Add some code for generating a shader constant usage profile.

Source/Core/VideoCommon/Src/PixelShaderGen.cpp

@@ -299,9 +299,9 @@ void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u
 	}
 	out.Write("\n");
 
-	out.Write("uniform float4 " I_COLORS"[4] : register(c%d);\n", C_COLORS);
+	out.Write("uniform float4 " I_COLORS"[4] : register(c%d);\n", C_COLORS); // TODO: first element not used??
 	out.Write("uniform float4 " I_KCOLORS"[4] : register(c%d);\n", C_KCOLORS);
-	out.Write("uniform float4 " I_ALPHA"[1] : register(c%d);\n", C_ALPHA);
+	out.Write("uniform float4 " I_ALPHA"[1] : register(c%d);\n", C_ALPHA); // TODO: Why is this an array...-.-
 	out.Write("uniform float4 " I_TEXDIMS"[8] : register(c%d);\n", C_TEXDIMS);
 	out.Write("uniform float4 " I_ZBIAS"[2] : register(c%d);\n", C_ZBIAS);
 	out.Write("uniform float4 " I_INDTEXSCALE"[2] : register(c%d);\n", C_INDTEXSCALE);
@@ -390,6 +390,8 @@ void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u
 		"float3 ldir, h;\n"
 		"float dist, dist2, attn;\n");
 /// TODO
+		out.SetConstantsUsed(C_PLIGHTS, C_PLIGHTS+39); // TODO: Can be optimized further
+		out.SetConstantsUsed(C_PMATERIALS, C_PMATERIALS+3);
 ///		p = GenerateLightingShader(p, components, I_PMATERIALS, I_PLIGHTS, "colors_", "colors_");
 	}
 
@@ -405,6 +407,7 @@ void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u
 	}
 	else
 	{
+		out.SetConstantsUsed(C_TEXDIMS, C_TEXDIMS+numTexgen-1);
 		for (unsigned int i = 0; i < numTexgen; ++i)
 		{
 			// optional perspective divides
@@ -450,7 +453,10 @@ void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u
 				SetUidField(tevindref.bi4, texmap);
 			}
 			if (texcoord < numTexgen)
+			{
+				out.SetConstantsUsed(C_INDTEXSCALE+i/2,C_INDTEXSCALE+i/2);
 				out.Write("tempcoord = uv%d.xy * " I_INDTEXSCALE"[%d].%s;\n", texcoord, i/2, (i&1)?"zw":"xy");
+			}
 			else
 				out.Write("tempcoord = float2(0.0f, 0.0f);\n");
 
@@ -504,12 +510,14 @@ void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u
 		WriteAlphaTest<T, type>(out, ApiType, dstAlphaMode);
 
 	// the screen space depth value = far z + (clip z / clip w) * z range
+	out.SetConstantsUsed(C_ZBIAS+1, C_ZBIAS+1);
 	out.Write("float zCoord = " I_ZBIAS"[1].x + (clipPos.z / clipPos.w) * " I_ZBIAS"[1].y;\n");
 
 	// Note: depth textures are disabled if early depth test is enabled
 	if (bpmem.ztex2.op != ZTEXTURE_DISABLE && !bpmem.zcontrol.early_ztest && bpmem.zmode.testenable)
 	{
 		// use the texture input of the last texture stage (textemp), hopefully this has been read and is in correct format...
+		out.SetConstantsUsed(C_ZBIAS, C_ZBIAS+1);
 		out.Write("zCoord = dot(" I_ZBIAS"[0].xyzw, textemp.xyzw) + " I_ZBIAS"[1].w %s;\n",
 									(bpmem.ztex2.op == ZTEXTURE_ADD) ? "+ zCoord" : "");
 
