSCALE);
 	WRITE(p, "uniform float4 "I_INDTEXMTX"[6] : register(c%d);\n", C_INDTEXMTX);
@@ -1103,15 +1099,9 @@ static void WriteStage(char *&p, int n, API_TYPE ApiType)
 void SampleTexture(char *&p, const char *destination, const char *texcoords, const char *texswap, int texmap, API_TYPE ApiType)
 {
 	if (ApiType == API_D3D11)
-		WRITE(p, "%s=Tex%d.Sample(samp%d, %s.xy * "I_TEXDIMS"[%d].xy).%s;\n", destination, texmap,texmap, texcoords, texmap, texswap);
-	else if (ApiType & API_D3D9)
-	{
-		// D3D9 uses different pixel to texel mapping, so we need to offset our sampling address by half a pixel (assuming native and virtual texture dimensions match each other, otherwise some math is involved).
-		// Read the MSDN article "Directly Mapping Texels to Pixels (Direct3D 9)" for further info.
-		WRITE(p, "%s=tex2D(samp%d, (%s.xy + 0.5f*"I_VTEXSCALE"[%d].%s) * "I_TEXDIMS"[%d].xy).%s;\n", destination, texmap, texcoords, texmap/2, (texmap&1)?"zw":"xy", texmap, texswap);
-	}
+		WRITE(p, "%s=Tex%d.Sample(samp%d,%s.xy * "I_TEXDIMS"[%d].xy).%s;\n", destination, texmap,texmap, texcoords, texmap, texswap);
 	else
-		WRITE(p, "%s=tex2D(samp%d, %s.xy * "I_TEXDIMS"[%d].xy).%s;\n", destination, texmap, texcoords, texmap, texswap);
+		WRITE(p, "%s=tex2D(samp%d,%s.xy * "I_TEXDIMS"[%d].xy).%s;\n", destination, texmap, texcoords, texmap, texswap);
 }
 
 static const char *tevAlphaFuncsTable[] =
diff --git a/Source/Core/VideoCommon/Src/PixelShaderGen.h b/Source/Core/VideoCommon/Src/PixelShaderGen.h
index 7374251c96..31242a916e 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderGen.h
+++ b/Source/Core/VideoCommon/Src/PixelShaderGen.h
@@ -24,27 +24,26 @@
 #define I_KCOLORS     "k"
 #define I_ALPHA       "alphaRef"
 #define I_TEXDIMS     "texdim"
-#define I_VTEXSCALE   "vtexscale"
 #define I_ZBIAS       "czbias"
 #define I_INDTEXSCALE "cindscale"
 #define I_INDTEXMTX   "cindmtx"
 #define I_FOG         "cfog"
 #define I_PLIGHTS	  "cLights"
-#define I_PMATERIALS  "cmtrl"
+#define I_PMATERIALS   "cmtrl"
 
 #define C_COLORMATRIX	0						// 0
 #define C_COLORS		0						// 0
 #define C_KCOLORS		(C_COLORS + 4)			// 4
 #define C_ALPHA			(C_KCOLORS + 4)			// 8
 #define C_TEXDIMS		(C_ALPHA + 1)			// 9
-#define C_VTEXSCALE		(C_TEXDIMS + 8)			//17 - virtual texture scaling factor (e.g. custom textures, scaled EFB copies)
-#define C_ZBIAS			(C_VTEXSCALE + 4)		//21
-#define C_INDTEXSCALE	(C_ZBIAS + 2)			//23
-#define C_INDTEXMTX		(C_INDTEXSCALE + 2)		//25
-#define C_FOG			(C_INDTEXMTX + 6)		//31
-#define C_PLIGHTS		(C_FOG + 3)				//34
-#define C_PMATERIALS	(C_PLIGHTS + 40)		//74
-#define C_PENVCONST_END (C_PMATERIALS + 4)		//78
+#define C_ZBIAS			(C_TEXDIMS + 8)			//17
+#define C_INDTEXSCALE	(C_ZBIAS + 2)			//19
+#define C_INDTEXMTX		(C_INDTEXSCALE + 2)		//21
+#define C_FOG			(C_INDTEXMTX + 6)		//27
+
+#define C_PLIGHTS		(C_FOG + 3)
+#define C_PMATERIALS	(C_PLIGHTS + 40)
+#define C_PENVCONST_END (C_PMATERIALS + 4)
 #define PIXELSHADERUID_MAX_VALUES 70
 #define PIXELSHADERUID_MAX_VALUES_SAFE 120
 
diff --git a/Source/Core/VideoCommon/Src/PixelShaderManager.cpp b/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
index 3e0f6d73c4..2521f80500 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
+++ b/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
@@ -36,11 +36,9 @@ static bool s_bFogRangeAdjustChanged;
 static int nLightsChanged[2]; // min,max
 static float lastRGBAfull[2][4][4];
 static u8 s_nTexDimsChanged;
-static u8 s_nVirtualTexScalesChanged;
 static u8 s_nIndTexScaleChanged;
 static u32 lastAlpha;
 static u32 lastTexDims[8]; // width | height << 16 | wrap_s << 28 | wrap_t << 30
-static float lastVirtualTexScales[16]; // even fields: width ratio;   odd fields: height ratio
 static u32 lastZBias;
 static int nMaterialsChanged;
 
