GSdx: replaced switch with ifs in the shaders, kingdom hearts should be fixed

git-svn-id: http://pcsx2.googlecode.com/svn/trunk@1362 96395faa-99c1-11dd-bbfe-3dabce05a288

plugins/GSdx/GS.cpp

@@ -126,9 +126,11 @@ static INT32 GSopen(void* dsp, char* title, int mt, int renderer)
 	case 3: s_gs = new GSRendererHW10(s_basemem, !!mt, s_irq); break;
 	case 4: s_gs = new GSRendererSW(s_basemem, !!mt, s_irq, new GSDevice10()); break;
 	case 5: s_gs = new GSRendererNull(s_basemem, !!mt, s_irq, new GSDevice10()); break;
+	#if 0
 	case 6: s_gs = new GSRendererOGL(s_basemem, !!mt, s_irq); break;
 	case 7: s_gs = new GSRendererSW(s_basemem, !!mt, s_irq, new GSDeviceOGL()); break;
 	case 8: s_gs = new GSRendererNull(s_basemem, !!mt, s_irq, new GSDeviceOGL()); break;
+	#endif
 	case 9: s_gs = new GSRendererSW(s_basemem, !!mt, s_irq, new GSDeviceNull()); break;
 	case 10: s_gs = new GSRendererNull(s_basemem, !!mt, s_irq, new GSDeviceNull()); break;
 	}

plugins/GSdx/GSDevice10.cpp

@@ -221,7 +221,7 @@ bool GSDevice10::Create(GSWnd* wnd, bool vsync)
 
 	//
 
-	Reset(1, 1, true);
+	Reset(1, 1, false);
 
 	//
 

plugins/GSdx/GSDevice10.h

@@ -26,7 +26,6 @@
 
 class GSDevice10 : public GSDevice
 {
-private:
 	ID3D10Buffer* m_vb;
 	size_t m_vb_stride;
 	ID3D10InputLayout* m_layout;

plugins/GSdx/GSDevice7.cpp

@@ -75,7 +75,7 @@ bool GSDevice7::Create(GSWnd* wnd, bool vsync)
 		return false;
 	}
 
-	Reset(1, 1, true);
+	Reset(1, 1, false);
 
 	return true;
 }

plugins/GSdx/GSDeviceNull.cpp

@@ -29,7 +29,7 @@ bool GSDeviceNull::Create(GSWnd* wnd, bool vsync)
 		return false;
 	}
 
-	Reset(1, 1, true);
+	Reset(1, 1, false);
 
 	return true;
 }

plugins/GSdx/GSDeviceOGL.cpp

@@ -27,11 +27,21 @@
 GSDeviceOGL::GSDeviceOGL()
 	: m_hDC(NULL)
 	, m_hGLRC(NULL)
+	, m_vbo(0)
+	, m_fbo(0)
+	, m_topology(-1)
 {
+	m_vertices.stride = 0;
+	m_vertices.start = 0;
+	m_vertices.count = 0;
+	m_vertices.limit = 0;
 }
 
 GSDeviceOGL::~GSDeviceOGL()
 {
+	if(m_vbo) glDeleteBuffers(1, &m_vbo);
+	if(m_fbo) glDeleteFramebuffers(1, &m_fbo);
+	
 	#ifdef _WINDOWS
 
 	if(m_hGLRC) {wglMakeCurrent(NULL, NULL); wglDeleteContext(m_hGLRC);}
@@ -75,24 +85,40 @@ bool GSDeviceOGL::Create(GSWnd* wnd, bool vsync)
 	if(!m_hDC) return false;
 
 	if(!SetPixelFormat(m_hDC, ChoosePixelFormat(m_hDC, &pfd), &pfd)) 
+	{
 		return false;
+	}
 
 	m_hGLRC = wglCreateContext(m_hDC);
 
 	if(!m_hGLRC) return false;
 
 	if(!wglMakeCurrent(m_hDC, m_hGLRC))
+	{
 		return false;
+	}
+
+	if(WGLEW_EXT_swap_control)
+	{
+		wglSwapIntervalEXT(vsync ? 1 : 0);
+	}
 
 	#endif
 
+	if(glewInit() != GLEW_OK)
+	{
+		return false;
+	}
+
 	// const char* exts = (const char*)glGetString(GL_EXTENSIONS);
 
-	// TODO
+	glGenBuffers(1, &m_vbo);
+	glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+	// TODO: setup layout?
 