@@ -521,7 +529,10 @@ void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u
 	out.Write("depth = zCoord;\n");
 
 	if (dstAlphaMode == DSTALPHA_ALPHA_PASS)
+	{
+		out.SetConstantsUsed(C_ALPHA, C_ALPHA);
 		out.Write("  ocol0 = float4(prev.rgb, " I_ALPHA"[0].a);\n");
+	}
 	else
 	{
 		WriteFog<T, type>(out);
@@ -532,6 +543,7 @@ void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u
 	// single pass
 	if (dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
 	{
+		out.SetConstantsUsed(C_ALPHA, C_ALPHA);
 		// Colors will be blended against the alpha from ocol1...
 		out.Write("  ocol1 = ocol0;\n");
 		// ...and the alpha from ocol0 will be written to the framebuffer.
@@ -639,6 +651,7 @@ static void WriteStage(T& out, int n, API_TYPE ApiType)
 			if (bpmem.tevind[n].mid <= 3)
 			{
 				int mtxidx = 2*(bpmem.tevind[n].mid-1);
+				out.SetConstantsUsed(C_INDTEXMTX+mtxidx, C_INDTEXMTX+mtxidx);
 				out.Write("float2 indtevtrans%d = float2(dot(" I_INDTEXMTX"[%d].xyz, indtevcrd%d), dot(" I_INDTEXMTX"[%d].xyz, indtevcrd%d));\n",
 							n, mtxidx, n, mtxidx+1, n);
 			}
@@ -646,12 +659,14 @@ static void WriteStage(T& out, int n, API_TYPE ApiType)
 			{ // s matrix
 				_assert_(bpmem.tevind[n].mid >= 5);
 				int mtxidx = 2*(bpmem.tevind[n].mid-5);
+				out.SetConstantsUsed(C_INDTEXMTX+mtxidx, C_INDTEXMTX+mtxidx);
 				out.Write("float2 indtevtrans%d = " I_INDTEXMTX"[%d].ww * uv%d.xy * indtevcrd%d.xx;\n", n, mtxidx, texcoord, n);
 			}
 			else if (bpmem.tevind[n].mid <= 11 && bHasTexCoord)
 			{ // t matrix
 				_assert_(bpmem.tevind[n].mid >= 9);
 				int mtxidx = 2*(bpmem.tevind[n].mid-9);
+				out.SetConstantsUsed(C_INDTEXMTX+mtxidx, C_INDTEXMTX+mtxidx);
 				out.Write("float2 indtevtrans%d = " I_INDTEXMTX"[%d].ww * uv%d.xy * indtevcrd%d.yy;\n", n, mtxidx, texcoord, n);
 			}
 			else
@@ -757,6 +772,10 @@ static void WriteStage(T& out, int n, API_TYPE ApiType)
 		{
 			out.Write("ckonsttemp = konsttemp;\n");
 		}
+		if (kc > 7)
+			out.SetConstantsUsed(C_KCOLORS+((kc-0xc)%4),C_KCOLORS+((kc-0xc)%4));
+		if (ka > 7)
+			out.SetConstantsUsed(C_KCOLORS+((ka-0xc)%4),C_KCOLORS+((ka-0xc)%4));
 	}
 
 	if(cc.a == TEVCOLORARG_CPREV || cc.a == TEVCOLORARG_APREV
@@ -782,6 +801,8 @@ static void WriteStage(T& out, int n, API_TYPE ApiType)
 	|| cc.c == TEVCOLORARG_C0 || cc.c == TEVCOLORARG_A0
 	|| ac.a == TEVALPHAARG_A0 || ac.b == TEVALPHAARG_A0 || ac.c == TEVALPHAARG_A0)
 	{
+		// TODO: WTF?
+		out.SetConstantsUsed(C_COLORS+1,C_COLORS+1);
 		if(RegisterStates[1].AlphaNeedOverflowControl || RegisterStates[1].ColorNeedOverflowControl)
 		{
 			out.Write("cc0 = frac(c0 * (255.0f/256.0f)) * (256.0f/255.0f);\n");
@@ -800,6 +821,7 @@ static void WriteStage(T& out, int n, API_TYPE ApiType)
 	|| cc.c == TEVCOLORARG_C1 || cc.c == TEVCOLORARG_A1
 	|| ac.a == TEVALPHAARG_A1 || ac.b == TEVALPHAARG_A1 || ac.c == TEVALPHAARG_A1)
 	{
+		out.SetConstantsUsed(C_COLORS+2,C_COLORS+2);
 		if(RegisterStates[2].AlphaNeedOverflowControl || RegisterStates[2].ColorNeedOverflowControl)
 		{
 			out.Write("cc1 = frac(c1 * (255.0f/256.0f)) * (256.0f/255.0f);\n");
@@ -818,6 +840,7 @@ static void WriteStage(T& out, int n, API_TYPE ApiType)
 	|| cc.c == TEVCOLORARG_C2 || cc.c == TEVCOLORARG_A2
 	|| ac.a == TEVALPHAARG_A2 || ac.b == TEVALPHAARG_A2 || ac.c == TEVALPHAARG_A2)
 	{
+		out.SetConstantsUsed(C_COLORS+3,C_COLORS+3);
 		if(RegisterStates[3].AlphaNeedOverflowControl || RegisterStates[3].ColorNeedOverflowControl)
 		{
 			out.Write("cc2 = frac(c2 * (255.0f/256.0f)) * (256.0f/255.0f);\n");
@@ -834,6 +857,28 @@ static void WriteStage(T& out, int n, API_TYPE ApiType)
 	RegisterStates[cc.dest].ColorNeedOverflowControl = (cc.clamp == 0);
 	RegisterStates[cc.dest].AuxStored = false;
 
