VideoSoftware: Cleanup PE perf metrics; returning the proper value now.

Source/Plugins/Plugin_VideoSoftware/Src/BPMemLoader.cpp

@@ -91,7 +91,7 @@ void SWBPWritten(int address, int newvalue)
         SWPixelEngine::pereg.boxTop = newvalue & 0x3ff;
         break;
 	case BPMEM_CLEAR_PIXEL_PERF:
-		// TODO: Parameter?
+		// TODO: I didn't test if the value written to this register affects the amount of cleared registers
 		SWPixelEngine::pereg.perfZcompInputZcomplocLo = 0;
 		SWPixelEngine::pereg.perfZcompInputZcomplocHi = 0;
 		SWPixelEngine::pereg.perfZcompOutputZcomplocLo = 0;

Source/Plugins/Plugin_VideoSoftware/Src/Rasterizer.cpp

@@ -150,18 +150,15 @@ inline void Draw(s32 x, s32 y, s32 xi, s32 yi)
 
 	if (bpmem.zcontrol.early_ztest && bpmem.zmode.testenable && g_SWVideoConfig.bZComploc)
 	{
-		// TODO: Verify that perf regs are being incremented even if test is disabled
-		if (++SWPixelEngine::pereg.perfZcompInputZcomplocLo == 0)
-			SWPixelEngine::pereg.perfZcompInputZcomplocHi++;
-
+		// TODO: Test if perf regs are incremented even if test is disabled
+		SWPixelEngine::pereg.IncZInputQuadCount(true);
 		if (bpmem.zmode.testenable)
 		{
 			// early z
 			if (!EfbInterface::ZCompare(x, y, z))
 				return;
 		}
-		if (++SWPixelEngine::pereg.perfZcompOutputZcomplocLo == 0)
-			SWPixelEngine::pereg.perfZcompOutputZcomplocHi++;
+		SWPixelEngine::pereg.IncZOutputQuadCount(true);
 	}
 
 	RasterBlockPixel& pixel = rasterBlock.Pixel[xi][yi];

Source/Plugins/Plugin_VideoSoftware/Src/SWPixelEngine.cpp

@@ -81,17 +81,6 @@ void Read16(u16& _uReturnValue, const u32 _iAddress)
 
     if (address <= 0x2e)
         _uReturnValue = ((u16*)&pereg)[address >> 1];
-
-	if (address > 0x16)
-	{
-		ERROR_LOG(PIXELENGINE, "addr %#08x, ret %#04x; %#04x%04x, %#04x%04x, %#04x%04x, %#04x%04x, %#04x%04x, %#04x%04x\n", address, _uReturnValue,
-					pereg.perfZcompInputZcomplocHi, pereg.perfZcompInputZcomplocLo,
-					pereg.perfZcompOutputZcomplocHi, pereg.perfZcompOutputZcomplocLo,
-					pereg.perfZcompInputHi, pereg.perfZcompInputLo,
-					pereg.perfZcompOutputHi, pereg.perfZcompOutputLo,
-					pereg.perfBlendInputHi, pereg.perfBlendInputLo,
-					pereg.perfEfbCopyClocksHi, pereg.perfEfbCopyClocksLo);
-	}
 }
 
 void Write32(const u32 _iValue, const u32 _iAddress)

Source/Plugins/Plugin_VideoSoftware/Src/SWPixelEngine.h

@@ -158,6 +158,54 @@ namespace SWPixelEngine
 		u16 perfBlendInputHi;
 		u16 perfEfbCopyClocksLo;
 		u16 perfEfbCopyClocksHi;
+
+		// NOTE: hardware doesn't process individual pixels but quads instead. Current software renderer architecture works on pixels though, so we have this "quad" hack here to only increment the registers on every fourth rendered pixel
+		void IncZInputQuadCount(bool early_ztest)
+		{
+			static int quad = 0;
+			if (++quad != 3)
+				return;
+			quad = 0;
+
+			if (early_ztest)
+			{
+				if (++perfZcompInputZcomplocLo == 0)
+					perfZcompInputZcomplocHi++;
+			}
+			else
+			{
+				if (++perfZcompInputLo == 0)
+					perfZcompInputHi++;
+			}
+		}
+		void IncZOutputQuadCount(bool early_ztest)
+		{
+			static int quad = 0;
+			if (++quad != 3)
+				return;
+			quad = 0;
+
+			if (early_ztest)
+			{
+				if (++perfZcompOutputZcomplocLo == 0)
+					perfZcompOutputZcomplocHi++;
+			}
+			else
+			{
+				if (++perfZcompOutputLo == 0)
+					perfZcompOutputHi++;
+			}
+		}
+		void IncBlendInputQuadCount()
+		{
+			static int quad = 0;
+			if (++quad != 3)
+				return;
+			quad = 0;
+
+			if (++perfBlendInputLo == 0)
+				perfBlendInputHi++;
+		}
     };
 
     extern PEReg pereg;

Source/Plugins/Plugin_VideoSoftware/Src/Tev.cpp

@@ -789,14 +789,12 @@ void Tev::Draw()
 	if (late_ztest && bpmem.zmode.testenable)
 	{
 		// TODO: Check against hw if these values get incremented even if depth testing is disabled
-        if (++SWPixelEngine::pereg.perfZcompInputLo == 0)
-			SWPixelEngine::pereg.perfZcompInputHi++;
+        SWPixelEngine::pereg.IncZInputQuadCount(false);
 
-		if (!EfbInterface::ZCompare(Position[0], Position[1], Position[2]))
-			return;
+        if (!EfbInterface::ZCompare(Position[0], Position[1], Position[2]))
+            return;
 
-        if (++SWPixelEngine::pereg.perfZcompOutputLo == 0)
-			SWPixelEngine::pereg.perfZcompOutputHi++;
+        SWPixelEngine::pereg.IncZOutputQuadCount(false);
 	}
 
 #if ALLOW_TEV_DUMPS
@@ -820,9 +818,7 @@ void Tev::Draw()
 #endif
 
     INCSTAT(swstats.thisFrame.tevPixelsOut);
-
-	if (++SWPixelEngine::pereg.perfBlendInputLo == 0)
-		SWPixelEngine::pereg.perfBlendInputHi++;
+	SWPixelEngine::pereg.IncBlendInputQuadCount();
 
     EfbInterface::BlendTev(Position[0], Position[1], output);
 }