Overbright bits hack. FP16 FBO's with r_ext_framebuffer_object 2. Occlusion query for god rays.

reaction/code/renderer/qglextensions.h

@@ -14,17 +14,21 @@
 	HANDLE_EXT_FUNC(	PFNGLRENDERBUFFERSTORAGEEXTPROC,				glRenderbufferStorageEXT)		\
 	HANDLE_EXT_FUNC(	PFNGLBINDRENDERBUFFEREXTPROC,					glBindRenderbufferEXT)			\
 
-#define GL_ARB_occlusion_query_functions																\
-	HANDLE_EXT_FUNC(	PFNGLGENQUERIESARBPROC,							glGenQueriesARB)				\
-	HANDLE_EXT_FUNC(	PFNGLDELETEQUERIESARBPROC,						glDeleteQueriesARB)				\
-	HANDLE_EXT_FUNC(	PFNGLISQUERYARBPROC,							glIsQueryARB)					\
-	HANDLE_EXT_FUNC(	PFNGLBEGINQUERYARBPROC,							glBeginQueryARB)				\
-	HANDLE_EXT_FUNC(	PFNGLENDQUERYARBPROC,							glEndQueryARB)					\
-	HANDLE_EXT_FUNC(	PFNGLGETQUERYIVARBPROC,							glGetQueryivARB)				\
-	HANDLE_EXT_FUNC(	PFNGLGETQUERYOBJECTIVARBPROC,					glGetQueryObjectivARB)			\
-	HANDLE_EXT_FUNC(	PFNGLGETQUERYOBJECTUIVARBPROC,					glGetQueryObjectuivARB)			\
+#define GL_ARB_occlusion_query_functions																\
+	HANDLE_EXT_FUNC(	PFNGLGENQUERIESARBPROC,							glGenQueriesARB)				\
+	HANDLE_EXT_FUNC(	PFNGLDELETEQUERIESARBPROC,						glDeleteQueriesARB)				\
+	HANDLE_EXT_FUNC(	PFNGLISQUERYARBPROC,							glIsQueryARB)					\
+	HANDLE_EXT_FUNC(	PFNGLBEGINQUERYARBPROC,							glBeginQueryARB)				\
+	HANDLE_EXT_FUNC(	PFNGLENDQUERYARBPROC,							glEndQueryARB)					\
+	HANDLE_EXT_FUNC(	PFNGLGETQUERYIVARBPROC,							glGetQueryivARB)				\
+	HANDLE_EXT_FUNC(	PFNGLGETQUERYOBJECTIVARBPROC,					glGetQueryObjectivARB)			\
+	HANDLE_EXT_FUNC(	PFNGLGETQUERYOBJECTUIVARBPROC,					glGetQueryObjectuivARB)			\
 
+#define GL_MISSING_FUNCTIONS																			\
+	HANDLE_EXT_FUNC(	PFNGLBLENDEQUATIONPROC,							glBlendEquation)				\
 
+#define GL_ARB_color_buffer_float_functions																\
+	HANDLE_EXT_FUNC(	PFNGLCLAMPCOLORARBPROC,							glClampColorARB)				\
 
 #ifndef GL_EXT_framebuffer_multisample
 	#define GL_EXT_framebuffer_multisample 1
@@ -57,6 +61,7 @@
 	GL_EXT_framebuffer_multisample_functions																\
 	GL_EXT_framebuffer_blit_functions																		\
 	GL_ARB_occlusion_query_functions																		\
-
+	GL_ARB_color_buffer_float_functions																		\
+	GL_MISSING_FUNCTIONS																					\
 
 #endif

reaction/code/renderer/tr_backend.c

@@ -601,6 +601,7 @@ void RB_RenderDrawSurfList( drawSurf_t *drawSurfs, int numDrawSurfs ) {
 	float			originalTime;
 	float			depth[2];
 	fbo_t*			fbo = NULL;
+	qboolean		inQuery = qfalse;
 
