GSdx: it's hard to keep track of the leftover vertices properly, a bit of sps was still possible, psx sprites were fixed too

git-svn-id: http://pcsx2.googlecode.com/svn/trunk@5065 96395faa-99c1-11dd-bbfe-3dabce05a288

plugins/GSdx/GPUDrawScanline.cpp

@@ -83,7 +83,7 @@ void GPUDrawScanline::EndDraw(uint64 frame, uint64 ticks, int pixels)
 
 #ifndef ENABLE_JIT_RASTERIZER
 
-void GPUDrawScanline::SetupPrim(const GSVertexSW& vertex, const GSVertexSW& dscan)
+void GPUDrawScanline::SetupPrim(const GSVertexSW* vertex, const uint32* index, const GSVertexSW& dscan)
 {
 	GPUScanlineSelector sel = m_global.sel;
 
@@ -93,7 +93,7 @@ void GPUDrawScanline::SetupPrim(const GSVertexSW& vertex, const GSVertexSW& dsca
 	{
 		if(sel.sprite)
 		{
-			GSVector4i t = (GSVector4i(vertex.t) >> 8) - GSVector4i::x00000001();
+			GSVector4i t = (GSVector4i(vertex[index[1]].t) >> 8) - GSVector4i::x00000001();
 
 			t = t.ps32(t);
 			t = t.upl16(t);

plugins/GSdx/GPUDrawScanline.h

@@ -65,7 +65,7 @@ public:
 
 #ifndef ENABLE_JIT_RASTERIZER
 
-	void SetupPrim(const GSVertexSW& vertex, const GSVertexSW& dscan);
+	void SetupPrim(const GSVertexSW* vertex, const uint32* index, const GSVertexSW& dscan);
 	void DrawScanline(int pixels, int left, int top, const GSVertexSW& scan);
 	void DrawEdge(int pixels, int left, int top, const GSVertexSW& scan);
 	void DrawRect(const GSVector4i& r, const GSVertexSW& v);

plugins/GSdx/GPUSetupPrimCodeGenerator.cpp

@@ -27,6 +27,11 @@
 
 using namespace Xbyak;
 
+static const int _args = 0;
+static const int _vertex = _args + 4;
+static const int _index = _args + 8;
+static const int _dscan = _args + 12;
+
 GPUSetupPrimCodeGenerator::GPUSetupPrimCodeGenerator(void* param, uint32 key, void* code, size_t maxsize)
 	: GSCodeGenerator(code, maxsize)
 	, m_local(*(GPUScanlineLocalData*)param)
@@ -50,7 +55,12 @@ void GPUSetupPrimCodeGenerator::Generate()
 		{
 			// t = (GSVector4i(vertices[1].t) >> 8) - GSVector4i::x00000001();
 
-			cvttps2dq(xmm1, ptr[ecx + sizeof(GSVertexSW) * 1 + offsetof(GSVertexSW, t)]);
+			mov(ecx, ptr[esp + _index]);
+			mov(ecx, ptr[ecx + sizeof(uint32) * 1]);
+			shl(ecx, 6); // * sizeof(GSVertexSW)
+			add(ecx, ptr[esp + _vertex]);
+
+			cvttps2dq(xmm1, ptr[ecx + offsetof(GSVertexSW, t)]);
 			psrld(xmm1, 8);
 			psrld(xmm0, 31);
 			psubd(xmm1, xmm0);
@@ -86,6 +96,8 @@ void GPUSetupPrimCodeGenerator::Generate()
 
 	if(m_sel.tme || m_sel.iip && m_sel.tfx != 3)
 	{
+		mov(edx, dword[esp + _dscan]);
+
 		for(int i = 0; i < 3; i++)
 		{
 			movaps(Xmm(5 + i), ptr[&m_shift[i]]);

plugins/GSdx/GSDrawScanline.cpp

@@ -104,7 +104,7 @@ void GSDrawScanline::EndDraw(uint64 frame, uint64 ticks, int pixels)
 
 // FIXME: something's not right with the sky in burnout 3
 
-void GSDrawScanline::SetupPrim(const GSVertexSW& vertex, const GSVertexSW& dscan)
+void GSDrawScanline::SetupPrim(const GSVertexSW* vertex, const uint32* index, const GSVertexSW& dscan)
 {
 	GSScanlineSelector sel = m_global.sel;
 
