diff --git a/gsdx/GS.h b/gsdx/GS.h
index 7b2b295..7f4a4f4 100644
--- a/gsdx/GS.h
+++ b/gsdx/GS.h
@@ -700,9 +700,7 @@ REG64_(GIFReg, PRIM)
 	UINT32 FIX:1;
 	UINT32 _PAD1:21;
 	UINT32 _PAD2:32;
-REG_END2
-	bool TME_FST() {return (ai32[0] & 0x110) == 0x110;}
-REG_END2
+REG_END
 
 REG64_(GIFReg, PRMODE)
 	UINT32 _PRIM:3;
diff --git a/gsdx/GSBlock.cpp b/gsdx/GSBlock.cpp
index 605fef5..1a85542 100644
--- a/gsdx/GSBlock.cpp
+++ b/gsdx/GSBlock.cpp
@@ -26,7 +26,6 @@ const GSVector4i GSBlock::m_r16mask(0, 1, 4, 5, 2, 3, 6, 7, 8, 9, 12, 13, 10, 11
 const GSVector4i GSBlock::m_r8mask(0, 4, 2, 6, 8, 12, 10, 14, 1, 5, 3, 7, 9, 13, 11, 15);
 const GSVector4i GSBlock::m_r4mask(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15);
 
-const GSVector4i GSBlock::m_rgbx(0x00ffffff);
 const GSVector4i GSBlock::m_xxxa(0x00008000);
 const GSVector4i GSBlock::m_xxbx(0x00007c00);
 const GSVector4i GSBlock::m_xgxx(0x000003e0);
diff --git a/gsdx/GSBlock.h b/gsdx/GSBlock.h
index b7af49a..aba27d7 100644
--- a/gsdx/GSBlock.h
+++ b/gsdx/GSBlock.h
@@ -31,7 +31,6 @@ class GSBlock
 	static const GSVector4i m_r8mask;
 	static const GSVector4i m_r4mask;
 
-	static const GSVector4i m_rgbx;
 	static const GSVector4i m_xxxa;
 	static const GSVector4i m_xxbx;
 	static const GSVector4i m_xgxx;
@@ -1040,7 +1039,7 @@ public:
 	{
 		#if _M_SSE >= 0x200
 
-		GSVector4i rgbx = m_rgbx;
+		GSVector4i mask = GSVector4i::x00ffffff();
 
 		for(int i = 0; i < 4; i++, src += srcpitch * 2)
 		{
@@ -1048,19 +1047,19 @@ public:
 			GSVector4i v1 = GSVector4i::load(src + 16, src + srcpitch);
 			GSVector4i v2 = GSVector4i::load<false>(src + srcpitch + 8);
 
-			((GSVector4i*)dst)[i * 4 + 0] = v0.upl32(v0.srl<3>()).upl64(v0.srl<6>().upl32(v0.srl<9>())) & rgbx;
+			((GSVector4i*)dst)[i * 4 + 0] = v0.upl32(v0.srl<3>()).upl64(v0.srl<6>().upl32(v0.srl<9>())) & mask;
 
 			v0 = v0.srl<12>(v1);
 
-			((GSVector4i*)dst)[i * 4 + 1] = v0.upl32(v0.srl<3>()).upl64(v0.srl<6>().upl32(v0.srl<9>())) & rgbx;
+			((GSVector4i*)dst)[i * 4 + 1] = v0.upl32(v0.srl<3>()).upl64(v0.srl<6>().upl32(v0.srl<9>())) & mask;
 
 			v0 = v1.srl<8>(v2);
 
-			((GSVector4i*)dst)[i * 4 + 2] = v0.upl32(v0.srl<3>()).upl64(v0.srl<6>().upl32(v0.srl<9>())) & rgbx;
+			((GSVector4i*)dst)[i * 4 + 2] = v0.upl32(v0.srl<3>()).upl64(v0.srl<6>().upl32(v0.srl<9>())) & mask;
 
 			v0 = v2.srl<4>();
 
-			((GSVector4i*)dst)[i * 4 + 3] = v0.upl32(v0.srl<3>()).upl64(v0.srl<6>().upl32(v0.srl<9>())) & rgbx;
+			((GSVector4i*)dst)[i * 4 + 3] = v0.upl32(v0.srl<3>()).upl64(v0.srl<6>().upl32(v0.srl<9>())) & mask;
 		}
 
 		#else 
@@ -1215,7 +1214,7 @@ public:
 		const GSVector4i* s = (const GSVector4i*)src;
 
 		GSVector4i TA0(TEXA.TA0 << 24);
-		GSVector4i mask = m_rgbx;
+		GSVector4i mask = GSVector4i::x00ffffff();
 
 		for(int i = 0; i < 4; i++, dst += dstpitch * 2)
 		{
@@ -1880,7 +1879,7 @@ public:
 		const GSVector4i* s = (const GSVector4i*)src;
 
 		GSVector4i TA0(TEXA.TA0 << 24);
-		GSVector4i mask = m_rgbx;
+		GSVector4i mask = GSVector4i::x00ffffff();
 
 		for(int i = 0; i < 4; i++, dst += dstpitch * 2)
 		{
diff --git a/gsdx/GSCrc.cpp b/gsdx/GSCrc.cpp
index 6bfb457..96d37f5 100644
--- a/gsdx/GSCrc.cpp
+++ b/gsdx/GSCrc.cpp
@@ -51,10 +51,12 @@ CRC::Game CRC::m_games[] =
 	{0xc1274668, FFXII, EU, false},
 	{0xdc2a467e, FFXII, EU, false},
 	{0xca284668, FFXII, EU, false},
+	{0x280AD120, FFXII, JP, false},
 	{0x8BE3D7B2, ShadowHearts, Unknown, false},
 	{0xDEFA4763, ShadowHearts, US, false},
 	{0x21068223, Okami, US, false},
 	{0x891f223f, Okami, FR, false},
+	{0xC5DEFEA0, Okami, JP, false},
 	{0x053D2239, MetalGearSolid3, US, false},
 	{0x086273D2, MetalGearSolid3, FR, false},
 	{0x26A6E286, MetalGearSolid3, EU, false},
@@ -62,6 +64,7 @@ CRC::Game CRC::m_games[] =
 	{0x9F185CE1, MetalGearSolid3, Unknown, false},
 	{0x98D4BC93, MetalGearSolid3, EU, false},
 	{0x86BC3040, MetalGearSolid3, US, false}, //Subsistance disc 1
+	{0x0481AD8A, MetalGearSolid3, JP, false},
 	{0x278722BF, DBZBT2, US, false},
 	{0xFE961D28, DBZBT2, US, false},
 	{0x0393B6BE, DBZBT2, EU, false},
@@ -78,7 +81,9 @@ CRC::Game CRC::m_games[] =
 	{0x7D8F539A, SoTC, EU, false},
 	{0x3122B508, OnePieceGrandAdventure, US, false},
 	{0x6F8545DB, ICO, US, false},
+	{0xB01A4C95, ICO, JP, false},
 	{0x5C991F4E, ICO, Unknown, false},
+	{0xAEAD1CA3, GT4, JP, false},
 	{0x44A61C8F, GT4, Unknown, false},
 	{0x0086E35B, GT4, Unknown, false},
 	{0x77E61C8A, GT4, Unknown, false},
@@ -86,14 +91,17 @@ CRC::Game CRC::m_games[] =
 	{0xC1640D2C, WildArms5, US, false},
 	{0x0FCF8FE4, WildArms5, EU, false},
 	{0x2294D322, WildArms5, JP, false}, 
+	{0x565B6170, WildArms5, JP, false},
 	{0x8B029334, Manhunt2, Unknown, false},
 	{0x09F49E37, CrashBandicootWoC, Unknown, false},
 	{0x013E349D, ResidentEvil4, US, false},
 	{0x6BA2F6B9, ResidentEvil4, Unknown, false},
 	{0x60FA8C69, ResidentEvil4, JP, false},
 	{0x72E1E60E, Spartan, Unknown, false},
+	{0x5ED8FB53, AceCombat4, JP, false},
 	{0x1B9B7563, AceCombat4, Unknown, false},
 	{0xEC432B24, Drakengard2, Unknown, false},
+	{0xFC46EA61, Tekken5, JP, false},
 	{0x1F88EE37, Tekken5, Unknown, false},
 	{0x652050D2, Tekken5, Unknown, false}, 
 	{0x9E98B8AE, IkkiTousen, JP, false},
@@ -106,6 +114,7 @@ CRC::Game CRC::m_games[] =
 	{0xf0a6d880, HarvestMoon, US, true},
 	{0x75c01a04, NamcoXCapcom, US, false},
 	{0xBF6F101F, GiTS, US, false},
+	{0xA5768F53, GiTS, JP, false},
 	{0x6BF11378, Onimusha3, US, false},	
 	{0xF442260C, MajokkoALaMode2, JP, false},
 	{0x14FE77F7, TalesOfAbyss, US, false},
diff --git a/gsdx/GSDrawScanline.cpp b/gsdx/GSDrawScanline.cpp
index 22a85bb..ad9fd29 100644
--- a/gsdx/GSDrawScanline.cpp
+++ b/gsdx/GSDrawScanline.cpp
@@ -176,7 +176,8 @@ void GSDrawScanline::BeginDraw(const GSRasterizerData* data, Functions* f)
 		|| m_env.sel.abe != 255 
 		|| m_env.sel.ztst > 1 
 		|| m_env.sel.atst > 1
-		|| m_env.sel.date)
+		|| m_env.sel.date
+		|| m_env.sel.fge)
 		{
 			f->sr = NULL;
 		}
@@ -308,6 +309,8 @@ void GSDrawScanline::SetupPrim(const GSVertexSW* vertices, const GSVertexSW& dsc
 		
 		c = c.upl16(c.zwxy());
 