-	Reset(1, 1, true);
+	GSVector4i r = wnd->GetClientRect();
 
-	//
+	Reset(r.width(), r.height(), false);
 
 	return true;
 }
@@ -102,12 +128,12 @@ bool GSDeviceOGL::Reset(int w, int h, bool fs)
 	if(!__super::Reset(w, h, fs))
 		return false;
 
-	m_backbuffer = new GSTextureOGL(0); // ???
+	m_backbuffer = new GSTextureOGL(0, GSTexture::RenderTarget, w, h);
 
 	glCullFace(GL_FRONT_AND_BACK);
 	glDisable(GL_LIGHTING);
 	glDisable(GL_ALPHA_TEST);
-	glDisable(GL_SCISSOR_TEST);
+	glEnable(GL_SCISSOR_TEST);
 
 	glMatrixMode(GL_PROJECTION);
 	glLoadIdentity();
@@ -116,76 +142,134 @@ bool GSDeviceOGL::Reset(int w, int h, bool fs)
 
 	glHint(GL_PERSPECTIVE_CORRECTION_HINT, GL_NICEST);
 
-	// glViewport(0, 0, ?, ?);
+	// glViewport(0, 0, w, h);
+
+	if(m_fbo) {glDeleteFramebuffersEXT(1, &m_fbo); m_fbo = 0;}
+
+	glGenFramebuffers(1, &m_fbo);
+
+	if(m_fbo == 0) return false;
+
+	glBindFramebuffer(GL_FRAMEBUFFER, m_fbo);
 
 	return true;
 }
 
 void GSDeviceOGL::Flip()
 {
-	// TODO
+	#ifdef _WINDOWS
+
+	SwapBuffers(m_hDC);
+
+	#endif
 }
 
 void GSDeviceOGL::BeginScene()
 {
-	// TODO
 }
 
 void GSDeviceOGL::DrawPrimitive()
 {
-	// TODO
+	glDrawArrays(m_topology, m_vertices.count, m_vertices.start);
 }
 
 void GSDeviceOGL::EndScene()
 {
-	// TODO
+	m_vertices.start += m_vertices.count;
+	m_vertices.count = 0;
 }
 
 void GSDeviceOGL::ClearRenderTarget(GSTexture* t, const GSVector4& c)
 {
-	// TODO
+	// TODO: disable scissor, color mask
+
+    glClearColor(c.r, c.g, c.b, c.a);
+	glClear(GL_COLOR_BUFFER_BIT);
 }
 
 void GSDeviceOGL::ClearRenderTarget(GSTexture* t, uint32 c)
 {
-	// TODO
+	ClearRenderTarget(t, GSVector4(c) * (1.0f / 255));
 }
 
 void GSDeviceOGL::ClearDepth(GSTexture* t, float c)
 {
-	// TODO
+	// TODO: disable scissor, depth mask
+
+    glClearDepth(c);
+	glClear(GL_DEPTH_BUFFER_BIT);
 }
 
 void GSDeviceOGL::ClearStencil(GSTexture* t, uint8 c)
 {
-	// TODO
+	// TODO: disable scissor, depth (?) mask
+
+	glClearStencil((GLint)c);
+	glClear(GL_STENCIL_BUFFER_BIT);
 }
 
 GSTexture* GSDeviceOGL::Create(int type, int w, int h, int format)
 {
-	// TODO
+	GLuint texture = 0;
 
-	return NULL;
+	switch(type)
+	{
+	case GSTexture::RenderTarget:
+		glGenTextures(1, &texture);
+		glBindTexture(GL_TEXTURE_2D, texture);
+		glTexImage2D(GL_TEXTURE_2D, 0, format, w, h, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
+		break;
+	case GSTexture::DepthStencil:
+	    glGenRenderbuffersEXT(1, &texture);
+        glBindRenderbufferEXT(GL_RENDERBUFFER_EXT, texture);
+        glRenderbufferStorageEXT(GL_RENDERBUFFER_EXT, GL_DEPTH32F_STENCIL8, w, h);
+		// TODO: depth textures?
+		break;
+	case GSTexture::Texture:
+		glGenTextures(1, &texture);
+		glBindTexture(GL_TEXTURE_2D, texture);
+		glTexImage2D(GL_TEXTURE_2D, 0, format, w, h, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
+		break;
+	case GSTexture::Offscreen:
+		// TODO: ???
+		break;
+	}
+
+	if(!texture) return false;
+
+	GSTextureOGL* t = new GSTextureOGL(texture, type, w, h, format);
+
+	switch(type)
+	{
+	case GSTexture::RenderTarget:
+		ClearRenderTarget(t, 0);
+		break;
+	case GSTexture::DepthStencil:
+		ClearDepth(t, 0);
+		break;
+	}
+
+	return t;
 }
 
 GSTexture* GSDeviceOGL::CreateRenderTarget(int w, int h, int format)
 {
-	return __super::CreateRenderTarget(w, h, format ? format : 0); // TODO
+	return __super::CreateRenderTarget(w, h, format ? format : GL_RGBA8);
 }
 
 GSTexture* GSDeviceOGL::CreateDepthStencil(int w, int h, int format)
 {
-	return __super::CreateDepthStencil(w, h, format ? format : 0); // TODO
+	return __super::CreateDepthStencil(w, h, format ? format : GL_DEPTH32F_STENCIL8); // TODO: GL_DEPTH24_STENCIL8_EXT, GL_DEPTH24_STENCIL8
 }
 
