gsdx/GPUDrawScanline.cpp

@@ -22,8 +22,9 @@
 #include "StdAfx.h"
 #include "GPUDrawScanline.h"
 
-GPUDrawScanline::GPUDrawScanline(GPUState* state)
+GPUDrawScanline::GPUDrawScanline(GPUState* state, int id)
 	: m_state(state)
+	, m_id(id)
 {
 	Init();
 }
@@ -34,7 +35,7 @@ GPUDrawScanline::~GPUDrawScanline()
 
 // IDrawScanline
 
-bool GPUDrawScanline::SetupDraw(const GSRasterizerData* data)
+bool GPUDrawScanline::BeginDraw(const GSRasterizerData* data)
 {
 	GPUDrawingEnvironment& env = m_state->m_env;
 

gsdx/GPUDrawScanline.h

@@ -101,18 +101,20 @@ class GPUDrawScanline : public GSAlignedClass<16>, public IDrawScanline
 
 protected:
 	GPUState* m_state;
-	int m_filter;
-	int m_dither;
+	int m_id;
 
 public:
-	GPUDrawScanline(GPUState* state);
+	GPUDrawScanline(GPUState* state, int id);
 	virtual ~GPUDrawScanline();
 
 	// IDrawScanline
 
-	bool SetupDraw(const GSRasterizerData* data);
+	bool BeginDraw(const GSRasterizerData* data);
+	void EndDraw(const GSRasterizerStats& stats) {}
 	void SetupPrim(GS_PRIM_CLASS primclass, const GSVertexSW* vertices, const GSVertexSW& dscan);
 	void DrawScanline(int top, int left, int right, const GSVertexSW& v);
 	void DrawSolidRect(const GSVector4i& r, const GSVertexSW& v);
 	DrawScanlinePtr GetDrawScanlinePtr();
+
+	void PrintStats() {}
 };

gsdx/GPURendererSW.h

@@ -164,14 +164,15 @@ protected:
 		default: __assume(0);
 		}
 
-		int prims = m_rl.Draw(&data);
+		m_rl.Draw(&data);
+
+		GSRasterizerStats stats;
+
+		m_rl.GetStats(stats);
 	
-		m_perfmon.Put(GSPerfMon::Prim, prims);
 		m_perfmon.Put(GSPerfMon::Draw, 1);
-
-		int pixels = m_rl.GetPixels();
-
-		m_perfmon.Put(GSPerfMon::Fillrate, pixels); 
+		m_perfmon.Put(GSPerfMon::Prim, stats.prims);
+		m_perfmon.Put(GSPerfMon::Fillrate, stats.pixels); 
 
 		// TODO
 

gsdx/GSBlock.h

@@ -96,7 +96,7 @@ public:
 
 		#else
 
-		const DWORD* d = &columnTable32[(i & 3) << 1][0];
+		const BYTE* d = &columnTable32[(i & 3) << 1][0];
 
 		for(int j = 0; j < 2; j++, d += 8, src += srcpitch)
 		{
@@ -138,7 +138,7 @@ public:
 
 		#else
 
-		const DWORD* d = &columnTable16[(i & 3) << 1][0];
+		const BYTE* d = &columnTable16[(i & 3) << 1][0];
 
 		for(int j = 0; j < 2; j++, d += 16, src += srcpitch)
 		{
@@ -182,7 +182,7 @@ public:
 
 		#else
 
-		const DWORD* d = &columnTable8[(i & 3) << 2][0];
+		const BYTE* d = &columnTable8[(i & 3) << 2][0];
 
 		for(int j = 0; j < 4; j++, d += 16, src += srcpitch)
 		{
@@ -229,7 +229,7 @@ public:
 
 		#else
 
-		const DWORD* d = &columnTable4[(i & 3) << 2][0];
+		const WORD* d = &columnTable4[(i & 3) << 2][0];
 
 		for(int j = 0; j < 4; j++, d += 32, src += srcpitch)
 		{
@@ -307,7 +307,7 @@ public:
 
 		#else
 
-		const DWORD* d = &columnTable32[0][0];
+		const BYTE* d = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, d += 8, src += srcpitch)
 		{
@@ -341,7 +341,7 @@ public:
 
 		#else
 
-		const DWORD* d = &columnTable16[0][0];
+		const BYTE* d = &columnTable16[0][0];
 
 		for(int j = 0; j < 8; j++, d += 16, src += srcpitch)
 		{
@@ -368,7 +368,7 @@ public:
 
 		#else
 
-		const DWORD* d = &columnTable8[0][0];
+		const BYTE* d = &columnTable8[0][0];
 
 		for(int j = 0; j < 16; j++, d += 16, src += srcpitch)
 		{
@@ -395,7 +395,7 @@ public:
 