+		if(!tme) c = c.srl16(7);
+
 		m_env.c.rb = c.xxxx();
 		m_env.c.ga = c.zzzz();
 	}
@@ -371,20 +374,21 @@ void GSDrawScanline::SampleTexture(DWORD ltf, DWORD tlu, const GSVector4i& u, co
 		GSVector4i rb10 = c10 & mask;
 		GSVector4i rb11 = c11 & mask;
 
+		rb00 = rb00.lerp16<0>(rb01, uf);
+		rb10 = rb10.lerp16<0>(rb11, uf);
+		rb00 = rb00.lerp16<0>(rb10, vf);
+
+		c[0] = rb00;
+
 		GSVector4i ga00 = (c00 >> 8) & mask;
 		GSVector4i ga01 = (c01 >> 8) & mask;
 		GSVector4i ga10 = (c10 >> 8) & mask;
 		GSVector4i ga11 = (c11 >> 8) & mask;
 
-		rb00 = rb00.lerp16<0>(rb01, uf);
-		rb10 = rb10.lerp16<0>(rb11, uf);
-		rb00 = rb00.lerp16<0>(rb10, vf);
-
 		ga00 = ga00.lerp16<0>(ga01, uf);
 		ga10 = ga10.lerp16<0>(ga11, uf);
 		ga00 = ga00.lerp16<0>(ga10, vf);
 
-		c[0] = rb00;
 		c[1] = ga00;
 	}
 	else
@@ -409,49 +413,54 @@ void GSDrawScanline::SampleTexture(DWORD ltf, DWORD tlu, const GSVector4i& u, co
 	}
 }
 
-void GSDrawScanline::ColorTFX(DWORD tfx, const GSVector4i& rbf, const GSVector4i& gaf, GSVector4i& rbt, GSVector4i& gat)
+void GSDrawScanline::ColorTFX(DWORD iip, DWORD tfx, const GSVector4i& rbf, const GSVector4i& gaf, GSVector4i& rbt, GSVector4i& gat)
 {
+	GSVector4i rb = iip ? rbf : m_env.c.rb;
+	GSVector4i ga = iip ? gaf : m_env.c.ga;
+
 	GSVector4i af;
 
 	switch(tfx)
 	{
 	case TFX_MODULATE:
-		rbt = rbt.modulate16<1>(rbf).clamp8();
+		rbt = rbt.modulate16<1>(rb).clamp8();
 		break;
 	case TFX_DECAL:
 		break;
 	case TFX_HIGHLIGHT:
 	case TFX_HIGHLIGHT2:
-		af = gaf.yywwlh().srl16(7);
-		rbt = rbt.modulate16<1>(rbf).add16(af).clamp8();
-		gat = gat.modulate16<1>(gaf).add16(af).clamp8().mix16(gat);
+		af = ga.yywwlh().srl16(7);
+		rbt = rbt.modulate16<1>(rb).add16(af).clamp8();
+		gat = gat.modulate16<1>(ga).add16(af).clamp8().mix16(gat);
 		break;
 	case TFX_NONE:
-		rbt = rbf.srl16(7);
+		rbt = iip ? rb.srl16(7) : rb;
 		break;
 	default:
 		__assume(0);
 	}
 }
 
-void GSDrawScanline::AlphaTFX(DWORD tfx, DWORD tcc, const GSVector4i& gaf, GSVector4i& gat)
+void GSDrawScanline::AlphaTFX(DWORD iip, DWORD tfx, DWORD tcc, const GSVector4i& gaf, GSVector4i& gat)
 {
+	GSVector4i ga = iip ? gaf : m_env.c.ga;
+
 	switch(tfx)
 	{
 	case TFX_MODULATE:
-		gat = gat.modulate16<1>(gaf).clamp8(); // mul16hrs rounds and breaks fogging in resident evil 4 (only modulate16<0> uses mul16hrs, but watch out)
-		if(!tcc) gat = gat.mix16(gaf.srl16(7));
+		gat = gat.modulate16<1>(ga).clamp8(); // mul16hrs rounds and breaks fogging in resident evil 4 (only modulate16<0> uses mul16hrs, but watch out)
+		if(!tcc) gat = gat.mix16(ga.srl16(7));
 		break;
 	case TFX_DECAL: 
 		break;
 	case TFX_HIGHLIGHT: 
-		gat = gat.mix16(!tcc ? gaf.srl16(7) : gat.addus8(gaf.srl16(7)));
+		gat = gat.mix16(!tcc ? ga.srl16(7) : gat.addus8(ga.srl16(7)));
 		break;
 	case TFX_HIGHLIGHT2:
-		if(!tcc) gat = gat.mix16(gaf.srl16(7));
+		if(!tcc) gat = gat.mix16(ga.srl16(7));
 		break;
 	case TFX_NONE: 
-		gat = gaf.srl16(7);
+		gat = iip ? ga.srl16(7) : ga;
 		break; 
 	default: 
 		__assume(0);
@@ -716,11 +725,6 @@ void GSDrawScanline::DrawScanline(int top, int left, int right, const GSVertexSW
 		rb = vc.xxxx().add16(m_env.d[skip].rb);
 		ga = vc.zzzz().add16(m_env.d[skip].ga);
 	}
-	else
-	{
-		rb = m_env.c.rb;
-		ga = m_env.c.ga;
-	}
 
 	//
 
@@ -772,7 +776,7 @@ void GSDrawScanline::DrawScanline(int top, int left, int right, const GSVertexSW
 				SampleTexture(m_env.sel.ltf, m_env.sel.tlu, u, v, c);
 			}
 
-			AlphaTFX(m_env.sel.tfx, m_env.sel.tcc, ga, c[1]);
+			AlphaTFX(iip, m_env.sel.tfx, m_env.sel.tcc, ga, c[1]);
 
 			GSVector4i fm = m_env.fm;
 			GSVector4i zm = m_env.zm;
@@ -782,7 +786,7 @@ void GSDrawScanline::DrawScanline(int top, int left, int right, const GSVertexSW
 				continue;
 			}
 
-			ColorTFX(m_env.sel.tfx, rb, ga, c[0], c[1]);
+			ColorTFX(iip, m_env.sel.tfx, rb, ga, c[0], c[1]);
 
 			Fog(m_env.sel.fge, f, c[0], c[1]);
 
@@ -975,11 +979,6 @@ void GSDrawScanline::DrawScanlineEx(int top, int left, int right, const GSVertex
 		rb = vc.xxxx().add16(m_env.d[skip].rb);
 		ga = vc.zzzz().add16(m_env.d[skip].ga);
 	}
-	else
-	{
-		rb = m_env.c.rb;
-		ga = m_env.c.ga;
-	}
 
 	//
 
@@ -1031,7 +1030,7 @@ void GSDrawScanline::DrawScanlineEx(int top, int left, int right, const GSVertex
 				SampleTexture(ltf, tlu, u, v, c);
 			}
 
-			AlphaTFX(tfx, tcc, ga, c[1]);
+			AlphaTFX(iip, tfx, tcc, ga, c[1]);
 
 			GSVector4i fm = m_env.fm;
 			GSVector4i zm = m_env.zm;
@@ -1041,7 +1040,7 @@ void GSDrawScanline::DrawScanlineEx(int top, int left, int right, const GSVertex
 				continue;
 			}
 
-			ColorTFX(tfx, rb, ga, c[0], c[1]);
+			ColorTFX(iip, tfx, rb, ga, c[0], c[1]);
 
 			Fog(fge, f, c[0], c[1]);
 
@@ -1091,7 +1090,7 @@ void GSDrawScanline::DrawScanlineEx(int top, int left, int right, const GSVertex
 					rb = c[abea * 2 + 0];
 					ga = c[abea * 2 + 1];
 
-					if(abeb != 2)
+					if(abeb < 2)
 					{
 						rb = rb.sub16(c[abeb * 2 + 0]);
 						ga = ga.sub16(c[abeb * 2 + 1]);
diff --git a/gsdx/GSDrawScanline.h b/gsdx/GSDrawScanline.h
index f8b8035..d2a491d 100644
--- a/gsdx/GSDrawScanline.h
+++ b/gsdx/GSDrawScanline.h
@@ -153,8 +153,8 @@ class GSDrawScanline : public GSAlignedClass<16>, public IDrawScanline
 	__forceinline GSVector4i Wrap(const GSVector4i& t);
 
 	__forceinline void SampleTexture(DWORD ltf, DWORD tlu, const GSVector4i& u, const GSVector4i& v, GSVector4i* c);
-	__forceinline void ColorTFX(DWORD tfx, const GSVector4i& rbf, const GSVector4i& gaf, GSVector4i& rbt, GSVector4i& gat);
-	__forceinline void AlphaTFX(DWORD tfx, DWORD tcc, const GSVector4i& gaf, GSVector4i& gat);
+	__forceinline void ColorTFX(DWORD iip, DWORD tfx, const GSVector4i& rbf, const GSVector4i& gaf, GSVector4i& rbt, GSVector4i& gat);
+	__forceinline void AlphaTFX(DWORD iip, DWORD tfx, DWORD tcc, const GSVector4i& gaf, GSVector4i& gat);
 	__forceinline void Fog(DWORD fge, const GSVector4i& f, GSVector4i& rb, GSVector4i& ga);
 	__forceinline bool TestZ(DWORD zpsm, DWORD ztst, const GSVector4i& zs, const GSVector4i& zd, GSVector4i& test);
 	__forceinline bool TestAlpha(DWORD atst, DWORD afail, const GSVector4i& ga, GSVector4i& fm, GSVector4i& zm, GSVector4i& test);
diff --git a/gsdx/GSRenderer.h b/gsdx/GSRenderer.h
index 2cd876c..b02cd9b 100644
--- a/gsdx/GSRenderer.h
+++ b/gsdx/GSRenderer.h
@@ -456,108 +456,6 @@ protected:
 		__super::Reset();
 	}
 