 GSTexture* GSDeviceOGL::CreateTexture(int w, int h, int format)
 {
-	return __super::CreateTexture(w, h, format ? format : 0); // TODO
+	return __super::CreateTexture(w, h, format ? format : GL_RGBA8);
 }
 
 GSTexture* GSDeviceOGL::CreateOffscreen(int w, int h, int format)
 {
-	return __super::CreateOffscreen(w, h, format ? format : 0); // TODO
+	return __super::CreateOffscreen(w, h, format ? format : GL_RGBA8);
 }
 
 GSTexture* GSDeviceOGL::CopyOffscreen(GSTexture* src, const GSVector4& sr, int w, int h, int format)
@@ -209,3 +293,50 @@ void GSDeviceOGL::DoInterlace(GSTexture* st, GSTexture* dt, int shader, bool lin
 {
 	// TODO
 }
+
+void GSDeviceOGL::IASetVertexBuffer(const void* vertices, size_t stride, size_t count)
+{
+	ASSERT(m_vertices.count == 0);
+
+	glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+
+	if(count * stride > m_vertices.limit * m_vertices.stride)
+	{
+		m_vertices.start = 0;
+		m_vertices.count = 0;
+		m_vertices.limit = 0;
+	}
+
+	if(m_vertices.limit == 0)
+	{
+		m_vertices.limit = max(count * 3 / 2, 10000);
+
+		glBufferData(GL_ARRAY_BUFFER, m_vertices.limit * stride, NULL, GL_DYNAMIC_DRAW); // GL_STREAM_DRAW?
+	}
+
+	if(m_vertices.start + count > m_vertices.limit || stride != m_vertices.stride)
+	{
+		m_vertices.start = 0; // TODO: how to discard and not overwrite previous data?
+	}
+
+	if(GLvoid* v = glMapBuffer(GL_ARRAY_BUFFER, GL_WRITE_ONLY))
+	{
+		GSVector4i::storent((uint8*)v + m_vertices.start * stride, vertices, count * stride);
+
+		glUnmapBuffer(GL_ARRAY_BUFFER);
+	}
+
+	m_vertices.count = count;
+	m_vertices.stride = stride;
+}
+
+void GSDeviceOGL::IASetInputLayout()
+{
+	// TODO
+}
+
+void GSDeviceOGL::IASetPrimitiveTopology(int topology)
+{
+	m_topology = topology;
+}
+

plugins/GSdx/GSDeviceOGL.h

@@ -33,9 +33,23 @@ class GSDeviceOGL : public GSDevice
 	
 	#endif
 
+	GLuint m_vbo;
+	GLuint m_fbo;
+
+	struct
+	{
+		size_t stride, start, count, limit;
+	} m_vertices;
+
+	int m_topology;
+
+	//
+
 	static void OnStaticCgError(CGcontext ctx, CGerror err, void* p) {((GSDeviceOGL*)p)->OnCgError(ctx, err);}
 	void OnCgError(CGcontext ctx, CGerror err);
 
+	//
+
 	GSTexture* Create(int type, int w, int h, int format);
 
 	void DoMerge(GSTexture* st[2], GSVector4* sr, GSVector4* dr, GSTexture* dt, bool slbg, bool mmod, const GSVector4& c);
@@ -67,4 +81,8 @@ public:
 	GSTexture* CopyOffscreen(GSTexture* src, const GSVector4& sr, int w, int h, int format = 0);
 
 	void StretchRect(GSTexture* st, const GSVector4& sr, GSTexture* dt, const GSVector4& dr, int shader = 0, bool linear = true);
+
+	void IASetVertexBuffer(const void* vertices, size_t stride, size_t count);
+	void IASetInputLayout(); // TODO
+	void IASetPrimitiveTopology(int topology);
 };

plugins/GSdx/GSRendererHW10.cpp

@@ -336,16 +336,15 @@ void GSRendererHW10::Draw(int prim, GSTexture* rt, GSTexture* ds, GSTextureCache
 
 	GSTextureFX10::PSConstantBuffer ps_cb;
 
-	ps_cb.FogColorAREF = GSVector4((int)env.FOGCOL.FCR, (int)env.FOGCOL.FCG, (int)env.FOGCOL.FCB, (int)context->TEST.AREF) / 255;
+	ps_cb.FogColor_AREF = GSVector4((int)env.FOGCOL.FCR, (int)env.FOGCOL.FCG, (int)env.FOGCOL.FCB, (int)context->TEST.AREF) / 255;
-	ps_cb.TA = GSVector4((int)env.TEXA.TA0, (int)env.TEXA.TA1) / 255;
 
 	if(context->TEST.ATST == 2 || context->TEST.ATST == 5)
 	{
-		ps_cb.FogColorAREF.a -= 0.9f / 255;
+		ps_cb.FogColor_AREF.a -= 0.9f / 255;
 	}
 	else if(context->TEST.ATST == 3 || context->TEST.ATST == 6)
 	{
-		ps_cb.FogColorAREF.a += 0.9f / 255;
+		ps_cb.FogColor_AREF.a += 0.9f / 255;
 	}
 
 	if(tex)
@@ -355,28 +354,27 @@ void GSRendererHW10::Draw(int prim, GSTexture* rt, GSTexture* ds, GSTextureCache
 		int w = tex->m_texture->GetWidth();
 		int h = tex->m_texture->GetHeight();
 