@@ -147,12 +147,12 @@ void GSDrawScanline::SetupPrim(const GSVertexSW& vertex, const GSVertexSW& dscan
 		{
 			if(has_f)
 			{
-				m_local.p.f = GSVector4i(vertex.p).zzzzh().zzzz();
+				m_local.p.f = GSVector4i(vertex[index[1]].p).zzzzh().zzzz();
 			}
 
 			if(has_z)
 			{
-				m_local.p.z = vertex.t.u32[3]; // uint32 z is bypassed in t.w
+				m_local.p.z = vertex[index[1]].t.u32[3]; // uint32 z is bypassed in t.w
 			}
 		}
 	}
@@ -236,7 +236,17 @@ void GSDrawScanline::SetupPrim(const GSVertexSW& vertex, const GSVertexSW& dscan
 		}
 		else
 		{
-			GSVector4i c = GSVector4i(vertex.c);
+			int last = 0;
+			
+			switch(sel.prim)
+			{
+			case GS_POINT_CLASS: last = 0; break;
+			case GS_LINE_CLASS: last = 1; break;
+			case GS_TRIANGLE_CLASS: last = 2; break;
+			case GS_SPRITE_CLASS: last = 1; break;
+			}
+
+			GSVector4i c = GSVector4i(vertex[index[last]].c);
 
 			c = c.upl16(c.zwxy());
 

plugins/GSdx/GSDrawScanline.h

@@ -65,7 +65,7 @@ public:
 
 #ifndef ENABLE_JIT_RASTERIZER
 	
-	void SetupPrim(const GSVertexSW& vertex, const GSVertexSW& dscan);
+	void SetupPrim(const GSVertexSW* vertex, const uint32* index, const GSVertexSW& dscan);
 	void DrawScanline(int pixels, int left, int top, const GSVertexSW& scan);
 	void DrawEdge(int pixels, int left, int top, const GSVertexSW& scan);
 

plugins/GSdx/GSDrawingContext.h

@@ -41,12 +41,11 @@ public:
 	GIFRegFRAME		FRAME;
 	GIFRegZBUF		ZBUF;
 
-	__aligned(struct, 32)
+	struct
 	{
-		GSVector4i dx10;
-		GSVector4 dx9;
 		GSVector4 in;
-		GSVector4 ex;
+		GSVector4 ofex;
+		uint32 ofxy;
 	} scissor;
 
 	struct
@@ -83,25 +82,22 @@ public:
 
 	void UpdateScissor()
 	{
-		scissor.dx10 = GSVector4i(
+		scissor.ofex = GSVector4(
 			(int)((SCISSOR.SCAX0 << 4) + XYOFFSET.OFX),
 			(int)((SCISSOR.SCAY0 << 4) + XYOFFSET.OFY),
 			(int)((SCISSOR.SCAX1 << 4) + XYOFFSET.OFX),
 			(int)((SCISSOR.SCAY1 << 4) + XYOFFSET.OFY));
 
-		scissor.dx9 = GSVector4(scissor.dx10);
-
 		scissor.in = GSVector4(
 			(int)SCISSOR.SCAX0,
 			(int)SCISSOR.SCAY0,
 			(int)SCISSOR.SCAX1 + 1,
 			(int)SCISSOR.SCAY1 + 1);
 
-		scissor.ex = GSVector4(
+		uint16 ofx = (uint16)XYOFFSET.OFX - 15;
-			(int)SCISSOR.SCAX0,
+		uint16 ofy = (uint16)XYOFFSET.OFY - 15;
-			(int)SCISSOR.SCAY0,
+
-			(int)SCISSOR.SCAX1,
+		scissor.ofxy = ((ofy << 16) | ofx); // ceil(xy) => (xy - offset + 15) >> 4 => (xy - [offset - 15]) >> 4
-			(int)SCISSOR.SCAY1);
 	}
 