-	void VertexKick(bool skip)
-	{
-		if(m_vl.GetCount() < m_vprim)
-		{
-			return;
-		}
-
-		if(m_count > m_maxcount)
-		{
-			m_maxcount = max(10000, m_maxcount * 3/2);
-			m_vertices = (Vertex*)_aligned_realloc(m_vertices, sizeof(Vertex) * m_maxcount, 16);
-			m_maxcount -= 100;
-		}
-
-		Vertex* v = &m_vertices[m_count];
-
-		int count = 0;
-
-		switch(PRIM->PRIM)
-		{
-		case GS_POINTLIST:
-			m_vl.GetAt(0, v[0]);
-			m_vl.RemoveAll();
-			count = 1;
-			break;
-		case GS_LINELIST:
-			m_vl.GetAt(0, v[0]);
-			m_vl.GetAt(1, v[1]);
-			m_vl.RemoveAll();
-			count = 2;
-			break;
-		case GS_LINESTRIP:
-			m_vl.GetAt(0, v[0]);
-			m_vl.GetAt(1, v[1]);
-			m_vl.RemoveAt(0, 1);
-			count = 2;
-			break;
-		case GS_TRIANGLELIST:
-			m_vl.GetAt(0, v[0]);
-			m_vl.GetAt(1, v[1]);
-			m_vl.GetAt(2, v[2]);
-			m_vl.RemoveAll();
-			count = 3;
-			break;
-		case GS_TRIANGLESTRIP:
-			m_vl.GetAt(0, v[0]);
-			m_vl.GetAt(1, v[1]);
-			m_vl.GetAt(2, v[2]);
-			m_vl.RemoveAt(0, 2);
-			count = 3;
-			break;
-		case GS_TRIANGLEFAN:
-			m_vl.GetAt(0, v[0]);
-			m_vl.GetAt(1, v[1]);
-			m_vl.GetAt(2, v[2]);
-			m_vl.RemoveAt(1, 1);
-			count = 3;
-			break;
-		case GS_SPRITE:
-			m_vl.GetAt(0, v[0]);
-			m_vl.GetAt(1, v[1]);
-			m_vl.RemoveAll();
-			count = 2;
-			break;
-		case GS_INVALID:
-			ASSERT(0);
-			m_vl.RemoveAll();
-			return;
-		default:
-			__assume(0);
-		}
-
-		if(!skip)
-		{
-			(this->*m_fpDrawingKickHandlers[PRIM->PRIM])(v, count);
-
-			m_count += count;
-/*
-			if(m_context->TEST.DATE)
-			{
-				Flush();
-			}
-
-			if(m_env.COLCLAMP.CLAMP == 0)
-			{
-				Flush();
-			}
-*/
-		}
-	}
-
-	typedef void (GSRendererT<Device, Vertex>::*DrawingKickHandler)(Vertex* v, int& count);
-
-	DrawingKickHandler m_fpDrawingKickHandlers[8];
-
-	void DrawingKickNull(Vertex* v, int& count)
-	{
-		ASSERT(0);
-	}
-
-	virtual void Draw() = 0;
-
 	void ResetPrim()
 	{
 		m_vl.RemoveAll();
@@ -582,43 +480,140 @@ protected:
 				  PRIM->TME ? (int)m_context->TEX0.PSM : 0xff);
 			*/
 
-			/*
-			static bool first = true;
-			static GIFRegPRIM s_PRIM;
-			static GIFRegFRAME s_FRAME;
-			static GIFRegTEX0 s_TEX0;
-
-			if(first || PRIM->TME != s_PRIM.TME || m_context->FRAME.i64 != s_FRAME.i64 || m_context->TEX0.i64 != s_TEX0.i64)
-			{
-				first = false;
-				s_PRIM = *PRIM;
-				s_FRAME = m_context->FRAME;
-				s_TEX0 = m_context->TEX0;
-				printf("%05x %2d", s_FRAME.Block(), s_FRAME.PSM);
-				if(s_PRIM.TME) printf(" %05x %2d %05x %2d", s_TEX0.TBP0, s_TEX0.PSM, s_TEX0.CBP, s_TEX0.CPSM);
-				printf("\n");
-			}
-			*/
-
 			Draw();
 
 			m_count = 0;
 		}
 	}
 
+	void GrowVertexBuffer()
+	{
+		m_maxcount = max(10000, m_maxcount * 3/2);
+		m_vertices = (Vertex*)_aligned_realloc(m_vertices, sizeof(Vertex) * m_maxcount, 16);
+		m_maxcount -= 100;
+	}
+
+	template<DWORD prim> void VertexKick(bool skip)
+	{
+		(this->*m_fpAddVertexHandlers[PRIM->TME][PRIM->FST])();
+
+		DWORD count = 0;
+
+		switch(prim)
+		{
+		case GS_POINTLIST: count = 1; break;
+		case GS_LINELIST: count = 2; break;
+		case GS_LINESTRIP: count = 2; break;
+		case GS_TRIANGLELIST: count = 3; break;
+		case GS_TRIANGLESTRIP: count = 3; break;
+		case GS_TRIANGLEFAN: count = 3; break;
+		case GS_SPRITE: count = 2; break;
+		case GS_INVALID: count = 1; break;
+		default: __assume(0);
+		}
+
+		if(m_vl.GetCount() < count)
+		{
+			return;
+		}
+
+		if(m_count >= m_maxcount)
+		{
+			GrowVertexBuffer();
+		}
+
+		Vertex* v = &m_vertices[m_count];
+
+		switch(prim)
+		{
+		case GS_POINTLIST:
+			m_vl.GetAt(0, v[0]);
+			m_vl.RemoveAll();
+			break;
+		case GS_LINELIST:
+			m_vl.GetAt(0, v[0]);
+			m_vl.GetAt(1, v[1]);
+			m_vl.RemoveAll();
+			break;
+		case GS_LINESTRIP:
+			m_vl.GetAt(0, v[0]);
+			m_vl.GetAt(1, v[1]);
+			m_vl.RemoveAt(0, 1);
+			break;
+		case GS_TRIANGLELIST:
+			m_vl.GetAt(0, v[0]);
+			m_vl.GetAt(1, v[1]);
+			m_vl.GetAt(2, v[2]);
+			m_vl.RemoveAll();
+			break;
+		case GS_TRIANGLESTRIP:
+			m_vl.GetAt(0, v[0]);
+			m_vl.GetAt(1, v[1]);
+			m_vl.GetAt(2, v[2]);
+			m_vl.RemoveAt(0, 2);
+			break;
+		case GS_TRIANGLEFAN:
+			m_vl.GetAt(0, v[0]);
+			m_vl.GetAt(1, v[1]);
+			m_vl.GetAt(2, v[2]);
+			m_vl.RemoveAt(1, 1);
+			break;
+		case GS_SPRITE:
+			m_vl.GetAt(0, v[0]);
+			m_vl.GetAt(1, v[1]);
+			m_vl.RemoveAll();
+			break;
+		case GS_INVALID:
+			ASSERT(0);
+			m_vl.RemoveAll();
+			return;
+		default:
+			__assume(0);
+		}
+
+		if(!skip)
+		{
+			(this->*m_fpAddPrimHandlers[prim])(v, count);
+
+			m_count += count;
+		}
+	}
+
+	virtual void Draw() = 0;
+
+	typedef void (GSRendererT<Device, Vertex>::*AddVertexHandler)();
+	typedef void (GSRendererT<Device, Vertex>::*AddPrimHandler)(Vertex* v, DWORD& count);
+
+	AddVertexHandler m_fpAddVertexHandlers[2][2];
+	AddPrimHandler m_fpAddPrimHandlers[8];
+
+	void AddVertexNull() {ASSERT(0);}
+	void AddPrimNull(Vertex* v, DWORD& count) {ASSERT(0);}
+
 public:
 	GSRendererT(BYTE* base, bool mt, void (*irq)(), int nloophack, const GSRendererSettings& rs, bool psrr = true)
 		: GSRenderer<Device>(base, mt, irq, nloophack, rs, psrr)
 		, m_count(0)
-		, m_maxcount(10000)
+		, m_maxcount(0)
 		, m_vertices(NULL)
 	{
-		m_vertices = (Vertex*)_aligned_malloc(sizeof(Vertex) * m_maxcount, 16);
-		m_maxcount -= 100;
+		m_fpVertexKickHandlers[GS_POINTLIST] = (VertexKickHandler)&GSRendererT<Device, Vertex>::VertexKick<GS_POINTLIST>;
+		m_fpVertexKickHandlers[GS_LINELIST] = (VertexKickHandler)&GSRendererT<Device, Vertex>::VertexKick<GS_LINELIST>;
+		m_fpVertexKickHandlers[GS_LINESTRIP] = (VertexKickHandler)&GSRendererT<Device, Vertex>::VertexKick<GS_LINESTRIP>;
+		m_fpVertexKickHandlers[GS_TRIANGLELIST] = (VertexKickHandler)&GSRendererT<Device, Vertex>::VertexKick<GS_TRIANGLELIST>;
+		m_fpVertexKickHandlers[GS_TRIANGLESTRIP] = (VertexKickHandler)&GSRendererT<Device, Vertex>::VertexKick<GS_TRIANGLESTRIP>;
+		m_fpVertexKickHandlers[GS_TRIANGLEFAN] = (VertexKickHandler)&GSRendererT<Device, Vertex>::VertexKick<GS_TRIANGLEFAN>;
+		m_fpVertexKickHandlers[GS_SPRITE] = (VertexKickHandler)&GSRendererT<Device, Vertex>::VertexKick<GS_SPRITE>;
+		m_fpVertexKickHandlers[GS_INVALID] = (VertexKickHandler)&GSRendererT<Device, Vertex>::VertexKick<GS_INVALID>;
 