+		ps_cb.WH_TA = GSVector4((int)(1 << context->TEX0.TW), (int)(1 << context->TEX0.TH), env.TEXA.TA0, env.TEXA.TA1) / GSVector4(1, 255).xxyy();
+		ps_cb.HalfTexel = GSVector4(-0.5f, 0.5f).xxyy() / GSVector4(w, h).xyxy();
+
 		switch(context->CLAMP.WMS)
 		{
 		case 0: 
-			ps_cb.MinMax.x = w - 1;
-			ps_cb.MinMax.z = 0;
 			ps_ssel.tau = 1; 
 			break;
 		case 1: 
-			ps_cb.MinMax.x = 0;
-			ps_cb.MinMax.z = w - 1;
 			ps_ssel.tau = 0; 
 			break;
 		case 2: 
-			ps_cb.MinMax.x = (int)context->CLAMP.MINU * w / (1 << context->TEX0.TW);
+			ps_cb.MinMax.x = ((float)(int)context->CLAMP.MINU) / (1 << context->TEX0.TW);
-			ps_cb.MinMax.z = (int)context->CLAMP.MAXU * w / (1 << context->TEX0.TW);
+			ps_cb.MinMax.z = ((float)(int)context->CLAMP.MAXU) / (1 << context->TEX0.TW);
 			ps_cb.MinMaxF.x = ((float)(int)context->CLAMP.MINU + 0.5f) / (1 << context->TEX0.TW);
 			ps_cb.MinMaxF.z = ((float)(int)context->CLAMP.MAXU) / (1 << context->TEX0.TW);
 			ps_ssel.tau = 0; 
 			break;
 		case 3: 
-			ps_cb.MinMax.x = context->CLAMP.MINU;
+			ps_cb.MskFix.x = context->CLAMP.MINU;
-			ps_cb.MinMax.z = context->CLAMP.MAXU;
+			ps_cb.MskFix.z = context->CLAMP.MAXU;
 			ps_ssel.tau = 1; 
 			break;
 		default: 
@@ -386,25 +384,21 @@ void GSRendererHW10::Draw(int prim, GSTexture* rt, GSTexture* ds, GSTextureCache
 		switch(context->CLAMP.WMT)
 		{
 		case 0: 
-			ps_cb.MinMax.y = h - 1;
-			ps_cb.MinMax.w = 0;
 			ps_ssel.tav = 1; 
 			break;
 		case 1: 
-			ps_cb.MinMax.y = 0;
-			ps_cb.MinMax.w = h - 1;
 			ps_ssel.tav = 0; 
 			break;
 		case 2: 
-			ps_cb.MinMax.y = (int)context->CLAMP.MINV * h / (1 << context->TEX0.TH);
+			ps_cb.MinMax.y = ((float)(int)context->CLAMP.MINV) / (1 << context->TEX0.TH);
-			ps_cb.MinMax.w = (int)context->CLAMP.MAXV * h / (1 << context->TEX0.TH);
+			ps_cb.MinMax.w = ((float)(int)context->CLAMP.MAXV) / (1 << context->TEX0.TH);
 			ps_cb.MinMaxF.y = ((float)(int)context->CLAMP.MINV + 0.5f) / (1 << context->TEX0.TH);
 			ps_cb.MinMaxF.w = ((float)(int)context->CLAMP.MAXV) / (1 << context->TEX0.TH);
 			ps_ssel.tav = 0; 
 			break;
 		case 3: 
-			ps_cb.MinMax.y = context->CLAMP.MINV;
+			ps_cb.MskFix.y = context->CLAMP.MINV;
-			ps_cb.MinMax.w = context->CLAMP.MAXV;
+			ps_cb.MskFix.w = context->CLAMP.MAXV;
 			ps_ssel.tav = 1; 
 			break;
 		default: 

plugins/GSdx/GSRendererHW9.cpp

@@ -308,17 +308,15 @@ void GSRendererHW9::Draw(int prim, GSTexture* rt, GSTexture* ds, GSTextureCache:
 
 	GSTextureFX9::PSConstantBuffer ps_cb;
 
