GSdx: just squeezing a few more fps.

git-svn-id: http://pcsx2.googlecode.com/svn/trunk@1534 96395faa-99c1-11dd-bbfe-3dabce05a288

plugins/GSdx/GPURenderer.cpp

@@ -162,7 +162,7 @@ void GPURenderer::VSync()
 	
 	GetClientRect(m_hWnd, r);
 
-	m_dev->Present(r.fit(m_aspectratio), 0);
+	m_dev->Present(r.fit(m_aspectratio), 0, true);
 }
 
 bool GPURenderer::MakeSnapshot(const string& path)

plugins/GSdx/GS.cpp

@@ -303,7 +303,7 @@ EXPORT_C GSsetFrameSkip(int frameskip)
 
 EXPORT_C GSsetFrameLimit(int limit)
 {
-	s_gs->SetFrameLimit(limit);
+	s_gs->SetFrameLimit(limit != 0);
 }
 
 #ifdef _WINDOWS

plugins/GSdx/GS.h

@@ -815,7 +815,7 @@ REG64_(GIFReg, TEX1)
 	uint32 _PAD2:9;
 	uint32 L:2;
 	uint32 _PAD3:11;
-	uint32 K:12;
+	uint32 K:12; // TODO: 1:7:4 (signed? 16x scaled?)
 	uint32 _PAD4:20;
 REG_END2
 

plugins/GSdx/GSClut.cpp

@@ -314,43 +314,56 @@ void GSClut::GetAlphaMinMax32(int& amin, int& amax)
 	{
 		m_read.adirty = false;
 
-		// uint32 bpp = GSLocalMemory::m_psm[m_read.TEX0.PSM].trbpp;
+		if(GSLocalMemory::m_psm[m_read.TEX0.CPSM].trbpp == 24 && m_read.TEXA.AEM == 0)
-		uint32 cbpp = GSLocalMemory::m_psm[m_read.TEX0.CPSM].trbpp;
-		uint32 pal = GSLocalMemory::m_psm[m_read.TEX0.PSM].pal;
-
-		if(cbpp == 24 && m_read.TEXA.AEM == 0)
 		{
 			m_read.amin = m_read.TEXA.TA0;
 			m_read.amax = m_read.TEXA.TA0;
 		}
 		else
 		{
-			int amin = 255;
-			int amax = 0;
-
 			const GSVector4i* p = (const GSVector4i*)m_buff32;
 
-			for(int i = 0, j = pal >> 4; i < j; i++)
+			GSVector4i amin, amax;
+
+			if(GSLocalMemory::m_psm[m_read.TEX0.PSM].pal == 256)
+			{
+				amin = GSVector4i::xffffffff();
+				amax = GSVector4i::zero();
+
+				for(int i = 0; i < 16; i++)
 				{
 					GSVector4i v0 = (p[i * 4 + 0] >> 24).ps32(p[i * 4 + 1] >> 24);
 					GSVector4i v1 = (p[i * 4 + 2] >> 24).ps32(p[i * 4 + 3] >> 24);
+					GSVector4i v2 = v0.pu16(v1);
 
-				GSVector4i v2 = v0.min_i16(v1);
+					amin = amin.min_u8(v2);
-				GSVector4i v3 = v0.max_i16(v1);
+					amax = amax.max_u8(v2);
+				}
+			}
+			else
+			{
+				ASSERT(GSLocalMemory::m_psm[m_read.TEX0.PSM].pal == 16);
 
-				v2 = v2.min_i16(v2.zwxy());
+				GSVector4i v0 = (p[0] >> 24).ps32(p[1] >> 24);
-				v3 = v3.max_i16(v3.zwxy());
+				GSVector4i v1 = (p[2] >> 24).ps32(p[3] >> 24);
-				v2 = v2.min_i16(v2.zwxyl());
+				GSVector4i v2 = v0.pu16(v1);
-				v3 = v3.max_i16(v3.zwxyl());
-				v2 = v2.min_i16(v2.yxwzl());
-				v3 = v3.max_i16(v3.yxwzl());
 
-				amin = min(amin, v2.extract16<0>());
+				amin = v2;
-				amax = max(amax, v3.extract16<0>());
+				amax = v2;
 			}
 
-			m_read.amin = amin;
+			amin = amin.min_u8(amin.zwxy());
-			m_read.amax = amax;
+			amax = amax.max_u8(amax.zwxy());
+			amin = amin.min_u8(amin.zwxyl());
+			amax = amax.max_u8(amax.zwxyl());
+			amin = amin.min_u8(amin.yxwzl());
+			amax = amax.max_u8(amax.yxwzl());
+
+			GSVector4i v0 = amin.upl8(amax).u8to16();
+			GSVector4i v1 = v0.yxwz();
+
+			m_read.amin = v0.min_i16(v1).extract16<0>();
+			m_read.amax = v0.max_i16(v1).extract16<1>();
 		}
 	}
 

plugins/GSdx/GSCrc.cpp

@@ -89,6 +89,7 @@ CRC::Game CRC::m_games[] =
 	{0x7D8F539A, SoTC, EU, 0},
 	{0x3122B508, OnePieceGrandAdventure, US, 0},
 	{0x8DF14A24, OnePieceGrandAdventure, Unknown, 0},
+	{0xB049DD5E, OnePieceGrandBattle, US, 0},
 	{0x5D02CC5B, OnePieceGrandBattle, Unknown, 0},
 	{0x6F8545DB, ICO, US, 0},
 	{0xB01A4C95, ICO, JP, 0},

plugins/GSdx/GSDevice.cpp

@@ -75,14 +75,14 @@ bool GSDevice::Reset(int w, int h, int mode)
 	return true;
 }
 
-void GSDevice::Present(const GSVector4i& r, int shader)
+void GSDevice::Present(const GSVector4i& r, int shader, bool limit)
 {
 	GSVector4i cr = m_wnd->GetClientRect();
 
 	int w = std::max(cr.width(), 1);
 	int h = std::max(cr.height(), 1);
 
-	if(!m_backbuffer || m_backbuffer->GetWidth() != w || m_backbuffer->GetHeight() != h)
+	if(!m_backbuffer || m_backbuffer->m_size.x != w || m_backbuffer->m_size.y != h)
 	{
 		if(!Reset(w, h, DontCare))
 		{
@@ -99,7 +99,7 @@ void GSDevice::Present(const GSVector4i& r, int shader)
 		StretchRect(m_current, m_backbuffer, GSVector4(r), s_shader[shader]);
 	}
 
-	Flip();
+	Flip(limit);
 }
 
 GSTexture* GSDevice::Fetch(int type, int w, int h, int format)
@@ -240,7 +240,7 @@ bool GSDevice::ResizeTexture(GSTexture** t, int w, int h)
 
 	GSTexture* t2 = *t;
 
-	if(t2 == NULL || t2->GetWidth() != w || t2->GetHeight() != h)
+	if(t2 == NULL || t2->m_size.x != w || t2->m_size.y != h)
 	{
 		delete t2;
 
@@ -251,8 +251,3 @@ bool GSDevice::ResizeTexture(GSTexture** t, int w, int h)
 
 	return t2 != NULL;
 }
-
-void GSDevice::SetVSync(bool vsync)
-{
-	m_vsync = vsync;
-}

plugins/GSdx/GSDevice.h

@@ -79,8 +79,8 @@ public:
 	virtual bool Create(GSWnd* wnd, bool vsync);
 	virtual bool Reset(int w, int h, int mode);
 	virtual bool IsLost(bool update = false) {return false;}
-	virtual void Present(const GSVector4i& r, int shader);
+	virtual void Present(const GSVector4i& r, int shader, bool limit);
-	virtual void Flip() {}
+	virtual void Flip(bool limit) {}
 
 	virtual void BeginScene() {}
 	virtual void DrawPrimitive() {};
@@ -103,8 +103,6 @@ public:
 	virtual void StretchRect(GSTexture* st, GSTexture* dt, const GSVector4& dr, int shader = 0, bool linear = true);
 	virtual void StretchRect(GSTexture* st, const GSVector4& sr, GSTexture* dt, const GSVector4& dr, int shader = 0, bool linear = true) {}
 
-	virtual void SetVSync(bool vsync);
-
 	GSTexture* GetCurrent();
 
 	void Merge(GSTexture* st[2], GSVector4* sr, GSVector4* dr, const GSVector2i& fs, bool slbg, bool mmod, const GSVector4& c);

plugins/GSdx/GSDevice10.cpp

@@ -231,9 +231,9 @@ bool GSDevice10::Reset(int w, int h, int mode)
 	return true;
 }
 
-void GSDevice10::Flip()
+void GSDevice10::Flip(bool limit)
 {
-	m_swapchain->Present(m_vsync ? 1 : 0, 0);
+	m_swapchain->Present(m_vsync && limit ? 1 : 0, 0);
 }
 
 void GSDevice10::BeginScene()
@@ -456,7 +456,7 @@ void GSDevice10::StretchRect(GSTexture* st, const GSVector4& sr, GSTexture* dt,
 
 	// rs
 
-	RSSet(ds.x, ds.y);
+	RSSet(ds);
 
 	//
 
@@ -665,9 +665,9 @@ void GSDevice10::PSSetSamplerState(ID3D10SamplerState* ss0, ID3D10SamplerState*
 	}
 }
 
-void GSDevice10::RSSet(int width, int height, const GSVector4i* scissor)
+void GSDevice10::RSSet(const GSVector2i& size, const GSVector4i* scissor)
 {
-	if(m_viewport.x != width || m_viewport.y != height)
+	if(m_viewport != size)
 	{
 		D3D10_VIEWPORT vp;
 
@@ -675,17 +675,17 @@ void GSDevice10::RSSet(int width, int height, const GSVector4i* scissor)
 		
 		vp.TopLeftX = 0;
 		vp.TopLeftY = 0;
-		vp.Width = width;
+		vp.Width = size.x;
-		vp.Height = height;
+		vp.Height = size.y;
 		vp.MinDepth = 0.0f;
 		vp.MaxDepth = 1.0f;
 
 		m_dev->RSSetViewports(1, &vp);
 
-		m_viewport = GSVector2i(width, height);
+		m_viewport = size;
 	}
 
-	GSVector4i r = scissor ? *scissor : GSVector4i(0, 0, width, height);
+	GSVector4i r = scissor ? *scissor : GSVector4i(size).zwxy();
 
 	if(!m_scissor.eq(r))
 	{

plugins/GSdx/GSDevice10.h

@@ -97,7 +97,7 @@ public:
 
 	bool Create(GSWnd* wnd, bool vsync);
 	bool Reset(int w, int h, int mode);
-	void Flip();
+	void Flip(bool limit);
 
 	void BeginScene();
 	void DrawPrimitive();
@@ -130,7 +130,7 @@ public:
 	void PSSetShaderResources(GSTexture* sr0, GSTexture* sr1);
 	void PSSetShader(ID3D10PixelShader* ps, ID3D10Buffer* ps_cb);
 	void PSSetSamplerState(ID3D10SamplerState* ss0, ID3D10SamplerState* ss1);
-	void RSSet(int width, int height, const GSVector4i* scissor = NULL);
+	void RSSet(const GSVector2i& size, const GSVector4i* scissor = NULL);
 	void OMSetDepthStencilState(ID3D10DepthStencilState* dss, uint8 sref);
 	void OMSetBlendState(ID3D10BlendState* bs, float bf);
 	void OMSetRenderTargets(GSTexture* rt, GSTexture* ds);

plugins/GSdx/GSDevice11.cpp

@@ -268,9 +268,9 @@ bool GSDevice11::Reset(int w, int h, int mode)
 	return true;
 }
 
-void GSDevice11::Flip()
+void GSDevice11::Flip(bool limit)
 {
-	m_swapchain->Present(m_vsync ? 1 : 0, 0);
+	m_swapchain->Present(m_vsync && limit ? 1 : 0, 0);
 }
 
 void GSDevice11::BeginScene()
@@ -493,7 +493,7 @@ void GSDevice11::StretchRect(GSTexture* st, const GSVector4& sr, GSTexture* dt,
 