-		for(int i = 0; i < countof(m_fpDrawingKickHandlers); i++)
+		m_fpAddVertexHandlers[0][0] = &GSRendererT<Device, Vertex>::AddVertexNull;
+		m_fpAddVertexHandlers[0][1] = &GSRendererT<Device, Vertex>::AddVertexNull;
+		m_fpAddVertexHandlers[1][0] = &GSRendererT<Device, Vertex>::AddVertexNull;
+		m_fpAddVertexHandlers[1][1] = &GSRendererT<Device, Vertex>::AddVertexNull;
+
+		for(int i = 0; i < countof(m_fpAddPrimHandlers); i++)
 		{
-			m_fpDrawingKickHandlers[i] = &GSRendererT<Device, Vertex>::DrawingKickNull;
+			m_fpAddPrimHandlers[i] = &GSRendererT<Device, Vertex>::AddPrimNull;
 		}
 	}
 
diff --git a/gsdx/GSRendererHW10.cpp b/gsdx/GSRendererHW10.cpp
index 2e0ec4c..6dc61bf 100644
--- a/gsdx/GSRendererHW10.cpp
+++ b/gsdx/GSRendererHW10.cpp
@@ -27,23 +27,18 @@
 GSRendererHW10::GSRendererHW10(BYTE* base, bool mt, void (*irq)(), int nloophack, const GSRendererSettings& rs)
 	: GSRendererHW<Device, Vertex, TextureCache>(base, mt, irq, nloophack, rs, true)
 {
-	for(int i = 0; i < countof(m_fpDrawingKickHandlers); i++)
-	{
-		m_fpDrawingKickHandlers[i] = (DrawingKickHandler)&GSRendererHW10::DrawingKick;
-	}
+	m_fpAddVertexHandlers[0][0] = (AddVertexHandler)&GSRendererHW10::AddVertex<0, 0>;
+	m_fpAddVertexHandlers[0][1] = (AddVertexHandler)&GSRendererHW10::AddVertex<0, 0>;
+	m_fpAddVertexHandlers[1][0] = (AddVertexHandler)&GSRendererHW10::AddVertex<1, 0>;
+	m_fpAddVertexHandlers[1][1] = (AddVertexHandler)&GSRendererHW10::AddVertex<1, 1>;
 
-	m_fpDrawingKickHandlers[GS_POINTLIST] = (DrawingKickHandler)&GSRendererHW10::DrawingKickPoint;
-
-	#if _M_SSE >= 0x401
-
-	m_fpDrawingKickHandlers[GS_LINELIST] = (DrawingKickHandler)&GSRendererHW10::DrawingKickLine;
-	m_fpDrawingKickHandlers[GS_LINESTRIP] = (DrawingKickHandler)&GSRendererHW10::DrawingKickLine;
-	m_fpDrawingKickHandlers[GS_TRIANGLELIST] = (DrawingKickHandler)&GSRendererHW10::DrawingKickTriangle;
-	m_fpDrawingKickHandlers[GS_TRIANGLESTRIP] = (DrawingKickHandler)&GSRendererHW10::DrawingKickTriangle;
-	m_fpDrawingKickHandlers[GS_TRIANGLEFAN] = (DrawingKickHandler)&GSRendererHW10::DrawingKickTriangle;
-	m_fpDrawingKickHandlers[GS_SPRITE] = (DrawingKickHandler)&GSRendererHW10::DrawingKickSprite;
-
-	#endif
+	m_fpAddPrimHandlers[GS_POINTLIST] = (AddPrimHandler)&GSRendererHW10::AddPrim<GS_POINT_CLASS>;
+	m_fpAddPrimHandlers[GS_LINELIST] = (AddPrimHandler)&GSRendererHW10::AddPrim<GS_LINE_CLASS>;
+	m_fpAddPrimHandlers[GS_LINESTRIP] = (AddPrimHandler)&GSRendererHW10::AddPrim<GS_LINE_CLASS>;
+	m_fpAddPrimHandlers[GS_TRIANGLELIST] = (AddPrimHandler)&GSRendererHW10::AddPrim<GS_TRIANGLE_CLASS>;
+	m_fpAddPrimHandlers[GS_TRIANGLESTRIP] = (AddPrimHandler)&GSRendererHW10::AddPrim<GS_TRIANGLE_CLASS>;
+	m_fpAddPrimHandlers[GS_TRIANGLEFAN] = (AddPrimHandler)&GSRendererHW10::AddPrim<GS_TRIANGLE_CLASS>;
+	m_fpAddPrimHandlers[GS_SPRITE] = (AddPrimHandler)&GSRendererHW10::AddPrim<GS_SPRITE_CLASS>;
 }
 
 bool GSRendererHW10::Create(LPCTSTR title)
@@ -86,112 +81,76 @@ bool GSRendererHW10::Create(LPCTSTR title)
 	return true;
 }
 
-void GSRendererHW10::VertexKick(bool skip)
+template<DWORD tme, DWORD fst>
+void GSRendererHW10::AddVertex()
 {
-	Vertex& v = m_vl.AddTail();
+	Vertex& dst = m_vl.AddTail();
 
-	v.m128i[0] = m_v.m128i[0];
-	v.m128i[1] = m_v.m128i[1];
+	dst.m128i[0] = m_v.m128i[0];
+	dst.m128i[1] = m_v.m128i[1];
 
-	if(PRIM->TME_FST())
+	if(tme && fst)
 	{
-		GSVector4::storel(&v.ST, m_v.GetUV());
+		GSVector4::storel(&dst.ST, m_v.GetUV());
 	}
-
-	__super::VertexKick(skip);
 }
 
-int GSRendererHW10::ScissorTest(const GSVector4i& p0, const GSVector4i& p1)
+template<int primclass>
+void GSRendererHW10::AddPrim(Vertex* v, DWORD& count)
 {
 	GSVector4i scissor = m_context->scissor.dx10;
 
-	GSVector4i v0 = p0 < scissor;
-	GSVector4i v1 = p1 > scissor.zwxy();
+	#if _M_SSE >= 0x401
 
-	return (v0 | v1).mask() & 0xff;
-}
+	GSVector4i pmin, pmax, v0, v1, v2;
 
-void GSRendererHW10::DrawingKickPoint(Vertex* v, int& count)
-{
-	GSVector4i v0 = GSVector4i::load((int)v[0].p.xy).upl16();
+	switch(primclass)
+	{
+	case GS_POINT_CLASS:
+		v0 = GSVector4i::load((int)v[0].p.xy).upl16();
+		pmin = v0;
+		pmax = v0;
+		break;
+	case GS_LINE_CLASS:
+	case GS_SPRITE_CLASS:
+		v0 = GSVector4i::load((int)v[0].p.xy);
+		v1 = GSVector4i::load((int)v[1].p.xy);
+		pmin = v0.min_u16(v1).upl16();
+		pmax = v0.max_u16(v1).upl16();
+		break;
+	case GS_TRIANGLE_CLASS:
+		v0 = GSVector4i::load((int)v[0].p.xy);
+		v1 = GSVector4i::load((int)v[1].p.xy);
+		v2 = GSVector4i::load((int)v[2].p.xy);
+		pmin = v0.min_u16(v1).min_u16(v2).upl16();
+		pmax = v0.max_u16(v1).max_u16(v2).upl16();
+		break;
+	}
 
-	GSVector4i p0 = v0;
-	GSVector4i p1 = v0;
+	GSVector4i test = (pmax < scissor) | (pmin > scissor.zwxy());
 
-	if(ScissorTest(p0, p1))
+	if(test.mask() & 0xff)
 	{
 		count = 0;
 		return;
 	}
-}
 
-#if _M_SSE >= 0x401
+	#else
 
-void GSRendererHW10::DrawingKickLine(Vertex* v, int& count)
-{
-	GSVector4i v0 = GSVector4i::load((int)v[0].p.xy);
-	GSVector4i v1 = GSVector4i::load((int)v[1].p.xy);
-
-	GSVector4i p0 = v0.max_u16(v1).upl16();
-	GSVector4i p1 = v0.min_u16(v1).upl16();
-
-	if(ScissorTest(p0, p1))
+	switch(primclass)
 	{
-		count = 0;
-		return;
-	}
-}
-
-void GSRendererHW10::DrawingKickTriangle(Vertex* v, int& count)
-{
-	GSVector4i v0 = GSVector4i::load((int)v[0].p.xy);
-	GSVector4i v1 = GSVector4i::load((int)v[1].p.xy);
-	GSVector4i v2 = GSVector4i::load((int)v[2].p.xy);
-
-	GSVector4i p0 = v0.max_u16(v1).max_u16(v2).upl16();
-	GSVector4i p1 = v0.min_u16(v1).min_u16(v2).upl16();
-
-	if(ScissorTest(p0, p1))
-	{
-		count = 0;
-		return;
-	}
-}
-
-void GSRendererHW10::DrawingKickSprite(Vertex* v, int& count)
-{
-	GSVector4i v0 = GSVector4i::load((int)v[0].p.xy);
-	GSVector4i v1 = GSVector4i::load((int)v[1].p.xy);
-
-	GSVector4i p0 = v0.max_u16(v1).upl16();
-	GSVector4i p1 = v0.min_u16(v1).upl16();
-
-	if(ScissorTest(p0, p1))
-	{
-		count = 0;
-		return;
-	}
-}
-
-#endif
-
-void GSRendererHW10::DrawingKick(Vertex* v, int& count)
-{
-	GSVector4i scissor = m_context->scissor.dx10;
-
-	switch(count)
-	{
-	case 1:
-		if(v[0].p.x < scissor.x
+	case GS_POINT_CLASS:
+		if(v[0].p.x < scissor.x 
 		|| v[0].p.x > scissor.z
-		|| v[0].p.y < scissor.y
+		|| v[0].p.y < scissor.y 
 		|| v[0].p.y > scissor.w)
 		{
 			count = 0;
 			return;
 		}
 		break;
-	case 2:
+	case GS_LINE_CLASS:
+	case GS_SPRITE_CLASS:
 		if(v[0].p.x < scissor.x && v[1].p.x < scissor.x
 		|| v[0].p.x > scissor.z && v[1].p.x > scissor.z
 		|| v[0].p.y < scissor.y && v[1].p.y < scissor.y
@@ -201,7 +160,7 @@ void GSRendererHW10::DrawingKick(Vertex* v, int& count)
 			return;
 		}
 		break;
-	case 3:
+	case GS_TRIANGLE_CLASS:
 		if(v[0].p.x < scissor.x && v[1].p.x < scissor.x && v[2].p.x < scissor.x
 		|| v[0].p.x > scissor.z && v[1].p.x > scissor.z && v[2].p.x > scissor.z
 		|| v[0].p.y < scissor.y && v[1].p.y < scissor.y && v[2].p.y < scissor.y
@@ -211,9 +170,9 @@ void GSRendererHW10::DrawingKick(Vertex* v, int& count)
 			return;
 		}
 		break;
-	default:
-		__assume(0);
 	}
+
+	#endif
 }
 