@@ -63,7 +61,6 @@ void PixelShaderManager::Init()
 {
 	lastAlpha = 0;
 	memset(lastTexDims, 0, sizeof(lastTexDims));
-	memset(lastVirtualTexScales, 0, sizeof(lastVirtualTexScales));
 	lastZBias = 0;
 	memset(lastRGBAfull, 0, sizeof(lastRGBAfull));
 	Dirty();
@@ -73,7 +70,6 @@ void PixelShaderManager::Dirty()
 {
 	s_nColorsChanged[0] = s_nColorsChanged[1] = 15;
 	s_nTexDimsChanged = 0xFF;
-	s_nVirtualTexScalesChanged = 0xFF;
 	s_nIndTexScaleChanged = 0xFF;
 	s_nIndTexMtxChanged = 15;
 	s_bAlphaChanged = s_bZBiasChanged = s_bZTextureTypeChanged = s_bDepthRangeChanged = true;
@@ -87,7 +83,7 @@ void PixelShaderManager::Shutdown()
 
 }
 
-void PixelShaderManager::SetConstants(API_TYPE api_type)
+void PixelShaderManager::SetConstants()
 {
     for (int i = 0; i < 2; ++i)
 	{
@@ -113,16 +109,6 @@ void PixelShaderManager::SetConstants(API_TYPE api_type)
         s_nTexDimsChanged = 0;
     }
 
-	if ((api_type & API_D3D9) && s_nVirtualTexScalesChanged)
-	{
-		for (int i = 0; i < 8; i += 2)
-		{
-			if (s_nVirtualTexScalesChanged & (3<<i))
-				SetPSVirtualTexScalePair(i/2);
-		}
-		s_nVirtualTexScalesChanged = 0;
-	}
-
     if (s_bAlphaChanged)
 	{
 		SetPSConstant4f(C_ALPHA, (lastAlpha&0xff)/255.0f, ((lastAlpha>>8)&0xff)/255.0f, 0, ((lastAlpha>>16)&0xff)/255.0f);
@@ -352,13 +338,6 @@ void PixelShaderManager::SetPSTextureDims(int texid)
 	SetPSConstant4fv(C_TEXDIMS + texid, fdims);
 }
 
-void PixelShaderManager::SetPSVirtualTexScalePair(int texpairid)
-{
-	PRIM_LOG("vtexscale%d: %f %f %f %f\n", texpairid, lastVirtualTexScales[texpairid*4], lastVirtualTexScales[texpairid*4+1],
-														lastVirtualTexScales[texpairid*4+2], lastVirtualTexScales[texpairid*4+3]);
-	SetPSConstant4fv(C_VTEXSCALE + texpairid, &lastVirtualTexScales[texpairid*4]);
-}
-
 // This one is high in profiles (0.5%). TODO: Move conversion out, only store the raw color value
 // and update it when the shader constant is set, only.
 void PixelShaderManager::SetColorChanged(int type, int num, bool high)
@@ -397,25 +376,14 @@ void PixelShaderManager::SetDestAlpha(const ConstantAlpha& alpha)
     }
 }
 
-void PixelShaderManager::SetTexDims(int texmapid, u32 width, u32 height, u32 virtual_width, u32 virtual_height, u32 wraps, u32 wrapt, API_TYPE api_type)
+void PixelShaderManager::SetTexDims(int texmapid, u32 width, u32 height, u32 wraps, u32 wrapt)
 {
     u32 wh = width | (height << 16) | (wraps << 28) | (wrapt << 30);
-
-	bool refresh = lastTexDims[texmapid] != wh;
-	if (api_type & API_D3D9)
+    if (lastTexDims[texmapid] != wh)
 	{
-		refresh |= (lastVirtualTexScales[texmapid*2] != (float)width / (float)virtual_width);
-		refresh |= (lastVirtualTexScales[texmapid*2+1] != (float)height / (float)virtual_height);
-	}
-
-	if (refresh)
-	{
-		lastTexDims[texmapid] = wh;
-		lastVirtualTexScales[texmapid*2] = (float)width / (float)virtual_width;
-		lastVirtualTexScales[texmapid*2+1] = (float)height / (float)virtual_height;
+        lastTexDims[texmapid] = wh;
 		s_nTexDimsChanged |= 1 << texmapid;
-		s_nVirtualTexScalesChanged |= 1 << texmapid;
-	}
+    }
 }
 
 void PixelShaderManager::SetZTextureBias(u32 bias)
diff --git a/Source/Core/VideoCommon/Src/PixelShaderManager.h b/Source/Core/VideoCommon/Src/PixelShaderManager.h
index 5cbe86cae4..2d1c01cad6 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderManager.h
+++ b/Source/Core/VideoCommon/Src/PixelShaderManager.h
@@ -26,20 +26,18 @@
 class PixelShaderManager
 {
 	static void SetPSTextureDims(int texid);
-	static void SetPSVirtualTexScalePair(int texpairid);
-
 public:
 	static void Init();
 	static void Dirty();
 	static void Shutdown();
 