 		#else
 
-		const DWORD* d = &columnTable4[0][0];
+		const WORD* d = &columnTable4[0][0];
 
 		for(int j = 0; j < 16; j++, d += 32, src += srcpitch)
 		{
@@ -434,7 +434,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable32[(i & 3) << 1][0];
+		const BYTE* s = &columnTable32[(i & 3) << 1][0];
 
 		for(int j = 0; j < 2; j++, s += 8, dst += dstpitch)
 		{
@@ -492,7 +492,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable16[(i & 3) << 1][0];
+		const BYTE* s = &columnTable16[(i & 3) << 1][0];
 
 		for(int j = 0; j < 2; j++, s += 16, dst += dstpitch)
 		{
@@ -573,7 +573,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable8[(i & 3) << 2][0];
+		const BYTE* s = &columnTable8[(i & 3) << 2][0];
 
 		for(int j = 0; j < 4; j++, s += 16, dst += dstpitch)
 		{
@@ -660,7 +660,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable4[(i & 3) << 2][0];
+		const WORD* s = &columnTable4[(i & 3) << 2][0];
 
 		for(int j = 0; j < 4; j++, s += 32, dst += dstpitch)
 		{
@@ -738,7 +738,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable32[0][0];
+		const BYTE* s = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, s += 8, dst += dstpitch)
 		{
@@ -765,7 +765,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable16[0][0];
+		const BYTE* s = &columnTable16[0][0];
 
 		for(int j = 0; j < 8; j++, s += 16, dst += dstpitch)
 		{
@@ -792,7 +792,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable8[0][0];
+		const BYTE* s = &columnTable8[0][0];
 
 		for(int j = 0; j < 16; j++, s += 16, dst += dstpitch)
 		{
@@ -819,7 +819,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable4[0][0];
+		const WORD* s = &columnTable4[0][0];
 
 		for(int j = 0; j < 16; j++, s += 32, dst += dstpitch)
 		{
@@ -935,7 +935,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable32[0][0];
+		const BYTE* s = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, s += 8, dst += dstpitch)
 		{
@@ -980,7 +980,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable32[0][0];
+		const BYTE* s = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, s += 8, dst += dstpitch)
 		{
@@ -1023,7 +1023,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable32[0][0];
+		const BYTE* s = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, s += 8, dst += dstpitch)
 		{
@@ -1624,7 +1624,7 @@ public:
 
 		#else 
 
-		const DWORD* d = &columnTable32[0][0];
+		const BYTE* d = &columnTable32[0][0];
 
 		for(int j = 0, diff = srcpitch - 8 * 3; j < 8; j++, src += diff, d += 8)
 		{
@@ -1689,7 +1689,7 @@ public:
 
 		#else
 
-		const DWORD* d = &columnTable32[0][0];
+		const BYTE* d = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, src += srcpitch, dst += 8)
 		{
@@ -1816,7 +1816,7 @@ public:
 
 		#else
 
-		const DWORD* d = &columnTable32[0][0];
+		const BYTE* d = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, d += 8, src += srcpitch)
 		{
@@ -1944,7 +1944,7 @@ public:
 
 		#else
 
-		const DWORD* d = &columnTable32[0][0];
+		const BYTE* d = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, d += 8, src += srcpitch)
 		{
@@ -2004,7 +2004,7 @@ public:
 
 		DWORD TA0 = TEXA.TA0 << 24;
 
-		const DWORD* s = &columnTable32[0][0];
+		const BYTE* s = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, s += 8, dst += dstpitch)
 		{
@@ -2082,7 +2082,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable8[0][0];
+		const BYTE* s = &columnTable8[0][0];
 
 		for(int j = 0; j < 16; j++, s += 16, dst += dstpitch)
 		{
@@ -2169,14 +2169,14 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable4[0][0];
+		const WORD* s = &columnTable4[0][0];
 
 		for(int j = 0; j < 16; j++, s += 32, dst += dstpitch)
 		{
 			for(int i = 0; i < 16; i++)
 			{
-				DWORD a0 = s[i * 2 + 0];
-				DWORD a1 = s[i * 2 + 1];
+				BYTE a0 = s[i * 2 + 0];
+				BYTE a1 = s[i * 2 + 1];
 
 				BYTE c0 = (src[a0 >> 1] >> ((a0 & 1) << 2)) & 0x0f;
 				BYTE c1 = (src[a1 >> 1] >> ((a1 & 1) << 2)) & 0x0f;
@@ -2228,7 +2228,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable32[0][0];
+		const BYTE* s = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, s += 8, dst += dstpitch)
 		{
@@ -2281,7 +2281,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable32[0][0];
+		const BYTE* s = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, s += 8, dst += dstpitch)
 		{
@@ -2334,7 +2334,7 @@ public:
 
 		#else
 
-		const DWORD* s = &columnTable32[0][0];
+		const BYTE* s = &columnTable32[0][0];
 
 		for(int j = 0; j < 8; j++, s += 8, dst += dstpitch)
 		{

gsdx/GSDrawScanline.h

@@ -100,29 +100,29 @@ __declspec(align(16)) struct GSScanlineParam
 	const DWORD* clut;
 	DWORD tw;
 