 void GSRendererHW10::Draw(int prim, Texture& rt, Texture& ds, GSTextureCache<Device>::GSTexture* tex)
diff --git a/gsdx/GSRendererHW10.h b/gsdx/GSRendererHW10.h
index 4eee5be..b2aa13e 100644
--- a/gsdx/GSRendererHW10.h
+++ b/gsdx/GSRendererHW10.h
@@ -35,22 +35,14 @@ class GSRendererHW10 : public GSRendererHW<GSDevice10, GSVertexHW10, GSTextureCa
 protected:
 	GSTextureFX10 m_tfx;
 
-	void VertexKick(bool skip);
-	void DrawingKick(Vertex* v, int& count); // TODO
+	template<DWORD tme, DWORD fst> 
+	void AddVertex();
+	
+	template<int primclass> 
+	void AddPrim(Vertex* v, DWORD& count);
+
 	bool WrapZ(DWORD maxz);
 
-	__forceinline int ScissorTest(const GSVector4i& p0, const GSVector4i& p1);
-
-	void DrawingKickPoint(Vertex* v, int& count);
-
-	#if _M_SSE >= 0x401
-
-	void DrawingKickLine(Vertex* v, int& count);
-	void DrawingKickTriangle(Vertex* v, int& count);
-	void DrawingKickSprite(Vertex* v, int& count);
-
-	#endif
-
 	void Draw(int prim, Texture& rt, Texture& ds, GSTextureCache<Device>::GSTexture* tex);
 
 	struct
diff --git a/gsdx/GSRendererHW9.cpp b/gsdx/GSRendererHW9.cpp
index e7dad6c..709d051 100644
--- a/gsdx/GSRendererHW9.cpp
+++ b/gsdx/GSRendererHW9.cpp
@@ -30,13 +30,18 @@ GSRendererHW9::GSRendererHW9(BYTE* base, bool mt, void (*irq)(), int nloophack,
 	m_fba.enabled = !!AfxGetApp()->GetProfileInt(_T("Settings"), _T("fba"), TRUE);
 	m_logz = !!AfxGetApp()->GetProfileInt(_T("Settings"), _T("logz"), FALSE);
 
-	m_fpDrawingKickHandlers[GS_POINTLIST] = (DrawingKickHandler)&GSRendererHW9::DrawingKickPoint;
-	m_fpDrawingKickHandlers[GS_LINELIST] = (DrawingKickHandler)&GSRendererHW9::DrawingKickLine;
-	m_fpDrawingKickHandlers[GS_LINESTRIP] = (DrawingKickHandler)&GSRendererHW9::DrawingKickLine;
-	m_fpDrawingKickHandlers[GS_TRIANGLELIST] = (DrawingKickHandler)&GSRendererHW9::DrawingKickTriangle;
-	m_fpDrawingKickHandlers[GS_TRIANGLESTRIP] = (DrawingKickHandler)&GSRendererHW9::DrawingKickTriangle;
-	m_fpDrawingKickHandlers[GS_TRIANGLEFAN] = (DrawingKickHandler)&GSRendererHW9::DrawingKickTriangle;
-	m_fpDrawingKickHandlers[GS_SPRITE] = (DrawingKickHandler)&GSRendererHW9::DrawingKickSprite;
+	m_fpAddVertexHandlers[0][0] = (AddVertexHandler)&GSRendererHW9::AddVertex<0, 0>;
+	m_fpAddVertexHandlers[0][1] = (AddVertexHandler)&GSRendererHW9::AddVertex<0, 0>;
+	m_fpAddVertexHandlers[1][0] = (AddVertexHandler)&GSRendererHW9::AddVertex<1, 0>;
+	m_fpAddVertexHandlers[1][1] = (AddVertexHandler)&GSRendererHW9::AddVertex<1, 1>;
+
+	m_fpAddPrimHandlers[GS_POINTLIST] = (AddPrimHandler)&GSRendererHW9::AddPrim<GS_POINT_CLASS>;
+	m_fpAddPrimHandlers[GS_LINELIST] = (AddPrimHandler)&GSRendererHW9::AddPrim<GS_LINE_CLASS>;
+	m_fpAddPrimHandlers[GS_LINESTRIP] = (AddPrimHandler)&GSRendererHW9::AddPrim<GS_LINE_CLASS>;
+	m_fpAddPrimHandlers[GS_TRIANGLELIST] = (AddPrimHandler)&GSRendererHW9::AddPrim<GS_TRIANGLE_CLASS>;
+	m_fpAddPrimHandlers[GS_TRIANGLESTRIP] = (AddPrimHandler)&GSRendererHW9::AddPrim<GS_TRIANGLE_CLASS>;
+	m_fpAddPrimHandlers[GS_TRIANGLEFAN] = (AddPrimHandler)&GSRendererHW9::AddPrim<GS_TRIANGLE_CLASS>;
+	m_fpAddPrimHandlers[GS_SPRITE] = (AddPrimHandler)&GSRendererHW9::AddPrim<GS_SPRITE_CLASS>;
 }
 
 bool GSRendererHW9::Create(LPCTSTR title)
@@ -80,123 +85,93 @@ bool GSRendererHW9::Create(LPCTSTR title)
 	return true;
 }
 
-void GSRendererHW9::VertexKick(bool skip)
+template<DWORD tme, DWORD fst>
+void GSRendererHW9::AddVertex()
 {
-	Vertex& v = m_vl.AddTail();
+	Vertex& dst = m_vl.AddTail();
 
-	v.p.x = (float)m_v.XYZ.X;
-	v.p.y = (float)m_v.XYZ.Y;
-	v.p.z = (float)m_v.XYZ.Z;
+	dst.p.x = (float)m_v.XYZ.X;
+	dst.p.y = (float)m_v.XYZ.Y;
+	dst.p.z = (float)m_v.XYZ.Z;
 
-	v.c0 = m_v.RGBAQ.ai32[0];
-	v.c1 = m_v.FOG.ai32[1];
+	dst.c0 = m_v.RGBAQ.ai32[0];
+	dst.c1 = m_v.FOG.ai32[1];
 
-	if(PRIM->TME)
+	if(tme)
 	{
-		if(PRIM->FST)
+		if(fst)
 		{
-			GSVector4::storel(&v.t, m_v.GetUV());
+			GSVector4::storel(&dst.t, m_v.GetUV());
 		}
 		else
 		{
-			v.t.x = m_v.ST.S;
-			v.t.y = m_v.ST.T;
-			v.p.w = m_v.RGBAQ.Q;
+			dst.t.x = m_v.ST.S;
+			dst.t.y = m_v.ST.T;
+			dst.p.w = m_v.RGBAQ.Q;
 		}
 	}
-
-	__super::VertexKick(skip);
 }
 
-int GSRendererHW9::ScissorTest(const GSVector4& p0, const GSVector4& p1)
+template<int primclass>
+void GSRendererHW9::AddPrim(Vertex* v, DWORD& count)
 {
 	GSVector4 scissor = m_context->scissor.dx9;
 
-	GSVector4 v0 = p0 < scissor;
-	GSVector4 v1 = p1 > scissor.zwxy();
+	GSVector4 pmin, pmax;
 
-	return (v0 | v1).mask() & 3;
-}
-
-void GSRendererHW9::DrawingKickPoint(Vertex* v, int& count)
-{
-	GSVector4 p0 = v[0].p;
-	GSVector4 p1 = v[0].p;
-
-	if(ScissorTest(p0, p1))
+	switch(primclass)
 	{
-		count = 0;
-		return;
+	case GS_POINT_CLASS:
+		pmin = v[0].p;
+		pmax = v[0].p;
+		break;
+	case GS_LINE_CLASS:
+	case GS_SPRITE_CLASS:
+		pmin = v[0].p.minv(v[1].p);
+		pmax = v[0].p.maxv(v[1].p);
+		break;
+	case GS_TRIANGLE_CLASS:
+		pmin = v[0].p.minv(v[1].p).minv(v[2].p);
+		pmax = v[0].p.maxv(v[1].p).maxv(v[2].p);
+		break;
 	}
-}
 
