some work on zscale and zoffset, don't know if this is correct but at least it seem to fix one or two games.

thanks a lot to chaoscode he make possible to test this on nvidia. please test this a lot and give me feedback :)

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@4468 8ced0084-cf51-0410-be5f-012b33b47a6e

7 files changed, 56 insertions, 9 deletions

diff --git a/Source/Core/VideoCommon/Src/PixelShaderManager.cpp b/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
index c6a3e3da08..f298629536 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
+++ b/Source/Core/VideoCommon/Src/PixelShaderManager.cpp
@@ -31,7 +31,7 @@ static bool s_bZTextureTypeChanged;
 static bool s_bDepthRangeChanged;
 static bool s_bFogColorChanged;
 static bool s_bFogParamChanged;
-static float lastDepthRange[2]; // 0 = far z, 1 = far - near
+static float lastDepthRange[4]; // 0 = far z, 1 = far - near, 2 = scale , 3 = offset
 static float lastRGBAfull[2][4][4];
 static float lastCustomTexScale[8][2];
 static u8 s_nTexDimsChanged;
@@ -54,6 +54,10 @@ void PixelShaderManager::Init()
 	lastZBias = 0;
 	s_texturemask = 0;
 	memset(lastRGBAfull, 0, sizeof(lastRGBAfull));
+	lastDepthRange[0]=16777216.0f;
+	lastDepthRange[1]=16777216.0f;
+	lastDepthRange[2]=16777216.0f;
+	lastDepthRange[3]=0.0f;
 	Dirty();
 }
 
@@ -132,7 +136,7 @@ void PixelShaderManager::SetConstants()
 	if (s_bZBiasChanged || s_bDepthRangeChanged) 
 	{
         //ERROR_LOG("pixel=%x,%x, bias=%x\n", bpmem.zcontrol.pixel_format, bpmem.ztex2.type, lastZBias);
-        SetPSConstant4f(C_ZBIAS+1, lastDepthRange[0] / 16777215.0f, lastDepthRange[1] / 16777215.0f, 0, (float)( (((int)lastZBias<<8)>>8))/16777215.0f);
+        SetPSConstant4f(C_ZBIAS+1, (lastDepthRange[0] + lastDepthRange[3]) / lastDepthRange[2], (lastDepthRange[1] + lastDepthRange[3]) / lastDepthRange[2], 0, (float)( (((int)lastZBias<<8)>>8))/16777215.0f);
 		s_bZBiasChanged = s_bDepthRangeChanged = false;
     }
 
@@ -326,6 +330,24 @@ void PixelShaderManager::SetViewport(float* viewport)
 	}
 }
 
+void PixelShaderManager::SetZScale(float data)
+{
+	if (lastDepthRange[2] != data)
+	{
+		lastDepthRange[2] = data;
+		s_bDepthRangeChanged = true;
+	}
+}
+
+void PixelShaderManager::SetZOffset(float data)
+{
+	if (lastDepthRange[3] != data)
+	{
+		lastDepthRange[3] = data;
+		s_bDepthRangeChanged = true;
+	}
+}
+
 void PixelShaderManager::SetIndTexScaleChanged(u8 stagemask)
 {
     s_nIndTexScaleChanged |= stagemask;
diff --git a/Source/Core/VideoCommon/Src/PixelShaderManager.h b/Source/Core/VideoCommon/Src/PixelShaderManager.h
index 2e8935b69c..5625dbdb05 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderManager.h
+++ b/Source/Core/VideoCommon/Src/PixelShaderManager.h
@@ -54,6 +54,8 @@ public:
 	static void SetFogParamChanged();
 	static void SetColorMatrix(const float* pmatrix, const float* pfConstAdd);
 	static u32 GetTextureMask();
+	static void SetZScale(float data);
+	static void SetZOffset(float data);
 };
 