 	bool DepthRead() const

plugins/GSdx/GSRasterizer.cpp

@@ -210,21 +210,47 @@ void GSRasterizer::Draw(GSRasterizerData* data)
 template<bool scissor_test>
 void GSRasterizer::DrawPoint(const GSVertexSW* vertex, int vertex_count, const uint32* index, int index_count)
 {
-	for(int i = 0, count = index != NULL ? index_count : vertex_count; i < count; i++)
+	if(index != NULL)
 	{
-		const GSVertexSW& v = vertex[index != NULL ? index[i] : i];
+		for(int i = 0; i < index_count; i++, index++)
-
-		GSVector4i p(v.p);
-
-		if(!scissor_test || m_scissor.left <= p.x && p.x < m_scissor.right && m_scissor.top <= p.y && p.y < m_scissor.bottom)
 		{
-			if(IsOneOfMyScanlines(p.y))
+			const GSVertexSW& v = vertex[*index];
+
+			GSVector4i p(v.p);
+
+			if(!scissor_test || m_scissor.left <= p.x && p.x < m_scissor.right && m_scissor.top <= p.y && p.y < m_scissor.bottom)
 			{
-				m_pixels++;
+				if(IsOneOfMyScanlines(p.y))
+				{
+					m_pixels++;
 
-				m_ds->SetupPrim(v, v);
+					m_ds->SetupPrim(vertex, index, GSVertexSW::zero());
 
-				m_ds->DrawScanline(1, p.x, p.y, v);
+					m_ds->DrawScanline(1, p.x, p.y, v);
+				}
+			}
+		}
+	}
+	else
+	{
+		uint32 tmp_index[1] = {0};
+
+		for(int i = 0; i < vertex_count; i++, vertex++)
+		{
+			const GSVertexSW& v = vertex[0];
+
+			GSVector4i p(v.p);
+
+			if(!scissor_test || m_scissor.left <= p.x && p.x < m_scissor.right && m_scissor.top <= p.y && p.y < m_scissor.bottom)
+			{
+				if(IsOneOfMyScanlines(p.y))
+				{
+					m_pixels++;
+
+					m_ds->SetupPrim(vertex, tmp_index, GSVertexSW::zero());
+
+					m_ds->DrawScanline(1, p.x, p.y, v);
+				}
 			}
 		}
 	}
@@ -246,7 +272,7 @@ void GSRasterizer::DrawLine(const GSVertexSW* vertex, const uint32* index)
 		DrawEdge(v0, v1, dv, i, 0);
 		DrawEdge(v0, v1, dv, i, 1);
 
-		Flush(v1, GSVertexSW::zero(), true);
+		Flush(vertex, index, GSVertexSW::zero(), true);
 
 		return;
 	}
@@ -289,7 +315,7 @@ void GSRasterizer::DrawLine(const GSVertexSW* vertex, const uint32* index)
 
 					scan += dscan * (l - scan.p).xxxx();
 
-					m_ds->SetupPrim(v1, dscan);
+					m_ds->SetupPrim(vertex, index, dscan);
 
 					m_ds->DrawScanline(pixels, left, p.y, scan);
 				}
@@ -329,7 +355,7 @@ void GSRasterizer::DrawLine(const GSVertexSW* vertex, const uint32* index)
 
 		m_edge.count = e - m_edge.buff;
 
-		Flush(v1, GSVertexSW::zero());
+		Flush(vertex, index, GSVertexSW::zero());
 	}
 }
 
@@ -472,7 +498,7 @@ void GSRasterizer::DrawTriangle(const GSVertexSW* vertex, const uint32* index)
 		}
 	}
 
-	Flush(vertex[index[2]], dscan);
+	Flush(vertex, index, dscan);
 
 	if(m_ds->HasEdge())
 	{
@@ -487,7 +513,7 @@ void GSRasterizer::DrawTriangle(const GSVertexSW* vertex, const uint32* index)
 		DrawEdge(v0, v2, dv[1], orientation & 2, side & 2);
 		DrawEdge(v1, v2, dv[2], orientation & 4, side & 4);
 
-		Flush(vertex[index[2]], GSVertexSW::zero(), true);
+		Flush(vertex, index, GSVertexSW::zero(), true);
 	}
 }
 
@@ -609,7 +635,7 @@ void GSRasterizer::DrawSprite(const GSVertexSW* vertex, const uint32* index, boo
 	if((m & 2) == 0) scan.t += dedge.t * prestep.yyyy();
 	if((m & 1) == 0) scan.t += dscan.t * prestep.xxxx();
 
-	m_ds->SetupPrim(v1, dscan);
+	m_ds->SetupPrim(vertex, index, dscan);
 
 	while(1)
 	{
@@ -826,7 +852,7 @@ void GSRasterizer::AddScanline(GSVertexSW* e, int pixels, int left, int top, con
 	e->p.i16[2] = (int16)top;
 }
 