 	// rs
 
-	RSSet(ds.x, ds.y);
+	RSSet(ds);
 
 	//
 
@@ -702,9 +702,9 @@ void GSDevice11::PSSetSamplerState(ID3D11SamplerState* ss0, ID3D11SamplerState*
 	}
 }
 
-void GSDevice11::RSSet(int width, int height, const GSVector4i* scissor)
+void GSDevice11::RSSet(const GSVector2i& size, const GSVector4i* scissor)
 {
-	if(m_viewport.x != width || m_viewport.y != height)
+	if(m_viewport != size)
 	{
 		D3D11_VIEWPORT vp;
 
@@ -712,17 +712,17 @@ void GSDevice11::RSSet(int width, int height, const GSVector4i* scissor)
 		
 		vp.TopLeftX = 0;
 		vp.TopLeftY = 0;
-		vp.Width = width;
+		vp.Width = size.x;
-		vp.Height = height;
+		vp.Height = size.y;
 		vp.MinDepth = 0.0f;
 		vp.MaxDepth = 1.0f;
 
 		m_ctx->RSSetViewports(1, &vp);
 
-		m_viewport = GSVector2i(width, height);
+		m_viewport = size;
 	}
 
-	GSVector4i r = scissor ? *scissor : GSVector4i(0, 0, width, height);
+	GSVector4i r = scissor ? *scissor : GSVector4i(size).zwxy();
 
 	if(!m_scissor.eq(r))
 	{

plugins/GSdx/GSDevice11.h

@@ -100,7 +100,7 @@ public:
 
 	bool Create(GSWnd* wnd, bool vsync);
 	bool Reset(int w, int h, int mode);
-	void Flip();
+	void Flip(bool limit);
 
 	void BeginScene();
 	void DrawPrimitive();
@@ -133,7 +133,7 @@ public:
 	void PSSetShaderResources(GSTexture* sr0, GSTexture* sr1);
 	void PSSetShader(ID3D11PixelShader* ps, ID3D11Buffer* ps_cb);
 	void PSSetSamplerState(ID3D11SamplerState* ss0, ID3D11SamplerState* ss1);
-	void RSSet(int width, int height, const GSVector4i* scissor = NULL);
+	void RSSet(const GSVector2i& size, const GSVector4i* scissor = NULL);
 	void OMSetDepthStencilState(ID3D11DepthStencilState* dss, uint8 sref);
 	void OMSetBlendState(ID3D11BlendState* bs, float bf);
 	void OMSetRenderTargets(GSTexture* rt, GSTexture* ds);

plugins/GSdx/GSDevice7.cpp

@@ -136,7 +136,7 @@ bool GSDevice7::Reset(int w, int h, int mode)
 	return true;
 }
 
-void GSDevice7::Present(const GSVector4i& r, int shader)
+void GSDevice7::Present(const GSVector4i& r, int shader, bool limit)
 {
 	HRESULT hr;
 
@@ -145,7 +145,7 @@ void GSDevice7::Present(const GSVector4i& r, int shader)
 	int w = std::max(cr.width(), 1);
 	int h = std::max(cr.height(), 1);
 
-	if(!m_backbuffer || m_backbuffer->GetWidth() != w || m_backbuffer->GetHeight() != h)
+	if(!m_backbuffer || m_backbuffer->m_size.x != w || m_backbuffer->m_size.y != h)
 	{
 		if(!Reset(w, h, DontCare))
 		{
@@ -177,7 +177,7 @@ void GSDevice7::Present(const GSVector4i& r, int shader)
 
 	MapWindowPoints((HWND)m_wnd->GetHandle(), HWND_DESKTOP, (POINT*)&r2, 2);
 	
-	if(m_vsync)
+	if(m_vsync && limit)
 	{
 		hr = m_dd->WaitForVerticalBlank(DDWAITVB_BLOCKBEGIN, NULL);
 	}

plugins/GSdx/GSDevice7.h

@@ -43,5 +43,5 @@ public:
 	bool Create(GSWnd* wnd, bool vsync);
 	bool Reset(int w, int h, int mode);
 	bool IsLost(bool update) {return m_lost;}
-	void Present(const GSVector4i& r, int shader);
+	void Present(const GSVector4i& r, int shader, bool limit);
 };

plugins/GSdx/GSDevice9.cpp

@@ -348,7 +348,7 @@ bool GSDevice9::IsLost(bool update)
 	return m_lost;
 }
 
-void GSDevice9::Flip()
+void GSDevice9::Flip(bool limit)
 {
 	m_dev->EndScene();
 
@@ -427,7 +427,7 @@ void GSDevice9::ClearRenderTarget(GSTexture* rt, uint32 c)
 
 void GSDevice9::ClearDepth(GSTexture* t, float c)
 {
-	GSTexture* rt = CreateRenderTarget(t->GetWidth(), t->GetHeight());
+	GSTexture* rt = CreateRenderTarget(t->m_size.x, t->m_size.y);
 
 	CComPtr<IDirect3DSurface9> rtsurface;
 	CComPtr<IDirect3DSurface9> dssurface;
@@ -448,7 +448,7 @@ void GSDevice9::ClearDepth(GSTexture* t, float c)
 
 void GSDevice9::ClearStencil(GSTexture* t, uint8 c)
 {
-	GSTexture* rt = CreateRenderTarget(t->GetWidth(), t->GetHeight());
+	GSTexture* rt = CreateRenderTarget(t->m_size.x, t->m_size.y);
 
 	CComPtr<IDirect3DSurface9> rtsurface;
 	CComPtr<IDirect3DSurface9> dssurface;
@@ -637,7 +637,7 @@ void GSDevice9::StretchRect(GSTexture* st, const GSVector4& sr, GSTexture* dt, c
 
 	// rs
 
-	RSSet(ds.x, ds.y);
+	RSSet(ds);
 
 	//
 
@@ -865,9 +865,9 @@ void GSDevice9::PSSetSamplerState(Direct3DSamplerState9* ss)
 	}
 }
 
-void GSDevice9::RSSet(int width, int height, const GSVector4i* scissor)
+void GSDevice9::RSSet(const GSVector2i& size, const GSVector4i* scissor)
 {
-	GSVector4i r = scissor ? *scissor : GSVector4i(0, 0, width, height);
+	GSVector4i r = scissor ? *scissor : GSVector4i(size).zwxy();
 
 	if(!m_scissor.eq(r))
 	{

plugins/GSdx/GSDevice9.h

@@ -135,7 +135,7 @@ public:
 	bool Create(GSWnd* wnd, bool vsync);
 	bool Reset(int w, int h, int mode);
 	bool IsLost(bool update);
-	void Flip();
+	void Flip(bool limit);
 
 	void BeginScene();
 	void DrawPrimitive();
@@ -167,7 +167,7 @@ public:
 	void PSSetShaderResources(GSTexture* sr0, GSTexture* sr1);
 	void PSSetShader(IDirect3DPixelShader9* ps, const float* ps_cb, int ps_cb_len);
 	void PSSetSamplerState(Direct3DSamplerState9* ss);
-	void RSSet(int width, int height, const GSVector4i* scissor = NULL);
+	void RSSet(const GSVector2i& size, const GSVector4i* scissor = NULL);
 	void OMSetDepthStencilState(Direct3DDepthStencilState9* dss);
 	void OMSetBlendState(Direct3DBlendState9* bs, uint32 bf);
 	void OMSetRenderTargets(GSTexture* rt, GSTexture* ds);

plugins/GSdx/GSDeviceOGL.cpp

@@ -185,16 +185,16 @@ bool GSDeviceOGL::Reset(int w, int h, int mode)
 	return true;
 }
 
-void GSDeviceOGL::Present(const GSVector4i& r, int shader)
+void GSDeviceOGL::Present(const GSVector4i& r, int shader, bool limit)
 {
 	glBindFramebuffer(GL_FRAMEBUFFER, 0); CheckError();
 
 	// TODO: m_current => backbuffer
 
-	Flip();
+	Flip(limit);
 }
 
-void GSDeviceOGL::Flip()
+void GSDeviceOGL::Flip(bool limit)
 {
 	#ifdef _WINDOWS
 
@@ -438,16 +438,16 @@ void GSDeviceOGL::PSSetSamplerState(SamplerStateOGL* ss)
 	}
 }
 
-void GSDeviceOGL::RSSet(int width, int height, const GSVector4i* scissor)
+void GSDeviceOGL::RSSet(const GSVector2i& size, const GSVector4i* scissor)
 {
-	if(m_viewport.x != width || m_viewport.y != height)
+	if(m_viewport != size)
 	{
-		glViewport(0, 0, width, height); CheckError();
+		glViewport(0, 0, size.x, size.y); CheckError();
 
-		m_viewport = GSVector2i(width, height);
+		m_viewport = size;
 	}
 
-	GSVector4i r = scissor ? *scissor : GSVector4i(0, 0, width, height);
+	GSVector4i r = scissor ? *scissor : GSVector4i(size).zwxy();
 
 	if(!m_scissor.eq(r))
 	{

plugins/GSdx/GSDeviceOGL.h

@@ -111,8 +111,8 @@ public:
 
 	bool Create(GSWnd* wnd, bool vsync);
 	bool Reset(int w, int h, int mode);
-	void Present(const GSVector4i& r, int shader);
+	void Present(const GSVector4i& r, int shader, bool limit);
-	void Flip();
+	void Flip(bool limit);
 
 	void BeginScene();
 	void DrawPrimitive();
@@ -139,7 +139,7 @@ public:
 	void IASetPrimitiveTopology(int topology);
 
 	void PSSetSamplerState(SamplerStateOGL* ss);
-	void RSSet(int width, int height, const GSVector4i* scissor);
+	void RSSet(const GSVector2i& size, const GSVector4i* scissor);
 	void OMSetDepthStencilState(DepthStencilStateOGL* dss);
 	void OMSetBlendState(BlendStateOGL* bs, float bf);
 	void OMSetRenderTargets(GSTexture* rt, GSTexture* ds);

plugins/GSdx/GSLocalMemory.cpp

@@ -66,14 +66,14 @@ int GSLocalMemory::rowOffset16SZ[2048];
 int GSLocalMemory::rowOffset8[2][2048];
 int GSLocalMemory::rowOffset4[2][2048];
 
-int GSLocalMemory::blockOffset32[256];
+short GSLocalMemory::blockOffset32[256];
-int GSLocalMemory::blockOffset32Z[256];
+short GSLocalMemory::blockOffset32Z[256];
-int GSLocalMemory::blockOffset16[256];
+short GSLocalMemory::blockOffset16[256];
-int GSLocalMemory::blockOffset16S[256];
+short GSLocalMemory::blockOffset16S[256];
-int GSLocalMemory::blockOffset16Z[256];
+short GSLocalMemory::blockOffset16Z[256];
-int GSLocalMemory::blockOffset16SZ[256];
+short GSLocalMemory::blockOffset16SZ[256];
-int GSLocalMemory::blockOffset8[256];
+short GSLocalMemory::blockOffset8[256];
-int GSLocalMemory::blockOffset4[256];
+short GSLocalMemory::blockOffset4[256];
 
