gl optimization and im2d shader override

src/gl/gl3device.cpp

@@ -98,6 +98,12 @@ struct UniformObject
 	RGBAf lightColor[MAX_LIGHTS];
 };
 
+struct GLShaderState
+{
+	RGBA matColor;
+	SurfaceProperties surfProps;
+};
+
 const char *shaderDecl330 = "#version 330\n";
 const char *shaderDecl100es =
 "#version 100\n"\
@@ -129,6 +135,7 @@ static GLuint whitetex;
 static UniformState uniformState;
 static UniformScene uniformScene;
 static UniformObject uniformObject;
+static GLShaderState shaderState;
 
 #ifndef RW_GL_USE_UBOS
 // State
@@ -878,13 +885,39 @@ setViewMatrix(float32 *mat)
 
 Shader *lastShaderUploaded;
 
+#define U(i) currentShader->uniformLocations[i]
+
+void
+setMaterial(const RGBA &color, const SurfaceProperties &surfaceprops)
+{
+	bool force = lastShaderUploaded != currentShader;
+	if(force || !equal(shaderState.matColor, color)){
+		rw::RGBAf col;
+		convColor(&col, &color);
+		glUniform4fv(U(u_matColor), 1, (GLfloat*)&col);
+		shaderState.matColor = color;
+	}
+
+	if(force ||
+	   shaderState.surfProps.ambient != surfaceprops.ambient ||
+	   shaderState.surfProps.specular != surfaceprops.specular ||
+	   shaderState.surfProps.diffuse != surfaceprops.diffuse){
+		float surfProps[4];
+		surfProps[0] = surfaceprops.ambient;
+		surfProps[1] = surfaceprops.specular;
+		surfProps[2] = surfaceprops.diffuse;
+		surfProps[3] = 0.0f;
+		glUniform4fv(U(u_surfProps), 1, surfProps);
+		shaderState.surfProps = surfaceprops;
+	}
+}
+
 void
 flushCache(void)
 {
 	flushGlRenderState();
 
 #ifndef RW_GL_USE_UBOS
-#define U(i) currentShader->uniformLocations[i]
 
 	// TODO: this is probably a stupid way to do it without UBOs
 	if(lastShaderUploaded != currentShader){

src/gl/gl3immed.cpp

@@ -23,6 +23,9 @@ uint32 im2DVbo, im2DIbo;
 #ifdef RW_GL_USE_VAOS
 uint32 im2DVao;
 #endif
+
+Shader *im2dOverrideShader;
+
 static int32 u_xform;
 
 #define STARTINDICES 10000
@@ -130,6 +133,9 @@ im2DRenderPrimitive(PrimitiveType primType, void *vertices, int32 numVertices)
 	xform[2] = -1.0f;
 	xform[3] = 1.0f;
 
+	if(im2dOverrideShader)
+		im2dOverrideShader->use();
+	else
 		im2dShader->use();
 #ifndef RW_GL_USE_VAOS
 	setAttribPointers(im2dattribDesc, 3);
@@ -170,6 +176,9 @@ im2DRenderIndexedPrimitive(PrimitiveType primType,
 	xform[2] = -1.0f;
 	xform[3] = 1.0f;
 
+	if(im2dOverrideShader)
+		im2dOverrideShader->use();
+	else
 		im2dShader->use();
 #ifndef RW_GL_USE_VAOS
 	setAttribPointers(im2dattribDesc, 3);

src/gl/gl3matfx.cpp

@@ -72,20 +72,11 @@ void
 matfxDefaultRender(InstanceDataHeader *header, InstanceData *inst)
 {
 	Material *m;
-	RGBAf col;
-	GLfloat surfProps[4];
 	m = inst->material;
 
 	defaultShader->use();
 
-	convColor(&col, &m->color);
+	setMaterial(m->color, m->surfaceProps);
-	glUniform4fv(U(u_matColor), 1, (GLfloat*)&col);
-
-	surfProps[0] = m->surfaceProps.ambient;
-	surfProps[1] = m->surfaceProps.specular;
-	surfProps[2] = m->surfaceProps.diffuse;
-	surfProps[3] = 0.0f;
-	glUniform4fv(U(u_surfProps), 1, surfProps);
 
 	setTexture(0, m->texture);
 
@@ -128,8 +119,6 @@ void
 matfxEnvRender(InstanceDataHeader *header, InstanceData *inst, MatFX::Env *env)
 {
 	Material *m;
-	RGBAf col;
-	GLfloat surfProps[4];
 	m = inst->material;
 
 	if(env->tex == nil || env->coefficient == 0.0f){
@@ -143,14 +132,7 @@ matfxEnvRender(InstanceDataHeader *header, InstanceData *inst, MatFX::Env *env)
 	setTexture(1, env->tex);
 	uploadEnvMatrix(env->frame);
 
-	convColor(&col, &m->color);
+	setMaterial(m->color, m->surfaceProps);
-	glUniform4fv(U(u_matColor), 1, (GLfloat*)&col);
-
-	surfProps[0] = m->surfaceProps.ambient;
-	surfProps[1] = m->surfaceProps.specular;
-	surfProps[2] = m->surfaceProps.diffuse;
-	surfProps[3] = 0.0f;
-	glUniform4fv(U(u_surfProps), 1, surfProps);
 
 	float fxparams[2];
 	fxparams[0] = env->coefficient;

src/gl/gl3render.cpp

@@ -144,14 +144,7 @@ defaultRenderCB(Atomic *atomic, InstanceDataHeader *header)
 	while(n--){
 		m = inst->material;
 