+	GSLocalMemory::Offset* fbo;
+	GSLocalMemory::Offset* zbo;
+
 	DWORD fm, zm;
 };
 
 class GSDrawScanline : public GSAlignedClass<16>, public IDrawScanline
 {
-	GSScanlineEnvironment m_env;
-
-	struct Offset
+	struct ActiveDrawScanlinePtr
 	{
-		GSVector4i row[1024];
-		int* col[4];
-		DWORD hash;
+		UINT64 frame;
+		UINT64 frames;
+		__int64 ticks;
+		__int64 pixels;
+		DrawScanlinePtr dsf;
 	};
 
-	CRBMapC<DWORD, Offset*> m_omap;
-	Offset* m_fbo;
-	Offset* m_zbo;
+	GSScanlineEnvironment m_env;
 
-	void SetupOffset(Offset*& co, DWORD bp, DWORD bw, DWORD psm);
-	void FreeOffsets();
-
-	DrawScanlinePtr m_ds[4][4][4][2], m_dsf;
-	CRBMap<DWORD, DrawScanlinePtr> m_dsmap, m_dsmap2;
+	DrawScanlinePtr m_ds[4][4][4][2];
+	CRBMap<DWORD, DrawScanlinePtr> m_dsmap;
+	CRBMap<DWORD, ActiveDrawScanlinePtr*> m_dsmap_active;
+	ActiveDrawScanlinePtr* m_dsf;
 
 	void Init();
 
@@ -153,18 +153,31 @@ class GSDrawScanline : public GSAlignedClass<16>, public IDrawScanline
 	__forceinline GSVector4i ReadZBufX(int psm, const GSVector4i& addr) const;
 	__forceinline void WriteFrameAndZBufX(int fpsm, const GSVector4i& fa, const GSVector4i& fm, const GSVector4i& f, int zpsm, const GSVector4i& za, const GSVector4i& zm, const GSVector4i& z, int pixels);
 
+	template<class T, bool masked> 
+	void DrawSolidRectT(const GSVector4i* row, int* col, const GSVector4i& r, DWORD c, DWORD m);
+
+	template<class T, bool masked> 
+	__forceinline void FillRect(const GSVector4i* row, int* col, const GSVector4i& r, const GSVector4i& c, const GSVector4i& m);
+
+	template<class T, bool masked> 
+	__forceinline void FillBlock(const GSVector4i* row, int* col, const GSVector4i& r, const GSVector4i& c, const GSVector4i& m);
+
 protected:
 	GSState* m_state;
+	int m_id;
 
 public:
-	GSDrawScanline(GSState* state);
+	GSDrawScanline(GSState* state, int id);
 	virtual ~GSDrawScanline();
 
 	// IDrawScanline
 
-	bool SetupDraw(const GSRasterizerData* data);
+	bool BeginDraw(const GSRasterizerData* data);
+	void EndDraw(const GSRasterizerStats& stats);
 	void SetupPrim(GS_PRIM_CLASS primclass, const GSVertexSW* vertices, const GSVertexSW& dscan);
 	void DrawScanline(int top, int left, int right, const GSVertexSW& v);
 	void DrawSolidRect(const GSVector4i& r, const GSVertexSW& v);
 	DrawScanlinePtr GetDrawScanlinePtr();
+
+	void PrintStats();
 };

gsdx/GSLocalMemory.cpp

@@ -386,6 +386,67 @@ GSLocalMemory::GSLocalMemory()
 GSLocalMemory::~GSLocalMemory()
 {
 	VirtualFree(m_vm8, 0, MEM_RELEASE);
+
+	POSITION pos = m_omap.GetHeadPosition();
+
+	while(pos)
+	{
+		Offset* o = m_omap.GetNextValue(pos);
+
+		for(int i = 0; i < countof(o->col); i++)
+		{
+			_aligned_free(o->col);
+		}
+
+		_aligned_free(o);
+	}
+
+	m_omap.RemoveAll();
+}
+
+GSLocalMemory::Offset* GSLocalMemory::GetOffset(DWORD bp, DWORD bw, DWORD psm, Offset* o)
+{
+	if(bw == 0) {ASSERT(0); return NULL;}
+
+	ASSERT(m_psm[psm].trbpp > 8); // only for 16/24/32/8h/4hh/4hl formats where all columns are the same
+
+	DWORD hash = bp | (bw << 14) | (psm << 20);
+
+	if(!o || o->hash != hash)
+	{
+		CRBMap<DWORD, Offset*>::CPair* pair = m_omap.Lookup(hash);
+
+		if(pair)
+		{
+			o = pair->m_value;
+		}
+		else
+		{
+			o = (Offset*)_aligned_malloc(sizeof(Offset), 16);
+
+			o->hash = hash;
+
+			pixelAddress pa = m_psm[psm].pa;
+
+			for(int i = 0, j = 2048; i < j; i++)
+			{
+				o->row[i] = GSVector4i((int)pa(0, i, bp, bw));
+			}
+
+			int* p = (int*)_aligned_malloc(sizeof(int) * (2048 + 3) * 4, 16);
+
+			for(int i = 0; i < 4; i++)
+			{
+				o->col[i] = &p[2048 * i + ((4 - (i & 3)) & 3)];
+
+				memcpy(o->col[i], m_psm[psm].rowOffset[0], sizeof(int) * 2048);
+			}
+
+			m_omap.SetAt(hash, o);
+		}
+	}
+
+	return o;
 }
 
 bool GSLocalMemory::FillRect(const GSVector4i& r, DWORD c, DWORD psm, DWORD bp, DWORD bw)

gsdx/GSLocalMemory.h

@@ -75,6 +75,13 @@ public:
 