 //
 
@@ -159,42 +159,42 @@ GSLocalMemory::GSLocalMemory()
 
 	for(int x = 0; x < countof(blockOffset32); x++)
 	{
-		blockOffset32[x] = (int)BlockNumber32(x << 3, 0, 0, 32) - (int)BlockNumber32(0, 0, 0, 32);
+		blockOffset32[x] = (short)((int)BlockNumber32(x << 3, 0, 0, 32) - (int)BlockNumber32(0, 0, 0, 32));
 	}
 
 	for(int x = 0; x < countof(blockOffset32Z); x++)
 	{
-		blockOffset32Z[x] = (int)BlockNumber32Z(x << 3, 0, 0, 32) - (int)BlockNumber32Z(0, 0, 0, 32);
+		blockOffset32Z[x] = (short)((int)BlockNumber32Z(x << 3, 0, 0, 32) - (int)BlockNumber32Z(0, 0, 0, 32));
 	}
 
 	for(int x = 0; x < countof(blockOffset16); x++)
 	{
-		blockOffset16[x] = (int)BlockNumber16(x << 3, 0, 0, 32) - (int)BlockNumber16(0, 0, 0, 32);
+		blockOffset16[x] = (short)((int)BlockNumber16(x << 3, 0, 0, 32) - (int)BlockNumber16(0, 0, 0, 32));
 	}
 
 	for(int x = 0; x < countof(blockOffset16S); x++)
 	{
-		blockOffset16S[x] = (int)BlockNumber16S(x << 3, 0, 0, 32) - (int)BlockNumber16S(0, 0, 0, 32);
+		blockOffset16S[x] = (short)((int)BlockNumber16S(x << 3, 0, 0, 32) - (int)BlockNumber16S(0, 0, 0, 32));
 	}
 
 	for(int x = 0; x < countof(blockOffset16Z); x++)
 	{
-		blockOffset16Z[x] = (int)BlockNumber16Z(x << 3, 0, 0, 32) - (int)BlockNumber16Z(0, 0, 0, 32);
+		blockOffset16Z[x] = (short)((int)BlockNumber16Z(x << 3, 0, 0, 32) - (int)BlockNumber16Z(0, 0, 0, 32));
 	}
 
 	for(int x = 0; x < countof(blockOffset16SZ); x++)
 	{
-		blockOffset16SZ[x] = (int)BlockNumber16SZ(x << 3, 0, 0, 32) - (int)BlockNumber16SZ(0, 0, 0, 32);
+		blockOffset16SZ[x] = (short)((int)BlockNumber16SZ(x << 3, 0, 0, 32) - (int)BlockNumber16SZ(0, 0, 0, 32));
 	}
 
 	for(int x = 0; x < countof(blockOffset8); x++)
 	{
-		blockOffset8[x] = (int)BlockNumber8(x << 3, 0, 0, 32) - (int)BlockNumber8(0, 0, 0, 32);
+		blockOffset8[x] = (short)((int)BlockNumber8(x << 3, 0, 0, 32) - (int)BlockNumber8(0, 0, 0, 32));
 	}
 
 	for(int x = 0; x < countof(blockOffset4); x++)
 	{
-		blockOffset4[x] = (int)BlockNumber4(x << 3, 0, 0, 32) - (int)BlockNumber4(0, 0, 0, 32);
+		blockOffset4[x] = (short)((int)BlockNumber4(x << 3, 0, 0, 32) - (int)BlockNumber4(0, 0, 0, 32));
 	}
 
 	for(int i = 0; i < countof(m_psm); i++)
@@ -417,9 +417,6 @@ GSLocalMemory::GSLocalMemory()
 	m_psm[PSM_PSMZ16].rtxbP = &GSLocalMemory::ReadTextureBlock16Z;
 	m_psm[PSM_PSMZ16S].rtxbP = &GSLocalMemory::ReadTextureBlock16SZ;
 
-	m_psm[PSM_PSMT8].pal = m_psm[PSM_PSMT8H].pal = 256;
-	m_psm[PSM_PSMT4].pal = m_psm[PSM_PSMT4HL].pal = m_psm[PSM_PSMT4HH].pal = 16;
-
 	m_psm[PSM_PSMCT16].bpp = m_psm[PSM_PSMCT16S].bpp = 16;
 	m_psm[PSM_PSMT8].bpp = 8;
 	m_psm[PSM_PSMT4].bpp = 4;
@@ -432,6 +429,15 @@ GSLocalMemory::GSLocalMemory()
 	m_psm[PSM_PSMZ24].trbpp = 24;
 	m_psm[PSM_PSMZ16].trbpp = m_psm[PSM_PSMZ16S].trbpp = 16;
 
+	m_psm[PSM_PSMT8].pal = m_psm[PSM_PSMT8H].pal = 256;
+	m_psm[PSM_PSMT4].pal = m_psm[PSM_PSMT4HL].pal = m_psm[PSM_PSMT4HH].pal = 16;
+
+	for(int i = 0; i < countof(m_psm); i++) m_psm[i].fmt = 3;
+	m_psm[PSM_PSMCT32].fmt = m_psm[PSM_PSMZ32].fmt = 0;
+	m_psm[PSM_PSMCT24].fmt = m_psm[PSM_PSMZ24].fmt = 1;
+	m_psm[PSM_PSMCT16].fmt = m_psm[PSM_PSMZ16].fmt = 2;
+	m_psm[PSM_PSMCT16S].fmt = m_psm[PSM_PSMZ16S].fmt = 2;
+
 	m_psm[PSM_PSMCT16].bs = m_psm[PSM_PSMCT16S].bs = GSVector2i(16, 8);
 	m_psm[PSM_PSMT8].bs = GSVector2i(16, 16);
 	m_psm[PSM_PSMT4].bs = GSVector2i(32, 16);
@@ -489,7 +495,7 @@ GSLocalMemory::BlockOffset* GSLocalMemory::GetBlockOffset(uint32 bp, uint32 bw,
 
 	for(int i = 0; i < 256; i++)
 	{
-		o->row[i] = (int)bn(0, i << 3, bp, bw);
+		o->row[i] = (short)bn(0, i << 3, bp, bw);
 	}
 
 	o->col = m_psm[psm].blockOffset;

plugins/GSdx/GSLocalMemory.h

@@ -62,11 +62,10 @@ public:
 			readImage ri;
 			readTexture rtx, rtxNP, rtxP;
 			readTextureBlock rtxb, rtxbP;
-			short bpp, trbpp;
+			uint16 bpp, trbpp, pal, fmt; 
-			int pal; 
 			GSVector2i bs, pgs;
 			int* rowOffset[8];
-			int* blockOffset;
+			short* blockOffset;
 		};
 
 		uint8 dummy[128];
@@ -83,8 +82,8 @@ public:
 
 	struct BlockOffset
 	{
-		int row[256]; // yn (n = 0 8 16 ...)
+		short row[256]; // yn (n = 0 8 16 ...)
-		int* col; // blockOffset*
+		short* col; // blockOffset*
 		uint32 hash;
 	};
 
@@ -123,14 +122,14 @@ protected:
 	static int rowOffset8[2][2048];
 	static int rowOffset4[2][2048];
 
-	static int blockOffset32[256];
+	static short blockOffset32[256];
-	static int blockOffset32Z[256];
+	static short blockOffset32Z[256];
-	static int blockOffset16[256];
+	static short blockOffset16[256];
-	static int blockOffset16S[256];
+	static short blockOffset16S[256];
-	static int blockOffset16Z[256];
+	static short blockOffset16Z[256];
-	static int blockOffset16SZ[256];
+	static short blockOffset16SZ[256];
-	static int blockOffset8[256];
+	static short blockOffset8[256];
-	static int blockOffset4[256];
+	static short blockOffset4[256];
 
 	__forceinline static uint32 Expand24To32(uint32 c, const GIFRegTEXA& TEXA)
 	{
@@ -701,7 +700,7 @@ public:
 				uint32 rb = s[x] & 0x00f800f8;
 				uint32 ga = s[x] & 0x8000f800;
 
-				d[o[x]] = (ga >> 16) | (rb >> 9) | (ga >> 6) | (rb >> 3);
+				d[o[x]] = (uint16)((ga >> 16) | (rb >> 9) | (ga >> 6) | (rb >> 3));
 			}
 		}
 	}

plugins/GSdx/GSRenderer.cpp

@@ -325,7 +325,7 @@ void GSRenderer::VSync(int field)
 
 	// present
 
-	m_dev->Present(m_wnd.GetClientRect().fit(m_aspectratio), m_shader);
+	m_dev->Present(m_wnd.GetClientRect().fit(m_aspectratio), m_shader, m_framelimit);
 
 	// snapshot
 
@@ -435,12 +435,6 @@ void GSRenderer::KeyEvent(GSKeyEventData* e)
 	}
 }
 
-
-void GSRenderer::SetFrameLimit(bool limit)
-{
-	m_dev->SetVSync(m_vsync && limit);
-}
-
 void GSRenderer::GetTextureMinMax(GSVector4i& r, bool linear)
 {
 	const GSDrawingContext* context = m_context;
@@ -579,28 +573,25 @@ void GSRenderer::GetAlphaMinMax()
 
 	if(PRIM->TME && context->TEX0.TCC)
 	{
-		uint32 bpp = GSLocalMemory::m_psm[context->TEX0.PSM].trbpp;
+		switch(GSLocalMemory::m_psm[context->TEX0.PSM].fmt)
-		uint32 cbpp = GSLocalMemory::m_psm[context->TEX0.CPSM].trbpp;
-		uint32 pal = GSLocalMemory::m_psm[context->TEX0.PSM].pal;
-
-		if(bpp == 32)
 		{
+		case 0:
 			a.y = 0;
 			a.w = 0xff;
-		}
+			break;
-		else if(bpp == 24)
+		case 1:
-		{
 			a.y = env.TEXA.AEM ? 0 : env.TEXA.TA0;
 			a.w = env.TEXA.TA0;
-		}
+			break;
-		else if(bpp == 16)
+		case 2:
-		{
 			a.y = env.TEXA.AEM ? 0 : min(env.TEXA.TA0, env.TEXA.TA1);
 			a.w = max(env.TEXA.TA0, env.TEXA.TA1);
-		}
+			break;
-		else
+		case 3:
-		{
 			m_mem.m_clut.GetAlphaMinMax32(a.y, a.w);
+			break;
+		default:
+			__assume(0);
 		}
 
 		switch(context->TEX0.TFX)
@@ -714,17 +705,25 @@ bool GSRenderer::TryAlphaTest(uint32& fm, uint32& zm)
 
 bool GSRenderer::IsLinear()
 {
-	float qmin = m_vt.m_min.t.z;
+	const GIFRegTEX1& TEX1 = m_context->TEX1;
-	float qmax = m_vt.m_max.t.z;
 
-	if(PRIM->FST)
+	bool mmin = TEX1.IsMinLinear();
+	bool mmag = TEX1.IsMagLinear();
+
+	if(mmag == mmin) return mmag;
+
+	float LODmin = (float)TEX1.K;
+	float LODmax = (float)TEX1.K;
+
+	if(!TEX1.LCM && !PRIM->FST) // if FST => assume Q = 1.0f (should not, but Q is very often bogus, 0 or DEN)
 	{
-		// assume Q = 1.0f => LOD > 0 (should not, but Q is very often bogus, 0 or DEN)
+		float f = (float)(1 << TEX1.L) / log(2.0f);
 
-		qmin = qmax = 1.0f;
+		LODmin += log(1.0f / abs(m_vt.m_min.t.z)) * f;
+		LODmax += log(1.0f / abs(m_vt.m_max.t.z)) * f;
 	}
 
-	return m_context->TEX1.IsLinear(qmin, qmax);
+	return LODmax <= 0 ? mmag : LODmin > 0 ? mmin : mmag || mmin;
 }
 
 bool GSRenderer::IsOpaque()

plugins/GSdx/GSRenderer.h

@@ -77,7 +77,6 @@ public:
 	virtual void VSync(int field);
 	virtual bool MakeSnapshot(const string& path);
 	virtual void KeyEvent(GSKeyEventData* e);
-	virtual void SetFrameLimit(bool limit);
 
 	virtual bool CanUpscale() 
 	{
@@ -125,7 +124,7 @@ protected:
 				  PRIM->TME ? (int)m_context->TEX0.PSM : 0xff);
 			*/
 