-	ps_cb.FogColorAREF = GSVector4((int)env.FOGCOL.FCR, (int)env.FOGCOL.FCG, (int)env.FOGCOL.FCB, (int)context->TEST.AREF) / 255;
+	ps_cb.FogColor_AREF = GSVector4((int)env.FOGCOL.FCR, (int)env.FOGCOL.FCG, (int)env.FOGCOL.FCB, (int)context->TEST.AREF) / 255;
-	ps_cb.TA0 = (float)(int)env.TEXA.TA0 / 255;
-	ps_cb.TA1 = (float)(int)env.TEXA.TA1 / 255;
 
 	if(context->TEST.ATST == 2 || context->TEST.ATST == 5)
 	{
-		ps_cb.FogColorAREF.a -= 0.9f / 255;
+		ps_cb.FogColor_AREF.a -= 0.9f / 255;
 	}
 	else if(context->TEST.ATST == 3 || context->TEST.ATST == 6)
 	{
-		ps_cb.FogColorAREF.a += 0.9f / 255;
+		ps_cb.FogColor_AREF.a += 0.9f / 255;
 	}
 
 	if(tex)
@@ -328,8 +326,8 @@ void GSRendererHW9::Draw(int prim, GSTexture* rt, GSTexture* ds, GSTextureCache:
 		int w = tex->m_texture->GetWidth();
 		int h = tex->m_texture->GetHeight();
 
-		ps_cb.WH = GSVector2(w, h);
+		ps_cb.WH_TA = GSVector4(w, h, env.TEXA.TA0, env.TEXA.TA1) / GSVector4(1, 255).xxyy();
-		ps_cb.HalfTexel = GSVector4(-0.5f / w, -0.5f / h, +0.5f / w, +0.5f / h);
+		ps_cb.HalfTexel = GSVector4(-0.5f, 0.5f).xxyy() / GSVector4(w, h).xyxy();
 
 		switch(context->CLAMP.WMS)
 		{

plugins/GSdx/GSSettingsDlg.cpp

@@ -33,9 +33,11 @@ GSSetting GSSettingsDlg::g_renderers[] =
 	{3, "Direct3D10 (Hardware)", NULL},
 	{4, "Direct3D10 (Software)", NULL},
 	{5, "Direct3D10 (Null)", NULL},
+	#if 0
 	{6, "OpenGL (Hardware)", NULL},
 	{7, "OpenGL (Software)", NULL},
 	{8, "OpenGL (Null)", NULL},
+	#endif
 	{9, "Null (Software)", NULL},
 	{10, "Null (Null)", NULL},
 };

plugins/GSdx/GSTextureFX10.h

@@ -77,10 +77,12 @@ public:
 
 	__declspec(align(16)) struct PSConstantBuffer
 	{
-		GSVector4 FogColorAREF;
+		GSVector4 FogColor_AREF;
-		GSVector4 TA;
+		GSVector4 HalfTexel;
-		GSVector4i MinMax;
+		GSVector4 WH_TA;
+		GSVector4 MinMax;
 		GSVector4 MinMaxF;
+		GSVector4i MskFix;
 
 		struct PSConstantBuffer() {memset(this, 0, sizeof(*this));}
 
@@ -93,13 +95,17 @@ public:
 			GSVector4i b1 = b[1];
 			GSVector4i b2 = b[2];
 			GSVector4i b3 = b[3];
+			GSVector4i b4 = b[4];
+			GSVector4i b5 = b[5];
 
-			if(!((a[0] == b0) & (a[1] == b1) & (a[2] == b2) & (a[3] == b3)).alltrue())
+			if(!((a[0] == b0) & (a[1] == b1) & (a[2] == b2) & (a[3] == b3) & (a[4] == b4) & (a[5] == b5)).alltrue())
 			{
 				a[0] = b0;
 				a[1] = b1;
 				a[2] = b2;
 				a[3] = b3;
+				a[4] = b4;
+				a[5] = b5;
 
 				return true;
 			}

plugins/GSdx/GSTextureFX9.h

@@ -53,11 +53,9 @@ public:
 
 	struct PSConstantBuffer
 	{
-		GSVector4 FogColorAREF;
+		GSVector4 FogColor_AREF;
 		GSVector4 HalfTexel;
-		GSVector2 WH;
+		GSVector4 WH_TA;
-		float TA0;
-		float TA1;
 		GSVector4 MinMax;
 		GSVector4 MinMaxF;
 		GSVector4i MskFix;

plugins/GSdx/GSTextureOGL.cpp

@@ -22,19 +22,29 @@
 #include "stdafx.h"
 #include "GSTextureOGL.h"
 
-GSTextureOGL::GSTextureOGL(GLuint texture)
+GSTextureOGL::GSTextureOGL(GLuint texture, int type, int width, int height, int format)
 	: m_texture(texture)
-	, m_type(None)
+	, m_type(type)
-	, m_width(0)
+	, m_width(width)
-	, m_height(0)
+	, m_height(height)
-	, m_format(0)
+	, m_format(format)
 {
-	// TODO: m_type, m_format, fb/ds?
+}
 