-void GSRasterizer::Flush(const GSVertexSW& vertex, const GSVertexSW& dscan, bool edge)
+void GSRasterizer::Flush(const GSVertexSW* vertex, const uint32* index, const GSVertexSW& dscan, bool edge)
 {
 	// TODO: on win64 this could be the place where xmm6-15 are preserved (not by each DrawScanline)
 
@@ -834,7 +860,7 @@ void GSRasterizer::Flush(const GSVertexSW& vertex, const GSVertexSW& dscan, bool
 
 	if(count > 0)
 	{
-		m_ds->SetupPrim(vertex, dscan);
+		m_ds->SetupPrim(vertex, index, dscan);
 
 		const GSVertexSW* RESTRICT e = m_edge.buff;
 		const GSVertexSW* RESTRICT ee = e + count;

plugins/GSdx/GSRasterizer.h

@@ -67,7 +67,7 @@ public:
 class IDrawScanline : public GSAlignedClass<32>
 {
 public:
-	typedef void (__fastcall *SetupPrimPtr)(const GSVertexSW& vertex, const GSVertexSW& dscan);
+	typedef void (*SetupPrimPtr)(const GSVertexSW* vertex, const uint32* index, const GSVertexSW& dscan);
 	typedef void (__fastcall *DrawScanlinePtr)(int pixels, int left, int top, const GSVertexSW& scan);
 	typedef void (IDrawScanline::*DrawRectPtr)(const GSVector4i& r, const GSVertexSW& v); // TODO: jit
 
@@ -86,14 +86,14 @@ public:
 
 #ifdef ENABLE_JIT_RASTERIZER
 
-	__forceinline void SetupPrim(const GSVertexSW& vertex, const GSVertexSW& dscan) {m_sp(vertex, dscan);}
+	__forceinline void SetupPrim(const GSVertexSW* vertex, const uint32* index, const GSVertexSW& dscan) {m_sp(vertex, index, dscan);}
 	__forceinline void DrawScanline(int pixels, int left, int top, const GSVertexSW& scan) {m_ds(pixels, left, top, scan);}
 	__forceinline void DrawEdge(int pixels, int left, int top, const GSVertexSW& scan) {m_de(pixels, left, top, scan);}
 	__forceinline void DrawRect(const GSVector4i& r, const GSVertexSW& v) {(this->*m_dr)(r, v);}
 
 #else
 
-	virtual void SetupPrim(const GSVertexSW& vertex, const GSVertexSW& dscan) = 0;
+	virtual void SetupPrim(const GSVertexSW* vertex, const uint32* index, const GSVertexSW& dscan) = 0;
 	virtual void DrawScanline(int pixels, int left, int top, const GSVertexSW& scan) = 0;
 	virtual void DrawEdge(int pixels, int left, int top, const GSVertexSW& scan) = 0;
 	virtual void DrawRect(const GSVector4i& r, const GSVertexSW& v) = 0;
@@ -140,7 +140,7 @@ protected:
 	void DrawEdge(const GSVertexSW& v0, const GSVertexSW& v1, const GSVertexSW& dv, int orientation, int side);
 
 	__forceinline void AddScanline(GSVertexSW* e, int pixels, int left, int top, const GSVertexSW& scan);
-	__forceinline void Flush(const GSVertexSW& vertex, const GSVertexSW& dscan, bool edge = false);
+	__forceinline void Flush(const GSVertexSW* vertex, const uint32* index, const GSVertexSW& dscan, bool edge = false);
 
 public:
 	GSRasterizer(IDrawScanline* ds, int id, int threads, GSPerfMon* perfmon);

plugins/GSdx/GSRendererSW.cpp

@@ -206,8 +206,6 @@ void GSRendererSW::Draw()
 
 	if(!GetScanlineGlobalData(sd)) return;
 
-	//
-
 	data->primclass = m_vt->m_primclass;
 	data->buff = (uint8*)_aligned_malloc(sizeof(GSVertexSW) * m_vertex.next + sizeof(uint32) * m_index.tail, 32);
 	data->vertex = (GSVertexSW*)data->buff;
@@ -218,6 +216,11 @@ void GSRendererSW::Draw()
 	memcpy(data->vertex, m_vertex.buff, sizeof(GSVertexSW) * m_vertex.next);
 	memcpy(data->index, m_index.buff, sizeof(uint32) * m_index.tail);
 