-void GSRendererHW9::DrawingKickLine(Vertex* v, int& count)
-{
-	GSVector4 p0 = v[0].p.maxv(v[1].p);
-	GSVector4 p1 = v[0].p.minv(v[1].p);
+	GSVector4 test = (pmax < scissor) | (pmin > scissor.zwxy());
 
-	if(ScissorTest(p0, p1))
+	if(test.mask() & 3)
 	{
 		count = 0;
 		return;
 	}
 
-	if(PRIM->IIP == 0)
+	switch(primclass)
 	{
-		v[0].c0 = v[1].c0;
+	case GS_POINT_CLASS:
+		break;
+	case GS_LINE_CLASS:
+		if(PRIM->IIP == 0) {v[0].c0 = v[1].c0;}
+		break;
+	case GS_SPRITE_CLASS:
+		if(PRIM->IIP == 0) {v[0].c0 = v[1].c0;}
+		v[0].p.z = v[1].p.z;
+		v[0].p.w = v[1].p.w;
 		v[0].c1 = v[1].c1;
+		v[2] = v[1];
+		v[3] = v[1];
+		v[1].p.y = v[0].p.y;
+		v[1].t.y = v[0].t.y;
+		v[2].p.x = v[0].p.x;
+		v[2].t.x = v[0].t.x;
+		v[4] = v[1];
+		v[5] = v[2];
+		count += 4;
+		break;
+	case GS_TRIANGLE_CLASS:
+		if(PRIM->IIP == 0) {v[0].c0 = v[1].c0 = v[2].c0;}
+		break;
 	}
 }
 
-void GSRendererHW9::DrawingKickTriangle(Vertex* v, int& count)
-{
-	GSVector4 p0 = v[0].p.maxv(v[1].p).maxv(v[2].p);
-	GSVector4 p1 = v[0].p.minv(v[1].p).minv(v[2].p);
-
-	if(ScissorTest(p0, p1))
-	{
-		count = 0;
-		return;
-	}
-
-	if(PRIM->IIP == 0)
-	{
-		v[0].c0 = v[2].c0;
-		v[0].c1 = v[2].c1;
-	}
-}
-
-void GSRendererHW9::DrawingKickSprite(Vertex* v, int& count)
-{
-	GSVector4 p0 = v[0].p.maxv(v[1].p);
-	GSVector4 p1 = v[0].p.minv(v[1].p);
-
-	if(ScissorTest(p0, p1))
-	{
-		count = 0;
-		return;
-	}
-
-	if(PRIM->IIP == 0)
-	{
-		v[0].c0 = v[1].c0;
-	}
-
-	v[0].p.z = v[1].p.z;
-	v[0].p.w = v[1].p.w;
-	v[0].c1 = v[1].c1;
-	v[2] = v[1];
-	v[3] = v[1];
-	v[1].p.y = v[0].p.y;
-	v[1].t.y = v[0].t.y;
-	v[2].p.x = v[0].p.x;
-	v[2].t.x = v[0].t.x;
-	v[4] = v[1];
-	v[5] = v[2];
-
-	count += 4;
-}
-
 void GSRendererHW9::Draw(int prim, Texture& rt, Texture& ds, GSTextureCache<Device>::GSTexture* tex)
 {
 	GSDrawingEnvironment& env = m_env;
diff --git a/gsdx/GSRendererHW9.h b/gsdx/GSRendererHW9.h
index c976c07..382e39c 100644
--- a/gsdx/GSRendererHW9.h
+++ b/gsdx/GSRendererHW9.h
@@ -36,16 +36,14 @@ protected:
 	GSTextureFX9 m_tfx;
 	bool m_logz;
 
-	void VertexKick(bool skip);
+	template<DWORD tme, DWORD fst>
+	void AddVertex();
+
+	template<int primclass>
+	void AddPrim(Vertex* v, DWORD& count);
+
 	bool WrapZ(float maxz);
 
-	__forceinline int ScissorTest(const GSVector4& p0, const GSVector4& p1);
-
-	void DrawingKickPoint(Vertex* v, int& count);
-	void DrawingKickLine(Vertex* v, int& count);
-	void DrawingKickTriangle(Vertex* v, int& count);
-	void DrawingKickSprite(Vertex* v, int& count);
-
 	void Draw(int prim, Texture& rt, Texture& ds, GSTextureCache<Device>::GSTexture* tex);
 
 	struct
diff --git a/gsdx/GSRendererNull.h b/gsdx/GSRendererNull.h
index be83a60..0c2d3a6 100644
--- a/gsdx/GSRendererNull.h
+++ b/gsdx/GSRendererNull.h
@@ -27,14 +27,11 @@
 template<class Device> class GSRendererNull : public GSRendererT<Device, GSVertexNull>
 {
 protected:
-	void VertexKick(bool skip)
+	void AddVertex()
 	{
-		m_vl.AddTail();
-
-		__super::VertexKick(skip);
 	}
 
-	void DrawingKick(GSVertexNull* v, int& count)
+	void AddPrim(GSVertexNull* v, DWORD& count)
 	{
 		m_perfmon.Put(GSPerfMon::Prim, 1);
 	}
@@ -52,9 +49,14 @@ public:
 	GSRendererNull(BYTE* base, bool mt, void (*irq)(), int nloophack, const GSRendererSettings& rs)
 		: GSRendererT<Device, GSVertexNull>(base, mt, irq, nloophack, rs)
 	{
-		for(int i = 0; i < countof(m_fpDrawingKickHandlers); i++)
+		m_fpAddVertexHandlers[0][0] = (AddVertexHandler)&GSRendererNull::AddVertex;
+		m_fpAddVertexHandlers[0][1] = (AddVertexHandler)&GSRendererNull::AddVertex;
+		m_fpAddVertexHandlers[1][0] = (AddVertexHandler)&GSRendererNull::AddVertex;
+		m_fpAddVertexHandlers[1][1] = (AddVertexHandler)&GSRendererNull::AddVertex;
+
+		for(int i = 0; i < countof(m_fpAddPrimHandlers); i++)
 		{
-			m_fpDrawingKickHandlers[i] = (DrawingKickHandler)&GSRendererNull::DrawingKick;
+			m_fpAddPrimHandlers[i] = (AddPrimHandler)&GSRendererNull::AddPrim;
 		}
 	}
 };
\ No newline at end of file
diff --git a/gsdx/GSRendererSW.h b/gsdx/GSRendererSW.h
index 3beeace..514baea 100644
--- a/gsdx/GSRendererSW.h
+++ b/gsdx/GSRendererSW.h
@@ -117,7 +117,7 @@ protected:
 		return true;
 	}
 
-	void VertexKick(bool skip)
+	template<DWORD tme, DWORD fst> void AddVertex()
 	{
 		GSVertexSW v;
 
@@ -125,15 +125,14 @@ protected:
 		GSVector4i o((int)m_context->XYOFFSET.OFX, (int)m_context->XYOFFSET.OFY);
 
 		v.p = GSVector4(p - o) * g_pos_scale;
-		v.p.z = (float)min(m_v.XYZ.Z, 0xffffff00); // max value which can survive the DWORD=>float=>DWORD conversion
 
-		v.c = GSVector4((DWORD)m_v.RGBAQ.ai32[0]) * 128.0f;
+		v.c = GSVector4(GSVector4i::load((int)m_v.RGBAQ.ai32[0]).u8to32() << 7);
 
-		if(PRIM->TME)
+		if(tme)
 		{
 			float q;
 
-			if(PRIM->FST)
+			if(fst)
 			{
 				v.t = GSVector4(GSVector4i(m_v.UV.U, m_v.UV.V) << (16 - 4));
 				q = 1.0f;
@@ -148,13 +147,14 @@ protected:
 			v.t = v.t.xyxy(GSVector4::load(q));
 		}
 
-		m_vl.AddTail() = v;
+		GSVertexSW& dst = m_vl.AddTail();
 
-		__super::VertexKick(skip);
+		dst = v;
+
+		dst.p.z = (float)min(m_v.XYZ.Z, 0xffffff00); // max value which can survive the DWORD => float => DWORD conversion
 	}
 
-	template<int primclass>
-	void DrawingKick(GSVertexSW* v, int& count)
+	template<int primclass> void AddPrim(GSVertexSW* v, DWORD& count)
 	{
 		GSVector4 pmin, pmax;
 
@@ -213,10 +213,10 @@ protected:
 			break;
 		case GS_TRIANGLE_CLASS:
 			if(PRIM->IIP == 0) {v[0].c = v[2].c; v[1].c = v[2].c;}
-			m_vtrace.min.t = m_vtrace.min.t.minv(v[0].t).minv(v[1].t).minv(v[2].t);
-			m_vtrace.max.t = m_vtrace.max.t.maxv(v[0].t).maxv(v[1].t).maxv(v[2].t);
-			m_vtrace.min.c = m_vtrace.min.c.minv(v[0].c).minv(v[1].c).minv(v[2].c);
-			m_vtrace.max.c = m_vtrace.max.c.maxv(v[0].c).maxv(v[1].c).maxv(v[2].c);
+			m_vtrace.min.t = m_vtrace.min.t.minv(v[0].t).minv(v[1].t.minv(v[2].t));
+			m_vtrace.max.t = m_vtrace.max.t.maxv(v[0].t).maxv(v[1].t.maxv(v[2].t));
+			m_vtrace.min.c = m_vtrace.min.c.minv(v[0].c).minv(v[1].c.minv(v[2].c));
+			m_vtrace.max.c = m_vtrace.max.c.maxv(v[0].c).maxv(v[1].c.maxv(v[2].c));
 			break;
 		case GS_SPRITE_CLASS:
 			m_vtrace.min.t = m_vtrace.min.t.minv(v[0].t).minv(v[1].t);
@@ -457,40 +457,92 @@ protected:
 					{
 						p.sel.fst = 1;
 