+/*	if (cc.d == TEVCOLORARG_C0 || cc.d == TEVCOLORARG_A0 || ac.d == TEVALPHAARG_A0)
+	{
+		out.SetConstantsUsed(C_COLORS+1,C_COLORS+1);
+		// TODO: 11 bit signed overflow..
+	}
+	if (cc.d == TEVCOLORARG_C1 || cc.d == TEVCOLORARG_A1 || ac.d == TEVALPHAARG_A1)
+	{
+		out.SetConstantsUsed(C_COLORS+2,C_COLORS+2);
+		// TODO: 11 bit signed overflow..
+	}
+	if (cc.d == TEVCOLORARG_C2 || cc.d == TEVCOLORARG_A2 || ac.d == TEVALPHAARG_A2)
+	{
+		out.SetConstantsUsed(C_COLORS+3,C_COLORS+3);
+		// TODO: 11 bit signed overflow..
+	}*/
+
+	// TODO: Are there enums for this?
+	if (cc.dest >= 1 && cc.dest <= 3)
+		out.SetConstantsUsed(C_COLORS+cc.dest, C_COLORS+cc.dest);
+	if (ac.dest >= 1 && ac.dest <= 3)
+		out.SetConstantsUsed(C_COLORS+ac.dest, C_COLORS+ac.dest);
+
 	out.Write("// color combine\n");
 	if (cc.clamp)
 		out.Write("%s = saturate(", tevCOutputTable[cc.dest]);
@@ -935,6 +980,7 @@ static void WriteStage(T& out, int n, API_TYPE ApiType)
 template<class T, GenOutput type>
 void SampleTexture(T& out, const char *destination, const char *texcoords, const char *texswap, int texmap, API_TYPE ApiType)
 {
+	out.SetConstantsUsed(C_TEXDIMS+texmap,C_TEXDIMS+texmap);
 	if (ApiType == API_D3D11)
 		out.Write("%s=Tex%d.Sample(samp%d,%s.xy * " I_TEXDIMS"[%d].xy).%s;\n", destination, texmap,texmap, texcoords, texmap, texswap);
 	else
@@ -970,6 +1016,8 @@ static void WriteAlphaTest(T& out, API_TYPE ApiType, DSTALPHA_MODE dstAlphaMode)
 		I_ALPHA"[0].g"
 	};
 
+	out.SetConstantsUsed(C_ALPHA, C_ALPHA);
+
 	// using discard then return works the same in cg and dx9 but not in dx11
 	out.Write("if(!( ");
 
@@ -1034,6 +1082,7 @@ static void WriteFog(T& out)
 
 	SetUidField(fog.proj, bpmem.fog.c_proj_fsel.proj);
 