-		convColor(&col, &m->color);
+		setMaterial(m->color, m->surfaceProps);
-		glUniform4fv(U(u_matColor), 1, (GLfloat*)&col);
-
-		surfProps[0] = m->surfaceProps.ambient;
-		surfProps[1] = m->surfaceProps.specular;
-		surfProps[2] = m->surfaceProps.diffuse;
-		surfProps[3] = 0.0f;
-		glUniform4fv(U(u_surfProps), 1, surfProps);
 
 		setTexture(0, m->texture);
 

src/gl/gl3shader.cpp

@@ -85,12 +85,13 @@ printShaderSource(const char **src)
 	bool printline;
 	int line = 1;
 	for(f = 0; src[f]; f++){
+		int fileline = 1;
 		char c;
 		file = src[f];
 		printline = true;
 		while(c = *file++, c != '\0'){
 			if(printline)
-				printf("%.4d: ", line++);
+				printf("%.4d/%d:%.4d: ", line++, f, fileline++);
 			putchar(c);
 			printline = c == '\n';
 		}
@@ -243,6 +244,10 @@ Shader::create(const char **vsrc, const char **fsrc)
 		glUniform1i(loc, i);
 	}
 
+	// reset program
+	if(currentShader)
+		glUseProgram(currentShader->program);
+
 	return sh;
 }
 

src/gl/gl3skin.cpp

@@ -282,8 +282,6 @@ void
 skinRenderCB(Atomic *atomic, InstanceDataHeader *header)
 {
 	Material *m;
-	RGBAf col;
-	GLfloat surfProps[4];
 
 	setWorldMatrix(atomic->getFrame()->getLTM());
 	lightingCB(atomic);
@@ -306,14 +304,7 @@ skinRenderCB(Atomic *atomic, InstanceDataHeader *header)
 	while(n--){
 		m = inst->material;
 
-		convColor(&col, &m->color);
+		setMaterial(m->color, m->surfaceProps);
-		glUniform4fv(U(u_matColor), 1, (GLfloat*)&col);
-
-		surfProps[0] = m->surfaceProps.ambient;
-		surfProps[1] = m->surfaceProps.specular;
-		surfProps[2] = m->surfaceProps.diffuse;
-		surfProps[3] = 0.0f;
-		glUniform4fv(U(u_surfProps), 1, surfProps);
 
 		setTexture(0, m->texture);
 

src/gl/rwgl3.h

@@ -170,6 +170,8 @@ extern const char *shaderDecl;	// #version stuff
 extern const char *header_vert_src;
 extern const char *header_frag_src;
 
+extern Shader *im2dOverrideShader;
+
 // per Scene
 void setProjectionMatrix(float32*);
 void setViewMatrix(float32*);
@@ -180,6 +182,7 @@ int32 setLights(WorldLights *lightData);
 
 // per Mesh
 void setTexture(int32 n, Texture *tex);
+void setMaterial(const RGBA &color, const SurfaceProperties &surfaceprops);
 
 void setAlphaBlend(bool32 enable);
 bool32 getAlphaBlend(void);

tools/playground/main.cpp

@@ -175,6 +175,65 @@ setupClump(rw::Clump *clump)
 	}
 }
 
+#define MUL(x, y) ((x)*(y)/255)
+
+int
+calcVCfx(int fb, int col, int a, int iter)
+{
+	int prev = fb;
+	int col2 = col*2;
+	if(col2 > 255) col2 = 255;
+	for(int i = 0; i < iter; i++){
+		int tmp = MUL(fb, 255-a) + MUL(MUL(prev, col2), a);
+		tmp += MUL(prev, col);
+		tmp += MUL(prev, col);
+		prev = tmp > 255 ? 255 : tmp;
+	}
+	return prev;
+}
+
+int
+calcIIIfx(int fb, int col, int a, int iter)
+{
+	int prev = fb;
+	for(int i = 0; i < iter; i++){
+		int tmp = MUL(fb, 255-a) + MUL(MUL(prev, col), a);
+		prev = tmp > 255 ? 255 : tmp;
+	}
+	return prev;
+}
+
+void
+postfxtest(void)
+{
+	rw::Image *img = rw::Image::create(256, 256, 32);
+	img->allocate();
+	int x, y;
+	int iter;
+	static char filename[100];
+	for(iter = 0; iter < 10; iter++){
+		for(y = 0; y < 256; y++)
+			for(x = 0; x < 256; x++){
+				int res = calcVCfx(y, x, 30, iter);
+//				int res = calcIIIfx(y, x, 30, iter);
+				if(0 && res == y){
+					img->pixels[y*img->stride + x*img->bpp + 0] = 255;
+					img->pixels[y*img->stride + x*img->bpp + 1] = 0;
+					img->pixels[y*img->stride + x*img->bpp + 2] = 0;
+				}else{
+					img->pixels[y*img->stride + x*img->bpp + 0] = res;
+					img->pixels[y*img->stride + x*img->bpp + 1] = res;
+					img->pixels[y*img->stride + x*img->bpp + 2] = res;
+				}
+				img->pixels[y*img->stride + x*img->bpp + 3] = 255;
+			}
+		sprintf(filename, "vcfx_%02d.bmp", iter);
+//		sprintf(filename, "iiifx_%02d.bmp", iter);
+		rw::writeBMP(img, filename);
+	}
+	exit(0);
+}
+
 bool
 InitRW(void)
 {
@@ -184,6 +243,8 @@ InitRW(void)
 
 	rw::d3d::isP8supported = false;
 
+	postfxtest();
+
 	initFont();
 
 	rw::RGBA foreground = { 255, 255, 0, 255 };