 	GSClut m_clut;
 
+	struct Offset
+	{
+		GSVector4i row[2048];
+		int* col[4];
+		DWORD hash;
+	};
+
 protected:
 	static DWORD pageOffset32[32][32][64];
 	static DWORD pageOffset32Z[32][32][64];
@@ -108,10 +115,16 @@ protected:
 
 	friend class GSClut;
 
+	//
+
+	CRBMapC<DWORD, Offset*> m_omap;
+
 public:
 	GSLocalMemory();
 	virtual ~GSLocalMemory();
 
+	Offset* GetOffset(DWORD bp, DWORD bw, DWORD psm, Offset* o = NULL);
+
 	// address
 
 	static DWORD PageNumber32(int x, int y, DWORD bp, DWORD bw)

gsdx/GSRasterizer.cpp

@@ -26,7 +26,6 @@ GSRasterizer::GSRasterizer(IDrawScanline* ds, int id, int threads)
 	: m_ds(ds)
 	, m_id(id)
 	, m_threads(threads)
-	, m_pixels(0)
 {
 }
 
@@ -35,42 +34,51 @@ GSRasterizer::~GSRasterizer()
 	delete m_ds;
 }
 
-int GSRasterizer::Draw(const GSRasterizerData* data)
+void GSRasterizer::Draw(const GSRasterizerData* data)
 {
-	bool solid = m_ds->SetupDraw(data);
+	bool solid = m_ds->BeginDraw(data);
 
 	const GSVector4i scissor = data->scissor;
 	const GSVertexSW* vertices = data->vertices;
 	const int count = data->count;
 
-	int prims = 0;
+	m_stats.Reset();
+
+	__int64 start = __rdtsc();
 
 	switch(data->primclass)
 	{
 	case GS_POINT_CLASS:
-		prims = count;
+		m_stats.prims = count;
 		for(int i = 0; i < count; i++) DrawPoint(&vertices[i], scissor);
 		break;
 	case GS_LINE_CLASS:
 		ASSERT(!(count & 1));
-		prims = count / 2;
+		m_stats.prims = count / 2;
 		for(int i = 0; i < count; i += 2) DrawLine(&vertices[i], scissor);
 		break;
 	case GS_TRIANGLE_CLASS:
 		ASSERT(!(count % 3));
-		prims = count / 3;
+		m_stats.prims = count / 3;
 		for(int i = 0; i < count; i += 3) DrawTriangle(&vertices[i], scissor);
 		break;
 	case GS_SPRITE_CLASS:
 		ASSERT(!(count & 1));
-		prims = count / 2;
+		m_stats.prims = count / 2;
 		for(int i = 0; i < count; i += 2) DrawSprite(&vertices[i], scissor, solid);
 		break;
 	default:
 		__assume(0);
 	}
 
-	return prims;
+	m_stats.ticks = __rdtsc() - start;
+
+	m_ds->EndDraw(m_stats);
+}
+
+void GSRasterizer::GetStats(GSRasterizerStats& stats)
+{
+	stats = m_stats;
 }
 
 void GSRasterizer::DrawPoint(const GSVertexSW* v, const GSVector4i& scissor)
@@ -85,7 +93,7 @@ void GSRasterizer::DrawPoint(const GSVertexSW* v, const GSVector4i& scissor)
 		{
 			m_ds->DrawScanline(p.y, p.x, p.x + 1, *v);
 
-			m_pixels++;
+			m_stats.pixels++;
 		}
 	}
 }
@@ -224,7 +232,7 @@ void GSRasterizer::DrawTriangleTop(GSVertexSW* v, const GSVector4i& scissor)
 
 	if(top < scissor.y) top = scissor.y;
 	if(bottom > scissor.w) bottom = scissor.w;
-			
+
 	if(top < bottom)
 	{
 		float py = (float)top - l.p.y;
@@ -316,7 +324,7 @@ void GSRasterizer::DrawTriangleTopBottom(GSVertexSW* v, const GSVector4i& scisso
 	GSVector4 dr;
 
 	bool b = (longest.p > GSVector4::zero()).mask() & 1;
-	
+
 	if(b)
 	{
 		dl = v01 / v01.p.yyyy();
@@ -428,7 +436,7 @@ void GSRasterizer::DrawTriangleSection(int top, int bottom, GSVertexSW& l, const
 
 				if(pixels > 0)
 				{
-					m_pixels += pixels;
+					m_stats.pixels += pixels;
 