+	for(size_t i = 0; i < m_index.tail; i++)
+	{
+		ASSERT(((GSVertexSW*)m_vertex.buff + m_index.buff[i])->_pad.u32[0] == 0x12345678);
+	}
+
 	//
 
 	const GSDrawingContext* context = m_context;

plugins/GSdx/GSSetupPrimCodeGenerator.x86.avx.cpp

@@ -27,10 +27,17 @@
 
 using namespace Xbyak;
 
+static const int _args = 0;
+static const int _vertex = _args + 4;
+static const int _index = _args + 8;
+static const int _dscan = _args + 12;
+
 void GSSetupPrimCodeGenerator::Generate()
 {
 	if((m_en.z || m_en.f) && m_sel.prim != GS_SPRITE_CLASS || m_en.t || m_en.c && m_sel.iip)
 	{
+		mov(edx, dword[esp + _dscan]);
+
 		for(int i = 0; i < 5; i++)
 		{
 			vmovaps(Xmm(3 + i), ptr[&m_shift[i]]);
@@ -107,7 +114,12 @@ void GSSetupPrimCodeGenerator::Depth()
 	}
 	else
 	{
-		// GSVector4 p = vertex.p;
+		// GSVector4 p = vertex[index[1]].p;
+
+		mov(ecx, ptr[esp + _index]);
+		mov(ecx, ptr[ecx + sizeof(uint32) * 1]);
+		shl(ecx, 6); // * sizeof(GSVertexSW)
+		add(ecx, ptr[esp + _vertex]);
 
 		vmovaps(xmm0, ptr[ecx + offsetof(GSVertexSW, p)]);
 
@@ -312,7 +324,25 @@ void GSSetupPrimCodeGenerator::Color()
 	}
 	else
 	{
-		// GSVector4i c = GSVector4i(vertex.c);
+		// GSVector4i c = GSVector4i(vertex[index[last].c);
+
+		int last = 0;
+
+		switch(m_sel.prim)
+		{
+		case GS_POINT_CLASS: last = 0; break;
+		case GS_LINE_CLASS: last = 1; break;
+		case GS_TRIANGLE_CLASS: last = 2; break;
+		case GS_SPRITE_CLASS: last = 1; break;
+		}
+
+		if(!(m_sel.prim == GS_SPRITE_CLASS && (m_en.z || m_en.f))) // if this is a sprite, the last vertex was already loaded in Depth()
+		{
+			mov(ecx, ptr[esp + _index]);
+			mov(ecx, ptr[ecx + sizeof(uint32) * last]);
+			shl(ecx, 6); // * sizeof(GSVertexSW)
+			add(ecx, ptr[esp + _vertex]);
+		}
 
 		vcvttps2dq(xmm0, ptr[ecx + offsetof(GSVertexSW, c)]);
 

plugins/GSdx/GSSetupPrimCodeGenerator.x86.cpp

@@ -27,10 +27,17 @@
 
 using namespace Xbyak;
 
+static const int _args = 0;
+static const int _vertex = _args + 4;
+static const int _index = _args + 8;
+static const int _dscan = _args + 12;
+
 void GSSetupPrimCodeGenerator::Generate()
 {
 	if((m_en.z || m_en.f) && m_sel.prim != GS_SPRITE_CLASS || m_en.t || m_en.c && m_sel.iip)
 	{
+		mov(edx, dword[esp + _dscan]);
+
 		for(int i = 0; i < 5; i++)
 		{
 			movaps(Xmm(3 + i), ptr[&m_shift[i]]);
@@ -112,7 +119,12 @@ void GSSetupPrimCodeGenerator::Depth()
 	}
 	else
 	{
-		// GSVector4 p = vertex.p;
+		// GSVector4 p = vertex[index[1]].p;
+
+		mov(ecx, ptr[esp + _index]);
+		mov(ecx, ptr[ecx + sizeof(uint32) * 1]);
+		shl(ecx, 6); // * sizeof(GSVertexSW)
+		add(ecx, ptr[esp + _vertex]);
 
 		movaps(xmm0, ptr[ecx + offsetof(GSVertexSW, p)]);
 