 	// save original time for entity shader offsets
 	originalTime = backEnd.refdef.floatTime;
@@ -665,6 +666,10 @@ void RB_RenderDrawSurfList( drawSurf_t *drawSurfs, int numDrawSurfs ) {
 			{
 				R_FBO_Bind(fbo);
 				qglDepthRange(depth[0], depth[1]);
+				if (inQuery) {
+					inQuery = qfalse;
+					qglEndQueryARB(GL_SAMPLES_PASSED_ARB);
+				}
 			}
 
 			if ( entityNum != ENTITYNUM_WORLD ) {
@@ -687,11 +692,16 @@ void RB_RenderDrawSurfList( drawSurf_t *drawSurfs, int numDrawSurfs ) {
 					// if we're rendering to a fbo
 					if (fbo) {
 						VectorCopy(backEnd.currentEntity->e.origin, backEnd.sunFlarePos);
-						backEnd.hasSunFlare = qtrue;
-						sunflare = qtrue;
 						// switch FBO
 						R_FBO_Bind(tr.fbo.fbos[1]);
 						qglDepthRange(1.f, 1.f);
+						if (glRefConfig.occlusionQuery && !inQuery && !backEnd.hasSunFlare) {
+							inQuery = qtrue;
+							tr.sunFlareQueryActive[tr.sunFlareQueryIndex] = qtrue;
+							qglBeginQueryARB(GL_SAMPLES_PASSED_ARB, tr.sunFlareQuery[tr.sunFlareQueryIndex]);
+						}
+						backEnd.hasSunFlare = qtrue;
+						sunflare = qtrue;
 					} else {
 						depthRange = qtrue;
 					}
@@ -785,6 +795,10 @@ void RB_RenderDrawSurfList( drawSurf_t *drawSurfs, int numDrawSurfs ) {
 	}
 
 	R_FBO_Bind(fbo);
+	if (inQuery) {
+		inQuery = qfalse;
+		qglEndQueryARB(GL_SAMPLES_PASSED_ARB);
+	}
 	
 	// go back to the world modelview matrix
 

reaction/code/renderer/tr_fbo.c

@@ -1,9 +1,5 @@
 #include "tr_local.h"
 
-#ifndef ARRAY_SIZE
-#	define ARRAY_SIZE(arr)				(sizeof(arr) / sizeof(arr[0]))
-#endif
-
 #define MAX_FBO_COLOR_BUFFERS		8
 
 /*
@@ -426,6 +422,18 @@ static void R_FBO_CreateDefaultBuffers(void)
 		NULL,
 		NULL);
 	
+	tr.fbo.tiny[0] = R_FBO_CreateSimple(
+		"tiny0",
+		R_FBO_CreateColorBuffer("*tinyBuffer0", glConfig.vidWidth/4, glConfig.vidHeight/4, qfalse, qfalse, GL_CLAMP_TO_EDGE),
+		NULL,
+		NULL);
+	
+	tr.fbo.tiny[1] = R_FBO_CreateSimple(
+		"tiny1",
+		R_FBO_CreateColorBuffer("*tinyBuffer1", glConfig.vidWidth/4, glConfig.vidHeight/4, qfalse, qfalse, GL_CLAMP_TO_EDGE),
+		NULL,
+		NULL);
+	
 	ri.Printf(PRINT_DEVELOPER, "...created %d FBOs\n", tr.fbo.numFBOs);
 }
 

reaction/code/renderer/tr_image.c

@@ -501,11 +501,19 @@ static void Upload32( unsigned *data,
 	GLenum		internalFormat = GL_RGB;
 	float		rMax = 0, gMax = 0, bMax = 0;
 
