Some video SW clean up and bug fixes. Did a hacky fix for bad light data. Add fog rendering. Add some offsets to make rasterized pixel locations and texture coordinates more accurate.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@5361 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Plugins/Plugin_VideoSoftware/Src/Clipper.cpp

@@ -339,6 +339,9 @@ namespace Clipper
 		Vertices[0] = lineV0;
         Vertices[1] = lineV1;
 
+		// point to a valid vertex to store to when clipping
+		Vertices[2] = &ClippedVertices[17];
+
 		ClipLine(indices);
 
 		if(indices[0] != SKIP_FLAG)

Source/Plugins/Plugin_VideoSoftware/Src/DebugUtil.cpp

@@ -105,7 +105,7 @@ void DumpActiveTextures()
         u32 texmap = bpmem.tevindref.getTexMap(stageNum);
 
 		s32 maxLod = GetMaxTextureLod(texmap);
-		for (s32 mip = 0; mip < maxLod; ++mip)
+		for (s32 mip = 0; mip <= maxLod; ++mip)
 		{
 			SaveTexture(StringFromFormat("%star%i_ind%i_map%i_mip%i.tga", File::GetUserPath(D_DUMPTEXTURES_IDX), stats.thisFrame.numDrawnObjects, stageNum, texmap, mip).c_str(), texmap, mip);
 		}
@@ -120,7 +120,7 @@ void DumpActiveTextures()
         int texmap = order.getTexMap(stageOdd);
 
         s32 maxLod = GetMaxTextureLod(texmap);
-		for (s32 mip = 0; mip < maxLod; ++mip)
+		for (s32 mip = 0; mip <= maxLod; ++mip)
 		{
 			SaveTexture(StringFromFormat("%star%i_stage%i_map%i_mip%i.tga", File::GetUserPath(D_DUMPTEXTURES_IDX), stats.thisFrame.numDrawnObjects, stageNum, texmap, mip).c_str(), texmap, mip);
 		}

Source/Plugins/Plugin_VideoSoftware/Src/HwRasterizer.cpp

@@ -85,13 +85,13 @@ namespace HwRasterizer
     void DrawColorVertex(OutputVertexData *v)
     {
         glColor3ub(v->color[0][0], v->color[0][1], v->color[0][2]);
-        glVertex3f(v->screenPosition[0] / efbHalfWidth - 1.0f, 1.0f - v->screenPosition[1] / efbHalfHeight, v->screenPosition[2]);
+        glVertex3f(v->screenPosition.x / efbHalfWidth - 1.0f, 1.0f - v->screenPosition.y / efbHalfHeight, v->screenPosition.z);
     }
 
     void DrawTextureVertex(OutputVertexData *v)
     {
-        glTexCoord2f(v->texCoords[0][0] * texWidth, v->texCoords[0][1] * texHeight);
+        glTexCoord2f(v->texCoords[0].x * texWidth, v->texCoords[0].y * texHeight);
-        glVertex3f(v->screenPosition[0] / efbHalfWidth - 1.0f, 1.0f - v->screenPosition[1] / efbHalfHeight, v->screenPosition[2]);
+        glVertex3f(v->screenPosition.x / efbHalfWidth - 1.0f, 1.0f - v->screenPosition.y / efbHalfHeight, v->screenPosition.z);
     }
 
     void DrawTriangleFrontFace(OutputVertexData *v0, OutputVertexData *v1, OutputVertexData *v2)

Source/Plugins/Plugin_VideoSoftware/Src/Rasterizer.cpp

@@ -49,6 +49,11 @@ Slope WSlope;
 Slope ColorSlopes[2][4];
 Slope TexSlopes[8][3];
 
+s32 vertex0X;
+s32 vertex0Y;
+float vertexOffsetX;
+float vertexOffsetY;
+
 s32 scissorLeft = 0;
 s32 scissorTop = 0;
 s32 scissorRight = 0;
@@ -108,7 +113,10 @@ inline void Draw(s32 x, s32 y, s32 xi, s32 yi)
 {
     INCSTAT(stats.thisFrame.rasterizedPixels);
 
-	float zFloat = 1.0f + ZSlope.GetValue(x, y);
+	float dx = vertexOffsetX + (float)(x - vertex0X);
+	float dy = vertexOffsetY + (float)(y - vertex0Y);
+
+	float zFloat = 1.0f + ZSlope.GetValue(dx, dy);
 	if (zFloat < 0.0f || zFloat > 1.0f)
 		return;
 