-			if(GSUtil::EncodePSM(m_context->FRAME.PSM) != 3 && GSUtil::EncodePSM(m_context->ZBUF.PSM) != 3)
+			if(GSLocalMemory::m_psm[m_context->FRAME.PSM].fmt < 3 && GSLocalMemory::m_psm[m_context->ZBUF.PSM].fmt < 3)
 			{
 				// FIXME: berserk fpsm = 27 (8H)
 

plugins/GSdx/GSRendererDX.h

@@ -82,9 +82,21 @@ public:
 
 		GSTextureFX::OMDepthStencilSelector om_dssel;
 
+		if(context->TEST.ZTE)
+		{
+			om_dssel.ztst = context->TEST.ZTST;
+			om_dssel.zwe = !context->ZBUF.ZMSK;
+		}
+		else
+		{
+			om_dssel.ztst = ZTST_ALWAYS;
+			om_dssel.zwe = 0;
+		}
+/*
 		om_dssel.zte = context->TEST.ZTE;
 		om_dssel.ztst = context->TEST.ZTST;
 		om_dssel.zwe = !context->ZBUF.ZMSK;
+*/
 		om_dssel.date = context->FRAME.PSM != PSM_PSMCT24 ? context->TEST.DATE : 0;
 		om_dssel.fba = m_fba ? context->FBA.FBA : 0;
 
@@ -115,22 +127,17 @@ public:
 			}
 		}
 
-		om_bsel.wr = (context->FRAME.FBMSK & 0x000000ff) != 0x000000ff;
+		om_bsel.wrgba = ~GSVector4i::load((int)context->FRAME.FBMSK).eq8(GSVector4i::xffffffff()).mask();
-		om_bsel.wg = (context->FRAME.FBMSK & 0x0000ff00) != 0x0000ff00;
-		om_bsel.wb = (context->FRAME.FBMSK & 0x00ff0000) != 0x00ff0000;
-		om_bsel.wa = (context->FRAME.FBMSK & 0xff000000) != 0xff000000;
 
 		// vs
 
 		GSTextureFX::VSSelector vs_sel;
 
-		vs_sel.bppz = 0;
 		vs_sel.tme = PRIM->TME;
 		vs_sel.fst = PRIM->FST;
 		vs_sel.logz = m_logz ? 1 : 0;
-		vs_sel.prim = primclass;
 
-		if(om_dssel.zte && om_dssel.ztst > 0 && om_dssel.zwe)
+		if(om_dssel.ztst >= ZTST_ALWAYS && om_dssel.zwe)
 		{
 			if(context->ZBUF.PSM == PSM_PSMZ24)
 			{
@@ -139,7 +146,7 @@ public:
 					ASSERT(m_vt.m_min.p.z > 0xffffff);
 
 					vs_sel.bppz = 1;
-					om_dssel.ztst = 1;
+					om_dssel.ztst = ZTST_ALWAYS;
 				}
 			}
 			else if(context->ZBUF.PSM == PSM_PSMZ16 || context->ZBUF.PSM == PSM_PSMZ16S)
@@ -149,36 +156,23 @@ public:
 					ASSERT(m_vt.m_min.p.z > 0xffff); // sfex capcom logo
 
 					vs_sel.bppz = 2;
-					om_dssel.ztst = 1;
+					om_dssel.ztst = ZTST_ALWAYS;
 				}
 			}
 		}
 
 		GSTextureFX::VSConstantBuffer vs_cb;
 
-		float sx = 2.0f * rt->m_scale.x / (rt->GetWidth() * 16);
+		float sx = 2.0f * rt->m_scale.x / (rt->m_size.x << 4);
-		float sy = 2.0f * rt->m_scale.y / (rt->GetHeight() * 16);
+		float sy = 2.0f * rt->m_scale.y / (rt->m_size.y << 4);
 		float ox = (float)(int)context->XYOFFSET.OFX;
 		float oy = (float)(int)context->XYOFFSET.OFY;
-		float ox2 = 2.0f * m_pixelcenter.x / rt->GetWidth();
+		float ox2 = 2.0f * m_pixelcenter.x / rt->m_size.x;
-		float oy2 = 2.0f * m_pixelcenter.y / rt->GetHeight();
+		float oy2 = 2.0f * m_pixelcenter.y / rt->m_size.y;
 
 		vs_cb.VertexScale = GSVector4(sx, -sy, 1.0f / UINT_MAX, 0.0f);
 		vs_cb.VertexOffset = GSVector4(ox * sx + ox2 + 1, -(oy * sy + oy2 + 1), 0.0f, -1.0f);
 
-		if(PRIM->TME)
-		{
-			if(PRIM->FST)
-			{
-				vs_cb.TextureScale.x = 1.0f / (16 << context->TEX0.TW);
-				vs_cb.TextureScale.y = 1.0f / (16 << context->TEX0.TH);
-			}
-			else
-			{
-				vs_cb.TextureScale = GSVector2(1.0f, 1.0f);
-			}
-		}
-
 		// gs
 
 		GSTextureFX::GSSelector gs_sel;
@@ -224,7 +218,6 @@ public:
 
 		if(tex)
 		{
-			ps_sel.fst = PRIM->FST;
 			ps_sel.wms = context->CLAMP.WMS;
 			ps_sel.wmt = context->CLAMP.WMT;
 			ps_sel.fmt = tex->m_fmt;
@@ -234,46 +227,34 @@ public:
 			ps_sel.ltf = m_filter == 2 ? IsLinear() : m_filter;
 			ps_sel.rt = tex->m_target;
 
-			int w = tex->m_texture->GetWidth();
+			int w = tex->m_texture->m_size.x;
-			int h = tex->m_texture->GetHeight();
+			int h = tex->m_texture->m_size.y;
 
 			int tw = (int)(1 << context->TEX0.TW);
 			int th = (int)(1 << context->TEX0.TH);
 
-			ps_cb.WH = GSVector4(tw, th, w, h);
+			GSVector4 WH(tw, th, w, h);
-			ps_cb.HalfTexel = GSVector4(-0.5f, 0.5f).xxyy() / GSVector4(w, h).xyxy();
+
-			ps_cb.MinF_TA.z = (float)(int)env.TEXA.TA0 / 255;
+			if(PRIM->FST)
-			ps_cb.MinF_TA.w = (float)(int)env.TEXA.TA1 / 255;
+			{
+				vs_cb.TextureScale = GSVector4(1.0f / 16) / WH.xyxy();
+
+				ps_sel.fst = 1;
+			}
+
+			ps_cb.WH = WH;
+			ps_cb.HalfTexel = GSVector4(-0.5f, 0.5f).xxyy() / WH.zwzw();
+			ps_cb.MskFix = GSVector4i(context->CLAMP.MINU, context->CLAMP.MINV, context->CLAMP.MAXU, context->CLAMP.MAXV);
+
+			GSVector4 clamp(ps_cb.MskFix);
+			GSVector4 ta(env.TEXA & GSVector4i::x000000ff());
+
+			ps_cb.MinMax = clamp / WH.xyxy();
+			ps_cb.MinF_TA = (clamp + 0.5f).xyxy(ta) / WH.xyxy(GSVector4(255, 255));
 
 			ps_ssel.tau = (context->CLAMP.WMS + 3) >> 1;
 			ps_ssel.tav = (context->CLAMP.WMT + 3) >> 1;
 			ps_ssel.ltf = ps_sel.ltf;
-
-			switch(ps_sel.wms)
-			{
-			case CLAMP_REGION_CLAMP: 
-				ps_cb.MinMax.x = (float)(int)context->CLAMP.MINU / tw;
-				ps_cb.MinMax.z = (float)(int)context->CLAMP.MAXU / tw;
-				ps_cb.MinF_TA.x = ((float)(int)context->CLAMP.MINU + 0.5f) / tw;
-				break;
-			case CLAMP_REGION_REPEAT: 
-				ps_cb.MskFix.x = context->CLAMP.MINU;
-				ps_cb.MskFix.z = context->CLAMP.MAXU;
-				break;
-			}
-
-			switch(ps_sel.wmt)
-			{
-			case CLAMP_REGION_CLAMP: 
-				ps_cb.MinMax.y = (float)(int)context->CLAMP.MINV / th;
-				ps_cb.MinMax.w = (float)(int)context->CLAMP.MAXV / th;
-				ps_cb.MinF_TA.y = ((float)(int)context->CLAMP.MINV + 0.5f) / th;
-				break;
-			case CLAMP_REGION_REPEAT: 
-				ps_cb.MskFix.y = context->CLAMP.MINV;
-				ps_cb.MskFix.w = context->CLAMP.MAXV;
-				break;
-			}
 		}
 		else
 		{
@@ -282,10 +263,7 @@ public:
 
 		// rs
 
-		int w = rt->GetWidth();
+		GSVector4i scissor = GSVector4i(GSVector4(rt->m_scale).xyxy() * context->scissor.in).rintersect(GSVector4i(rt->GetSize()).zwxy());
-		int h = rt->GetHeight();
-
-		GSVector4i scissor = GSVector4i(GSVector4(rt->m_scale).xyxy() * context->scissor.in).rintersect(GSVector4i(0, 0, w, h));
 
 		//
 
@@ -296,7 +274,7 @@ public:
 		m_tfx->SetupVS(vs_sel, &vs_cb);
 		m_tfx->SetupGS(gs_sel);
 		m_tfx->SetupPS(ps_sel, &ps_cb, ps_ssel, tex ? tex->m_texture : NULL, tex ? tex->m_palette : NULL);
-		m_tfx->SetupRS(w, h, scissor);
+		m_tfx->SetupRS(rt->m_size, scissor);
 
 		// draw
 

plugins/GSdx/GSRendererDX10.cpp

@@ -86,9 +86,11 @@ void GSRendererDX10::VertexKick(bool skip)
 	{
 		GSVector4i scissor = m_context->scissor.dx10;
 
+		GSVector4i pmin, pmax;
+
 		#if _M_SSE >= 0x401
 
-		GSVector4i pmin, pmax, v0, v1, v2;
+		GSVector4i v0, v1, v2;
 
 		switch(prim)
 		{
@@ -116,52 +118,53 @@ void GSRendererDX10::VertexKick(bool skip)
 			break;
 		}
 
-		GSVector4i test = (pmax < scissor) | (pmin > scissor.zwxy());
-
-		if(test.mask() & 0xff)
-		{
-			return;
-		}
-
 		#else
 
 		switch(prim)
 		{
 		case GS_POINTLIST:
-			if(v[0].p.x < scissor.x 
+			pmin.x = v[0].p.x;
-			|| v[0].p.x > scissor.z
+			pmin.y = v[0].p.y;
-			|| v[0].p.y < scissor.y 
+			pmax.x = v[0].p.x;
-			|| v[0].p.y > scissor.w)
+			pmax.y = v[0].p.y;
-			{
-				return;
-			}
 			break;
 		case GS_LINELIST:
 		case GS_LINESTRIP:
 		case GS_SPRITE:
-			if(v[0].p.x < scissor.x && v[1].p.x < scissor.x
+			pmin.x = std::min<uint16>(v[0].p.x, v[1].p.x);
-			|| v[0].p.x > scissor.z && v[1].p.x > scissor.z
+			pmin.y = std::min<uint16>(v[0].p.y, v[1].p.y);
-			|| v[0].p.y < scissor.y && v[1].p.y < scissor.y
+			pmax.x = std::max<uint16>(v[0].p.x, v[1].p.x);
-			|| v[0].p.y > scissor.w && v[1].p.y > scissor.w)
+			pmax.y = std::max<uint16>(v[0].p.y, v[1].p.y);
-			{
-				return;
-			}
 			break;
 		case GS_TRIANGLELIST:
 		case GS_TRIANGLESTRIP:
 		case GS_TRIANGLEFAN:
-			if(v[0].p.x < scissor.x && v[1].p.x < scissor.x && v[2].p.x < scissor.x
+			pmin.x = std::min<uint16>(std::min<uint16>(v[0].p.x, v[1].p.x), v[2].p.x);
-			|| v[0].p.x > scissor.z && v[1].p.x > scissor.z && v[2].p.x > scissor.z
+			pmin.y = std::min<uint16>(std::min<uint16>(v[0].p.y, v[1].p.y), v[2].p.y);
-			|| v[0].p.y < scissor.y && v[1].p.y < scissor.y && v[2].p.y < scissor.y
+			pmax.x = std::max<uint16>(std::max<uint16>(v[0].p.x, v[1].p.x), v[2].p.x);
-			|| v[0].p.y > scissor.w && v[1].p.y > scissor.w && v[2].p.y > scissor.w)
+			pmax.y = std::max<uint16>(std::max<uint16>(v[0].p.y, v[1].p.y), v[2].p.y);
-			{
-				return;
-			}
 			break;
 		}
 
