Almost there.

Source/Core/VideoCommon/Src/PixelShaderGen.cpp

@@ -1274,7 +1274,7 @@ static void WriteAlphaTest(char *&p, API_TYPE ApiType,DSTALPHA_MODE dstAlphaMode
 	WRITE(p, tevAlphaFuncsTable[compindex],alphaRef[1]);//lookup the second component from the alpha function table
 	WRITE(p, ")) {\n");
 
-	WRITE(p, "ocol0 = 0;\n");
+	WRITE(p, "ocol0 = float4(0);\n");
 	if (dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
 		WRITE(p, "ocol1 = 0;\n");
 	if (DepthTextureEnable)

Source/Core/VideoCommon/Src/VertexShaderGen.cpp

@@ -291,7 +291,7 @@ const char *GenerateVertexShaderCode(u32 components, API_TYPE ApiType)
 		}
 		else
 		{
-			WRITE(p, "int posmtx = fposmtx;\n");
+			WRITE(p, "int posmtx = int(fposmtx);\n");
 		}
 
 		WRITE(p, "float4 pos = float4(dot("I_TRANSFORMMATRICES"[posmtx], rawpos), dot("I_TRANSFORMMATRICES"[posmtx+1], rawpos), dot("I_TRANSFORMMATRICES"[posmtx+2], rawpos), 1);\n");		

Source/Plugins/Plugin_VideoOGL/Src/PixelShaderCache.cpp

@@ -38,8 +38,8 @@ namespace OGL
 {
 
 static int s_nMaxPixelInstructions;
-static GLuint s_ColorMatrixProgram = 0;
+static FRAGMENTSHADER s_ColorMatrixProgram;
-static GLuint s_DepthMatrixProgram = 0;
+static FRAGMENTSHADER s_DepthMatrixProgram;
 PixelShaderCache::PSCache PixelShaderCache::PixelShaders;
 PIXELSHADERUID PixelShaderCache::s_curuid;
 bool PixelShaderCache::s_displayCompileAlert;
@@ -56,12 +56,12 @@ bool (*pCompilePixelShader)(FRAGMENTSHADER&, const char*);
 
 GLuint PixelShaderCache::GetDepthMatrixProgram()
 {
-    return s_DepthMatrixProgram;
+    return s_DepthMatrixProgram.glprogid;
 }
 
 GLuint PixelShaderCache::GetColorMatrixProgram()
 {
-    return s_ColorMatrixProgram;
+    return s_ColorMatrixProgram.glprogid;
 }
 