@@ -133,7 +141,7 @@ inline void Draw(s32 x, s32 y, s32 xi, s32 yi)
 	for (unsigned int i = 0; i < bpmem.genMode.numcolchans; i++)
 	{
 		for(int comp = 0; comp < 4; comp++)
-			tev.Color[i][comp] = (u8)ColorSlopes[i][comp].GetValue(x, y);
+			tev.Color[i][comp] = (u8)ColorSlopes[i][comp].GetValue(dx, dy);
 	}
 
 	// tex coords
@@ -159,7 +167,19 @@ inline void Draw(s32 x, s32 y, s32 xi, s32 yi)
     tev.Draw();
 }
 
-void InitSlope(Slope *slope, float f1, float f2, float f3, float DX31, float DX12, float DY12, float DY31, float X1, float Y1)
+void InitTriangle(float X1, float Y1, s32 xi, s32 yi)
+{
+	vertex0X = xi;
+    vertex0Y = yi;
+
+	// adjust a little less than 0.5
+	const float adjust = 0.495f;
+
+	vertexOffsetX = ((float)xi - X1) + adjust;
+	vertexOffsetY = ((float)yi - Y1) + adjust;
+}
+
+void InitSlope(Slope *slope, float f1, float f2, float f3, float DX31, float DX12, float DY12, float DY31)
 {
 	float DF31 = f3 - f1;
     float DF21 = f2 - f1;
@@ -169,8 +189,6 @@ void InitSlope(Slope *slope, float f1, float f2, float f3, float DX31, float DX1
     slope->dfdx = -a / c;
     slope->dfdy = -b / c;
     slope->f0 = f1;
-    slope->x0 = X1;
-    slope->y0 = Y1;
 }
 
 inline void CalculateLOD(s32 &lod, bool &linear, u32 texmap, u32 texcoord)
@@ -210,7 +228,7 @@ inline void CalculateLOD(s32 &lod, bool &linear, u32 texmap, u32 texcoord)
 	bias >>= 1;
 	lod += bias;
 
-	linear = (lod >= 0 && (tm0.min_filter & 4) || lod < 0 && tm0.mag_filter);
+	linear = (lod > 0 && (tm0.min_filter & 4) || lod <= 0 && tm0.mag_filter);
 
 	// order of checks matters
 	// should be:
@@ -227,10 +245,10 @@ void BuildBlock(s32 blockX, s32 blockY)
 		{
 			RasterBlockPixel& pixel = rasterBlock.Pixel[xi][yi];
 
-			s32 x = xi + blockX;
+			float dx = vertexOffsetX + (float)(xi + blockX - vertex0X);
-			s32 y = yi + blockY;			
+			float dy = vertexOffsetY + (float)(yi + blockY - vertex0Y);
 
-			float invW = 1.0f / WSlope.GetValue(x, y);
+			float invW = 1.0f / WSlope.GetValue(dx, dy);
 			pixel.InvW = invW;
 
 			// tex coords
@@ -239,14 +257,14 @@ void BuildBlock(s32 blockX, s32 blockY)
 				float projection;
 				if (xfregs.texMtxInfo[i].projection)
 				{
-					float q = TexSlopes[i][2].GetValue(x, y) * invW;
+					float q = TexSlopes[i][2].GetValue(dx, dy) * invW;
 					projection = invW / q;
 				}
 				else
 					projection = invW;
 
-				pixel.Uv[i][0] = TexSlopes[i][0].GetValue(x, y) * projection;
+				pixel.Uv[i][0] = TexSlopes[i][0].GetValue(dx, dy) * projection;
-				pixel.Uv[i][1] = TexSlopes[i][1].GetValue(x, y) * projection;
+				pixel.Uv[i][1] = TexSlopes[i][1].GetValue(dx, dy) * projection;
 			}
 		}
 	}
@@ -288,14 +306,15 @@ void DrawTriangleFrontFace(OutputVertexData *v0, OutputVertexData *v1, OutputVer
 
     // adapted from http://www.devmaster.net/forums/showthread.php?t=1884
 
-    // 28.4 fixed-pou32 coordinates. rounded to nearest
+    // 28.4 fixed-pou32 coordinates. rounded to nearest and adjusted to match hardware output
-    const s32 Y1 = iround(16.0f * v0->screenPosition[1]);
+	// could also take floor and adjust -8
-    const s32 Y2 = iround(16.0f * v1->screenPosition[1]);
+	const s32 Y1 = iround(16.0f * v0->screenPosition[1]) - 9;
-    const s32 Y3 = iround(16.0f * v2->screenPosition[1]);
+	const s32 Y2 = iround(16.0f * v1->screenPosition[1]) - 9;
+	const s32 Y3 = iround(16.0f * v2->screenPosition[1]) - 9;
 