+	out.SetConstantsUsed(C_FOG, C_FOG+1);
 	if (bpmem.fog.c_proj_fsel.proj == 0)
 	{
 		// perspective
@@ -1053,6 +1102,7 @@ static void WriteFog(T& out)
 	SetUidField(fog.RangeBaseEnabled, bpmem.fogRange.Base.Enabled);
 	if(bpmem.fogRange.Base.Enabled)
 	{
+		out.SetConstantsUsed(C_FOG+2, C_FOG+2);
 		out.Write("  float x_adjust = (2.0f * (clipPos.x / " I_FOG"[2].y)) - 1.0f - " I_FOG"[2].x;\n");
 		out.Write("  x_adjust = sqrt(x_adjust * x_adjust + " I_FOG"[2].z * " I_FOG"[2].z) / " I_FOG"[2].z;\n");
 		out.Write("  ze *= x_adjust;\n");
@@ -1082,3 +1132,9 @@ void GeneratePixelShaderCode(PixelShaderCode& object, DSTALPHA_MODE dstAlphaMode
 {
 	GeneratePixelShader<PixelShaderCode, GO_ShaderCode>(object, dstAlphaMode, ApiType, components);
 }
+
+void GetPixelShaderConstantProfile(PixelShaderConstantProfile& object, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u32 components)
+{
+	GeneratePixelShader<PixelShaderConstantProfile, GO_ShaderCode>(object, dstAlphaMode, ApiType, components);
+}
+

Source/Core/VideoCommon/Src/PixelShaderGen.h

@@ -72,7 +72,6 @@ struct pixel_shader_uid_data
         u32 numtevstages : 4;
 		u32 numindstages : 3;
 	} genMode;
-	u32 fogc_proj_fselfsel : 3;
 	struct
 	{
 		u32 unknown : 1;
@@ -177,11 +176,11 @@ struct pixel_shader_uid_data
 
 typedef ShaderUid<pixel_shader_uid_data> PixelShaderUid;
 typedef ShaderCode<pixel_shader_uid_data> PixelShaderCode;
-//typedef ShaderConstantProfile<pixel_shader_uid_data> PixelShaderConstantProfile;
+typedef ShaderConstantProfile<pixel_shader_uid_data> PixelShaderConstantProfile;
 
 
 void GeneratePixelShaderCode(PixelShaderCode& object, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u32 components);
 void GetPixelShaderUid(PixelShaderUid& object, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u32 components);
-//void GetPixelShaderConstantProfile(PixelShaderConstantProfile& object, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u32 components);
+void GetPixelShaderConstantProfile(PixelShaderConstantProfile& object, DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType, u32 components);
 
 #endif // GCOGL_PIXELSHADER_H

Source/Core/VideoCommon/Src/PixelShaderManager.cpp

@@ -83,39 +83,58 @@ void PixelShaderManager::Shutdown()
 
 }
 
-void PixelShaderManager::SetConstants()
+void PixelShaderManager::SetConstants(u32 components)
 {
-    for (int i = 0; i < 2; ++i)
+	PixelShaderConstantProfile constant_profile(C_PENVCONST_END);
+	/// TODO: dst alpha/api/components type parameter...
+	GetPixelShaderConstantProfile(constant_profile, DSTALPHA_DUAL_SOURCE_BLEND, API_OPENGL, components);
+
+	static int saved_updates = 0;
+	static int necessary_updates = 0;
+
+
+#define IncStuff() { \
+	saved_updates++; \
+	printf("Saved a constant update at line %d! Saved %d against %d now!\n", __LINE__, saved_updates, necessary_updates); }
+
+	for (int i = 0; i < 2; ++i)
 	{
-        if (s_nColorsChanged[i])
+		if (s_nColorsChanged[i])
 		{
-            int baseind = i ? C_KCOLORS : C_COLORS;
+			int baseind = i ? C_KCOLORS : C_COLORS;
-            for (int j = 0; j < 4; ++j)
+			for (int j = 0; j < 4; ++j)
 			{
-                if (s_nColorsChanged[i] & (1 << j))
+				if ((s_nColorsChanged[i] & (1 << j)) && constant_profile.ConstantIsUsed(baseind+j))
-                    SetPSConstant4fv(baseind+j, &lastRGBAfull[i][j][0]);
+				{
-            }
+					SetPSConstant4fv(baseind+j, &lastRGBAfull[i][j][0]);
-            s_nColorsChanged[i] = 0;
+					s_nColorsChanged[i] &= ~(1<<j);
-        }
+					++necessary_updates;
-    }
+				} else if ((s_nColorsChanged[i] & (1 << j))) IncStuff();
+			}
+		}
+	}
 
     if (s_nTexDimsChanged)
 	{
         for (int i = 0; i < 8; ++i)
 		{
-            if (s_nTexDimsChanged & (1<<i))
+            if (s_nTexDimsChanged & (1<<i) && constant_profile.ConstantIsUsed(C_TEXDIMS+i))
+			{
+					++necessary_updates;
 				SetPSTextureDims(i);
+				s_nTexDimsChanged &= ~(1<<i);
+			}else if (s_nTexDimsChanged & (1<<i)) IncStuff();
         }
-        s_nTexDimsChanged = 0;
     }
 