-	// Makro - render target
+	// Makro - no data, this must be a render target
 	if (!data)
 	{
-		internalFormat = GL_RGBA8;
-		qglTexImage2D (GL_TEXTURE_2D, 0, internalFormat, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
+		if (r_ext_framebuffer_object->integer >= 2 && glRefConfig.textureFloat)
+		{
+			internalFormat = GL_RGBA16F_ARB;
+			qglTexImage2D (GL_TEXTURE_2D, 0, internalFormat, width, height, 0, GL_RGBA, GL_FLOAT, NULL);
+		}
+		else
+		{
+			internalFormat = GL_RGBA8;
+			qglTexImage2D (GL_TEXTURE_2D, 0, internalFormat, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
+		}
 		GL_CheckErrors();
 		*format = internalFormat;
 		*pUploadWidth = width;
@@ -777,19 +785,45 @@ done:
 		ri.Hunk_FreeTempMemory( resampledBuffer );
 }
 
+static image_t *R_AllocImage(const char* name, int width, int height, qboolean mipmap, qboolean allowPicmip, int glWrapClampMode)
+{
+	image_t		*image;
+	long		hash;
+
+	image = tr.images[tr.numImages++] = ri.Hunk_Alloc( sizeof( image_t ), h_low );
+	Com_Memset(image, 0, sizeof(*image));
+
+	image->texnum = 1024 + tr.numImages;
+	image->mipmap = mipmap;
+	image->allowPicmip = allowPicmip;
+
+	strcpy (image->imgName, name);
+
+	image->width = width;
+	image->height = height;
+	image->wrapClampMode = glWrapClampMode;
+
+	hash = generateHashValue(name);
+	image->next = hashTable[hash];
+	hashTable[hash] = image;
+
+	return image;
+}
 
 /*
 ================
 R_CreateImage
 
 This is the only way any image_t are created
+
+Makro - except maybe for render targets
+at some point in the near future...
 ================
 */
 image_t *R_CreateImage( const char *name, const byte *pic, int width, int height, 
 					   qboolean mipmap, qboolean allowPicmip, int glWrapClampMode ) {
 	image_t		*image;
 	qboolean	isLightmap = qfalse;
-	long		hash;
 
 	if (strlen(name) >= MAX_QPATH ) {
 		ri.Error (ERR_DROP, "R_CreateImage: \"%s\" is too long\n", name);
@@ -802,18 +836,7 @@ image_t *R_CreateImage( const char *name, const byte *pic, int width, int height
 		ri.Error( ERR_DROP, "R_CreateImage: MAX_DRAWIMAGES hit\n");
 	}
 
-	image = tr.images[tr.numImages] = ri.Hunk_Alloc( sizeof( image_t ), h_low );
-	image->texnum = 1024 + tr.numImages;
-	tr.numImages++;
-
-	image->mipmap = mipmap;
-	image->allowPicmip = allowPicmip;
-
-	strcpy (image->imgName, name);
-
-	image->width = width;
-	image->height = height;
-	image->wrapClampMode = glWrapClampMode;
+	image = R_AllocImage(name, width, height, mipmap, allowPicmip, glWrapClampMode);
 
 	// lightmaps are always allocated on TMU 1
 	if ( qglActiveTextureARB && isLightmap ) {
@@ -845,10 +868,6 @@ image_t *R_CreateImage( const char *name, const byte *pic, int width, int height
 		GL_SelectTexture( 0 );
 	}
 
-	hash = generateHashValue(name);
-	image->next = hashTable[hash];
-	hashTable[hash] = image;
-
 	return image;
 }
 
@@ -1226,12 +1245,12 @@ void R_SetColorMappings( void ) {
 
 	// setup the overbright lighting
 	tr.overbrightBits = r_overBrightBits->integer;
-	if ( !glConfig.deviceSupportsGamma ) {
+	if ( !glConfig.deviceSupportsGamma && !glRefConfig.glslOverbright ) {
 		tr.overbrightBits = 0;		// need hardware gamma for overbright
 	}
 