-	static void SetConstants(API_TYPE api_type); // sets pixel shader constants
+	static void SetConstants(); // sets pixel shader constants
 
 	// constant management, should be called after memory is committed
 	static void SetColorChanged(int type, int index, bool high);
 	static void SetAlpha(const AlphaFunc& alpha);
 	static void SetDestAlpha(const ConstantAlpha& alpha);
-	static void SetTexDims(int texmapid, u32 width, u32 height, u32 virtual_width, u32 virtual_height, u32 wraps, u32 wrapt, API_TYPE api_type);
+	static void SetTexDims(int texmapid, u32 width, u32 height, u32 wraps, u32 wrapt);
 	static void SetZTextureBias(u32 bias);
 	static void SetViewportChanged();
 	static void SetIndMatrixChanged(int matrixidx);
diff --git a/Source/Core/VideoCommon/Src/RenderBase.cpp b/Source/Core/VideoCommon/Src/RenderBase.cpp
index 3b06bbb1c0..ef0fff57f7 100644
--- a/Source/Core/VideoCommon/Src/RenderBase.cpp
+++ b/Source/Core/VideoCommon/Src/RenderBase.cpp
@@ -43,6 +43,7 @@
 #include "XFMemory.h"
 #include "FifoPlayer/FifoRecorder.h"
 #include "AVIDump.h"
+#include "VertexShaderManager.h"
 
 #include <cmath>
 #include <string>
@@ -193,6 +194,7 @@ bool Renderer::CalculateTargetSize(int multiplier)
 	{
 		s_target_width  = newEFBWidth;
 		s_target_height = newEFBHeight;
+		VertexShaderManager::SetViewportChanged();
 		return true;
 	}
 	return false;
diff --git a/Source/Core/VideoCommon/Src/VertexShaderGen.cpp b/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
index 5076f9f044..3df779a7ba 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
+++ b/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
@@ -493,6 +493,13 @@ const char *GenerateVertexShaderCode(u32 components, API_TYPE api_type)
 	    //seems to get rather complicated
 	}
 
+	if (api_type & API_D3D9)
+	{
+		// D3D9 is addressing pixel centers instead of pixel boundaries in clip space.
+		// Thus we need to offset the final position by half a pixel
+		WRITE(p, "o.pos = o.pos + float4("I_DEPTHPARAMS".z, "I_DEPTHPARAMS".w, 0.f, 0.f);\n");
+	}
+
 	WRITE(p, "return o;\n}\n");
 
 
diff --git a/Source/Core/VideoCommon/Src/VertexShaderGen.h b/Source/Core/VideoCommon/Src/VertexShaderGen.h
index 0c522286b1..cb253a9b6c 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderGen.h
+++ b/Source/Core/VideoCommon/Src/VertexShaderGen.h
@@ -35,7 +35,7 @@
 #define I_TRANSFORMMATRICES     "ctrmtx"
 #define I_NORMALMATRICES        "cnmtx"
 #define I_POSTTRANSFORMMATRICES "cpostmtx"
-#define I_DEPTHPARAMS           "cDepth"
+#define I_DEPTHPARAMS           "cDepth" // farZ, zRange, scaled viewport width, scaled viewport height
 
 #define C_POSNORMALMATRIX        0
 #define C_PROJECTION            (C_POSNORMALMATRIX + 6)
diff --git a/Source/Core/VideoCommon/Src/VertexShaderManager.cpp b/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
index d7d25c5c85..b6afbc61c6 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
+++ b/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
@@ -306,7 +306,11 @@ void VertexShaderManager::SetConstants()
 	if (bViewportChanged)
 	{
 		bViewportChanged = false;
-		SetVSConstant4f(C_DEPTHPARAMS,xfregs.viewport.farZ / 16777216.0f,xfregs.viewport.zRange / 16777216.0f,0.0f,0.0f);
+		SetVSConstant4f(C_DEPTHPARAMS,
+						xfregs.viewport.farZ / 16777216.0f,
+						xfregs.viewport.zRange / 16777216.0f,
+						-1.f / (float)g_renderer->EFBToScaledX((int)ceil(2.0f * xfregs.viewport.wd)),
+						1.f / (float)g_renderer->EFBToScaledY((int)ceil(-2.0f * xfregs.viewport.ht)));
 		// This is so implementation-dependent that we can't have it here.
 		UpdateViewport(s_viewportCorrection);
 		bProjectionChanged = true;