-    if (s_bAlphaChanged)
+    if (s_bAlphaChanged && constant_profile.ConstantIsUsed(C_ALPHA))
 	{
+					++necessary_updates;
 		SetPSConstant4f(C_ALPHA, (lastAlpha&0xff)/255.0f, ((lastAlpha>>8)&0xff)/255.0f, 0, ((lastAlpha>>16)&0xff)/255.0f);
 		s_bAlphaChanged = false;
-    }
+    } else if (s_bAlphaChanged) IncStuff();
 
-	if (s_bZTextureTypeChanged)
+	if (s_bZTextureTypeChanged && constant_profile.ConstantIsUsed(C_ZBIAS))
 	{
         float ftemp[4];
         switch (bpmem.ztex2.type)
@@ -133,11 +152,12 @@ void PixelShaderManager::SetConstants()
 				ftemp[0] = 16711680.0f/16777215.0f; ftemp[1] = 65280.0f/16777215.0f; ftemp[2] = 255.0f/16777215.0f; ftemp[3] = 0;
                 break;
         }
+					++necessary_updates;
 		SetPSConstant4fv(C_ZBIAS, ftemp);
 		s_bZTextureTypeChanged = false;
-	}
+	} else if (s_bZTextureTypeChanged) IncStuff();
 
-	if (s_bZBiasChanged || s_bDepthRangeChanged)
+	if ((s_bZBiasChanged || s_bDepthRangeChanged) && constant_profile.ConstantIsUsed(C_ZBIAS+1))
 	{
         // reversed gxsetviewport(xorig, yorig, width, height, nearz, farz)
 		// [0] = width/2
@@ -148,9 +168,10 @@ void PixelShaderManager::SetConstants()
 		// [5] = 16777215 * farz
 
 		//ERROR_LOG("pixel=%x,%x, bias=%x\n", bpmem.zcontrol.pixel_format, bpmem.ztex2.type, lastZBias);
+					++necessary_updates;
 		SetPSConstant4f(C_ZBIAS+1, xfregs.viewport.farZ / 16777216.0f, xfregs.viewport.zRange / 16777216.0f, 0, (float)(lastZBias)/16777215.0f);
 		s_bZBiasChanged = s_bDepthRangeChanged = false;
-    }
+    }else if ((s_bZBiasChanged || s_bDepthRangeChanged)) IncStuff();
 
     // indirect incoming texture scales
     if (s_nIndTexScaleChanged)
@@ -158,7 +179,7 @@ void PixelShaderManager::SetConstants()
 		// set as two sets of vec4s, each containing S and T of two ind stages.
         float f[8];
 
-        if (s_nIndTexScaleChanged & 0x03)
+        if ((s_nIndTexScaleChanged & 0x03) && constant_profile.ConstantIsUsed(C_INDTEXSCALE))
 		{
             for (u32 i = 0; i < 2; ++i)
 			{
@@ -166,26 +187,30 @@ void PixelShaderManager::SetConstants()
                 f[2 * i + 1] = bpmem.texscale[0].getScaleT(i & 1);
                 PRIM_LOG("tex indscale%d: %f %f\n", i, f[2 * i], f[2 * i + 1]);
             }
+					++necessary_updates;
 			SetPSConstant4fv(C_INDTEXSCALE, f);
+			s_nIndTexScaleChanged &= ~0x03;
         }
+        else if ((s_nIndTexScaleChanged & 0x03)) IncStuff();
 