 void PixelShaderCache::Init()
@@ -105,7 +105,66 @@ void PixelShaderCache::Init()
     glGetProgramivARB(GL_FRAGMENT_PROGRAM_ARB, GL_MAX_PROGRAM_NATIVE_INSTRUCTIONS_ARB, (GLint *)&maxinst);
     glGetProgramivARB(GL_FRAGMENT_PROGRAM_ARB, GL_MAX_PROGRAM_NATIVE_ATTRIBS_ARB, (GLint *)&maxattribs);
     INFO_LOG(VIDEO, "pixel max_alu=%d, max_inst=%d, max_attrib=%d", s_nMaxPixelInstructions, maxinst, maxattribs);
-
+	if(g_ActiveConfig.bUseGLSL)
+	{
+		char pmatrixprog[2048];
+		sprintf(pmatrixprog, "#extension GL_ARB_texture_rectangle : enable\n"
+			"uniform sampler2DRect samp0;\n"
+			"uniform vec4 "I_COLORS"[7];\n"
+			"void main(){\n"
+			"vec4 Temp0, Temp1;\n"
+			"vec4 K0 = vec4(0.5, 0.5, 0.5, 0.5);\n"
+			"Temp0 = texture2DRect(samp0, gl_TexCoord[0].xy);\n"
+			"Temp0 = Temp0 * "I_COLORS"[%d];\n"
+			"Temp0 = Temp0 + K0;\n"
+			"Temp0 = floor(Temp0);\n"
+			"Temp0 = Temp0 * "I_COLORS"[%d];\n"
+			"Temp1.x = dot(Temp0, "I_COLORS"[%d]);\n"
+			"Temp1.y = dot(Temp0, "I_COLORS"[%d]);\n"
+			"Temp1.z = dot(Temp0, "I_COLORS"[%d]);\n"
+			"Temp1.w = dot(Temp0, "I_COLORS"[%d]);\n"
+			"gl_FragData[0] = Temp1 + "I_COLORS"[%d];\n"
+			"}\n", C_COLORS+5, C_COLORS+6, C_COLORS, C_COLORS+1, C_COLORS+2, C_COLORS+3, C_COLORS+4);
+		if(!pCompilePixelShader(s_ColorMatrixProgram, pmatrixprog))
+		{
+			ERROR_LOG(VIDEO, "Failed to create color matrix fragment program");
+			s_ColorMatrixProgram.Destroy();
+		}
+		sprintf(pmatrixprog, "#extension GL_ARB_texture_rectangle : enable\n"
+			"uniform sampler2DRect samp0;\n"
+			"uniform vec4 "I_COLORS"[5];\n"
+			"void main(){\n"
+			"vec4 R0, R1, R2;\n"
+			"vec4 K0 = vec4(255.99998474121, 0.003921568627451, 256.0, 0.0);\n"
+			"vec4 K1 = vec4(15.0, 0.066666666666, 0.0, 0.0);\n"
+			"R2 = texture2DRect(samp0, gl_TexCoord[0].xy);\n"
+			"R0.x = R2.x * K0.x;\n"
+			"R0.x = floor(R0).x;\n"
+			"R0.yzw = (R0 - R0.x).yzw;\n"
+			"R0.yzw = (R0 * K0.z).yzw;\n"
+			"R0.y = floor(R0).y;\n"
+			"R0.zw = (R0 - R0.y).zw;\n"
+			"R0.zw = (R0 * K0.z).zw;\n"
+			"R0.z = floor(R0).z;\n"
+			"R0.w = R0.x;\n"
+			"R0 = R0 * K0.y;\n"
+			"R0.w = (R0 * K1.x).w;\n"
+			"R0.w = floor(R0).w;\n"
+			"R0.w = (R0 * K1.y).w;\n"
+			"R1.x = dot(R0, "I_COLORS"[%d]);\n"
+			"R1.y = dot(R0, "I_COLORS"[%d]);\n"
+			"R1.z = dot(R0, "I_COLORS"[%d]);\n"
+			"R1.w = dot(R0, "I_COLORS"[%d]);\n"
+			"gl_FragData[0] = R1 * "I_COLORS"[%d];\n"
+			"}\n", C_COLORS, C_COLORS+1, C_COLORS+2, C_COLORS+3, C_COLORS+4);
+		if(!pCompilePixelShader(s_DepthMatrixProgram, pmatrixprog))
+		{
+			ERROR_LOG(VIDEO, "Failed to create depth matrix fragment program");
+			s_DepthMatrixProgram.Destroy();
+		}
+	}
+	else
+	{
 		char pmatrixprog[2048];
 		sprintf(pmatrixprog, "!!ARBfp1.0"
 				"TEMP R0;\n"
@@ -121,16 +180,15 @@ void PixelShaderCache::Init()
 				"DP4 R1.z, R0, program.env[%d];\n"
 				"DP4 R1.w, R0, program.env[%d];\n"
 				"ADD result.color, R1, program.env[%d];\n"
-            "END\n",C_COLORMATRIX+5,C_COLORMATRIX+6, C_COLORMATRIX, C_COLORMATRIX+1, C_COLORMATRIX+2, C_COLORMATRIX+3, C_COLORMATRIX+4);
+				"END\n",C_COLORS+5,C_COLORS+6, C_COLORS, C_COLORS+1, C_COLORS+2, C_COLORS+3, C_COLORS+4);
-    glGenProgramsARB(1, &s_ColorMatrixProgram);
+		glGenProgramsARB(1, &s_ColorMatrixProgram.glprogid);
-    SetCurrentShader(s_ColorMatrixProgram);
+		SetCurrentShader(s_ColorMatrixProgram.glprogid);
 		glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, (GLsizei)strlen(pmatrixprog), pmatrixprog);
 