 
diff --git a/Source/Core/VideoCommon/Src/VertexShaderGen.cpp b/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
index a49f83a71e..bbd5a34f64 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
+++ b/Source/Core/VideoCommon/Src/VertexShaderGen.cpp
@@ -198,7 +198,7 @@ const char *GenerateVertexShader(u32 components, bool D3D)
                     //"half3 norm2 = normalize(_norm2);\n");
     }
     else {
-        WRITE(p, "float4 pos = float4(dot("I_POSNORMALMATRIX".T0, rawpos), dot("I_POSNORMALMATRIX".T1, rawpos), dot("I_POSNORMALMATRIX".T2, rawpos), 1);\n");
+        WRITE(p, "float4 pos = float4(dot("I_POSNORMALMATRIX".T0, rawpos), dot("I_POSNORMALMATRIX".T1, rawpos), dot("I_POSNORMALMATRIX".T2, rawpos), 1.0f);\n");
         if (components & VB_HAS_NRM0)
             WRITE(p, "half3 _norm0 = half3(dot("I_POSNORMALMATRIX".N0.xyz, rawnorm0), dot("I_POSNORMALMATRIX".N1.xyz, rawnorm0), dot("I_POSNORMALMATRIX".N2.xyz, rawnorm0));\n"
                      "half3 norm0 = normalize(_norm0);\n");
@@ -243,7 +243,7 @@ const char *GenerateVertexShader(u32 components, bool D3D)
                 if (components & (VB_HAS_COL0<<j) )
                     WRITE(p, "lacc = color%d;\n", j);
                 else
-					WRITE(p, "lacc = half4(0.0f,0.0f,0.0f,0.0f);\n");
+					WRITE(p, "lacc = half4(0,0,0,0);\n");
             }
             else // from color
                 WRITE(p, "lacc = "I_MATERIALS".C%d;\n", j);
@@ -338,7 +338,7 @@ const char *GenerateVertexShader(u32 components, bool D3D)
         case XF_SRCNORMAL_INROW:
             if (components & VB_HAS_NRM0) {
                 _assert_( texinfo.inputform == XF_TEXINPUT_ABC1 );
-                WRITE(p, "float4 coord = float4(rawnorm0.xyz, 1.0);\n");
+                WRITE(p, "float4 coord = float4(rawnorm0.xyz, 1.0f);\n");
             }
             else WRITE(p, "float4 coord = float4(0.0f, 0.0f, 1.0f, 1.0f);\n");  // avoid errors
             break;
@@ -348,14 +348,14 @@ const char *GenerateVertexShader(u32 components, bool D3D)
         case XF_SRCBINORMAL_T_INROW:
             if (components & VB_HAS_NRM1) {
                 _assert_( texinfo.inputform == XF_TEXINPUT_ABC1 );
-                WRITE(p, "float4 coord = float4(rawnorm1.xyz, 1.0);\n");
+                WRITE(p, "float4 coord = float4(rawnorm1.xyz, 1.0f);\n");
             }
             else WRITE(p, "float4 coord = float4(0.0f, 0.0f, 1.0f, 1.0f);\n");  // avoid errors
             break;
         case XF_SRCBINORMAL_B_INROW:
             if (components & VB_HAS_NRM2) {
                 _assert_( texinfo.inputform == XF_TEXINPUT_ABC1 );
-                WRITE(p, "float4 coord = float4(rawnorm2.xyz, 1.0);\n");
+                WRITE(p, "float4 coord = float4(rawnorm2.xyz, 1.0f);\n");
             }
             else WRITE(p, "float4 coord = float4(0.0f, 0.0f, 1.0f, 1.0f);\n");  // avoid errors
             break;
@@ -479,7 +479,7 @@ char* GenerateLightShader(char* p, int index, const LitChannel& chan, const char
             WRITE(p, "attn = max(0.0f, dot("I_LIGHTS".lights[%d].cosatt.xyz, half3(1, attn, attn*attn))) / dot("I_LIGHTS".lights[%d].distatt.xyz, half3(1,dist,dist2));\n", index, index);
         }
         else if (chan.attnfunc == 1) { // specular
-            WRITE(p, "attn = dot(norm0, "I_LIGHTS".lights[%d].pos.xyz) > 0 ? max(0.0f, dot(norm0, "I_LIGHTS".lights[%d].dir.xyz)) : 0;\n", index, index);
+            WRITE(p, "attn = dot(norm0, "I_LIGHTS".lights[%d].pos.xyz) > 0.0f ? max(0.0f, dot(norm0, "I_LIGHTS".lights[%d].dir.xyz)) : 0.0f;\n", index, index);
             WRITE(p, "ldir = half3(1,attn,attn*attn);\n");
             WRITE(p, "attn = max(0.0f, dot("I_LIGHTS".lights[%d].cosatt.xyz, ldir)) / dot("I_LIGHTS".lights[%d].distatt.xyz, ldir);\n", index, index);
         }
diff --git a/Source/Core/VideoCommon/Src/VertexShaderManager.cpp b/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
index e278cb45d9..d913d11df3 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
+++ b/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
@@ -62,7 +62,8 @@ void VertexShaderManager::Init()
 
