Video_Software: Implement PE pixel metrics

Source/Plugins/Plugin_VideoSoftware/Src/BPMemLoader.cpp

@@ -90,6 +90,15 @@ void SWBPWritten(int address, int newvalue)
         SWPixelEngine::pereg.boxBottom = newvalue >> 10;
         SWPixelEngine::pereg.boxTop = newvalue & 0x3ff;
         break;
+	case BPMEM_CLEAR_PIXEL_PERF:
+		// TODO: Parameter?
+		SWPixelEngine::pereg.perfZcompInputZcomploc = 0;
+		SWPixelEngine::pereg.perfZcompOutputZcomploc = 0;
+		SWPixelEngine::pereg.perfZcompInput = 0;
+		SWPixelEngine::pereg.perfZcompOutput = 0;
+		SWPixelEngine::pereg.perfBlendInput = 0;
+		SWPixelEngine::pereg.perfEfbCopyClocks = 0;
+		break;
     case BPMEM_LOADTLUT0: // This one updates bpmem.tlutXferSrc, no need to do anything here.
 		break;
 	case BPMEM_LOADTLUT1: // Load a Texture Look Up Table

Source/Plugins/Plugin_VideoSoftware/Src/Rasterizer.cpp

@@ -23,6 +23,7 @@
 #include "BPMemLoader.h"
 #include "XFMemLoader.h"
 #include "Tev.h"
+#include "SWPixelEngine.h"
 #include "SWStatistics.h"
 #include "SWVideoConfig.h"
 
@@ -125,11 +126,17 @@ inline void Draw(s32 x, s32 y, s32 xi, s32 yi)
 	if (z < 0 || z > 0x00ffffff)
 		return;
 
-	if (bpmem.zcontrol.zcomploc && bpmem.zmode.testenable)
+	if (bpmem.zcontrol.zcomploc)
 	{
-		// early z
-		if (!EfbInterface::ZCompare(x, y, z))
-			return;
+		// TODO: Verify that perf regs are being incremented even if test is disabled
+		SWPixelEngine::pereg.perfZcompInputZcomploc++;
+		if (bpmem.zmode.testenable)
+		{
+			// early z
+			if (!EfbInterface::ZCompare(x, y, z))
+				return;
+		}
+		SWPixelEngine::pereg.perfZcompOutputZcomploc++;
 	}
 
 	RasterBlockPixel& pixel = rasterBlock.Pixel[xi][yi];

Source/Plugins/Plugin_VideoSoftware/Src/SWPixelEngine.h

@@ -38,6 +38,21 @@ namespace SWPixelEngine
 	    PE_BBOX_RIGHT	 = 0x012, // Flip Right
 	    PE_BBOX_TOP		 = 0x014, // Flip Top
 	    PE_BBOX_BOTTOM	 = 0x016, // Flip Bottom
+
+		// NOTE: Order not verified
+		// These indicate the number of quads that are being used as input/output for each particular stage
+		PE_PERF_ZCOMP_INPUT_ZCOMPLOC_L  = 0x18,
+		PE_PERF_ZCOMP_INPUT_ZCOMPLOC_H  = 0x1a,
+		PE_PERF_ZCOMP_OUTPUT_ZCOMPLOC_L = 0x1c,
+		PE_PERF_ZCOMP_OUTPUT_ZCOMPLOC_H = 0x1e,
+		PE_PERF_ZCOMP_INPUT_L           = 0x20,
+		PE_PERF_ZCOMP_INPUT_H           = 0x22,
+		PE_PERF_ZCOMP_OUTPUT_L          = 0x24,
+		PE_PERF_ZCOMP_OUTPUT_H          = 0x26,
+		PE_PERF_BLEND_INPUT_L           = 0x28,
+		PE_PERF_BLEND_INPUT_H           = 0x2a,
+		PE_PERF_EFB_COPY_CLOCKS_L       = 0x2c,
+		PE_PERF_EFB_COPY_CLOCKS_H       = 0x2e,
     };
 
     union UPEZConfReg
@@ -125,10 +140,18 @@ namespace SWPixelEngine
         UPECtrlReg ctrl;
         u16 unk0;
         u16 token;
+
         u16 boxLeft;
         u16 boxRight;
         u16 boxTop;
         u16 boxBottom;
+
+		u16 perfZcompInputZcomploc;
+		u16 perfZcompOutputZcomploc;
+		u16 perfZcompInput;
+		u16 perfZcompOutput;
+		u16 perfBlendInput;
+		u16 perfEfbCopyClocks;
     };
 
     extern PEReg pereg;

Source/Plugins/Plugin_VideoSoftware/Src/Tev.cpp

@@ -20,6 +20,7 @@
 #include "Tev.h"
 #include "EfbInterface.h"
 #include "TextureSampler.h"
+#include "SWPixelEngine.h"
 #include "SWStatistics.h"
 #include "SWVideoConfig.h"
 #include "DebugUtil.h"
@@ -784,11 +785,16 @@ void Tev::Draw()
 		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]))
-            return;
-    }
+    if (!bpmem.zcontrol.zcomploc)
+	{
+        SWPixelEngine::pereg.perfZcompInput++;
+		if (bpmem.zmode.testenable)
+	    {
+	        if (!EfbInterface::ZCompare(Position[0], Position[1], Position[2]))
+	            return;
+	    }
+        SWPixelEngine::pereg.perfZcompOutput++;
+	}
 
 #if ALLOW_TEV_DUMPS
 	if (g_SWVideoConfig.bDumpTevStages)
@@ -812,6 +818,8 @@ void Tev::Draw()
 
     INCSTAT(swstats.thisFrame.tevPixelsOut);
 
+	SWPixelEngine::pereg.perfBlendInput++;
+
     EfbInterface::BlendTev(Position[0], Position[1], output);
 }