-        if (s_nIndTexScaleChanged & 0x0c) {
+        if ((s_nIndTexScaleChanged & 0x0c) && constant_profile.ConstantIsUsed(C_INDTEXSCALE+1)) {
             for (u32 i = 2; i < 4; ++i) {
                 f[2 * i] = bpmem.texscale[1].getScaleS(i & 1);
                 f[2 * i + 1] = bpmem.texscale[1].getScaleT(i & 1);
                 PRIM_LOG("tex indscale%d: %f %f\n", i, f[2 * i], f[2 * i + 1]);
             }
+					++necessary_updates;
 			SetPSConstant4fv(C_INDTEXSCALE+1, &f[4]);
+			s_nIndTexScaleChanged &= ~0x0c;
         }
-
+        else if ((s_nIndTexScaleChanged & 0x0c)) IncStuff();
-        s_nIndTexScaleChanged = 0;
     }
 
     if (s_nIndTexMtxChanged)
 	{
         for (int i = 0; i < 3; ++i)
 		{
-            if (s_nIndTexMtxChanged & (1 << i))
+            if ((s_nIndTexMtxChanged & (1 << i)) && (constant_profile.ConstantIsUsed(C_INDTEXMTX+2*i) || constant_profile.ConstantIsUsed(C_INDTEXMTX+2*i+1)))
 			{
                 int scale = ((u32)bpmem.indmtx[i].col0.s0 << 0) |
 					        ((u32)bpmem.indmtx[i].col1.s1 << 2) |
@@ -195,6 +220,8 @@ void PixelShaderManager::SetConstants()
                 // xyz - static matrix
                 // TODO w - dynamic matrix scale / 256...... somehow / 4 works better
                 // rev 2972 - now using / 256.... verify that this works
+					++necessary_updates;
+					++necessary_updates;
 				SetPSConstant4f(C_INDTEXMTX + 2 * i,
                     bpmem.indmtx[i].col0.ma * fscale,
 					bpmem.indmtx[i].col1.mc * fscale,
@@ -210,19 +237,22 @@ void PixelShaderManager::SetConstants()
                 	i, 1024.0f*fscale,
                 	bpmem.indmtx[i].col0.ma * fscale, bpmem.indmtx[i].col1.mc * fscale, bpmem.indmtx[i].col2.me * fscale,
                 	bpmem.indmtx[i].col0.mb * fscale, bpmem.indmtx[i].col1.md * fscale, bpmem.indmtx[i].col2.mf * fscale);
-            }
+
+				s_nIndTexMtxChanged &= ~(1 << i);
+            }else if ((s_nIndTexMtxChanged & (1 << i))) {IncStuff();IncStuff();}
         }
-        s_nIndTexMtxChanged = 0;
     }
 
-    if (s_bFogColorChanged)
+    if (s_bFogColorChanged && constant_profile.ConstantIsUsed(C_FOG))
 	{
+					++necessary_updates;
 		SetPSConstant4f(C_FOG, bpmem.fog.color.r / 255.0f, bpmem.fog.color.g / 255.0f, bpmem.fog.color.b / 255.0f, 0);
 		s_bFogColorChanged = false;
-    }
+    }else if (s_bFogColorChanged) IncStuff();
 
-    if (s_bFogParamChanged)
+    if (s_bFogParamChanged && constant_profile.ConstantIsUsed(C_FOG+1))
 	{
+					++necessary_updates;
 		if(!g_ActiveConfig.bDisableFog)
 		{
 			//downscale magnitude to 0.24 bits
@@ -235,10 +265,11 @@ void PixelShaderManager::SetConstants()
 			SetPSConstant4f(C_FOG + 1, 0.0, 1.0, 0.0, 1.0);
 
         s_bFogParamChanged = false;
-    }
+    }else if ( s_bFogParamChanged) IncStuff();
 
-	if (s_bFogRangeAdjustChanged)
+	if (s_bFogRangeAdjustChanged && constant_profile.ConstantIsUsed(C_FOG+2))
 	{
+					++necessary_updates;
 		if(!g_ActiveConfig.bDisableFog && bpmem.fogRange.Base.Enabled == 1)
 		{
 			//bpmem.fogRange.Base.Center : center of the viewport in x axis. observation: bpmem.fogRange.Base.Center = realcenter + 342;
@@ -257,8 +288,9 @@ void PixelShaderManager::SetConstants()
 			SetPSConstant4f(C_FOG + 2, 0.0f, 1.0f, 1.0f, 0.0f); // Need to update these values for older hardware that fails to divide by zero in shaders.
 
 		s_bFogRangeAdjustChanged = false;
-	}
+	}else if ( s_bFogRangeAdjustChanged) IncStuff();
 