 	memset(&xfregs, 0, sizeof(xfregs));
 	memset(xfmem, 0, sizeof(xfmem));
-
+	xfregs.ZScale = 16777216.0f;
+	xfregs.Zoffset = 0.0f;
 	ResetView();    
 }
 
@@ -451,6 +452,18 @@ void VertexShaderManager::SetViewport(float* _Viewport)
     bViewportChanged = true;
 }
 
+void VertexShaderManager::SetZScale(float data)
+{
+    xfregs.ZScale = data;
+    bViewportChanged = true;
+}
+
+void VertexShaderManager::SetZOffset(float data)
+{
+    xfregs.Zoffset = data;
+    bViewportChanged = true;
+}
+
 void VertexShaderManager::SetViewportChanged()
 {
     bViewportChanged = true;
diff --git a/Source/Core/VideoCommon/Src/VertexShaderManager.h b/Source/Core/VideoCommon/Src/VertexShaderManager.h
index be7a506716..a7aa81ccfe 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderManager.h
+++ b/Source/Core/VideoCommon/Src/VertexShaderManager.h
@@ -38,6 +38,8 @@ public:
 	static void SetTexMatrixChangedA(u32 Value);
 	static void SetTexMatrixChangedB(u32 Value);
 	static void SetMaterialColor(int index, u32 data);
+	static void SetZScale(float data);
+	static void SetZOffset(float data);
 
 	static void TranslateView(float x, float y);
 	static void RotateView(float x, float y);
diff --git a/Source/Core/VideoCommon/Src/XFMemory.h b/Source/Core/VideoCommon/Src/XFMemory.h
index 200fdadb11..0e6529a4d5 100644
--- a/Source/Core/VideoCommon/Src/XFMemory.h
+++ b/Source/Core/VideoCommon/Src/XFMemory.h
@@ -228,6 +228,8 @@ struct XFRegisters
     bool bEnableDualTexTransform;
 	float rawViewport[6];
 	float rawProjection[7];
+	float ZScale;
+	float Zoffset;
 };
 
 
diff --git a/Source/Core/VideoCommon/Src/XFStructs.cpp b/Source/Core/VideoCommon/Src/XFStructs.cpp
index 3fc7984f9e..5fd25d31d1 100644
--- a/Source/Core/VideoCommon/Src/XFStructs.cpp
+++ b/Source/Core/VideoCommon/Src/XFStructs.cpp
@@ -174,12 +174,18 @@ void LoadXFReg(u32 transferSize, u32 baseAddress, u32 *pData)
 			// paper mario writes 16777216.0f, 1677721.75
 			// Killer 7 writes 16777216.0f here
 			case XFMEM_SETZSCALE:
+				VertexManager::Flush();
+				VertexShaderManager::SetZScale((float)data);
+				PixelShaderManager::SetZScale((float)data);
 				INFO_LOG(VIDEO, "Set ZScale : %x=%x\n", address, data);
 				break;
 
 			// paper mario writes 16777216.0f, 5033165.0f
 			// Killer 7 alterns this between 16777216.0f and 16710107.0f
             case XFMEM_SETZOFFSET:
+				VertexManager::Flush();
+				VertexShaderManager::SetZOffset((float)data);
+				PixelShaderManager::SetZOffset((float)data);
 				INFO_LOG(VIDEO, "Set ZOffset : %x=%x\n", address, data);
 				break;