 	// never overbright in windowed mode
-	if ( !glConfig.isFullscreen ) 
+	if ( !glConfig.isFullscreen && !glRefConfig.glslOverbright ) 
 	{
 		tr.overbrightBits = 0;
 	}
@@ -1268,6 +1287,9 @@ void R_SetColorMappings( void ) {
 
 	shift = tr.overbrightBits;
 
+	if (glRefConfig.glslOverbright)
+		shift = 0;
+
 	for ( i = 0; i < 256; i++ ) {
 		if ( g == 1 ) {
 			inf = i;

reaction/code/renderer/tr_init.c

@@ -410,7 +410,7 @@ void RB_TakeScreenshotJPEG( int x, int y, int width, int height, char *fileName
 
 	buffer = ri.Hunk_AllocateTempMemory(glConfig.vidWidth*glConfig.vidHeight*4);
 
-	qglReadPixels( x, y, width, height, GL_RGBA, GL_UNSIGNED_BYTE, buffer ); 
+	qglReadPixels( x, y, width, height, GL_RGB, GL_UNSIGNED_BYTE, buffer ); 
 
 	// gamma correct
 	if ( glConfig.deviceSupportsGamma ) {
@@ -430,6 +430,7 @@ RB_TakeScreenshotCmd
 */
 const void *RB_TakeScreenshotCmd( const void *data ) {
 	const screenshotCommand_t	*cmd;
+	fbo_t						*fbo = R_FBO_Bind(NULL);
 	
 	cmd = (const screenshotCommand_t *)data;
 	
@@ -437,6 +438,8 @@ const void *RB_TakeScreenshotCmd( const void *data ) {
 		RB_TakeScreenshotJPEG( cmd->x, cmd->y, cmd->width, cmd->height, cmd->fileName);
 	else
 		RB_TakeScreenshot( cmd->x, cmd->y, cmd->width, cmd->height, cmd->fileName);
+
+	R_FBO_Bind(fbo);
 	
 	return (const void *)(cmd + 1);	
 }
@@ -708,14 +711,19 @@ RB_TakeVideoFrameCmd
 const void *RB_TakeVideoFrameCmd( const void *data )
 {
 	const videoFrameCommand_t	*cmd;
-	int												frameSize;
-	int												i;
+	int							frameSize;
+	int							i;
+	fbo_t						*fbo;
 	
 	cmd = (const videoFrameCommand_t *)data;
+
+	fbo = R_FBO_Bind(NULL);
 	
 	qglReadPixels( 0, 0, cmd->width, cmd->height, GL_RGBA,
 			GL_UNSIGNED_BYTE, cmd->captureBuffer );
 
+	R_FBO_Bind(fbo);
+
 	// gamma correct
 	if( glConfig.deviceSupportsGamma )
 		R_GammaCorrect( cmd->captureBuffer, cmd->width * cmd->height * 4 );
@@ -1070,12 +1078,16 @@ void R_InitQueries(void)
 {
 	if (!glRefConfig.occlusionQuery)
 		return;
+
+	qglGenQueriesARB(ARRAY_SIZE(tr.sunFlareQuery), tr.sunFlareQuery);
 }
 
 void R_ShutDownQueries(void)
 {
 	if (!glRefConfig.occlusionQuery)
 		return;
+
+	qglDeleteQueriesARB(ARRAY_SIZE(tr.sunFlareQuery), tr.sunFlareQuery);
 }
 
 /*

reaction/code/renderer/tr_local.h

@@ -43,6 +43,10 @@ long myftol( float f );
 
 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
 
+#ifndef ARRAY_SIZE
+#	define ARRAY_SIZE(arr)				(sizeof(arr) / sizeof(arr[0]))
+#endif
+
 // everything that is needed by the backend needs
 // to be double buffered to allow it to run in
 // parallel on a dual cpu machine
@@ -2166,6 +2170,11 @@ typedef struct {
 	qboolean	framebufferBlit;
 	qboolean	framebufferMultisample;
 	