+	// TODO: use constant profile here!
 	if (g_ActiveConfig.bEnablePixelLighting && g_ActiveConfig.backend_info.bSupportsPixelLighting)  // config check added because the code in here was crashing for me inside SetPSConstant4f
 	{
 		if (nLightsChanged[0] >= 0)
@@ -353,7 +385,8 @@ void PixelShaderManager::SetPSTextureDims(int texid)
 	SetPSConstant4fv(C_TEXDIMS + texid, fdims);
 }
 
-// This one is high in profiles (0.5%). TODO: Move conversion out, only store the raw color value
+// This one is high in profiles (0.5%).
+// TODO: Move conversion out, only store the raw color value
 // and update it when the shader constant is set, only.
 void PixelShaderManager::SetColorChanged(int type, int num, bool high)
 {

Source/Core/VideoCommon/Src/PixelShaderManager.h

@@ -34,7 +34,7 @@ public:
 	static void Shutdown();
 	static void DoState(PointerWrap &p);
 
-	static void SetConstants(); // sets pixel shader constants
+	static void SetConstants(u32 components); // sets pixel shader constants
 
 	// constant management, should be called after memory is committed
 	static void SetColorChanged(int type, int index, bool high);

Source/Core/VideoCommon/Src/ShaderGenCommon.h

@@ -23,6 +23,8 @@
 #include <string.h>
 #include "CommonTypes.h"
 
+#include <vector>
+
 template<class uid_data>
 class ShaderUid
 {
@@ -36,6 +38,7 @@ public:
 	void Write(const char* fmt, ...) {}
 	const char* GetBuffer() { return NULL; }
 	void SetBuffer(char* buffer) { }
+	inline void SetConstantsUsed(unsigned int first_index, unsigned int last_index) {}
 
 	bool operator == (const ShaderUid& obj) const
 	{
@@ -55,7 +58,7 @@ public:
 		return false;
 	}
 
-	uid_data& GetUidData() { return data; }
+	inline uid_data& GetUidData() { return data; }
 
 private:
 	union
@@ -86,16 +89,46 @@ public:
 	const char* GetBuffer() { return buf; }
 	void SetBuffer(char* buffer) { buf = buffer; write_ptr = buffer; }
 	uid_data& GetUidData() { return *(uid_data*)NULL; }
+	inline void SetConstantsUsed(unsigned int first_index, unsigned int last_index) {}
 
 private:
 	const char* buf;
 	char* write_ptr;
 };
 
+template<class uid_data>
+class ShaderConstantProfile
+{
+public:
+	ShaderConstantProfile(int num_constants) { constant_usage.resize(num_constants); }
+
+	void Write(const char* fmt, ...) {}
+	const char* GetBuffer() { return NULL; }
+	void SetBuffer(char* buffer) { }
+	uid_data& GetUidData() { return *(uid_data*)NULL; }
+
+	// has room for optimization (if it matters at all...)
+	void NumConstants() { return constant_usage.size(); }
+
+	inline void SetConstantsUsed(unsigned int first_index, unsigned int last_index)
+	{
+		for (unsigned int i = first_index; i < last_index+1; ++i)
+			constant_usage[i] = true;
+	}
+
+	inline bool ConstantIsUsed(unsigned int index)
+	{
+		return constant_usage[index];
+	}
+private:
+	std::vector<bool> constant_usage; // TODO: Is vector<bool> appropriate here?
+};
+
 enum GenOutput
 {
 	GO_ShaderCode,
 	GO_ShaderUid,
+	GO_ShaderConstantProfile,
 };
 
 #endif // _SHADERGENCOMMON_H

Source/Plugins/Plugin_VideoOGL/Src/VertexManager.cpp

@@ -180,7 +180,7 @@ void VertexManager::vFlush()
 
 	// set global constants
 	VertexShaderManager::SetConstants();
-	PixelShaderManager::SetConstants();
+	PixelShaderManager::SetConstants(g_nativeVertexFmt->m_components);
 
 	bool useDstAlpha = !g_ActiveConfig.bDstAlphaPass && bpmem.dstalpha.enable && bpmem.blendmode.alphaupdate
 		&& bpmem.zcontrol.pixel_format == PIXELFMT_RGBA6_Z24;