 					GSVertexSW scan = l;
 
@@ -496,7 +504,7 @@ void GSRasterizer::DrawSprite(const GSVertexSW* vertices, const GSVector4i& scis
 		{
 			m_ds->DrawSolidRect(r, scan);
 
-			m_pixels += (r.z - r.x) * (r.w - r.y);
+			m_stats.pixels += (r.z - r.x) * (r.w - r.y);
 		}
 
 		return;
@@ -527,7 +535,7 @@ void GSRasterizer::DrawSprite(const GSVertexSW* vertices, const GSVector4i& scis
 		{
 			(m_ds->*dsf)(top, left, right, scan);
 
-			m_pixels += right - left;
+			m_stats.pixels += right - left;
 		}
 	}
 }
@@ -563,13 +571,11 @@ GSRasterizerMT::~GSRasterizerMT()
 	}
 }
 
-int GSRasterizerMT::Draw(const GSRasterizerData* data)
+void GSRasterizerMT::Draw(const GSRasterizerData* data)
 {
-	int prims = 0;
-
 	if(m_id == 0)
 	{
-		prims = __super::Draw(data);
+		__super::Draw(data);
 	}
 	else
 	{
@@ -577,8 +583,6 @@ int GSRasterizerMT::Draw(const GSRasterizerData* data)
 
 		InterlockedBitTestAndSet(m_sync, m_id);
 	}
-
-	return prims;
 }
 
 DWORD WINAPI GSRasterizerMT::StaticThreadProc(LPVOID lpParam)
@@ -631,17 +635,19 @@ void GSRasterizerList::FreeRasterizers()
 	}
 }
 
-int GSRasterizerList::Draw(const GSRasterizerData* data)
+void GSRasterizerList::Draw(const GSRasterizerData* data)
 {
 	*m_sync = 0;
 
-	int prims = 0;
+	m_stats.Reset();
+
+	__int64 start = __rdtsc();
 
 	POSITION pos = GetTailPosition();
 
 	while(pos)
 	{
-		prims += GetPrev(pos)->Draw(data);
+		GetPrev(pos)->Draw(data);
 	}
 
 	while(*m_sync)
@@ -649,19 +655,22 @@ int GSRasterizerList::Draw(const GSRasterizerData* data)
 		_mm_pause();
 	}
 
-	return prims;
-}
+	m_stats.ticks = __rdtsc() - start;
 
-int GSRasterizerList::GetPixels()
-{
-	int pixels = 0;
-	
-	POSITION pos = GetHeadPosition();
+	pos = GetHeadPosition();
 
 	while(pos)
 	{
-		pixels += GetNext(pos)->GetPixels();
-	}
+		GSRasterizerStats s;
 
-	return pixels;
+		GetNext(pos)->GetStats(s);
+
+		m_stats.pixels += s.pixels;
+		m_stats.prims = max(m_stats.prims, s.prims);
+	}
+}
+
+void GSRasterizerList::GetStats(GSRasterizerStats& stats)
+{
+	stats = m_stats;
 }

gsdx/GSRasterizer.h

@@ -24,6 +24,7 @@
 #include "GS.h"
 #include "GSVertexSW.h"
 
+// 
 #define FAST_DRAWSCANLINE
 
 __declspec(align(16)) class GSRasterizerData
@@ -36,11 +37,21 @@ public:
 	const void* param;
 };
 
+struct GSRasterizerStats
+{
+	__int64 ticks;
+	int prims;
+	int pixels;
+
+	GSRasterizerStats() {Reset();}
+	void Reset() {ticks = 0; pixels = prims = 0;}
+};
+
 class IRasterizer
 {
 public:
-	virtual int Draw(const GSRasterizerData* data) = 0;
-	virtual int GetPixels() = 0;
+	virtual void Draw(const GSRasterizerData* data) = 0;
+	virtual void GetStats(GSRasterizerStats& stats) = 0;
 };
 
 class IDrawScanline
@@ -50,11 +61,14 @@ public:
 
 	virtual ~IDrawScanline() {}
 
-	virtual bool SetupDraw(const GSRasterizerData* data) = 0;
+	virtual bool BeginDraw(const GSRasterizerData* data) = 0;
+	virtual void EndDraw(const GSRasterizerStats& stats) = 0;
 	virtual void SetupPrim(GS_PRIM_CLASS primclass, const GSVertexSW* vertices, const GSVertexSW& dscan) = 0;
 	virtual void DrawScanline(int top, int left, int right, const GSVertexSW& v) = 0;
 	virtual void DrawSolidRect(const GSVector4i& r, const GSVertexSW& v) = 0;
 	virtual DrawScanlinePtr GetDrawScanlinePtr() = 0;
+
+	virtual void PrintStats() = 0;
 };
 
 class GSRasterizer : public IRasterizer
@@ -63,7 +77,7 @@ protected:
 	IDrawScanline* m_ds;
 	int m_id;
 	int m_threads;
-	int m_pixels;
+	GSRasterizerStats m_stats;
 
 	void DrawPoint(const GSVertexSW* v, const GSVector4i& scissor);
 	void DrawLine(const GSVertexSW* v, const GSVector4i& scissor);
@@ -82,8 +96,10 @@ public:
 
