Optimized vertex and pixel shader generation. Medium FPS boost.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@500 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Plugins/Plugin_VideoDX9/Src/PixelShader.cpp

@@ -302,35 +302,33 @@ const char *GeneratePixelShader()
 	int numSamplers = 8;
 
 	char *p = text;
-	WRITE(p,"//Pixel Shader for TEV stages\n");
+	WRITE(p,"//Pixel Shader for TEV stages\n\
-	WRITE(p,"//%i TEV stages, %i texgens, %i IND stages, %i COL channels\n",
+//%i TEV stages, %i texgens, %i IND stages, %i COL channels\n",
 		bpmem.genMode.numtevstages,bpmem.genMode.numtexgens,bpmem.genMode.numindstages,bpmem.genMode.numcolchans);
 
 	//write kcolor declarations
-	for (int i = 0; i < 4; i++)
+	for (int i = 0; i < 4; i++) {
-		WRITE(p,"float4 k%i : register(c%i);\n",i,PS_CONST_KCOLORS+i);
+		if(i < 3) {
+			WRITE(p,"float4 k%i : register(c%i);\n\
+                     float4 color%i : register(c%i);\n",i,PS_CONST_KCOLORS+i, i,PS_CONST_COLORS+i+1);
+		} else {
+			WRITE(p,"float4 k%i : register(c%i);\n",i,PS_CONST_KCOLORS+i);
+		}
+	}
 
-	for (int i = 0; i < 3; i++)
+	WRITE(p,"float constalpha : register(c%i);\n\
-		WRITE(p,"float4 color%i : register(c%i);\n",i,PS_CONST_COLORS+i+1);
+float2 alphaRef : register(c%i);\n\n\
+sampler samp[%i] : register(s0);\n\n\
+float4 main(in float4 colors[2] : COLOR0",PS_CONST_CONSTALPHA,PS_CONST_ALPHAREF,numSamplers);
 
-	WRITE(p,"float constalpha : register(c%i);\n",PS_CONST_CONSTALPHA);
-	WRITE(p,"float2 alphaRef : register(c%i);\n",PS_CONST_ALPHAREF);
-
-	WRITE(p,"\n");
-
-	WRITE(p,"sampler samp[%i] : register(s0);\n",numSamplers);
-
-	WRITE(p,"\n");
-
-	WRITE(p,"float4 main(in float4 colors[2] : COLOR0");
 	if (numTexgen)
 		WRITE(p,", float4 uv[%i] : TEXCOORD0",numTexgen);
 	else
 		WRITE(p,", float4 uv[1] : TEXCOORD0"); //HACK
-	WRITE(p,") : COLOR\n");
+	WRITE(p,") : COLOR\n\
-	WRITE(p,"{\n");
+{\n\
-	WRITE(p,"float4 c0=color0,c1=color1,c2=color2,prev=float4(0.0f,0.0f,0.0f,0.0f),textemp,rastemp,konsttemp;\n");
+float4 c0=color0,c1=color1,c2=color2,prev=float4(0.0f,0.0f,0.0f,0.0f),textemp,rastemp,konsttemp;\n\
-	WRITE(p,"\n");
+\n");
 
 	for (int i = 0; i < numStages; i++)
 		WriteStage(p,i); //build the equation for this stage
@@ -342,8 +340,7 @@ const char *GeneratePixelShader()
 	else
 		WRITE(p,"  return prev;\n");
 
-	WRITE(p,"}\n");
+	WRITE(p,"}\n\0");
-	WRITE(p,"\0");
 	
 	return text;
 }

Source/Plugins/Plugin_VideoDX9/Src/ShaderManager.cpp

@@ -152,9 +152,15 @@ void VShaderCache::Shutdown()
 
 void VShaderCache::SetShader()
 {
+	static LPDIRECT3DVERTEXSHADER9 shader = NULL;
 	if (D3D::GetShaderVersion() < 2)
 		return; // we are screwed
 
+	if(shader) {
+		//D3D::dev->SetVertexShader(shader);
+		return;
+	}
+	
 	static LPDIRECT3DVERTEXSHADER9 lastShader = 0;
 	DVSTARTPROFILE();
 
@@ -176,7 +182,7 @@ void VShaderCache::SetShader()
 	}
 
 	const char *code = GenerateVertexShader();
-	LPDIRECT3DVERTEXSHADER9 shader = D3D::CompileVShader(code, int(strlen(code)));
+	shader = D3D::CompileVShader(code, int(strlen(code)));
 	if (shader)
 	{
 		//Make an entry in the table
@@ -185,6 +191,7 @@ void VShaderCache::SetShader()
 		entry.frameCount=frameCount;
 		vshaders[currentHash] = entry;
 	}
+
 	D3D::dev->SetVertexShader(shader);
 
 	INCSTAT(stats.numVertexShadersCreated);

Source/Plugins/Plugin_VideoDX9/Src/VertexShader.cpp

@@ -20,6 +20,46 @@
 #include "VertexShader.h"
 #include "BPStructs.h"
 
+static const char *genericVS = "// Generic Vertex Shader\
+\n\
+struct VS_INPUT {\n\
+float4 pos : POSITION;\n\
+float3 normal : NORMAL;\n\
+float4 colors[2] : COLOR0;\n\
+float3 uv[8] : TEXCOORD0;\n\
+};\n\
+\n\
+struct VS_OUTPUT {\n\
+float4 pos : POSITION;\n\
+float4 colors[2] : COLOR0;\n\
+//numtexgen\n\
+float4 uv[5] : TEXCOORD0;\n\
+};\n\
+\n\
+uniform matrix matWorldViewProj : register(c0);\n\
+\n\
+VS_OUTPUT main(const VS_INPUT input)\n\
+{\n\
+VS_OUTPUT output;\n\
+\n\
+output.pos = mul(matWorldViewProj, input.pos);\n\
+// texgen\n\
+output.uv[0] = float4(input.uv[0].xy,0,input.uv[0].z);\n\
+output.uv[1] = float4(input.uv[1].xy,0,input.uv[1].z);\n\
+output.uv[2] = float4(input.uv[2].xy,0,input.uv[2].z);\n\
+output.uv[3] = float4(input.uv[3].xy,0,input.uv[3].z);\n\
+output.uv[4] = float4(input.uv[4].xy,0,input.uv[4].z);\n\
+\n\
+for (int i=0; i<2; i++)\n    output.colors[i] = input.colors[i];\n\
+return output;\n\
+}\0";
+
+const char *GenerateVertexShader() {
+	return genericVS;
+}
+
+/*
+
 char text2[65536];
 #define WRITE p+=sprintf
 
@@ -91,7 +131,7 @@ const char *GenerateVertexShader()
 /*
  *	xform->vertexshader ideas
 
-*/
+*//*
 
 void WriteTexgen(char *&p, int n)
 {
@@ -102,4 +142,4 @@ void WriteTexgen(char *&p, int n)
 void WriteLight(int color, int component)
 {
 
-}
+} */