 		GLenum err = GL_REPORT_ERROR();
 		if (err != GL_NO_ERROR) {
 			ERROR_LOG(VIDEO, "Failed to create color matrix fragment program");
-        glDeleteProgramsARB(1, &s_ColorMatrixProgram);
+			s_ColorMatrixProgram.Destroy();
-        s_ColorMatrixProgram = 0;
 		}
 
 		sprintf(pmatrixprog, "!!ARBfp1.0\n"
@@ -177,26 +235,24 @@ void PixelShaderCache::Init()
 				"DP4 R1.z, R0, program.env[%d];\n"
 				"DP4 R1.w, R0, program.env[%d];\n"
 				"ADD result.color, R1, program.env[%d];\n"
-            "END\n", C_COLORMATRIX, C_COLORMATRIX+1, C_COLORMATRIX+2, C_COLORMATRIX+3, C_COLORMATRIX+4);
+				"END\n", C_COLORS, C_COLORS+1, C_COLORS+2, C_COLORS+3, C_COLORS+4);
-    glGenProgramsARB(1, &s_DepthMatrixProgram);
+		glGenProgramsARB(1, &s_DepthMatrixProgram.glprogid);
-    SetCurrentShader(s_DepthMatrixProgram);
+		SetCurrentShader(s_DepthMatrixProgram.glprogid);
 		glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, (GLsizei)strlen(pmatrixprog), pmatrixprog);
 
 		err = GL_REPORT_ERROR();
 		if (err != GL_NO_ERROR) {
 			ERROR_LOG(VIDEO, "Failed to create depth matrix fragment program");
-        glDeleteProgramsARB(1, &s_DepthMatrixProgram);
+			s_DepthMatrixProgram.Destroy();
-        s_DepthMatrixProgram = 0;
+		}
 	}
 
 }
 
 void PixelShaderCache::Shutdown()
 {
-    glDeleteProgramsARB(1, &s_ColorMatrixProgram);
+	s_ColorMatrixProgram.Destroy();
-    s_ColorMatrixProgram = 0;
+	s_DepthMatrixProgram.Destroy();
-    glDeleteProgramsARB(1, &s_DepthMatrixProgram);
-    s_DepthMatrixProgram = 0;
     PSCache::iterator iter = PixelShaders.begin();
     for (; iter != PixelShaders.end(); iter++)
         iter->second.Destroy();
@@ -235,7 +291,6 @@ FRAGMENTSHADER* PixelShaderCache::SetShader(DSTALPHA_MODE dstAlphaMode, u32 comp
     // Make an entry in the table
     PSCacheEntry& newentry = PixelShaders[uid];
     last_entry = &newentry;
-    newentry.shader.bGLSL = g_ActiveConfig.bUseGLSL;
     const char *code = GeneratePixelShaderCode(dstAlphaMode, g_ActiveConfig.bUseGLSL ? API_GLSL : API_OPENGL, components);
 
     if (g_ActiveConfig.bEnableShaderDebugging && code)
@@ -320,7 +375,6 @@ bool CompileGLSLPixelShader(FRAGMENTSHADER& ps, const char* pstrprogram)
 				FILE *fp = fopen(szTemp, "wb");
 				fwrite(pstrprogram, strlen(pstrprogram), 1, fp);
 				fclose(fp);
-		
 		if(strstr(infoLog, "warning") != NULL || strstr(infoLog, "error") != NULL)
 			exit(0);
 		delete[] infoLog;
@@ -339,6 +393,7 @@ bool CompileGLSLPixelShader(FRAGMENTSHADER& ps, const char* pstrprogram)
 
 	(void)GL_REPORT_ERROR();
 	ps.glprogid = result;
+	ps.bGLSL = true;
 	return true;
 }
 void PixelShaderCache::SetPSSampler(const char * name, unsigned int Tex)
@@ -462,6 +517,7 @@ bool CompileCGPixelShader(FRAGMENTSHADER& ps, const char* pstrprogram)
     }
 
     glGenProgramsARB(1, &ps.glprogid);
+    ps.bGLSL = false;
     PixelShaderCache::SetCurrentShader(ps.glprogid);
     glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, (GLsizei)strlen(pcompiledprog), pcompiledprog);
 

Source/Plugins/Plugin_VideoOGL/Src/TextureCache.cpp

@@ -43,6 +43,7 @@
 #include "ImageWrite.h"
 #include "MemoryUtil.h"
 #include "PixelShaderCache.h"
+#include "ProgramShaderCache.h"
 #include "PixelShaderManager.h"
 #include "Render.h"
 #include "Statistics.h"
@@ -299,6 +300,12 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 
 		glViewport(0, 0, virtual_width, virtual_height);
 