 	// IRasterizer
 
-	int Draw(const GSRasterizerData* data);
-	int GetPixels() {int pixels = m_pixels; m_pixels = 0; return pixels;}
+	void Draw(const GSRasterizerData* data);
+	void GetStats(GSRasterizerStats& stats);
+
+	void PrintStats() {m_ds->PrintStats();}
 };
 
 class GSRasterizerMT : public GSRasterizer
@@ -104,12 +120,13 @@ public:
 
 	// IRasterizer
 
-	int Draw(const GSRasterizerData* data);
+	void Draw(const GSRasterizerData* data);
 };
 
 class GSRasterizerList : public CAtlList<GSRasterizerMT*>, public IRasterizer
 {
 	long* m_sync;
+	GSRasterizerStats m_stats;
 
 	void FreeRasterizers();
 
@@ -125,12 +142,14 @@ public:
 
 		for(int i = 0; i < threads; i++) 
 		{
-			AddTail(new GSRasterizerMT(new DS(parent), i, threads, m_sync));
+			AddTail(new GSRasterizerMT(new DS(parent, i), i, threads, m_sync));
 		}
 	}
 
 	// IRasterizer
 
-	int Draw(const GSRasterizerData* data);
-	int GetPixels();
+	void Draw(const GSRasterizerData* data);
+	void GetStats(GSRasterizerStats& stats);
+
+	void PrintStats() {GetHead()->PrintStats();}
 };

gsdx/GSRendererSW.h

@@ -35,6 +35,8 @@ protected:
 	GSTextureCacheSW* m_tc;
 	Texture m_texture[2];
 	bool m_reset;
+	GSLocalMemory::Offset* m_fbo;
+	GSLocalMemory::Offset* m_zbo;
 
 	__declspec(align(16)) struct VertexTrace
 	{
@@ -83,6 +85,8 @@ protected:
 
 			m_reset = false;
 		}
+
+		if((m_perfmon.GetFrame() & 255) == 0) m_rl.PrintStats();
 	}
 
 	void ResetDevice() 
@@ -273,6 +277,12 @@ protected:
 
 		p.vm = m_mem.m_vm32;
 
+		m_fbo = m_mem.GetOffset(context->FRAME.Block(), context->FRAME.FBW, context->FRAME.PSM, m_fbo);
+		m_zbo = m_mem.GetOffset(context->ZBUF.Block(), context->FRAME.FBW, context->ZBUF.PSM, m_zbo);
+
+		p.fbo = m_fbo;
+		p.zbo = m_zbo;
+
 		p.sel.dw = 0;
 
 		p.sel.fpsm = 3;
@@ -574,19 +584,19 @@ __int64 start = __rdtsc();
 		data.count = m_count;
 		data.param = &p;
 
-		int prims = m_rl.Draw(&data);
-		
-		m_perfmon.Put(GSPerfMon::Prim, prims);
+		m_rl.Draw(&data);
+
+		GSRasterizerStats stats;
+
+		m_rl.GetStats(stats);
+	
 		m_perfmon.Put(GSPerfMon::Draw, 1);
-		
-		int pixels = m_rl.GetPixels();
-
-		m_perfmon.Put(GSPerfMon::Fillrate, pixels);
-
+		m_perfmon.Put(GSPerfMon::Prim, stats.prims);
+		m_perfmon.Put(GSPerfMon::Fillrate, stats.pixels);
 /*
 __int64 diff = __rdtsc() - start;
 s_total += diff;
-//if(pixels > 50000)
+if(pixels >= 50000)
 fprintf(s_fp, "[%I64d, %d, %d, %d] %08x, diff = %I64d /prim = %I64d /pixel = %I64d \n", frame, PRIM->PRIM, prims, pixels, p.sel, diff, diff / prims, pixels > 0 ? diff / pixels : 0);
 */
 		// TODO
@@ -636,7 +646,7 @@ fprintf(s_fp, "[%I64d, %d, %d, %d] %08x, diff = %I64d /prim = %I64d /pixel = %I6
 		if(s_dump)
 		{
 			CString str;
-			str.Format(_T("c:\\temp1\\_%05d_f%I64d_rt1_%05x_%d_%d.bmp"), s_n++, m_perfmon.GetFrame(), m_context->FRAME.Block(), m_context->FRAME.PSM);
+			str.Format(_T("c:\\temp1\\_%05d_f%I64d_rt1_%05x_%d.bmp"), s_n++, m_perfmon.GetFrame(), m_context->FRAME.Block(), m_context->FRAME.PSM);
 			if(s_save) {m_mem.SaveBMP(str, m_context->FRAME.Block(), m_context->FRAME.FBW, m_context->FRAME.PSM, GetFrameSize(1).cx, 512);}//GetFrameSize(1).cy);
 			str.Format(_T("c:\\temp1\\_%05d_f%I64d_rz1_%05x_%d.bmp"), s_n-1, m_perfmon.GetFrame(), m_context->ZBUF.Block(), m_context->ZBUF.PSM);
 			if(s_savez) {m_mem.SaveBMP(str, m_context->ZBUF.Block(), m_context->FRAME.FBW, m_context->ZBUF.PSM, GetFrameSize(1).cx, 512);}
@@ -651,6 +661,8 @@ fprintf(s_fp, "[%I64d, %d, %d, %d] %08x, diff = %I64d /prim = %I64d /pixel = %I6
 public:
 	GSRendererSW(BYTE* base, bool mt, void (*irq)(), int nloophack, const GSRendererSettings& rs, int threads)
 		: GSRendererT(base, mt, irq, nloophack, rs)
+		, m_fbo(NULL)
+		, m_zbo(NULL)
 	{
 		m_rl.Create<GSDrawScanline>(this, threads);
 