-						GSVector4 w = v[0].t.zzzz().rcpnr();
-
-						for(int i = 0, j = m_count; i < j; i++)
+						if(v[0].t.z != 1.0f)
 						{
-							v[i].t *= w;
+							GSVector4 w = v[0].t.zzzz().rcpnr();
+
+							for(int i = 0, j = m_count; i < j; i++)
+							{
+								v[i].t *= w;
+							}
+
+							m_vtrace.min.t *= w;
+							m_vtrace.max.t *= w;
 						}
 					}
 					else if(prim == GS_SPRITE)
 					{
 						p.sel.fst = 1;
 
+						GSVector4 tmin = GSVector4(FLT_MAX);
+						GSVector4 tmax = GSVector4(-FLT_MAX);
+
 						for(int i = 0, j = m_count; i < j; i += 2)
 						{
 							GSVector4 w = v[i + 1].t.zzzz().rcpnr();
 
-							v[i + 0].t *= w;
-							v[i + 1].t *= w;
+							GSVector4 v0 = v[i + 0].t * w;
+							GSVector4 v1 = v[i + 1].t * w;
+
+							v[i + 0].t = v0;
+							v[i + 1].t = v1;
+
+							tmin = tmin.minv(v0).minv(v1);
+							tmax = tmax.maxv(v0).maxv(v1);
 						}
+
+						m_vtrace.max.t = tmax;
+						m_vtrace.min.t = tmin;
 					}
 				}
 
-				if(p.sel.fst && p.sel.ltf)
+				if(p.sel.fst)
 				{
 					// if q is constant we can do the half pel shift for bilinear sampling on the vertices
 
-					GSVertexSW* v = m_vertices;
+					if(p.sel.ltf)
+					{
+						GSVector4 half(0x8000, 0x8000);
 
-					GSVector4 half(0x8000, 0x8000);
+						GSVertexSW* v = m_vertices;
+
+						for(int i = 0, j = m_count; i < j; i++)
+						{
+							v[i].t -= half;
+						}
+
+						m_vtrace.min.t -= half;
+						m_vtrace.max.t += half;
+					}
+				}
+				/*
+				else
+				{
+					GSVector4 tmin = GSVector4(FLT_MAX);
+					GSVector4 tmax = GSVector4(-FLT_MAX);
+
+					GSVertexSW* v = m_vertices;
 
 					for(int i = 0, j = m_count; i < j; i++)
 					{
-						v[i].t -= half;
+						GSVector4 v0 = v[i].t * v[i].t.zzzz().rcpnr();
+
+						tmin = tmin.minv(v0);
+						tmax = tmax.maxv(v0);
 					}
+
+					if(p.sel.ltf)
+					{
+						GSVector4 half(0x8000, 0x8000);
+
+						tmin -= half;
+						tmax += half;
+					}
+
+					m_vtrace.min.t = tmin;
+					m_vtrace.max.t = tmax;
 				}
+				*/
 
 				CRect r;
 				
@@ -522,15 +574,11 @@ protected:
 					p.sel.abe = context->ALPHA.ai32[0];
 					p.sel.pabe = env.PABE.PABE;
 				}
-				else
-				{
-					// printf("opaque\n");
-				}
+			}
 
-				if(PRIM->AA1)
-				{
-					// TODO: automatic alpha blending (ABE=1, A=0 B=1 C=0 D=1)
-				}
+			if(PRIM->AA1)
+			{
+				// TODO: automatic alpha blending (ABE=1, A=0 B=1 C=0 D=1)
 			}
 
 			if(p.sel.date 
@@ -642,7 +690,7 @@ protected:
 		m_tc->InvalidateVideoMem(BITBLTBUF, r);
 	}
 
-	void MinMaxUV(int w, int h, CRect& r, bool fst)
+	void MinMaxUV(int w, int h, CRect& r, DWORD fst)
 	{
 		const GSDrawingContext* context = m_context;
 
@@ -694,16 +742,34 @@ protected:
 
 		if(fst)
 		{
-			GSVector4i v = GSVector4i(m_vtrace.min.t.xyxy(m_vtrace.max.t) / (m_vtrace.min.t.zzzz() * 0x10000));
+			GSVector4i uv = GSVector4i(m_vtrace.min.t.xyxy(m_vtrace.max.t)).sra32(16);
+/*
+			int tw = context->TEX0.TW;
+			int th = context->TEX0.TH;
 
+			GSVector4i u = uv & GSVector4i::xffffffff().srl32(32 - tw);
+			GSVector4i v = uv & GSVector4i::xffffffff().srl32(32 - th);
+
+			GSVector4i uu = uv.sra32(tw);
+			GSVector4i vv = uv.sra32(th);
+
+			int mask = (uu.upl32(vv) == uu.uph32(vv)).mask();
+*/
 			switch(wms)
 			{
-			case CLAMP_REPEAT: // TODO
+			case CLAMP_REPEAT:
+/*
+				if(mask & 0x000f)
+				{
+					if(vr.x < u.x) vr.x = u.x;
+					if(vr.z > u.z + 1) vr.z = u.z + 1;
+				}
+*/
 				break;
 			case CLAMP_CLAMP:
 			case CLAMP_REGION_CLAMP:
-				if(vr.x < v.x) vr.x = v.x;
-				if(vr.z > v.z + 1) vr.z = v.z + 1;
+				if(vr.x < uv.x) vr.x = uv.x;
+				if(vr.z > uv.z + 1) vr.z = uv.z + 1;
 				break;
 			case CLAMP_REGION_REPEAT: // TODO
 				break;
@@ -713,12 +779,19 @@ protected:
 
 			switch(wmt)
 			{
-			case CLAMP_REPEAT: // TODO
+			case CLAMP_REPEAT:
+/*
+				if(mask & 0xf000)
+				{
+					if(vr.y < v.y) vr.y = v.y;
+					if(vr.w > v.w + 1) vr.w = v.w + 1;
+				}
+*/
 				break;
 			case CLAMP_CLAMP:
 			case CLAMP_REGION_CLAMP:
-				if(vr.y < v.y) vr.y = v.y;
-				if(vr.w > v.w + 1) vr.w = v.w + 1;
+				if(vr.y < uv.y) vr.y = uv.y;
+				if(vr.w > uv.w + 1) vr.w = uv.w + 1;
 				break;
 			case CLAMP_REGION_REPEAT: // TODO
 				break;
@@ -740,13 +813,18 @@ public:
 
 		m_tc = new GSTextureCacheSW(this);
 
-		m_fpDrawingKickHandlers[GS_POINTLIST] = (DrawingKickHandler)&GSRendererSW::DrawingKick<GS_POINT_CLASS>;
-		m_fpDrawingKickHandlers[GS_LINELIST] = (DrawingKickHandler)&GSRendererSW::DrawingKick<GS_LINE_CLASS>;
-		m_fpDrawingKickHandlers[GS_LINESTRIP] = (DrawingKickHandler)&GSRendererSW::DrawingKick<GS_LINE_CLASS>;
-		m_fpDrawingKickHandlers[GS_TRIANGLELIST] = (DrawingKickHandler)&GSRendererSW::DrawingKick<GS_TRIANGLE_CLASS>;
-		m_fpDrawingKickHandlers[GS_TRIANGLESTRIP] = (DrawingKickHandler)&GSRendererSW::DrawingKick<GS_TRIANGLE_CLASS>;
-		m_fpDrawingKickHandlers[GS_TRIANGLEFAN] = (DrawingKickHandler)&GSRendererSW::DrawingKick<GS_TRIANGLE_CLASS>;
-		m_fpDrawingKickHandlers[GS_SPRITE] = (DrawingKickHandler)&GSRendererSW::DrawingKick<GS_SPRITE_CLASS>;
+		m_fpAddVertexHandlers[0][0] = (AddVertexHandler)&GSRendererSW::AddVertex<0, 0>;
+		m_fpAddVertexHandlers[0][1] = (AddVertexHandler)&GSRendererSW::AddVertex<0, 0>;
+		m_fpAddVertexHandlers[1][0] = (AddVertexHandler)&GSRendererSW::AddVertex<1, 0>;
+		m_fpAddVertexHandlers[1][1] = (AddVertexHandler)&GSRendererSW::AddVertex<1, 1>;
+
+		m_fpAddPrimHandlers[GS_POINTLIST] = (AddPrimHandler)&GSRendererSW::AddPrim<GS_POINT_CLASS>;
+		m_fpAddPrimHandlers[GS_LINELIST] = (AddPrimHandler)&GSRendererSW::AddPrim<GS_LINE_CLASS>;
+		m_fpAddPrimHandlers[GS_LINESTRIP] = (AddPrimHandler)&GSRendererSW::AddPrim<GS_LINE_CLASS>;
+		m_fpAddPrimHandlers[GS_TRIANGLELIST] = (AddPrimHandler)&GSRendererSW::AddPrim<GS_TRIANGLE_CLASS>;
+		m_fpAddPrimHandlers[GS_TRIANGLESTRIP] = (AddPrimHandler)&GSRendererSW::AddPrim<GS_TRIANGLE_CLASS>;
+		m_fpAddPrimHandlers[GS_TRIANGLEFAN] = (AddPrimHandler)&GSRendererSW::AddPrim<GS_TRIANGLE_CLASS>;
+		m_fpAddPrimHandlers[GS_SPRITE] = (AddPrimHandler)&GSRendererSW::AddPrim<GS_SPRITE_CLASS>;
 	}
 
 	virtual ~GSRendererSW()
diff --git a/gsdx/GSState.cpp b/gsdx/GSState.cpp
index 6d94f5a..09a75ab 100644
--- a/gsdx/GSState.cpp
+++ b/gsdx/GSState.cpp
@@ -137,8 +137,6 @@ void GSState::Reset()
 
 	m_context = &m_env.CTXT[0];
 