+	qboolean	glslOverbright;
+	
+	qboolean	textureFloat;
+	qboolean	colorBufferFloat;
+	
 	qboolean	occlusionQuery;
 
 	// These next three are all required for one chunk of code, so glsl is
@@ -2229,6 +2238,7 @@ typedef struct {
 	
 	fbo_t					*full[2];	// full resolution, shared zbuffer
 	fbo_t					*quarter[2];	// quarter resolution, no zbuffer
+	fbo_t					*tiny[2];	// 1/16 resolution, no zbuffer
 
 	int						numFBOs;
 	fbo_t					*fbos[1024];
@@ -2352,6 +2362,10 @@ typedef struct {
 	int						numSkins;
 	skin_t					*skins[MAX_SKINS];
 
+	GLuint					sunFlareQuery[2];
+	int						sunFlareQueryIndex;
+	qboolean				sunFlareQueryActive[2];
+
 	float					sinTable[FUNCTABLE_SIZE];
 	float					squareTable[FUNCTABLE_SIZE];
 	float					triangleTable[FUNCTABLE_SIZE];

reaction/code/renderer/tr_postprocess.c

@@ -158,6 +158,35 @@ static void RB_RadialBlur(int passes, float stretch, float x, float y, float w,
 	RB_Color4f(1.f, 1.f, 1.f, 1.f);
 }
 
+static qboolean RB_UpdateSunFlareVis(void)
+{
+	GLuint sampleCount = 0;
+	if (!glRefConfig.occlusionQuery)
+		return qtrue;
+
+	tr.sunFlareQueryIndex ^= 1;
+	if (!tr.sunFlareQueryActive[tr.sunFlareQueryIndex])
+		return qtrue;
+
+	/* debug code */
+	if (0)
+	{
+		int iter;
+		for (iter=0 ; ; ++iter)
+		{
+			GLint available = 0;
+			qglGetQueryObjectivARB(tr.sunFlareQuery[tr.sunFlareQueryIndex], GL_QUERY_RESULT_AVAILABLE_ARB, &available);
+			if (available)
+				break;
+		}
+
+		ri.Printf(PRINT_DEVELOPER, "Waited %d iterations\n", iter);
+	}
+	
+	qglGetQueryObjectuivARB(tr.sunFlareQuery[tr.sunFlareQueryIndex], GL_QUERY_RESULT_ARB, &sampleCount);
+	return sampleCount > 0;
+}
+
 /*
 ================
 RB_GodRays
@@ -186,6 +215,9 @@ static void RB_GodRays(void)
 	if (dot < cutoff)
 		return;
 
+	if (!RB_UpdateSunFlareVis())
+		return;
+
 	VectorCopy(backEnd.sunFlarePos, pos);
 	pos[3] = 1.f;
 
@@ -202,8 +234,8 @@ static void RB_GodRays(void)
 	// viewport dimensions
 	w = glConfig.vidWidth;
 	h = glConfig.vidHeight;
-	w2 = w * 0.5f;
-	h2 = h * 0.5f;
+	w2 = glConfig.vidWidth / 2;
+	h2 = glConfig.vidHeight / 2;
 
 	// initialize quarter buffers
 	{
@@ -212,7 +244,7 @@ static void RB_GodRays(void)
 		RB_SetGL2D_Level(1);
 		
 		GL_State(GLS_DEPTHTEST_DISABLE | GLS_SRCBLEND_ONE | GLS_DSTBLEND_ZERO );
-		RB_Color4f(1.f, 1.f, 1.f, 1.f);
+		RB_Color4f(mul, mul, mul, 1.f);
 
 		// first, downsample the main framebuffers
 		R_FBO_Bind(tr.fbo.quarter[0]);
@@ -245,15 +277,67 @@ static void RB_GodRays(void)
 	