gsdx/GSState.cpp

@@ -416,12 +416,7 @@ void GSState::GIFPackedRegHandlerSTQ(GIFPackedReg* r)
 
 void GSState::GIFPackedRegHandlerUV(GIFPackedReg* r)
 {
-	#if _M_SSE >= 0x401
-
-	GSVector4i v = GSVector4i::loadl(r);
-	m_v.UV.ai32[0] = (UINT32)GSVector4i::store(v.pu32(v)) & 0x3fff3fff;
-
-	#elif _M_SSE >= 0x200
+	#if _M_SSE >= 0x200
 
 	GSVector4i v = GSVector4i::loadl(r) & GSVector4i::x00003fff();
 	m_v.UV.ai32[0] = (UINT32)GSVector4i::store(v.ps32(v));

gsdx/GSTables.cpp

@@ -22,21 +22,21 @@
 #include "StdAfx.h"
 #include "GSTables.h"
 
-const DWORD blockTable32[4][8] = {
+const BYTE blockTable32[4][8] = {
 	{  0,  1,  4,  5, 16, 17, 20, 21},
 	{  2,  3,  6,  7, 18, 19, 22, 23},
 	{  8,  9, 12, 13, 24, 25, 28, 29},
 	{ 10, 11, 14, 15, 26, 27, 30, 31}
 };
 
-const DWORD blockTable32Z[4][8] = {
+const BYTE blockTable32Z[4][8] = {
 	{ 24, 25, 28, 29,  8,  9, 12, 13},
 	{ 26, 27, 30, 31, 10, 11, 14, 15},
 	{ 16, 17, 20, 21,  0,  1,  4,  5},
 	{ 18, 19, 22, 23,  2,  3,  6,  7}
 };
 
-const DWORD blockTable16[8][4] = {
+const BYTE blockTable16[8][4] = {
 	{  0,  2,  8, 10 },
 	{  1,  3,  9, 11 },
 	{  4,  6, 12, 14 },
@@ -47,7 +47,7 @@ const DWORD blockTable16[8][4] = {
 	{ 21, 23, 29, 31 }
 };
 
-const DWORD blockTable16S[8][4] = {
+const BYTE blockTable16S[8][4] = {
 	{  0,  2, 16, 18 },
 	{  1,  3, 17, 19 },
 	{  8, 10, 24, 26 },
@@ -58,7 +58,7 @@ const DWORD blockTable16S[8][4] = {
 	{ 13, 15, 29, 31 }
 };
 
-const DWORD blockTable16Z[8][4] = {
+const BYTE blockTable16Z[8][4] = {
 	{ 24, 26, 16, 18 },
 	{ 25, 27, 17, 19 },
 	{ 28, 30, 20, 22 },
@@ -69,7 +69,7 @@ const DWORD blockTable16Z[8][4] = {
 	{ 13, 15,  5,  7 }
 };
 
-const DWORD blockTable16SZ[8][4] = {
+const BYTE blockTable16SZ[8][4] = {
 	{ 24, 26,  8, 10 },
 	{ 25, 27,  9, 11 },
 	{ 16, 18,  0,  2 },
@@ -80,14 +80,14 @@ const DWORD blockTable16SZ[8][4] = {
 	{ 21, 23,  5,  7 }
 };
 
-const DWORD blockTable8[4][8] = {
+const BYTE blockTable8[4][8] = {
 	{  0,  1,  4,  5, 16, 17, 20, 21},
 	{  2,  3,  6,  7, 18, 19, 22, 23},
 	{  8,  9, 12, 13, 24, 25, 28, 29},
 	{ 10, 11, 14, 15, 26, 27, 30, 31}
 };
 
-const DWORD blockTable4[8][4] = {
+const BYTE blockTable4[8][4] = {
 	{  0,  2,  8, 10 },
 	{  1,  3,  9, 11 },
 	{  4,  6, 12, 14 },
@@ -98,7 +98,7 @@ const DWORD blockTable4[8][4] = {
 	{ 21, 23, 29, 31 }
 };
 