-    const s32 X1 = iround(16.0f * v0->screenPosition[0]);
+	const s32 X1 = iround(16.0f * v0->screenPosition[0]) - 9;
-    const s32 X2 = iround(16.0f * v1->screenPosition[0]);
+	const s32 X2 = iround(16.0f * v1->screenPosition[0]) - 9;
-    const s32 X3 = iround(16.0f * v2->screenPosition[0]);
+	const s32 X3 = iround(16.0f * v2->screenPosition[0]) - 9;
 
     // Deltas
     const s32 DX12 = X1 - X2;
@@ -331,28 +350,30 @@ void DrawTriangleFrontFace(OutputVertexData *v0, OutputVertexData *v1, OutputVer
         return;
 
     // Setup slopes
-    float fltx1 = v0->screenPosition[0];
+    float fltx1 = v0->screenPosition.x;
-    float flty1 = v0->screenPosition[1];
+    float flty1 = v0->screenPosition.y;
-    float fltdx31 = v2->screenPosition[0] - fltx1;
+    float fltdx31 = v2->screenPosition.x - fltx1;
-    float fltdx12 = fltx1 - v1->screenPosition[0];
+    float fltdx12 = fltx1 - v1->screenPosition.x;
-    float fltdy12 = flty1 - v1->screenPosition[1];
+    float fltdy12 = flty1 - v1->screenPosition.y;
-    float fltdy31 = v2->screenPosition[1] - flty1;
+    float fltdy31 = v2->screenPosition.y - flty1;
+
+	InitTriangle(fltx1, flty1, (X1 + 0xF) >> 4, (Y1 + 0xF) >> 4);
 
     float w[3] = { 1.0f / v0->projectedPosition.w, 1.0f / v1->projectedPosition.w, 1.0f / v2->projectedPosition.w };
-    InitSlope(&WSlope, w[0], w[1], w[2], fltdx31, fltdx12, fltdy12, fltdy31, fltx1, flty1);
+    InitSlope(&WSlope, w[0], w[1], w[2], fltdx31, fltdx12, fltdy12, fltdy31);
 
-    InitSlope(&ZSlope, v0->screenPosition[2], v1->screenPosition[2], v2->screenPosition[2], fltdx31, fltdx12, fltdy12, fltdy31, fltx1, flty1);
+    InitSlope(&ZSlope, v0->screenPosition[2], v1->screenPosition[2], v2->screenPosition[2], fltdx31, fltdx12, fltdy12, fltdy31);
 
     for(unsigned int i = 0; i < bpmem.genMode.numcolchans; i++)
     {
         for(int comp = 0; comp < 4; comp++)
-            InitSlope(&ColorSlopes[i][comp], v0->color[i][comp], v1->color[i][comp], v2->color[i][comp], fltdx31, fltdx12, fltdy12, fltdy31, fltx1, flty1);
+            InitSlope(&ColorSlopes[i][comp], v0->color[i][comp], v1->color[i][comp], v2->color[i][comp], fltdx31, fltdx12, fltdy12, fltdy31);
     }
 
     for(unsigned int i = 0; i < bpmem.genMode.numtexgens; i++)
     {
         for(int comp = 0; comp < 3; comp++)
-            InitSlope(&TexSlopes[i][comp], v0->texCoords[i][comp] * w[0], v1->texCoords[i][comp] * w[1], v2->texCoords[i][comp] * w[2], fltdx31, fltdx12, fltdy12, fltdy31, fltx1, flty1);
+            InitSlope(&TexSlopes[i][comp], v0->texCoords[i][comp] * w[0], v1->texCoords[i][comp] * w[1], v2->texCoords[i][comp] * w[2], fltdx31, fltdx12, fltdy12, fltdy31);
     }
 
     // Start in corner of 8x8 block

Source/Plugins/Plugin_VideoSoftware/Src/Rasterizer.h

@@ -35,9 +35,8 @@ namespace Rasterizer
         float dfdx;
         float dfdy;
         float f0;
-        float x0;
+
-        float y0;
+		float GetValue(float dx, float dy) { return f0 + (dfdx * dx) + (dfdy * dy); }
-        float GetValue(s32 x, s32 y) { return f0 + (dfdx * (x - x0)) + (dfdy * (y - y0)); }
     };
 
 	struct RasterBlockPixel

Source/Plugins/Plugin_VideoSoftware/Src/Tev.cpp

@@ -705,6 +705,57 @@ void Tev::Draw()
         Position[2] = ztex & 0x00ffffff;
     }
 