-	glBindTexture(GL_TEXTURE_2D, texture);
+GSTextureOGL::~GSTextureOGL()
-
+{
-    glGetTexLevelParameteriv(GL_TEXTURE_2D, 0, GL_TEXTURE_WIDTH, &m_width);
+	if(m_texture)
-    glGetTexLevelParameteriv(GL_TEXTURE_2D, 0, GL_TEXTURE_HEIGHT, &m_height);
+	{
+		switch(m_type)
+		{
+		case DepthStencil:
+			glDeleteRenderbuffersEXT(1, &m_texture);
+			break;
+		default:
+			glDeleteTextures(1, &m_texture);
+			break;
+		}
+	}
 }
 
 int GSTextureOGL::GetType() const
@@ -59,7 +69,7 @@ int GSTextureOGL::GetFormat() const
 
 bool GSTextureOGL::Update(const GSVector4i& r, const void* data, int pitch)
 {
-	// TODO
+	// TODO: glTexSubImage2D looks like UpdateSubresource but does not take a pitch
 
 	return false;
 }

plugins/GSdx/GSTextureOGL.h

@@ -33,7 +33,8 @@ class GSTextureOGL : public GSTexture
 	int m_format;
 
 public:
-	explicit GSTextureOGL(GLuint texture);
+	GSTextureOGL(GLuint texture, int type, int width, int height, int format = 0);
+	virtual ~GSTextureOGL();
 
 	int GetType() const;
 	int GetWidth() const;

plugins/GSdx/res/tfx10.fx

@@ -155,9 +155,12 @@ cbuffer cb1
 {
 	float3 FogColor;
 	float AREF;
-	float4 TA;
+	float4 HalfTexel;
-	uint4 MinMax;
+	float2 WH;
+	float2 TA;
+	float4 MinMax;
 	float4 MinMaxF;
+	uint4 MskFix;
 };
 
 struct PS_INPUT
@@ -175,15 +178,15 @@ struct PS_OUTPUT
 
 #ifndef FST
 #define FST 0
-#define WMS 1
+#define WMS 3
-#define WMT 1
+#define WMT 3
 #define BPP 0
 #define AEM 0
 #define TFX 0
 #define TCC 1
-#define ATE 1
+#define ATE 0
 #define ATST 4
-#define FOG 1
+#define FOG 0
 #define CLR1 0
 #define FBA 0
 #define AOUT 0
@@ -207,58 +210,83 @@ float4 Extract16(uint i)
 	return f;
 }
 
-int4 wrapuv(int4 uv)
+float4 wrapuv(float4 uv)
 {
-	// TODO: bitwise ops don't work with render target textures that were not rendered at the native resolution.
-
 	if(WMS == WMT)
 	{
-		switch(WMS)
+		if(WMS == 0)
 		{
-		case 0: uv &= MinMax.xyxy; break;
+			uv = frac(uv);
-		case 1: uv = clamp(uv, 0, MinMax.zwzw); break;
+		}
-		case 2: uv = clamp(uv, MinMax.xyxy, MinMax.zwzw); break;
+		else if(WMS == 1)
-		case 3: uv = (uv & MinMax.xyxy) | MinMax.zwzw; break;
+		{
+			uv = saturate(uv);
+		}
+		else if(WMS == 2)
+		{
+			uv = clamp(uv, MinMax.xyxy, MinMax.zwzw);
+		}
+		else if(WMS == 3)
+		{
+			uv = (float4)(((int4)(uv * WH.xyxy) & MskFix.xyxy) | MskFix.zwzw) / WH.xyxy;
 		}
 	}
 	else
 	{
-		switch(WMS)
+		if(WMS == 0)
 		{
-		case 0: uv.xz &= MinMax.xx; break;
+			uv.xz = frac(uv.xz);
-		case 1: uv.xz = clamp(uv.xz, 0, MinMax.zz); break;
-		case 2: uv.xz = clamp(uv.xz, MinMax.xx, MinMax.zz); break;
-		case 3: uv.xz = (uv.xz & MinMax.xx) | MinMax.zz; break;
 		}
-		
+		else if(WMS == 1)
-		switch(WMT)
 		{
-		case 0: uv.yw &= MinMax.yy; break;
+			uv.xz = saturate(uv.xz);
-		case 1: uv.yw = clamp(uv.yw, 0, MinMax.ww); break;
+		}
-		case 2: uv.yw = clamp(uv.yw, MinMax.yy, MinMax.ww); break;
+		else if(WMS == 2)
-		case 3: uv.yw = (uv.yw & MinMax.yy) | MinMax.ww; break;
+		{
+			uv.xz = clamp(uv.xz, MinMax.xx, MinMax.zz);
+		}
+		else if(WMS == 3)
+		{
+			uv.xz = (float2)(((int2)(uv * WH.xyxy).xz & MskFix.xx) | MskFix.zz) / WH;
+		}
+	
+		if(WMT == 0)
+		{
+			uv.yw = frac(uv.yw);
+		}
+		else if(WMT == 1)
+		{
+			uv.yw = saturate(uv.yw);
+		}
+		else if(WMT == 2)
+		{
+			uv.yw = clamp(uv.yw, MinMax.yy, MinMax.ww);
+		}
+		else if(WMT == 3)
+		{
+			uv.yw = (float2)(((int2)(uv * WH.xyxy).yw & MskFix.yy) | MskFix.ww) / WH;
 		}
 	}
 	