@@ -327,7 +339,25 @@ void GSSetupPrimCodeGenerator::Color()
 	}
 	else
 	{
-		// GSVector4i c = GSVector4i(vertex.c);
+		// GSVector4i c = GSVector4i(vertex[index[last].c);
+
+		int last = 0;
+
+		switch(m_sel.prim)
+		{
+		case GS_POINT_CLASS: last = 0; break;
+		case GS_LINE_CLASS: last = 1; break;
+		case GS_TRIANGLE_CLASS: last = 2; break;
+		case GS_SPRITE_CLASS: last = 1; break;
+		}
+
+		if(!(m_sel.prim == GS_SPRITE_CLASS && (m_en.z || m_en.f))) // if this is a sprite, the last vertex was already loaded in Depth()
+		{
+			mov(ecx, ptr[esp + _index]);
+			mov(ecx, ptr[ecx + sizeof(uint32) * last]);
+			shl(ecx, 6); // * sizeof(GSVertexSW)
+			add(ecx, ptr[esp + _vertex]);
+		}
 
 		cvttps2dq(xmm0, ptr[ecx + offsetof(GSVertexSW, c)]);
 

plugins/GSdx/GSState.cpp

@@ -499,13 +499,6 @@ void GSState::GIFPackedRegHandlerUV(const GIFPackedReg* RESTRICT r)
 template<uint32 prim, uint32 adc>
 void GSState::GIFPackedRegHandlerXYZF2(const GIFPackedReg* RESTRICT r)
 {
-	if(adc) 
-	{
-		// not sure what the difference is between this and XYZF2 with ADC bit set
-
-		//printf("XYZF3 X %d Y %d Z %d F %d ADC %d\n", r->XYZF2.X, r->XYZF2.Y, r->XYZF2.Z, r->XYZF2.F, r->XYZF2.ADC); 
-	}
-
 	/*
 	m_v.XYZ.X = r->XYZF2.X;
 	m_v.XYZ.Y = r->XYZF2.Y;
@@ -525,12 +518,6 @@ void GSState::GIFPackedRegHandlerXYZF2(const GIFPackedReg* RESTRICT r)
 template<uint32 prim, uint32 adc>
 void GSState::GIFPackedRegHandlerXYZ2(const GIFPackedReg* RESTRICT r)
 {
-	if(adc)
-	{
-		// not sure what the difference is between this and XYZ2 with ADC bit set
-
-		//printf("XYZ3 X %d Y %d Z %d ADC %d\n", r->XYZ2.X, r->XYZ2.Y, r->XYZ2.Z, r->XYZ2.ADC); 
-	}
 /*
 	m_v.XYZ.X = r->XYZ2.X;
 	m_v.XYZ.Y = r->XYZ2.Y;
@@ -585,12 +572,17 @@ __forceinline void GSState::ApplyPRIM(const GIFRegPRIM& prim)
 	m_env.PRIM = (GSVector4i)prim;
 	m_env.PRMODE._PRIM = prim.PRIM;
 
-	m_context = &m_env.CTXT[PRIM->CTXT];
+	UpdateContext();
 
 	UpdateVertexKick();
 
 	ASSERT(m_index.tail == 0 || m_index.buff[m_index.tail - 1] + 1 == m_vertex.next);
 
+	if(m_index.tail == 0)
+	{
+		m_vertex.next = 0;
+	}
+
 	m_vertex.head = m_vertex.tail = m_vertex.next; // remove unused vertices from the end of the vertex buffer
 }
 
@@ -821,6 +813,8 @@ template<int i> void GSState::GIFRegHandlerXYOFFSET(const GIFReg* RESTRICT r)
 	m_env.CTXT[i].XYOFFSET = o;
 
 	m_env.CTXT[i].UpdateScissor();
+
+	UpdateScissor();
 }
 
 void GSState::GIFRegHandlerPRMODECONT(const GIFReg* RESTRICT r)
@@ -836,7 +830,7 @@ void GSState::GIFRegHandlerPRMODECONT(const GIFReg* RESTRICT r)
 
 	// if(PRIM->PRIM == 7) printf("Invalid PRMODECONT/PRIM\n");
 
-	m_context = &m_env.CTXT[PRIM->CTXT];
+	UpdateContext();
 
 	UpdateVertexKick();
 }
@@ -852,7 +846,7 @@ void GSState::GIFRegHandlerPRMODE(const GIFReg* RESTRICT r)
 	m_env.PRMODE = (GSVector4i)r->PRMODE;
 	m_env.PRMODE._PRIM = _PRIM;
 