+		if(g_ActiveConfig.bUseGLSL)
+		{
+			ProgramShaderCache::SetBothShaders((srcFormat == PIXELFMT_Z24) ? PixelShaderCache::GetDepthMatrixProgram() : PixelShaderCache::GetColorMatrixProgram(), 0);
+			PixelShaderCache::SetPSSampler("samp0", 0);
+		}
+		else
 			PixelShaderCache::SetCurrentShader((srcFormat == PIXELFMT_Z24) ? PixelShaderCache::GetDepthMatrixProgram() : PixelShaderCache::GetColorMatrixProgram());    
 		PixelShaderManager::SetColorMatrix(colmat); // set transformation
 		GL_REPORT_ERRORD();

Source/Plugins/Plugin_VideoOGL/Src/VertexManager.cpp

@@ -129,9 +129,6 @@ void VertexManager::vFlush()
 	//glBufferData(GL_ARRAY_BUFFER, s_pCurBufferPointer - LocalVBuffer, LocalVBuffer, GL_STREAM_DRAW);
 	GL_REPORT_ERRORD();
 
-	// setup the pointers
-	if (g_nativeVertexFmt)
-		g_nativeVertexFmt->SetupVertexPointers();
 	GL_REPORT_ERRORD();
 
 	u32 usedtextures = 0;
@@ -214,6 +211,21 @@ void VertexManager::vFlush()
 	VertexShaderManager::SetConstants();
 	PixelShaderManager::SetConstants();
 	
+	// setup the pointers
+	if (g_nativeVertexFmt)
+		g_nativeVertexFmt->SetupVertexPointers();
+	GL_REPORT_ERRORD();
+	if(g_ActiveConfig.bUseGLSL)
+		for (int i = 0; i < 8; i++)
+		{
+				if (usedtextures & (1 << i))
+				{
+						char tmp[16];
+						sprintf(tmp, "samp%d", i); // Bake this in to something so we don't have to sprintf?
+						PixelShaderCache::SetPSSampler(tmp, i);
+				}
+		}
+
 	Draw();
 
 	// run through vertex groups again to set alpha
@@ -224,6 +236,17 @@ void VertexManager::vFlush()
 		{
 			ProgramShaderCache::SetBothShaders(ps->glprogid, 0);
 			PixelShaderManager::SetConstants(); // Need to set these again
+			if (g_nativeVertexFmt)
+				g_nativeVertexFmt->SetupVertexPointers();
+			for (int i = 0; i < 8; i++)
+			{
+					if (usedtextures & (1 << i))
+					{
+							char tmp[16];
+							sprintf(tmp, "samp%d", i); // Bake this in to something so we don't have to sprintf?
+							PixelShaderCache::SetPSSampler(tmp, i);
+					}
+			}
 		}
 		else
 			if (ps) PixelShaderCache::SetCurrentShader(ps->glprogid);

Source/Plugins/Plugin_VideoOGL/Src/VertexShaderCache.cpp

@@ -125,7 +125,6 @@ VERTEXSHADER* VertexShaderCache::SetShader(u32 components)
     // Make an entry in the table
     VSCacheEntry& entry = vshaders[uid];
     last_entry = &entry;
-    entry.shader.bGLSL = g_ActiveConfig.bUseGLSL;
     const char *code = GenerateVertexShaderCode(components, g_ActiveConfig.bUseGLSL ? API_GLSL : API_OPENGL);
     GetSafeVertexShaderId(&entry.safe_uid, components);
 
@@ -223,6 +222,7 @@ bool CompileGLSLVertexShader(VERTEXSHADER& vs, const char* pstrprogram)
 
 	(void)GL_REPORT_ERROR();
 	vs.glprogid = result;
+	vs.bGLSL = true;
 	return true;
 }
 void SetVSConstant4fvByName(const char * name, unsigned int offset, const float *f, const unsigned int count = 1)
@@ -352,6 +352,7 @@ bool CompileCGVertexShader(VERTEXSHADER& vs, const char* pstrprogram)
         plocal = strstr(plocal + 13, "program.local");
     }
     glGenProgramsARB(1, &vs.glprogid);
+    vs.bGLSL = false;
     VertexShaderCache::SetCurrentShader(vs.glprogid);
 
     glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, (GLsizei)strlen(pcompiledprog), pcompiledprog);