 	return uv;
 }
 
-float2 clampuv(float2 tc)
+float2 clampuv(float2 uv)
 {
 	if(WMS == 2 && WMT == 2) 
 	{
-		tc = clamp(tc, MinMaxF.xy, MinMaxF.zw);
+		uv = clamp(uv, MinMaxF.xy, MinMaxF.zw);
 	}
 	else if(WMS == 2)
 	{
-		tc.x = clamp(tc.x, MinMaxF.x, MinMaxF.z);
+		uv.x = clamp(uv.x, MinMaxF.x, MinMaxF.z);
 	}
 	else if(WMT == 2)
 	{
-		tc.y = clamp(tc.y, MinMaxF.y, MinMaxF.w);
+		uv.y = clamp(uv.y, MinMaxF.y, MinMaxF.w);
 	}
 	
-	return tc;
+	return uv;
 }
 
 float4 sample(float2 tc, float w)
@@ -282,12 +310,11 @@ float4 sample(float2 tc, float w)
 	else
 	{
 		float w, h;
-
 		Texture.GetDimensions(w, h);
 		
-		float4 tc2 = (tc * float2(w, h)).xyxy + float4(-0.499, -0.499, 0.501, 0.501);
+		float4 uv2 = tc.xyxy + HalfTexel;
-		float2 dd = frac(tc2.xy);
+		float2 dd = frac(uv2.xy * float2(w, h)); 
-		int4 uv = wrapuv((int4)tc2);
+		float4 uv = wrapuv(uv2);
 
 		float4 t00, t01, t10, t11;
 
@@ -295,15 +322,15 @@ float4 sample(float2 tc, float w)
 		{
 			float4 a;
 
-			a.x = Texture.Load(int3(uv.xy, 0)).a;
+			a.x = Texture.Sample(TextureSampler, uv.xy).a;
-			a.y = Texture.Load(int3(uv.zy, 0)).a;
+			a.y = Texture.Sample(TextureSampler, uv.zy).a;
-			a.z = Texture.Load(int3(uv.xw, 0)).a;
+			a.z = Texture.Sample(TextureSampler, uv.xw).a;
-			a.w = Texture.Load(int3(uv.zw, 0)).a;
+			a.w = Texture.Sample(TextureSampler, uv.zw).a;
 
-			t00 = Palette.Load(a.x);
+			t00 = Palette.Sample(PaletteSampler, a.x);
-			t01 = Palette.Load(a.y);
+			t01 = Palette.Sample(PaletteSampler, a.y);
-			t10 = Palette.Load(a.z);
+			t10 = Palette.Sample(PaletteSampler, a.z);
-			t11 = Palette.Load(a.w);
+			t11 = Palette.Sample(PaletteSampler, a.w);
 		}
 		else if(BPP == 4) // 8HP + 16-bit palette
 		{
@@ -313,10 +340,10 @@ float4 sample(float2 tc, float w)
 		{
 			float4 r;
 
-			r.x = Texture.Load(int3(uv.xy, 0)).r;
+			r.x = Texture.Sample(TextureSampler, uv.xy).r;
-			r.y = Texture.Load(int3(uv.zy, 0)).r;
+			r.y = Texture.Sample(TextureSampler, uv.zy).r;
-			r.z = Texture.Load(int3(uv.xw, 0)).r;
+			r.z = Texture.Sample(TextureSampler, uv.xw).r;
-			r.w = Texture.Load(int3(uv.zw, 0)).r;
+			r.w = Texture.Sample(TextureSampler, uv.zw).r;
 			
 			uint4 i = r * 65535;
 
@@ -327,10 +354,10 @@ float4 sample(float2 tc, float w)
 		}
 		else
 		{
-			t00 = Texture.Load(int3(uv.xy, 0));
+			t00 = Texture.Sample(TextureSampler, uv.xy);
-			t01 = Texture.Load(int3(uv.zy, 0));
+			t01 = Texture.Sample(TextureSampler, uv.zy);
-			t10 = Texture.Load(int3(uv.xw, 0));
+			t10 = Texture.Sample(TextureSampler, uv.xw);
-			t11 = Texture.Load(int3(uv.zw, 0));
+			t11 = Texture.Sample(TextureSampler, uv.zw);
 		}
 
 		if(LTF)