-	m_context = &m_env.CTXT[PRIM->CTXT];
+	UpdateContext();
 
 	UpdateVertexKick();
 }
@@ -932,6 +926,8 @@ template<int i> void GSState::GIFRegHandlerSCISSOR(const GIFReg* RESTRICT r)
 	m_env.CTXT[i].SCISSOR = (GSVector4i)r->SCISSOR;
 
 	m_env.CTXT[i].UpdateScissor();
+
+	UpdateScissor();
 }
 
 template<int i> void GSState::GIFRegHandlerALPHA(const GIFReg* RESTRICT r)
@@ -1268,14 +1264,20 @@ void GSState::FlushPrim()
 		size_t stride = m_vertex.stride;
 		size_t head = m_vertex.head;
 		size_t tail = m_vertex.tail;
-		
+		size_t next = m_vertex.next;
+
 		if(tail > head)
 		{
 			switch(PRIM->PRIM)
 			{
+			case GS_POINTLIST:
+				break;
+			case GS_LINELIST:
 			case GS_LINESTRIP:
+			case GS_SPRITE:
 				if(tail > head + 0) memcpy(&buff[stride * 0], &m_vertex.buff[stride * (head + 0)], stride);
 				break;
+			case GS_TRIANGLELIST:
 			case GS_TRIANGLESTRIP:
 				if(tail > head + 0) memcpy(&buff[stride * 0], &m_vertex.buff[stride * (head + 0)], stride);
 				if(tail > head + 1) memcpy(&buff[stride * 1], &m_vertex.buff[stride * (head + 1)], stride);
@@ -1284,10 +1286,6 @@ void GSState::FlushPrim()
 				if(tail > head + 0) memcpy(&buff[stride * 0], &m_vertex.buff[stride * (head + 0)], stride);
 				if(tail > head + 1) memcpy(&buff[stride * 1], &m_vertex.buff[stride * (tail - 1)], stride);
 				break;
-			case GS_POINTLIST:
-			case GS_LINELIST:
-			case GS_TRIANGLELIST:
-			case GS_SPRITE:
 			case GS_INVALID:
 				break;
 			default:
@@ -1307,40 +1305,37 @@ void GSState::FlushPrim()
 			m_perfmon.Put(GSPerfMon::Prim, m_index.tail / GSUtil::GetVertexCount(PRIM->PRIM));
 		}
 
+		m_index.tail = 0;
+
 		m_vertex.head = 0;
 		m_vertex.tail = 0;
+		m_vertex.next = 0;
 
 		if(tail > head)
 		{
 			switch(PRIM->PRIM)
 			{
+			case GS_POINTLIST:
+				break;
+			case GS_LINELIST:
 			case GS_LINESTRIP:
+			case GS_SPRITE:
 				if(tail > head + 0) {memcpy(&m_vertex.buff[stride * 0], &buff[stride * 0], stride); m_vertex.tail++;}
 				break;
+			case GS_TRIANGLELIST:
 			case GS_TRIANGLESTRIP:
 			case GS_TRIANGLEFAN:
 				if(tail > head + 0) {memcpy(&m_vertex.buff[stride * 0], &buff[stride * 0], stride); m_vertex.tail++;}
 				if(tail > head + 1) {memcpy(&m_vertex.buff[stride * 1], &buff[stride * 1], stride); m_vertex.tail++;}
 				break;
-			case GS_POINTLIST:
-			case GS_LINELIST:
-			case GS_TRIANGLELIST:
-			case GS_SPRITE:
 			case GS_INVALID:
 				break;
 			default:
 				__assume(0);
 			}
-		}
 
-		m_vertex.next = m_vertex.tail;
+			m_vertex.next = next > head ? next - head : 0;
-		m_index.tail = 0;
+		}
-	}
-	else
-	{
-		m_vertex.head = 0;
-		m_vertex.tail = 0;
-		m_vertex.next = 0;
 	}
 }
 
@@ -2064,7 +2059,7 @@ int GSState::Defrost(const GSFreezeData* fd)
 
 	PRIM = !m_env.PRMODECONT.AC ? (GIFRegPRIM*)&m_env.PRMODE : &m_env.PRIM;
 
-	m_context = &m_env.CTXT[PRIM->CTXT];
+	UpdateContext();
 