+	// fog
+	if (bpmem.fog.c_proj_fsel.fsel)
+	{
+		float ze;
+
+		if (bpmem.fog.c_proj_fsel.proj == 0)
+		{
+			// perspective
+			// ze = A/(B - Zs)
+			s32 denom = bpmem.fog.b_magnitude - (Position[2] >> bpmem.fog.b_shift);
+			ze = bpmem.fog.a.GetA() / (float)denom;
+		} 
+		else 
+		{
+			// orthographic
+			// ze = a*Zs
+			ze = bpmem.fog.a.GetA() / (float)Position[2];
+		}
+
+		ze = (ze * (float)0xffffff) - bpmem.fog.c_proj_fsel.GetC();
+
+		// clamp 0 to 1
+		float fog = (ze<0.0f) ? 0.0f : ((ze>1.0f) ? 1.0f : ze);
+
+		switch (bpmem.fog.c_proj_fsel.fsel)
+		{
+			case 4: // exp
+				fog = 1.0f - pow(2.0f, -8.0f * fog);
+				break;
+			case 5: // exp2
+				fog = 1.0f - pow(2.0f, -8.0f * fog * fog);
+				break;
+			case 6: // backward exp
+				fog = 1.0f - fog;
+				fog = 1.0f - pow(2.0f, -8.0f * fog);
+				break;
+			case 7: // backward exp2
+				fog = 1.0f - fog;
+				fog = 1.0f - pow(2.0f, -8.0f * fog * fog);
+				break;
+		}
+
+		// lerp from output to fog color
+		u32 fogInt = (u32)(fog * 256);
+		u32 invFog = 256 - fogInt;
+
+		output[RED_C] = (output[RED_C] * invFog + fogInt * bpmem.fog.color.r) >> 8;
+		output[GRN_C] = (output[GRN_C] * invFog + fogInt * bpmem.fog.color.g) >> 8;
+		output[BLU_C] = (output[BLU_C] * invFog + fogInt * bpmem.fog.color.b) >> 8;
+	}
+
     if (!bpmem.zcontrol.zcomploc && bpmem.zmode.testenable)
     {
         if (!EfbInterface::ZCompare(Position[0], Position[1], Position[2]))

Source/Plugins/Plugin_VideoSoftware/Src/TextureSampler.cpp

@@ -157,12 +157,16 @@ void SampleMip(s32 s, s32 t, s32 mip, bool linear, u8 texmap, u8 *sample)
 		}
 	}
 
+    if (linear)
+    {
+		// offset linear sampling
+		s -= 64;
+		t -= 64;
+
 		// integer part of sample location
 		int imageS = s >> 7;
 		int imageT = t >> 7;
 
-    if (linear)
-    {
 		// linear sampling
 		int imageSPlus1 = imageS + 1;
         int fractS = s & 0x7f;
@@ -197,6 +201,10 @@ void SampleMip(s32 s, s32 t, s32 mip, bool linear, u8 texmap, u8 *sample)
     }
     else
     {
+		// integer part of sample location
+		int imageS = s >> 7;
+		int imageT = t >> 7;
+
         // nearest neighbor sampling
 		WrapCoord(imageS, tm0.wrap_s, imageWidth);
         WrapCoord(imageT, tm0.wrap_t, imageHeight);

Source/Plugins/Plugin_VideoSoftware/Src/TransformUnit.cpp

@@ -166,15 +166,9 @@ inline void TransformTexCoordRegular(const TexMtxInfo &texinfo, int coordNum, bo
 		else
 		{		
 			if (postInfo.normalize)
-			{
+				tempCoord = dst->normalized();
-				float length = sqrtf(dst[0] * dst[0] + dst[1] * dst[1] + dst[2] * dst[2]);
-				float invL = 1.0f / length;
-				tempCoord = *dst * invL;
-			}
 			else
-			{
 				tempCoord = *dst;
-			}
 
 			MultiplyVec3Mat34(tempCoord, postMat, *dst);
 		}
@@ -463,29 +457,32 @@ void TransformTexCoord(const InputVertexData *src, OutputVertexData *dst, bool s
                 float d1 = ldir * dst->normal[1];
                 float d2 = ldir * dst->normal[2];
 