 		#endif
 
+		GSVector4i test = (pmax < scissor) | (pmin > scissor.zwxy());
+
+		switch(prim)
+		{
+		case GS_TRIANGLELIST:
+		case GS_TRIANGLESTRIP:
+		case GS_TRIANGLEFAN:
+		case GS_SPRITE:
+			test |= pmin == pmax;
+			break;
+		}
+
+		if(test.mask() & 0xff)
+		{
+			return;
+		}
+
 		m_count += count;
 	}
 }
@@ -196,10 +199,9 @@ void GSRendererDX10::SetupDATE(GSTexture* rt, GSTexture* ds)
 
 	GSDevice10* dev = (GSDevice10*)m_dev;
 
-	int w = rt->GetWidth();
+	const GSVector2i& size = rt->m_size;
-	int h = rt->GetHeight();
 
-	if(GSTexture* t = dev->CreateRenderTarget(w, h))
+	if(GSTexture* t = dev->CreateRenderTarget(size.x, size.y))
 	{
 		// sfex3 (after the capcom logo), vf4 (first menu fading in), ffxii shadows, rumble roses shadows, persona4 shadows
 
@@ -215,7 +217,7 @@ void GSRendererDX10::SetupDATE(GSTexture* rt, GSTexture* ds)
 
 		// ia
 
-		GSVector4 s = GSVector4(rt->m_scale.x / w, rt->m_scale.y / h);
+		GSVector4 s = GSVector4(rt->m_scale.x / size.x, rt->m_scale.y / size.y);
 		GSVector4 o = GSVector4(-1.0f, 1.0f);
 
 		GSVector4 src = ((m_vt.m_min.p.xyxy(m_vt.m_max.p) + o.xxyy()) * s.xyxy()).sat(o.zzyy());
@@ -249,7 +251,7 @@ void GSRendererDX10::SetupDATE(GSTexture* rt, GSTexture* ds)
 
 		// rs
 
-		dev->RSSet(w, h);
+		dev->RSSet(size);
 
 		// set
 

plugins/GSdx/GSRendererDX11.cpp

@@ -86,9 +86,11 @@ void GSRendererDX11::VertexKick(bool skip)
 	{
 		GSVector4i scissor = m_context->scissor.dx10;
 
+		GSVector4i pmin, pmax;
+
 		#if _M_SSE >= 0x401
 
-		GSVector4i pmin, pmax, v0, v1, v2;
+		GSVector4i v0, v1, v2;
 
 		switch(prim)
 		{
@@ -116,52 +118,53 @@ void GSRendererDX11::VertexKick(bool skip)
 			break;
 		}
 
-		GSVector4i test = (pmax < scissor) | (pmin > scissor.zwxy());
-
-		if(test.mask() & 0xff)
-		{
-			return;
-		}
-
 		#else
 
 		switch(prim)
 		{
 		case GS_POINTLIST:
-			if(v[0].p.x < scissor.x 
+			pmin.x = v[0].p.x;
-			|| v[0].p.x > scissor.z
+			pmin.y = v[0].p.y;
-			|| v[0].p.y < scissor.y 
+			pmax.x = v[0].p.x;
-			|| v[0].p.y > scissor.w)
+			pmax.y = v[0].p.y;
-			{
-				return;
-			}
 			break;
 		case GS_LINELIST:
 		case GS_LINESTRIP:
 		case GS_SPRITE:
-			if(v[0].p.x < scissor.x && v[1].p.x < scissor.x
+			pmin.x = std::min<uint16>(v[0].p.x, v[1].p.x);
-			|| v[0].p.x > scissor.z && v[1].p.x > scissor.z
+			pmin.y = std::min<uint16>(v[0].p.y, v[1].p.y);
-			|| v[0].p.y < scissor.y && v[1].p.y < scissor.y
+			pmax.x = std::max<uint16>(v[0].p.x, v[1].p.x);
-			|| v[0].p.y > scissor.w && v[1].p.y > scissor.w)
+			pmax.y = std::max<uint16>(v[0].p.y, v[1].p.y);
-			{
-				return;
-			}
 			break;
 		case GS_TRIANGLELIST:
 		case GS_TRIANGLESTRIP:
 		case GS_TRIANGLEFAN:
-			if(v[0].p.x < scissor.x && v[1].p.x < scissor.x && v[2].p.x < scissor.x
+			pmin.x = std::min<uint16>(std::min<uint16>(v[0].p.x, v[1].p.x), v[2].p.x);
-			|| v[0].p.x > scissor.z && v[1].p.x > scissor.z && v[2].p.x > scissor.z
+			pmin.y = std::min<uint16>(std::min<uint16>(v[0].p.y, v[1].p.y), v[2].p.y);
-			|| v[0].p.y < scissor.y && v[1].p.y < scissor.y && v[2].p.y < scissor.y
+			pmax.x = std::max<uint16>(std::max<uint16>(v[0].p.x, v[1].p.x), v[2].p.x);
-			|| v[0].p.y > scissor.w && v[1].p.y > scissor.w && v[2].p.y > scissor.w)
+			pmax.y = std::max<uint16>(std::max<uint16>(v[0].p.y, v[1].p.y), v[2].p.y);
-			{
-				return;
-			}
 			break;
 		}
 
 		#endif
 
+		GSVector4i test = (pmax < scissor) | (pmin > scissor.zwxy());
+
+		switch(prim)
+		{
+		case GS_TRIANGLELIST:
+		case GS_TRIANGLESTRIP:
+		case GS_TRIANGLEFAN:
+		case GS_SPRITE:
+			test |= pmin == pmax;
+			break;
+		}
+
+		if(test.mask() & 0xff)
+		{
+			return;
+		}
+
 		m_count += count;
 	}
 }
@@ -196,10 +199,9 @@ void GSRendererDX11::SetupDATE(GSTexture* rt, GSTexture* ds)
 
 	GSDevice11* dev = (GSDevice11*)m_dev;
 
-	int w = rt->GetWidth();
+	const GSVector2i& size = rt->m_size;
-	int h = rt->GetHeight();
 
-	if(GSTexture* t = dev->CreateRenderTarget(w, h))
+	if(GSTexture* t = dev->CreateRenderTarget(size.x, size.y))
 	{
 		// sfex3 (after the capcom logo), vf4 (first menu fading in), ffxii shadows, rumble roses shadows, persona4 shadows
 
@@ -215,7 +217,7 @@ void GSRendererDX11::SetupDATE(GSTexture* rt, GSTexture* ds)
 
 		// ia
 
-		GSVector4 s = GSVector4(rt->m_scale.x / w, rt->m_scale.y / h);
+		GSVector4 s = GSVector4(rt->m_scale.x / size.x, rt->m_scale.y / size.y);
 		GSVector4 o = GSVector4(-1.0f, 1.0f);
 
 		GSVector4 src = ((m_vt.m_min.p.xyxy(m_vt.m_max.p) + o.xxyy()) * s.xyxy()).sat(o.zzyy());
@@ -249,7 +251,7 @@ void GSRendererDX11::SetupDATE(GSTexture* rt, GSTexture* ds)
 
 		// rs
 
-		dev->RSSet(w, h);
+		dev->RSSet(size);
 
 		// set
 

plugins/GSdx/GSRendererDX9.cpp

@@ -135,6 +135,16 @@ void GSRendererDX9::VertexKick(bool skip)
 
 		GSVector4 test = (pmax < scissor) | (pmin > scissor.zwxy());
 
+		switch(prim)
+		{
+		case GS_TRIANGLELIST:
+		case GS_TRIANGLESTRIP:
+		case GS_TRIANGLEFAN:
+		case GS_SPRITE:
+			test |= pmin == pmax;
+			break;
+		}
+
 		if(test.mask() & 3)
 		{
 			return;
@@ -206,10 +216,9 @@ void GSRendererDX9::SetupDATE(GSTexture* rt, GSTexture* ds)
 
 	GSDevice9* dev = (GSDevice9*)m_dev;
 
-	int w = rt->GetWidth();
+	const GSVector2i& size = rt->m_size;
-	int h = rt->GetHeight();
 
-	if(GSTexture* t = dev->CreateRenderTarget(w, h))
+	if(GSTexture* t = dev->CreateRenderTarget(size.x, size.y))
 	{
 		// sfex3 (after the capcom logo), vf4 (first menu fading in), ffxii shadows, rumble roses shadows, persona4 shadows
 
@@ -225,7 +234,7 @@ void GSRendererDX9::SetupDATE(GSTexture* rt, GSTexture* ds)
 
 		// ia
 
-		GSVector4 s = GSVector4(rt->m_scale.x / w, rt->m_scale.y / h);
+		GSVector4 s = GSVector4(rt->m_scale.x / size.x, rt->m_scale.y / size.y);
 		GSVector4 o = GSVector4(-1.0f, 1.0f);
 
 		GSVector4 src = ((m_vt.m_min.p.xyxy(m_vt.m_max.p) + o.xxyy()) * s.xyxy()).sat(o.zzyy());
@@ -255,7 +264,7 @@ void GSRendererDX9::SetupDATE(GSTexture* rt, GSTexture* ds)
 
 		// rs
 
-		dev->RSSet(w, h);
+		dev->RSSet(size);
 
 		//
 
@@ -282,7 +291,7 @@ void GSRendererDX9::UpdateFBA(GSTexture* rt)
 
 	// ia
 
-	GSVector4 s = GSVector4(rt->m_scale.x / rt->GetWidth(), rt->m_scale.y / rt->GetHeight());
+	GSVector4 s = GSVector4(rt->m_scale.x / rt->m_size.x, rt->m_scale.y / rt->m_size.y);
 	GSVector4 o = GSVector4(-1.0f, 1.0f);
 
 	GSVector4 src = ((m_vt.m_min.p.xyxy(m_vt.m_max.p) + o.xxyy()) * s.xyxy()).sat(o.zzyy());
@@ -310,7 +319,7 @@ void GSRendererDX9::UpdateFBA(GSTexture* rt)
 
 	// rs
 
-	dev->RSSet(rt->GetWidth(), rt->GetHeight());
+	dev->RSSet(rt->m_size);
 