plugins/GSdx/res/tfx9.fx

@@ -96,12 +96,12 @@ struct PS_INPUT
 #define AEM 0
 #define TFX 0
 #define TCC 1
-#define ATE 1
+#define ATE 0
 #define ATST 4
-#define FOG 1
+#define FOG 0
 #define CLR1 0
 #define RT 0
-#define LTF 1
+#define LTF 0
 #endif
 
 sampler Texture : register(s0);
@@ -113,63 +113,84 @@ float4 wrapuv(float4 uv)
 {
 	if(WMS == WMT)
 	{
-		switch(WMS)
+		if(WMS == 0)
+		{
+			uv = frac(uv);
+		}
+		else if(WMS == 1)
+		{
+			uv = saturate(uv);
+		}
+		else if(WMS == 2)
+		{
+			uv = clamp(uv, MinMax.xyxy, MinMax.zwzw);
+		}
+		else if(WMS == 3)
 		{
-		case 0: uv = frac(uv); break;
-		case 1: uv = saturate(uv); break;
-		case 2: uv = clamp(uv, MinMax.xyxy, MinMax.zwzw); break;
-		case 3: 
 			uv.x = tex1D(UMSKFIX, uv.x);
 			uv.y = tex1D(VMSKFIX, uv.y);
 			uv.z = tex1D(UMSKFIX, uv.z);
 			uv.w = tex1D(VMSKFIX, uv.w);
-			break;
 		}
 	}
 	else
 	{
-		switch(WMS)
+		if(WMS == 0)
+		{
+			uv.xz = frac(uv.xz);
+		}
+		else if(WMS == 1)
+		{
+			uv.xz = saturate(uv.xz);
+		}
+		else if(WMS == 2)
+		{
+			uv.xz = clamp(uv.xz, MinMax.xx, MinMax.zz);
+		}
+		else if(WMS == 3)
 		{
-		case 0: uv.xz = frac(uv.xz); break;
-		case 1: uv.xz = saturate(uv.xz); break;
-		case 2: uv.xz = clamp(uv.xz, MinMax.xx, MinMax.zz); break;
-		case 3:
 			uv.x = tex1D(UMSKFIX, uv.x);
 			uv.z = tex1D(UMSKFIX, uv.z);
-			break;
 		}
 
-		switch(WMT)
+		if(WMT == 0)
+		{
+			uv.yw = frac(uv.yw);
+		}
+		else if(WMT == 1)
+		{
+			uv.yw = saturate(uv.yw);
+		}
+		else if(WMT == 2)
+		{
+			uv.yw = clamp(uv.yw, MinMax.yy, MinMax.ww);
+		}
+		else if(WMT == 3)
 		{
-		case 0: uv.yw = frac(uv.yw); break;
-		case 1: uv.yw = saturate(uv.yw); break;
-		case 2: uv.yw = clamp(uv.yw, MinMax.yy, MinMax.ww); break;
-		case 3:
 			uv.y = tex1D(VMSKFIX, uv.y);
 			uv.w = tex1D(VMSKFIX, uv.w);
-			break;
 		}
 	}
 	
 	return uv;
 }
 
-float2 clampuv(float2 tc)
+float2 clampuv(float2 uv)
 {
 	if(WMS == 2 && WMT == 2) 
 	{
-		tc = clamp(tc, MinMaxF.xy, MinMaxF.zw);
+		uv = clamp(uv, MinMaxF.xy, MinMaxF.zw);
 	}
 	else if(WMS == 2)
 	{
-		tc.x = clamp(tc.x, MinMaxF.x, MinMaxF.z);
+		uv.x = clamp(uv.x, MinMaxF.x, MinMaxF.z);
 	}
 	else if(WMT == 2)
 	{
-		tc.y = clamp(tc.y, MinMaxF.y, MinMaxF.w);
+		uv.y = clamp(uv.y, MinMaxF.y, MinMaxF.w);
 	}
 	
-	return tc;
+	return uv;
 }
 
 float4 sample(float2 tc, float w)
@@ -192,9 +213,9 @@ float4 sample(float2 tc, float w)
 	}
 	else
 	{
-		float4 tc2 = tc.xyxy + HalfTexel;
+		float4 uv2 = tc.xyxy + HalfTexel;
-		float2 dd = frac(tc2.xy * WH); 
+		float2 dd = frac(uv2.xy * WH); 
-		float4 uv = wrapuv(tc2);
+		float4 uv = wrapuv(uv2);
 
 		float4 t00, t01, t10, t11;
 

plugins/GSdx/stdafx.h

@@ -115,10 +115,15 @@ typedef signed long long int64;
 
 // opengl
 
+#include <GL/glew.h>
 #include <GL/glut.h>
 #include <CG/cg.h>
 #include <CG/cgGL.h>
 
+#ifdef _WINDOWS
+#include <GL/wglew.h>
+#endif
+
 // sse
 
 #if !defined(_M_SSE) && (defined(_M_AMD64) || defined(_M_IX86_FP) && _M_IX86_FP >= 2)