 	// add result back on top of the main buffer
 	{
+		float mul = 1.f;
 		R_FBO_BindColorBuffer(R_FBO_Bind(tr.fbo.full[0]), 0);
 		RB_SetGL2D_Level(0);
 		GL_State(GLS_DEPTHTEST_DISABLE | GLS_SRCBLEND_ONE | GLS_DSTBLEND_ONE );
 		
-		RB_Color4f(1.f, 1.f, 1.f, 1.f);
+		RB_Color4f(mul, mul, mul, 1.f);
 		RB_DrawQuad(0.f, 0.f, w, h, 0.f, 0.f, 1.f, 1.f);
 	}
 }
 
+static void RB_BlurAxis(float w, float h, qboolean horizontal)
+{
+	float dx, dy;
+	float xmul, ymul;
+	float weights[3] = {
+		0.316216216f,
+		0.227027027f,
+		0.070270270f,
+	};
+	float offsets[3] = {
+		0.f,
+		1.3846153846f,
+		3.2307692308f,
+	};
+
+	xmul = horizontal;
+	ymul = 1.f - xmul;
+
+	xmul /= w;
+	ymul /= h;
+
+	GL_State(GLS_DEPTHTEST_DISABLE | GLS_SRCBLEND_ONE | GLS_DSTBLEND_ZERO );
+	
+	RB_Color4f(weights[0], weights[0], weights[0], 1.f);
+	RB_DrawQuad(0.f, 0.f, w, h, 0.f, 0.f, 1.f, 1.f);
+
+	GL_State(GLS_DEPTHTEST_DISABLE | GLS_SRCBLEND_ONE | GLS_DSTBLEND_ONE );
+	
+	RB_Color4f(weights[1], weights[1], weights[1], 1.f);
+	dx = offsets[1] * xmul;
+	dy = offsets[1] * ymul;
+	RB_DrawQuad(0.f, 0.f, w, h, dx, dy, 1.f+dx, 1.f+dy);
+	RB_DrawQuad(0.f, 0.f, w, h, -dx, -dy, 1.f-dx, 1.f-dy);
+
+	RB_Color4f(weights[2], weights[2], weights[2], 1.f);
+	dx = offsets[2] * xmul;
+	dy = offsets[2] * ymul;
+	RB_DrawQuad(0.f, 0.f, w, h, dx, dy, 1.f+dx, 1.f+dy);
+	RB_DrawQuad(0.f, 0.f, w, h, -dx, -dy, 1.f-dx, 1.f-dy);
+}
+
+static void RB_HBlur(float w, float h)
+{
+	RB_BlurAxis(w, h, qtrue);
+}
+
+static void RB_VBlur(float w, float h)
+{
+	RB_BlurAxis(w, h, qfalse);
+}
+
 void RB_FBO_Set2D(void)
 {
 	RB_DrawQuad = NULL;
@@ -321,6 +405,12 @@ void RB_FBO_Blit(void)
 	R_FBO_BindColorBuffer(fbo, 0);
 	GL_State(GLS_DEPTHTEST_DISABLE | GLS_SRCBLEND_ONE | GLS_DSTBLEND_ZERO );
 
+	if (glRefConfig.glslOverbright)
+	{
+		float mul = 1.f / tr.identityLight;
+		RB_Color4f(mul, mul, mul, 1.f);
+	}
+
 	RB_DrawQuad(0.f, 0.f, glConfig.vidWidth, glConfig.vidHeight, 0.f, 0.f, 1.f, 1.f);
 
 	R_FBO_Bind(fbo);
@@ -334,7 +424,7 @@ RB_PostProcess
 
 void RB_PostProcess(void)
 {
-	if (!glState.currentFBO)
+	if (!glState.currentFBO || backEnd.refdef.rdflags & RDF_NOWORLDMODEL)
 		return;
 
 	RB_FBO_Set2D();