 	//
 

plugins/GSdx/GSRendererSW.cpp

@@ -305,7 +305,7 @@ void GSRendererSW::GetScanlineParam(GSScanlineParam& p, GS_PRIM_CLASS primclass)
 
 	if(fwrite || ftest)
 	{
-		p.sel.fpsm = GSUtil::EncodePSM(context->FRAME.PSM);
+		p.sel.fpsm = GSLocalMemory::m_psm[context->FRAME.PSM].fmt;
 
 		if((primclass == GS_LINE_CLASS || primclass == GS_TRIANGLE_CLASS) && m_vt.m_eq.rgba != 0xffff)
 		{
@@ -433,15 +433,15 @@ void GSRendererSW::GetScanlineParam(GSScanlineParam& p, GS_PRIM_CLASS primclass)
 	}
 
 	bool zwrite = p.zm != 0xffffffff;
-	bool ztest = context->TEST.ZTE && context->TEST.ZTST > 1;
+	bool ztest = context->TEST.ZTE && context->TEST.ZTST > ZTST_ALWAYS;
 
 	p.sel.zwrite = zwrite;
 	p.sel.ztest = ztest;
 
 	if(zwrite || ztest)
 	{
-		p.sel.zpsm = GSUtil::EncodePSM(context->ZBUF.PSM);
+		p.sel.zpsm = GSLocalMemory::m_psm[context->ZBUF.PSM].fmt;
-		p.sel.ztst = ztest ? context->TEST.ZTST : 1;
+		p.sel.ztst = ztest ? context->TEST.ZTST : ZTST_ALWAYS;
 		p.sel.zoverflow = GSVector4i(m_vt.m_max.p).z == 0x80000000;
 	}
 }

plugins/GSdx/GSState.cpp

@@ -32,6 +32,7 @@ GSState::GSState(uint8* base, bool mt, void (*irq)())
 	, m_vprim(1)
 	, m_version(5)
 	, m_frameskip(0)
+	, m_framelimit(true)
 	, m_vkf(NULL)
 {
 	m_sssize = 0;
@@ -421,7 +422,7 @@ void GSState::GIFRegHandlerPRIM(GIFReg* r)
 
 	if(GSUtil::GetPrimClass(m_env.PRIM.PRIM) == GSUtil::GetPrimClass(r->PRIM.PRIM))
 	{
-		if(((m_env.PRIM.u64 ^ r->PRIM.u64) & ~7) != 0)
+		if((m_env.PRIM.u32[0] ^ r->PRIM.u32[0]) & 0x7f8) // all fields except PRIM
 		{
 			Flush();
 		}
@@ -1663,13 +1664,13 @@ void GSState::SetGameCRC(uint32 crc, int options)
 	m_game = CRC::Lookup(crc);
 }
 
-void GSState::SetFrameSkip(int frameskip)
+void GSState::SetFrameSkip(int skip)
 {
-	if(m_frameskip != frameskip)
+	if(m_frameskip != skip)
 	{
-		m_frameskip = frameskip;
+		m_frameskip = skip;
 
-		if(frameskip)
+		if(skip)
 		{
 			m_fpGIFPackedRegHandlers[GIF_REG_PRIM] = &GSState::GIFPackedRegHandlerNOP;
 			m_fpGIFPackedRegHandlers[GIF_REG_RGBA] = &GSState::GIFPackedRegHandlerNOP;
@@ -1722,6 +1723,11 @@ void GSState::SetFrameSkip(int frameskip)
 	}
 }
 
+void GSState::SetFrameLimit(bool limit)
+{
+	m_framelimit = limit;
+}
+
 // GSTransferBuffer
 
 GSState::GSTransferBuffer::GSTransferBuffer()

plugins/GSdx/GSState.h

@@ -210,6 +210,7 @@ public:
 	uint32 m_crc;
 	int m_options;
 	int m_frameskip;
+	bool m_framelimit;
 	CRC::Game m_game;
 	GSDump m_dump;
 
@@ -247,6 +248,7 @@ public:
 	int Defrost(const GSFreezeData* fd);
 	void GetLastTag(uint32* tag) {*tag = m_path3hack; m_path3hack = 0;}
 	virtual void SetGameCRC(uint32 crc, int options);
-	void SetFrameSkip(int frameskip);
+	void SetFrameSkip(int skip);
+	void SetFrameLimit(bool limit);
 };
 

plugins/GSdx/GSTexture.h

@@ -27,20 +27,19 @@ class GSTexture
 {
 public:
 	GSVector2 m_scale;
+	GSVector2i m_size;
 
 	struct GSMap {uint8* bits; int pitch;};
 
 	enum {None, RenderTarget, DepthStencil, Texture, Offscreen};
 
 public:
-	GSTexture() : m_scale(1, 1) {}
+	GSTexture() : m_scale(1, 1), m_size(0, 0) {}
 	virtual ~GSTexture() {}
 
 	virtual operator bool() {ASSERT(0); return false;}
 
 	virtual int GetType() const = 0;
-	virtual int GetWidth() const = 0;
-	virtual int GetHeight() const = 0;
 	virtual int GetFormat() const = 0;
 
 	virtual bool Update(const GSVector4i& r, const void* data, int pitch) = 0;
@@ -48,5 +47,8 @@ public:
 	virtual void Unmap() = 0;
 	virtual bool Save(const string& fn, bool dds = false) = 0;
 
-	GSVector2i GetSize() const {return GSVector2i(GetWidth(), GetHeight());}
+	int GetWidth() const {return m_size.x;}
+	int GetHeight() const {return m_size.y;}
+
+	GSVector2i GetSize() const {return m_size;}
 };

plugins/GSdx/GSTexture10.cpp

@@ -29,6 +29,9 @@ GSTexture10::GSTexture10(ID3D10Texture2D* texture)
 
 	m_texture->GetDevice(&m_dev);
 	m_texture->GetDesc(&m_desc);
+
+	m_size.x = (int)m_desc.Width;
+	m_size.y = (int)m_desc.Height;
 }
 
 int GSTexture10::GetType() const
@@ -40,16 +43,6 @@ int GSTexture10::GetType() const
 	return GSTexture::None;
 }
 
-int GSTexture10::GetWidth() const 
-{
-	return m_desc.Width;
-}
-
-int GSTexture10::GetHeight() const 
-{
-	return m_desc.Height;
-}
-
 int GSTexture10::GetFormat() const 
 {
 	return m_desc.Format;

plugins/GSdx/GSTexture10.h

@@ -36,8 +36,6 @@ public:
 	explicit GSTexture10(ID3D10Texture2D* texture);
 
 	int GetType() const;
-	int GetWidth() const;
-	int GetHeight() const;
 	int GetFormat() const;
 
 	bool Update(const GSVector4i& r, const void* data, int pitch);

plugins/GSdx/GSTexture11.cpp

@@ -31,6 +31,9 @@ GSTexture11::GSTexture11(ID3D11Texture2D* texture)
 	m_texture->GetDesc(&m_desc);
 
 	m_dev->GetImmediateContext(&m_ctx);
+
+	m_size.x = (int)m_desc.Width;
+	m_size.y = (int)m_desc.Height;
 }
 
 int GSTexture11::GetType() const
@@ -42,16 +45,6 @@ int GSTexture11::GetType() const
 	return GSTexture::None;
 }
 
-int GSTexture11::GetWidth() const 
-{
-	return m_desc.Width;
-}
-
-int GSTexture11::GetHeight() const 
-{
-	return m_desc.Height;
-}
-
 int GSTexture11::GetFormat() const 
 {
 	return m_desc.Format;

plugins/GSdx/GSTexture11.h

@@ -37,8 +37,6 @@ public:
 	explicit GSTexture11(ID3D11Texture2D* texture);
 
 	int GetType() const;
-	int GetWidth() const;
-	int GetHeight() const;
 	int GetFormat() const;
 
 	bool Update(const GSVector4i& r, const void* data, int pitch);

plugins/GSdx/GSTexture7.cpp

@@ -31,6 +31,9 @@ GSTexture7::GSTexture7(int type, IDirectDrawSurface7* system)
 	m_desc.dwSize = sizeof(m_desc);
 
 	system->GetSurfaceDesc(&m_desc);
+
+	m_size.x = (int)m_desc.dwWidth;
+	m_size.y = (int)m_desc.dwHeight;
 }
 
 GSTexture7::GSTexture7(int type, IDirectDrawSurface7* system, IDirectDrawSurface7* video)
@@ -43,6 +46,9 @@ GSTexture7::GSTexture7(int type, IDirectDrawSurface7* system, IDirectDrawSurface
 	m_desc.dwSize = sizeof(m_desc);
 
 	video->GetSurfaceDesc(&m_desc);
+
+	m_size.x = (int)m_desc.dwWidth;
+	m_size.y = (int)m_desc.dwHeight;
 }
 
 int GSTexture7::GetType() const
@@ -50,16 +56,6 @@ int GSTexture7::GetType() const
 	return m_type;
 }
 
-int GSTexture7::GetWidth() const 
-{
-	return m_desc.dwWidth;
-}
-
-int GSTexture7::GetHeight() const 
-{
-	return m_desc.dwHeight;
-}
-
 int GSTexture7::GetFormat() const 
 {
 	return (int)m_desc.ddpfPixelFormat.dwFourCC;

plugins/GSdx/GSTexture7.h

@@ -36,8 +36,6 @@ public:
 	GSTexture7(int type, IDirectDrawSurface7* system, IDirectDrawSurface7* video);
 
 	int GetType() const;
-	int GetWidth() const;
-	int GetHeight() const;
 	int GetFormat() const;
 
 	bool Update(const GSVector4i& r, const void* data, int pitch);

plugins/GSdx/GSTexture9.cpp

@@ -35,6 +35,9 @@ GSTexture9::GSTexture9(IDirect3DSurface9* surface)
 
 		ASSERT(m_texture != NULL);
 	}
+
+	m_size.x = (int)m_desc.Width;
+	m_size.y = (int)m_desc.Height;
 }
 
 GSTexture9::GSTexture9(IDirect3DTexture9* texture)
@@ -46,6 +49,9 @@ GSTexture9::GSTexture9(IDirect3DTexture9* texture)
 	texture->GetSurfaceLevel(0, &m_surface);
 
 	ASSERT(m_surface != NULL);
+
+	m_size.x = (int)m_desc.Width;
+	m_size.y = (int)m_desc.Height;
 }
 
 GSTexture9::~GSTexture9()
@@ -61,16 +67,6 @@ int GSTexture9::GetType() const
 	return GSTexture::None;
 }
 
-int GSTexture9::GetWidth() const 
-{
-	return m_desc.Width;
-}
-
-int GSTexture9::GetHeight() const 
-{
-	return m_desc.Height;
-}
-
 int GSTexture9::GetFormat() const 
 {
 	return m_desc.Format;

plugins/GSdx/GSTexture9.h

@@ -36,8 +36,6 @@ public:
 	virtual ~GSTexture9();
 
 	int GetType() const;
-	int GetWidth() const;
-	int GetHeight() const;
 	int GetFormat() const;
 
 	bool Update(const GSVector4i& r, const void* data, int pitch);

plugins/GSdx/GSTextureCache.cpp

@@ -445,11 +445,11 @@ void GSTextureCache::IncAge()
 {
 	int maxage = m_src.m_used ? 3 : 30;
 
-	for(hash_map<Source*, bool>::iterator i = m_src.m_surfaces.begin(); i != m_src.m_surfaces.end(); )
+	for(hash_set<Source*>::iterator i = m_src.m_surfaces.begin(); i != m_src.m_surfaces.end(); )
 	{
-		hash_map<Source*, bool>::iterator j = i++;
+		hash_set<Source*>::iterator j = i++;
 
-		Source* s = j->first;
+		Source* s = *j;
 
 		if(++s->m_age > maxage)
 		{
@@ -673,8 +673,8 @@ bool GSTextureCache::Source::Create(Target* dst)
 	}
 	else
 	{
-		sr.z /= st->GetWidth();
+		sr.z /= st->m_size.x;
-		sr.w /= st->GetHeight();
+		sr.w /= st->m_size.y;
 
 		m_renderer->m_dev->StretchRect(st, sr, dt, dr);
 	}
@@ -733,9 +733,12 @@ void GSTextureCache::Source::Update(const GIFRegTEX0& TEX0, const GIFRegTEXA& TE
 
 	GSVector2i bs = GSLocalMemory::m_psm[m_TEX0.PSM].bs;
 
+	int tw = std::max<int>(1 << m_TEX0.TW, bs.x);
+	int th = std::max<int>(1 << m_TEX0.TH, bs.y);
+
 	GSVector4i r = rect.ralign<GSVector4i::Outside>(bs); 
 