-	m_vprim = GSUtil::GetPrimVertexCount(PRIM->PRIM);
-
 	InvalidateTextureCache();
 }
 
@@ -518,8 +516,6 @@ void GSState::GIFRegHandlerPRIM(GIFReg* r)
 
 	m_context = &m_env.CTXT[PRIM->CTXT];
 
-	m_vprim = GSUtil::GetPrimVertexCount(PRIM->PRIM);
-
 	ResetPrim();
 }
 
@@ -677,27 +673,6 @@ void GSState::GIFRegHandlerPRMODECONT(GIFReg* r)
 	if(PRIM->PRIM == 7) TRACE(_T("Invalid PRMODECONT/PRIM\n"));
 
 	m_context = &m_env.CTXT[PRIM->CTXT];
-
-	m_vprim = GSUtil::GetPrimVertexCount(PRIM->PRIM);
-/*
-	if(m_env.PRMODECONT.AC != r->PRMODECONT.AC)
-	{
-		if(m_env.PRIM.ai32[0] != m_env.PRMODE.ai32[0])
-		{
-			Flush();
-		}
-
-		m_env.PRMODECONT.AC = r->PRMODECONT.AC;
-
-		PRIM = m_env.PRMODECONT.AC ? &m_env.PRIM : (GIFRegPRIM*)&m_env.PRMODE;
-
-		if(PRIM->PRIM == 7) TRACE(_T("Invalid PRMODECONT/PRIM\n"));
-
-		m_context = &m_env.CTXT[PRIM->CTXT];
-
-		m_vprim = GSUtil::GetPrimVertexCount(PRIM->PRIM);
-	}
-*/
 }
 
 void GSState::GIFRegHandlerPRMODE(GIFReg* r)
@@ -1584,8 +1559,6 @@ int GSState::Defrost(const GSFreezeData* fd)
 
 	m_context = &m_env.CTXT[PRIM->CTXT];
 
-	m_vprim = GSUtil::GetPrimVertexCount(PRIM->PRIM);
-
 	m_env.CTXT[0].UpdateScissor();
 	m_env.CTXT[1].UpdateScissor();
 
diff --git a/gsdx/GSState.h b/gsdx/GSState.h
index f9377dd..e2e0d35 100644
--- a/gsdx/GSState.h
+++ b/gsdx/GSState.h
@@ -125,6 +125,15 @@ class GSState : public GSAlignedClass<16>
 protected:
 	bool IsBadFrame(int& skip);
 
+	typedef void (GSState::*VertexKickHandler)(bool skip);
+
+	VertexKickHandler m_fpVertexKickHandlers[8];
+
+	void VertexKick(bool skip)
+	{
+		(this->*m_fpVertexKickHandlers[PRIM->PRIM])(skip);
+	}
+
 public:
 	GIFRegPRIM*		PRIM;
 	GSRegPMODE*		PMODE;
@@ -184,7 +193,6 @@ public:
 	virtual void Flush();
 	virtual void FlushPrim() = 0;
 	virtual void ResetPrim() = 0;
-	virtual void VertexKick(bool skip) = 0;
 	virtual void InvalidateVideoMem(const GIFRegBITBLTBUF& BITBLTBUF, CRect r) {}
 	virtual void InvalidateLocalMem(const GIFRegBITBLTBUF& BITBLTBUF, CRect r) {}
 	virtual void InvalidateTextureCache() {}
diff --git a/gsdx/GSUtil.cpp b/gsdx/GSUtil.cpp
index dacd4ab..bba238a 100644
--- a/gsdx/GSUtil.cpp
+++ b/gsdx/GSUtil.cpp
@@ -27,7 +27,6 @@
 static struct GSUtilMaps
 {
 	BYTE PrimClassField[8];
-	BYTE PrimVertexCount[8];
 	bool CompatibleBitsField[64][64];
 	bool SharedBitsField[64][64];
 
@@ -42,15 +41,6 @@ static struct GSUtilMaps
 		PrimClassField[GS_SPRITE] = GS_SPRITE_CLASS;
 		PrimClassField[GS_INVALID] = GS_INVALID_CLASS;
 
-		PrimVertexCount[GS_POINTLIST] = 1;
-		PrimVertexCount[GS_LINELIST] = 2;
-		PrimVertexCount[GS_LINESTRIP] = 2;
-		PrimVertexCount[GS_TRIANGLELIST] = 3;
-		PrimVertexCount[GS_TRIANGLESTRIP] = 3;
-		PrimVertexCount[GS_TRIANGLEFAN] = 3;
-		PrimVertexCount[GS_SPRITE] = 2;
-		PrimVertexCount[GS_INVALID] = 1;
-
 		memset(CompatibleBitsField, 0, sizeof(CompatibleBitsField));
 
 		CompatibleBitsField[PSM_PSMCT32][PSM_PSMCT24] = true;
@@ -87,11 +77,6 @@ GS_PRIM_CLASS GSUtil::GetPrimClass(DWORD prim)
 	return (GS_PRIM_CLASS)s_maps.PrimClassField[prim];
 }
 
-DWORD GSUtil::GetPrimVertexCount(DWORD prim)
-{
-	return s_maps.PrimVertexCount[prim];
-}
-
 bool GSUtil::HasSharedBits(DWORD spsm, DWORD dpsm)
 {
 	return s_maps.SharedBitsField[spsm][dpsm];
diff --git a/gsdx/GSUtil.h b/gsdx/GSUtil.h
index 4c3c495..973111f 100644
--- a/gsdx/GSUtil.h
+++ b/gsdx/GSUtil.h
@@ -27,7 +27,6 @@ class GSUtil
 {
 public:
 	static GS_PRIM_CLASS GetPrimClass(DWORD prim);
-	static DWORD GetPrimVertexCount(DWORD prim);
 
 	static bool HasSharedBits(DWORD spsm, DWORD dpsm);
 	static bool HasSharedBits(DWORD sbp, DWORD spsm, DWORD dbp, DWORD dpsm);
diff --git a/gsdx/GSVector.h b/gsdx/GSVector.h
index 4111418..9d7bcfc 100644
--- a/gsdx/GSVector.h
+++ b/gsdx/GSVector.h
@@ -452,7 +452,15 @@ public:
 
 	GSVector4i upl8() const
 	{
+		#if 0 // _M_SSE >= 0x401 // TODO: compiler bug
+
+		return GSVector4i(_mm_cvtepu8_epi16(m));
+		
+		#else
+
 		return GSVector4i(_mm_unpacklo_epi8(m, _mm_setzero_si128()));
+
+		#endif
 	}
 
 	GSVector4i uph8() const
@@ -462,7 +470,15 @@ public:
 
 	GSVector4i upl16() const
 	{
+		#if 0 //_M_SSE >= 0x401 // TODO: compiler bug
+
+		return GSVector4i(_mm_cvtepu16_epi32(m));
+		
+		#else
+
 		return GSVector4i(_mm_unpacklo_epi16(m, _mm_setzero_si128()));
+
+		#endif
 	}
 
 	GSVector4i uph16() const
@@ -472,7 +488,15 @@ public:
 
 	GSVector4i upl32() const
 	{
+		#if 0 //_M_SSE >= 0x401 // TODO: compiler bug
+
+		return GSVector4i(_mm_cvtepu32_epi64(m));
+		
+		#else
+
 		return GSVector4i(_mm_unpacklo_epi32(m, _mm_setzero_si128()));
+
+		#endif
 	}
 
 	GSVector4i uph32() const
@@ -492,6 +516,11 @@ public:
 
 	#if _M_SSE >= 0x401
 
+	// WARNING!!!
+	//
+	// MSVC (2008, 2010 ctp) believes that there is a "mem, reg" form of the pmovz/sx* instructions,
+	// turning these intrinsics into a minefield, don't spill regs when using them...
+
 	GSVector4i i8to16() const
 	{
 		return GSVector4i(_mm_cvtepi8_epi16(m));
diff --git a/gsdx/GSVertex.h b/gsdx/GSVertex.h
index 06fdf74..c1b52ac 100644
--- a/gsdx/GSVertex.h
+++ b/gsdx/GSVertex.h
@@ -48,7 +48,7 @@ __declspec(align(16)) struct GSVertex
 
 	GSVertex() {memset(this, 0, sizeof(*this));}
 
-	GSVector4 GetUV() {return GSVector4(GSVector4i::load(UV.ai32[0]).upl16());}
+	GSVector4 GetUV() const {return GSVector4(GSVector4i::load(UV.ai32[0]).upl16());}
 };
 
 struct GSVertexOld
diff --git a/gsdx/res/tfx10.fx b/gsdx/res/tfx10.fx
index 80bf578..7be092d 100644
--- a/gsdx/res/tfx10.fx
+++ b/gsdx/res/tfx10.fx
@@ -103,7 +103,6 @@ void gs_main(point VS_OUTPUT input[1], inout PointStream<VS_OUTPUT> stream)
 void gs_main(line VS_OUTPUT input[2], inout LineStream<VS_OUTPUT> stream)
 {
 	#if IIP == 0
-	input[0].t.z = input[1].t.z;
 	input[0].c = input[1].c;
 	#endif
 
@@ -117,9 +116,7 @@ void gs_main(line VS_OUTPUT input[2], inout LineStream<VS_OUTPUT> stream)
 void gs_main(triangle VS_OUTPUT input[3], inout TriangleStream<VS_OUTPUT> stream)
 {
 	#if IIP == 0
-	input[0].t.z = input[2].t.z;
 	input[0].c = input[2].c;
-	input[1].t.z = input[2].t.z;
 	input[1].c = input[2].c;
 	#endif