-                dst->texCoords[coordNum][0] = dst->texCoords[texinfo.embosssourceshift][0] + d1;
+                dst->texCoords[coordNum].x = dst->texCoords[texinfo.embosssourceshift].x + d1;
-                dst->texCoords[coordNum][1] = dst->texCoords[texinfo.embosssourceshift][1] + d2;
+                dst->texCoords[coordNum].y = dst->texCoords[texinfo.embosssourceshift].y + d2;
-                dst->texCoords[coordNum][2] = dst->texCoords[texinfo.embosssourceshift][2];
+                dst->texCoords[coordNum].z = dst->texCoords[texinfo.embosssourceshift].z;
             }
             break;
         case XF_TEXGEN_COLOR_STRGBC0:
             _assert_(texinfo.sourcerow == XF_SRCCOLORS_INROW);
             _assert_(texinfo.inputform == XF_TEXINPUT_AB11);
-            dst->texCoords[coordNum][0] = (float)dst->color[0][0] / 255.0f;
+            dst->texCoords[coordNum].x = (float)dst->color[0][0] / 255.0f;
-            dst->texCoords[coordNum][1] = (float)dst->color[0][1] / 255.0f;
+            dst->texCoords[coordNum].y = (float)dst->color[0][1] / 255.0f;
-            dst->texCoords[coordNum][2] = 1.0f;
+            dst->texCoords[coordNum].z = 1.0f;
             break;
         case XF_TEXGEN_COLOR_STRGBC1:
             _assert_(texinfo.sourcerow == XF_SRCCOLORS_INROW);
             _assert_(texinfo.inputform == XF_TEXINPUT_AB11);
-            dst->texCoords[coordNum][0] = (float)dst->color[1][0] / 255.0f;
+            dst->texCoords[coordNum].x = (float)dst->color[1][0] / 255.0f;
-            dst->texCoords[coordNum][1] = (float)dst->color[1][1] / 255.0f;
+            dst->texCoords[coordNum].y = (float)dst->color[1][1] / 255.0f;
-            dst->texCoords[coordNum][2] = 1.0f;
+            dst->texCoords[coordNum].z = 1.0f;
             break;
         default:
             ERROR_LOG(VIDEO, "Bad tex gen type %i", texinfo.texgentype);
         }
+    }
 
+	for (u32 coordNum = 0; coordNum < xfregs.numTexGens; coordNum++)
+    {
 		dst->texCoords[coordNum][0] *= (bpmem.texcoords[coordNum].s.scale_minus_1 + 1);
 		dst->texCoords[coordNum][1] *= (bpmem.texcoords[coordNum].t.scale_minus_1 + 1);
 	}

Source/Plugins/Plugin_VideoSoftware/Src/XFMemLoader.cpp

@@ -35,6 +35,30 @@ void XFWritten(u32 transferSize, u32 baseAddress)
 
     if (baseAddress <= 0x1026 && topAddress >= 0x1020)
         Clipper::SetViewOffset();
+
+	// fix lights so invalid values don't trash the lighting computations	
+	if (baseAddress <= 0x067f && topAddress >= 0x0604)
+	{
+		u32* x = xfregs.lights;
+
+		// go through all lights
+		for (int light = 0; light < 8; light++)
+		{
+			// skip to floating point values
+			x += 4;
+
+			for (int i = 0; i < 12; i++)
+			{
+				u32 xVal = *x;
+
+				// if the exponent is 255 then the number is inf or nan
+				if ((xVal & 0x7f800000) == 0x7f800000)
+					*x = 0;
+
+				x++;
+			}
+		}
+	}
 }
 
 void LoadXFReg(u32 transferSize, u32 baseAddress, u32 *pData)
@@ -66,8 +90,7 @@ void LoadIndexedXF(u32 val, int array)
     int size = ((val >> 12) & 0xF) + 1;
     //load stuff from array to address in xf mem
 
-    u32* xfmem = (u32*)&xfregs;
+	u32 *pData = (u32*)g_VideoInitialize.pGetMemoryPointer(arraybases[array] + arraystrides[array]*index);
 
-    for (int i = 0; i < size; i++)
+	LoadXFReg(size, address, pData);
-        xfmem[address + i] = Memory_Read_U32(arraybases[array] + arraystrides[array]*index + i*4);
 }

Source/Plugins/Plugin_VideoSoftware/Src/main.cpp

@@ -113,7 +113,6 @@ void Video_Prepare(void)
 // Run from the CPU thread (from VideoInterface.cpp)
 void Video_BeginField(u32 xfbAddr, FieldType field, u32 fbWidth, u32 fbHeight)
 {	
-    g_VideoInitialize.pCopiedToXFB(true);
 }
 
 // Run from the CPU thread (from VideoInterface.cpp)