-	if(r.eq(GSVector4i(0, 0, 1 << m_TEX0.TW, 1 << m_TEX0.TH)))
+	if(r.eq(GSVector4i(0, 0, tw, th)))
 	{
 		m_complete = true; // lame, but better than nothing
 	}
@@ -995,7 +998,7 @@ void GSTextureCache::Target::Update()
 
 void GSTextureCache::SourceMap::Add(Source* s, const GIFRegTEX0& TEX0, GSLocalMemory& mem)
 {
-	m_surfaces[s] = true;
+	m_surfaces.insert(s);
 
 	if(s->m_target)
 	{
@@ -1038,20 +1041,21 @@ void GSTextureCache::SourceMap::Add(Source* s, const GIFRegTEX0& TEX0, GSLocalMe
 
 			list<Source*>* m = &m_map[i << 5];
 
-			for(int j = 0; j < 32; j++)
+			unsigned long j;
-			{
+
-				if(p & (1 << j))
+			while(_BitScanForward(&j, p))
 			{
+				p ^= 1 << j;
+
 				m[j].push_front(s);
 			}
 		}
 	}
-	}
 }
 
 void GSTextureCache::SourceMap::RemoveAll()
 {
-	for_each(m_surfaces.begin(), m_surfaces.end(), delete_first());
+	for_each(m_surfaces.begin(), m_surfaces.end(), delete_object());
 
 	m_surfaces.clear();
 

plugins/GSdx/GSTextureCache.h

@@ -98,7 +98,7 @@ protected:
 
 	struct SourceMap
 	{
-		hash_map<Source*, bool> m_surfaces;
+		hash_set<Source*> m_surfaces;
 		list<Source*> m_map[MAX_PAGES];
 		uint32 m_pages[16];
 		bool m_used;

plugins/GSdx/GSTextureCacheSW.cpp

@@ -68,7 +68,7 @@ const GSTextureCacheSW::GSTexture* GSTextureCacheSW::Lookup(const GIFRegTEX0& TE
 	{
 		t = new GSTexture(m_state);
 
-		m_textures[t] = true;
+		m_textures.insert(t);
 
 		const GSLocalMemory::BlockOffset* bo = m_state->m_mem.GetBlockOffset(TEX0.TBP0, TEX0.TBW, TEX0.PSM);
 
@@ -100,16 +100,17 @@ const GSTextureCacheSW::GSTexture* GSTextureCacheSW::Lookup(const GIFRegTEX0& TE
 
 				list<GSTexture*>* m = &m_map[i << 5];
 
-				for(int j = 0; j < 32; j++)
+				unsigned long j;
-				{
+
-					if(p & (1 << j))
+				while(_BitScanForward(&j, p))
 				{
+					p ^= 1 << j;
+
 					m[j].push_front(t);
 				}
 			}
 		}
 	}
-	}
 
 	if(!t->Update(TEX0, TEXA, r))
 	{
@@ -125,7 +126,7 @@ const GSTextureCacheSW::GSTexture* GSTextureCacheSW::Lookup(const GIFRegTEX0& TE
 
 void GSTextureCacheSW::RemoveAll()
 {
-	for_each(m_textures.begin(), m_textures.end(), delete_first());
+	for_each(m_textures.begin(), m_textures.end(), delete_object());
 
 	m_textures.clear();
 
@@ -156,11 +157,11 @@ void GSTextureCacheSW::RemoveAt(GSTexture* t)
 
 void GSTextureCacheSW::IncAge()
 {
-	for(hash_map<GSTexture*, bool>::iterator i = m_textures.begin(); i != m_textures.end(); )
+	for(hash_set<GSTexture*>::iterator i = m_textures.begin(); i != m_textures.end(); )
 	{
-		hash_map<GSTexture*, bool>::iterator j = i++;
+		hash_set<GSTexture*>::iterator j = i++;
 
-		GSTexture* t = j->first;
+		GSTexture* t = *j;
 
 		if(++t->m_age > 30)
 		{
@@ -242,8 +243,8 @@ bool GSTextureCacheSW::GSTexture::Update(const GIFRegTEX0& TEX0, const GIFRegTEX
 
 	GSVector2i bs = psm.bs;
 
-	int tw = max(1 << TEX0.TW, bs.x);
+	int tw = std::max<int>(1 << TEX0.TW, bs.x);
-	int th = max(1 << TEX0.TH, bs.y);
+	int th = std::max<int>(1 << TEX0.TH, bs.y);
 
 	GSVector4i r = rect.ralign<GSVector4i::Outside>(bs);
 

plugins/GSdx/GSTextureCacheSW.h

@@ -46,7 +46,7 @@ public:
 
 protected:
 	GSState* m_state;
-	hash_map<GSTexture*, bool> m_textures;
+	hash_set<GSTexture*> m_textures;
 	list<GSTexture*> m_map[MAX_PAGES];
 	uint32 m_pages[16];
 

plugins/GSdx/GSTextureFX.h

@@ -45,12 +45,13 @@ public:
 	{
 		GSVector4 VertexScale;
 		GSVector4 VertexOffset;
-		GSVector2 TextureScale;
+		GSVector4 TextureScale;
-		float _pad[2];
 
 		struct VSConstantBuffer() 
 		{
-			memset(this, 0, sizeof(*this));
+			VertexScale = GSVector4::zero();
+			VertexOffset = GSVector4::zero();
+			TextureScale = GSVector4::zero();
 		}
 
 		__forceinline bool Update(const VSConstantBuffer* cb)
@@ -85,13 +86,12 @@ public:
 				uint32 tme:1;
 				uint32 fst:1;
 				uint32 logz:1;
-				uint32 prim:2;
 			};
 
 			uint32 key;
 		};
 
-		operator uint32() {return key & 0x7f;}
+		operator uint32() {return key & 0x1f;}
 
 		VSSelector() : key(0) {}
 	};
@@ -107,7 +107,12 @@ public:
 
 		struct PSConstantBuffer() 
 		{
-			memset(this, 0, sizeof(*this));
+			FogColor_AREF = GSVector4::zero();
+			HalfTexel = GSVector4::zero();
+			WH = GSVector4::zero();
+			MinMax = GSVector4::zero();
+			MinF_TA = GSVector4::zero();
+			MskFix = GSVector4i::zero();
 		}
 
 		__forceinline bool Update(const PSConstantBuffer* cb)
@@ -212,7 +217,6 @@ public:
 		{
 			struct
 			{
-				uint32 zte:1;
 				uint32 ztst:2;
 				uint32 zwe:1;
 				uint32 date:1;
@@ -222,7 +226,7 @@ public:
 			uint32 key;
 		};
 
-		operator uint32() {return key & 0x3f;}
+		operator uint32() {return key & 0x1f;}
 
 		OMDepthStencilSelector() : key(0) {}
 	};
@@ -244,6 +248,13 @@ public:
 				uint32 wa:1;
 			};
 
+			struct
+			{
+				uint32 _pad:1;
+				uint32 abcd:8;
+				uint32 wrgba:4;
+			};
+
 			uint32 key;
 		};
 
@@ -273,7 +284,7 @@ public:
 	virtual void SetupGS(GSSelector sel) = 0;
 	virtual void SetupPS(PSSelector sel, const PSConstantBuffer* cb, PSSamplerSelector ssel, GSTexture* tex, GSTexture* pal) = 0;
 	virtual void UpdatePS(PSSelector sel, const PSConstantBuffer* cb, PSSamplerSelector ssel) = 0;
-	virtual void SetupRS(int w, int h, const GSVector4i& scissor) = 0;
+	virtual void SetupRS(const GSVector2i& size, const GSVector4i& scissor) = 0;
 	virtual void SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix, GSTexture* rt, GSTexture* ds) = 0;
 	virtual void UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix) = 0;
 };

plugins/GSdx/GSTextureFX10.cpp

@@ -76,6 +76,13 @@ bool GSTextureFX10::Create(GSDevice* dev)
 
 	if(FAILED(hr)) return false;
 
+	// create layout
+
+	VSSelector sel;
+	VSConstantBuffer cb;
+
+	SetupVS(sel, &cb);
+
 	//
 
 	return true;
@@ -98,19 +105,17 @@ void GSTextureFX10::SetupVS(VSSelector sel, const VSConstantBuffer* cb)
 
 	if(i == m_vs.end())
 	{
-		string str[4];
+		string str[3];
 
 		str[0] = format("%d", sel.bppz);
 		str[1] = format("%d", sel.tme);
 		str[2] = format("%d", sel.fst);
-		str[3] = format("%d", sel.prim);
 
 		D3D10_SHADER_MACRO macro[] =
 		{
 			{"VS_BPPZ", str[0].c_str()},
 			{"VS_TME", str[1].c_str()},
 			{"VS_FST", str[2].c_str()},
-			{"VS_PRIM", str[3].c_str()},
 			{NULL, NULL},
 		};
 
@@ -297,9 +302,9 @@ void GSTextureFX10::UpdatePS(PSSelector sel, const PSConstantBuffer* cb, PSSampl
 	dev->PSSetSamplerState(ss0, ss1);
 }
 
-void GSTextureFX10::SetupRS(int w, int h, const GSVector4i& scissor)
+void GSTextureFX10::SetupRS(const GSVector2i& size, const GSVector4i& scissor)
 {
-	((GSDevice10*)m_dev)->RSSet(w, h, &scissor);
+	((GSDevice10*)m_dev)->RSSet(size, &scissor);
 }
 
 void GSTextureFX10::SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix, GSTexture* rt, GSTexture* ds)
@@ -312,11 +317,17 @@ void GSTextureFX10::SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel,
 void GSTextureFX10::UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix)
 {
 	GSDevice10* dev = (GSDevice10*)m_dev;
-
+/*
 	hash_map<uint32, CComPtr<ID3D10DepthStencilState> >::const_iterator i = m_om_dss.find(dssel);
 
 	if(i == m_om_dss.end())
 	{
+*/
+	CComPtr<ID3D10DepthStencilState>& om_dss = m_om_dss[dssel];
+
+	if(om_dss == NULL)
+	{
+
 		D3D10_DEPTH_STENCIL_DESC dsd;
 
 		memset(&dsd, 0, sizeof(dsd));
@@ -336,7 +347,7 @@ void GSTextureFX10::UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel,
 			dsd.BackFace.StencilDepthFailOp = D3D10_STENCIL_OP_KEEP;
 		}
 
-		if(!(dssel.zte && dssel.ztst == 1 && !dssel.zwe))
+		if(dssel.ztst != ZTST_ALWAYS || dssel.zwe)
 		{
 			static const D3D10_COMPARISON_FUNC ztst[] = 
 			{
@@ -346,11 +357,11 @@ void GSTextureFX10::UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel,
 				D3D10_COMPARISON_GREATER
 			};
 
-			dsd.DepthEnable = dssel.zte;
+			dsd.DepthEnable = true;
 			dsd.DepthWriteMask = dssel.zwe ? D3D10_DEPTH_WRITE_MASK_ALL : D3D10_DEPTH_WRITE_MASK_ZERO;
 			dsd.DepthFunc = ztst[dssel.ztst];
 		}
-
+/*
 		CComPtr<ID3D10DepthStencilState> dss;
 
 		(*dev)->CreateDepthStencilState(&dsd, &dss);
@@ -358,9 +369,13 @@ void GSTextureFX10::UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel,
 		m_om_dss[dssel] = dss;
 
 		i = m_om_dss.find(dssel);
+*/
+		(*dev)->CreateDepthStencilState(&dsd, &om_dss);
 	}
 
-	dev->OMSetDepthStencilState(i->second, 1);
+//	dev->OMSetDepthStencilState(i->second, 1);
+
+	dev->OMSetDepthStencilState(om_dss, 1);
 
 	hash_map<uint32, CComPtr<ID3D10BlendState> >::const_iterator j = m_om_bs.find(bsel);
 

plugins/GSdx/GSTextureFX10.h

@@ -34,7 +34,8 @@ class GSTextureFX10 : public GSTextureFX
 	CComPtr<ID3D10Buffer> m_ps_cb;
 	hash_map<uint32, CComPtr<ID3D10SamplerState> > m_ps_ss;
 	CComPtr<ID3D10SamplerState> m_palette_ss;
-	hash_map<uint32, CComPtr<ID3D10DepthStencilState> > m_om_dss;	
+	// hash_map<uint32, CComPtr<ID3D10DepthStencilState> > m_om_dss;
+	CComPtr<ID3D10DepthStencilState> m_om_dss[32];
 	hash_map<uint32, CComPtr<ID3D10BlendState> > m_om_bs;	
 