-const DWORD columnTable32[8][8] = {
+const BYTE columnTable32[8][8] = {
 	{  0,  1,  4,  5,  8,  9, 12, 13 },
 	{  2,  3,  6,  7, 10, 11, 14, 15 },
 	{ 16, 17, 20, 21, 24, 25, 28, 29 },
@@ -109,7 +109,7 @@ const DWORD columnTable32[8][8] = {
 	{ 50, 51, 54, 55, 58, 59, 62, 63 },
 };
 
-const DWORD columnTable16[8][16] = {
+const BYTE columnTable16[8][16] = {
 	{   0,   2,   8,  10,  16,  18,  24,  26, 
 	    1,   3,   9,  11,  17,  19,  25,  27 },
 	{   4,   6,  12,  14,  20,  22,  28,  30, 
@@ -128,7 +128,7 @@ const DWORD columnTable16[8][16] = {
 	  101, 103, 109, 111, 117, 119, 125, 127 },
 };
 
-const DWORD columnTable8[16][16] = {
+const BYTE columnTable8[16][16] = {
 	{   0,   4,  16,  20,  32,  36,  48,  52,	// column 0
 	    2,   6,  18,  22,  34,  38,  50,  54 },
 	{   8,  12,  24,  28,  40,  44,  56,  60,
@@ -163,7 +163,7 @@ const DWORD columnTable8[16][16] = {
 	  203, 207, 219, 223, 235, 239, 251, 255 },
 };
 
-const DWORD columnTable4[16][32] = {
+const WORD columnTable4[16][32] = {
 	{   0,   8,  32,  40,  64,  72,  96, 104,	// column 0
 	    2,  10,  34,  42,  66,  74,  98, 106,
 	    4,  12,  36,  44,  68,  76, 100, 108,
@@ -230,7 +230,7 @@ const DWORD columnTable4[16][32] = {
 	  407, 415, 439, 447, 471, 479, 503, 511 },
 };
 
-const DWORD clutTableT32I8[128] = 
+const BYTE clutTableT32I8[128] = 
 {
 	0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15, 
 	64, 65, 68, 69, 72, 73, 76, 77, 66, 67, 70, 71, 74, 75, 78, 79, 
@@ -242,13 +242,13 @@ const DWORD clutTableT32I8[128] =
 	112, 113, 116, 117, 120, 121, 124, 125, 114, 115, 118, 119, 122, 123, 126, 127
 };
 
-const DWORD clutTableT32I4[16] = 
+const BYTE clutTableT32I4[16] = 
 {
 	0, 1, 4, 5, 8, 9, 12, 13,
 	2, 3, 6, 7, 10, 11, 14, 15
 };
 
-const DWORD clutTableT16I8[32] = 
+const BYTE clutTableT16I8[32] = 
 {
 	0, 2, 8, 10, 16, 18, 24, 26,
 	4, 6, 12, 14, 20, 22, 28, 30,
@@ -256,7 +256,7 @@ const DWORD clutTableT16I8[32] =
 	5, 7, 13, 15, 21, 23, 29, 31
 };
 
-const DWORD clutTableT16I4[16] = 
+const BYTE clutTableT16I4[16] = 
 {
 	0, 2, 8, 10, 16, 18, 24, 26,
 	4, 6, 12, 14, 20, 22, 28, 30

gsdx/GSTables.h

@@ -21,19 +21,19 @@
 
 #pragma once
 
-extern const DWORD blockTable32[4][8];
-extern const DWORD blockTable32Z[4][8];
-extern const DWORD blockTable16[8][4];
-extern const DWORD blockTable16S[8][4];
-extern const DWORD blockTable16Z[8][4];
-extern const DWORD blockTable16SZ[8][4];
-extern const DWORD blockTable8[4][8];
-extern const DWORD blockTable4[8][4];
-extern const DWORD columnTable32[8][8];
-extern const DWORD columnTable16[8][16];
-extern const DWORD columnTable8[16][16];
-extern const DWORD columnTable4[16][32];
-extern const DWORD clutTableT32I8[128];
-extern const DWORD clutTableT32I4[16]; 
-extern const DWORD clutTableT16I8[32];
-extern const DWORD clutTableT16I4[16];
+extern const BYTE blockTable32[4][8];
+extern const BYTE blockTable32Z[4][8];
+extern const BYTE blockTable16[8][4];
+extern const BYTE blockTable16S[8][4];
+extern const BYTE blockTable16Z[8][4];
+extern const BYTE blockTable16SZ[8][4];
+extern const BYTE blockTable8[4][8];
+extern const BYTE blockTable4[8][4];
+extern const BYTE columnTable32[8][8];
+extern const BYTE columnTable16[8][16];
+extern const BYTE columnTable8[16][16];
+extern const WORD columnTable4[16][32];
+extern const BYTE clutTableT32I8[128];
+extern const BYTE clutTableT32I4[16]; 
+extern const BYTE clutTableT16I8[32];
+extern const BYTE clutTableT16I4[16];