 	UpdateVertexKick();
 
@@ -2080,6 +2075,8 @@ int GSState::Defrost(const GSFreezeData* fd)
 		m_env.CTXT[i].offset.fzb = m_mem.GetPixelOffset4(m_env.CTXT[i].FRAME, m_env.CTXT[i].ZBUF);
 	}
 
+	UpdateScissor();
+
 m_perfmon.SetFrame(5000);
 
 	return 0;
@@ -2094,6 +2091,19 @@ void GSState::SetGameCRC(uint32 crc, int options)
 
 //
 
+void GSState::UpdateContext()
+{
+	m_context = &m_env.CTXT[PRIM->CTXT];
+
+	UpdateScissor();
+}
+
+void GSState::UpdateScissor()
+{
+	m_scissor = m_context->scissor.ofex;
+	m_ofxy = m_context->scissor.ofxy;
+}
+
 void GSState::UpdateVertexKick() 
 {
 	uint32 prim = PRIM->PRIM;
@@ -2157,7 +2167,7 @@ __forceinline void GSState::VertexKick(uint32 skip)
 	tailptr[0] = v0;
 	tailptr[1] = v1;
 
-	m_vertex.xy[xy_tail & 3] = GSVector4(v1.upl32(v1.add16(GSVector4i::x000f()).srl16(4)).upl16());
+	m_vertex.xy[xy_tail & 3] = GSVector4(v1.upl32(v1.sub16(GSVector4i::load(m_ofxy)).sra16(4)).upl16()); // zw not sign extended, only useful for eq tests
 
 	#ifdef _DEBUG
 	memset(&tailptr[2], 0, m_vertex.stride - sizeof(GSVertex));
@@ -2221,9 +2231,7 @@ __forceinline void GSState::VertexKick(uint32 skip)
 			break;
 		}
 
-		GSVector4 scissor = m_context->scissor.dx9;
+		GSVector4 test = pmax < m_scissor | pmin > m_scissor.zwxy(); 
-
-		GSVector4 test = pmax < scissor | pmin > scissor.zwxy(); 
 		
 		switch(prim)
 		{

plugins/GSdx/GSState.h

@@ -131,6 +131,8 @@ protected:
 
 	GSVertex m_v;
 	float m_q;
+	GSVector4 m_scissor;
+	uint32 m_ofxy;
 	
 	struct 
 	{
@@ -168,6 +170,9 @@ protected:
 		InitConvertVertex2(T, GS_SPRITE) \
 		InitConvertVertex2(T, GS_INVALID) \
 
+	void UpdateContext();
+	void UpdateScissor();
+
 	virtual void UpdateVertexKick();
 
 	void GrowVertexBuffer();

plugins/GSdx/GSVertexTrace.x86.avx.cpp

@@ -26,12 +26,12 @@
 
 using namespace Xbyak;
 
-static const int _args = 0;
+static const int _args = 4;
-static const int _count = _args + 8; // rcx
+static const int _count = _args + 4; // rcx
-static const int _vertex = _args + 12; // rdx
+static const int _vertex = _args + 8; // rdx
-static const int _index = _args + 16; // r8
+static const int _index = _args + 12; // r8
-static const int _min = _args + 20; // r9
+static const int _min = _args + 16; // r9
-static const int _max = _args + 24; // _args + 4
+static const int _max = _args + 20; // _args + 4
 
 GSVertexTraceSW::CG::CG(const void* param, uint32 key, void* code, size_t maxsize)
 	: GSCodeGenerator(code, maxsize)

plugins/GSdx/GSVertexTrace.x86.cpp

@@ -26,12 +26,12 @@
 
 using namespace Xbyak;
 
-static const int _args = 0;
+static const int _args = 4;
-static const int _count = _args + 8; // rcx
+static const int _count = _args + 4; // rcx
-static const int _vertex = _args + 12; // rdx
+static const int _vertex = _args + 8; // rdx
-static const int _index = _args + 16; // r8
+static const int _index = _args + 12; // r8
-static const int _min = _args + 20; // r9
+static const int _min = _args + 16; // r9
-static const int _max = _args + 24; // _args + 4
+static const int _max = _args + 20; // _args + 4
 
 GSVertexTraceSW::CG::CG(const void* param, uint32 key, void* code, size_t maxsize)
 	: GSCodeGenerator(code, maxsize)