 	VSConstantBuffer m_vs_cb_cache;
@@ -50,7 +51,7 @@ public:
 	void SetupGS(GSSelector sel);
 	void SetupPS(PSSelector sel, const PSConstantBuffer* cb, PSSamplerSelector ssel, GSTexture* tex, GSTexture* pal);
 	void UpdatePS(PSSelector sel, const PSConstantBuffer* cb, PSSamplerSelector ssel);
-	void SetupRS(int w, int h, const GSVector4i& scissor);
+	void SetupRS(const GSVector2i& size, const GSVector4i& scissor);
 	void SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix, GSTexture* rt, GSTexture* ds);
 	void UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix);
 };

plugins/GSdx/GSTextureFX11.cpp

@@ -76,6 +76,13 @@ bool GSTextureFX11::Create(GSDevice* dev)
 
 	if(FAILED(hr)) return false;
 
+	// create layout
+
+	VSSelector sel;
+	VSConstantBuffer cb;
+
+	SetupVS(sel, &cb);
+
 	//
 
 	return true;
@@ -98,19 +105,17 @@ void GSTextureFX11::SetupVS(VSSelector sel, const VSConstantBuffer* cb)
 
 	if(i == m_vs.end())
 	{
-		string str[4];
+		string str[3];
 
 		str[0] = format("%d", sel.bppz);
 		str[1] = format("%d", sel.tme);
 		str[2] = format("%d", sel.fst);
-		str[3] = format("%d", sel.prim);
 
 		D3D11_SHADER_MACRO macro[] =
 		{
 			{"VS_BPPZ", str[0].c_str()},
 			{"VS_TME", str[1].c_str()},
 			{"VS_FST", str[2].c_str()},
-			{"VS_PRIM", str[3].c_str()},
 			{NULL, NULL},
 		};
 
@@ -301,9 +306,9 @@ void GSTextureFX11::UpdatePS(PSSelector sel, const PSConstantBuffer* cb, PSSampl
 	dev->PSSetSamplerState(ss0, ss1);
 }
 
-void GSTextureFX11::SetupRS(int w, int h, const GSVector4i& scissor)
+void GSTextureFX11::SetupRS(const GSVector2i& size, const GSVector4i& scissor)
 {
-	((GSDevice11*)m_dev)->RSSet(w, h, &scissor);
+	((GSDevice11*)m_dev)->RSSet(size, &scissor);
 }
 
 void GSTextureFX11::SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix, GSTexture* rt, GSTexture* ds)
@@ -340,7 +345,7 @@ void GSTextureFX11::UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel,
 			dsd.BackFace.StencilDepthFailOp = D3D11_STENCIL_OP_KEEP;
 		}
 
-		if(!(dssel.zte && dssel.ztst == 1 && !dssel.zwe))
+		if(dssel.ztst != ZTST_ALWAYS || dssel.zwe)
 		{
 			static const D3D11_COMPARISON_FUNC ztst[] = 
 			{
@@ -350,7 +355,7 @@ void GSTextureFX11::UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel,
 				D3D11_COMPARISON_GREATER
 			};
 
-			dsd.DepthEnable = dssel.zte;
+			dsd.DepthEnable = true;
 			dsd.DepthWriteMask = dssel.zwe ? D3D11_DEPTH_WRITE_MASK_ALL : D3D11_DEPTH_WRITE_MASK_ZERO;
 			dsd.DepthFunc = ztst[dssel.ztst];
 		}

plugins/GSdx/GSTextureFX11.h

@@ -50,7 +50,7 @@ public:
 	void SetupGS(GSSelector sel);
 	void SetupPS(PSSelector sel, const PSConstantBuffer* cb, PSSamplerSelector ssel, GSTexture* tex, GSTexture* pal);
 	void UpdatePS(PSSelector sel, const PSConstantBuffer* cb, PSSamplerSelector ssel);
-	void SetupRS(int w, int h, const GSVector4i& scissor);
+	void SetupRS(const GSVector2i& size, const GSVector4i& scissor);
 	void SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix, GSTexture* rt, GSTexture* ds);
 	void UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix);
 };

plugins/GSdx/GSTextureFX9.cpp

@@ -158,7 +158,7 @@ void GSTextureFX9::SetupPS(PSSelector sel, const PSConstantBuffer* cb, PSSampler
 	{
 		if(sel.wms == 3)
 		{
-			if(GSTexture* t = CreateMskFix(tex->GetWidth(), cb->MskFix.x, cb->MskFix.z))
+			if(GSTexture* t = CreateMskFix(tex->m_size.x, cb->MskFix.x, cb->MskFix.z))
 			{
 				(*dev)->SetTexture(2, *(GSTexture9*)t);
 			}
@@ -166,7 +166,7 @@ void GSTextureFX9::SetupPS(PSSelector sel, const PSConstantBuffer* cb, PSSampler
 
 		if(sel.wmt == 3)
 		{
-			if(GSTexture* t = CreateMskFix(tex->GetHeight(), cb->MskFix.y, cb->MskFix.w))
+			if(GSTexture* t = CreateMskFix(tex->m_size.y, cb->MskFix.y, cb->MskFix.w))
 			{
 				(*dev)->SetTexture(3, *(GSTexture9*)t);
 			}
@@ -265,9 +265,9 @@ void GSTextureFX9::UpdatePS(PSSelector sel, const PSConstantBuffer* cb, PSSample
 	dev->PSSetSamplerState(ss);
 }
 
-void GSTextureFX9::SetupRS(int w, int h, const GSVector4i& scissor)
+void GSTextureFX9::SetupRS(const GSVector2i& size, const GSVector4i& scissor)
 {
-	((GSDevice9*)m_dev)->RSSet(w, h, &scissor);
+	((GSDevice9*)m_dev)->RSSet(size, &scissor);
 }
 
 void GSTextureFX9::SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix, GSTexture* rt, GSTexture* ds)
@@ -303,7 +303,7 @@ void GSTextureFX9::UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel,
 			dss->StencilRef = 3;
 		}
 
-		if(!(dssel.zte && dssel.ztst == 1 && !dssel.zwe))
+		if(dssel.ztst != ZTST_ALWAYS || dssel.zwe)
 		{
 			static const D3DCMPFUNC ztst[] = 
 			{
@@ -313,7 +313,7 @@ void GSTextureFX9::UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel,
 				D3DCMP_GREATER
 			};
 
-			dss->DepthEnable = dssel.zte;
+			dss->DepthEnable = true;
 			dss->DepthWriteMask = dssel.zwe;
 			dss->DepthFunc = ztst[dssel.ztst];
 		}

plugins/GSdx/GSTextureFX9.h

@@ -48,7 +48,7 @@ public:
 	void SetupGS(GSSelector sel) {}
 	void SetupPS(PSSelector sel, const PSConstantBuffer* cb, PSSamplerSelector ssel, GSTexture* tex, GSTexture* pal);
 	void UpdatePS(PSSelector sel, const PSConstantBuffer* cb, PSSamplerSelector ssel);
-	void SetupRS(int w, int h, const GSVector4i& scissor);
+	void SetupRS(const GSVector2i& size, const GSVector4i& scissor);
 	void SetupOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix, GSTexture* rt, GSTexture* ds);
 	void UpdateOM(OMDepthStencilSelector dssel, OMBlendSelector bsel, uint8 afix);
 };

plugins/GSdx/GSTextureNull.h

@@ -32,9 +32,8 @@ public:
 	GSTextureNull(int type, int w, int h, int format);
 
 	int GetType() const {return m_desc.type;}
-	int GetWidth() const {return m_desc.w;}
-	int GetHeight() const {return m_desc.h;}
 	int GetFormat() const {return m_desc.format;}
+
 	bool Update(const GSVector4i& r, const void* data, int pitch) {return true;}
 	bool Map(GSMap& m, const GSVector4i* r) {return false;}
 	void Unmap() {}

plugins/GSdx/GSTextureOGL.cpp

@@ -26,10 +26,11 @@
 GSTextureOGL::GSTextureOGL(GLuint texture, int type, int width, int height, int format)
 	: m_texture(texture)
 	, m_type(type)
-	, m_width(width)
-	, m_height(height)
 	, m_format(format)
 {
+	m_size.x = width;
+	m_size.y = height;
+
 	// TODO: offscreen type should be just a memory array, also returned in Map
 
 	glGenBuffers(1, &m_pbo); GSDeviceOGL::CheckError();
@@ -61,16 +62,6 @@ int GSTextureOGL::GetType() const
 	return m_type;
 }
 
-int GSTextureOGL::GetWidth() const 
-{
-	return m_width;
-}
-
-int GSTextureOGL::GetHeight() const 
-{
-	return m_height;
-}
-
 int GSTextureOGL::GetFormat() const 
 {
 	return m_format;

plugins/GSdx/GSTextureOGL.h

@@ -29,8 +29,6 @@ class GSTextureOGL : public GSTexture
 	GLuint m_pbo;
 
 	int m_type;
-	int m_width;
-	int m_height;
 	int m_format;
 
 public:
@@ -38,8 +36,6 @@ public:
 	virtual ~GSTextureOGL();
 
 	int GetType() const;
-	int GetWidth() const;
-	int GetHeight() const;
 	int GetFormat() const;
 
 	bool Update(const GSVector4i& r, const void* data, int pitch);

plugins/GSdx/GSUtil.h

@@ -22,6 +22,7 @@
 #pragma once
 
 #include "GS.h"
+#include "GSLocalMemory.h"
 
 class GSUtil
 {
@@ -32,26 +33,6 @@ public:
 	static bool HasSharedBits(uint32 sbp, uint32 spsm, uint32 dbp, uint32 dpsm);
 	static bool HasCompatibleBits(uint32 spsm, uint32 dpsm);
 
-	static uint32 EncodePSM(uint32 psm)
-	{
-		switch(psm)
-		{
-		case PSM_PSMCT32: 
-		case PSM_PSMZ32: 
-			return 0;
-		case PSM_PSMCT24: 
-		case PSM_PSMZ24:
-			return 1;
-		case PSM_PSMCT16: 
-		case PSM_PSMCT16S: 
-		case PSM_PSMZ16:
-		case PSM_PSMZ16S:
-			return 2;
-		default:
-			return 3;
-		}
-	}
-
 	static bool CheckDirectX();
 	static bool CheckSSE();
 

plugins/GSdx/GSVector.h

@@ -28,6 +28,11 @@ public:
 	{
 		return x == v.x && y == v.y;
 	}
+
+	bool operator != (const GSVector2T& v) const
+	{
+		return x != v.x || y != v.y;
+	}
 };
 
 typedef GSVector2T<float> GSVector2;

plugins/GSdx/res/tfx.fx

@@ -12,7 +12,6 @@
 #define VS_BPPZ 0
 #define VS_TME 1
 #define VS_FST 1
-#define VS_PRIM 0
 #endif
 
 #ifndef GS_IIP
@@ -520,11 +519,6 @@ VS_OUTPUT vs_main(VS_INPUT input)
 		input.z = input.z & 0xffff;
 	}
 
-	if(VS_PRIM == 3) // sprite
-	{
-		//input.p.xy = (input.p.xy + 15) & ~15; // HACK
-	}
-
 	VS_OUTPUT output;
 	
 	// pos -= 0.05 (1/320 pixel) helps avoiding rounding problems (integral part of pos is usually 5 digits, 0.05 is about as low as we can go)

plugins/GSdx/stdafx.h

@@ -53,6 +53,7 @@
 #include <list>
 #include <map>
 #include <hash_map>
+#include <hash_set>
 #include <algorithm>
 
 using namespace std;