Make software overbright optional (cvar r_softOverbright) and reduce the number of FBOs and FBO blits when able.

7 files changed, 120 insertions, 161 deletions

diff --git a/src/rend2/tr_backend.c b/src/rend2/tr_backend.c
index 0c3321e9..55457a51 100644
--- a/src/rend2/tr_backend.c
+++ b/src/rend2/tr_backend.c
@@ -485,7 +485,7 @@ void RB_BeginDrawingView (void) {
 		// drawing more world check is in case of double renders, such as skyportals
 		if (backEnd.viewParms.targetFbo == NULL)
 		{
-			if (backEnd.framePostProcessed && (backEnd.refdef.rdflags & RDF_NOWORLDMODEL))
+			if (!tr.renderFbo || (backEnd.framePostProcessed && (backEnd.refdef.rdflags & RDF_NOWORLDMODEL)))
 			{
 				FBO_Bind(tr.screenScratchFbo);
 			}
@@ -959,7 +959,7 @@ void RE_StretchRaw (int x, int y, int w, int h, int cols, int rows, const byte *
 	// FIXME: HUGE hack
 	if (glRefConfig.framebufferObject)
 	{
-		if (backEnd.framePostProcessed)
+		if (!tr.renderFbo || backEnd.framePostProcessed)
 		{
 			FBO_Bind(tr.screenScratchFbo);
 		}
@@ -1108,7 +1108,7 @@ const void *RB_StretchPic ( const void *data ) {
 	// FIXME: HUGE hack
 	if (glRefConfig.framebufferObject)
 	{
-		if (backEnd.framePostProcessed)
+		if (!tr.renderFbo || backEnd.framePostProcessed)
 		{
 			FBO_Bind(tr.screenScratchFbo);
 		}
@@ -1223,58 +1223,22 @@ const void	*RB_DrawSurfs( const void *data ) {
 		qglColorMask(!backEnd.colorMask[0], !backEnd.colorMask[1], !backEnd.colorMask[2], !backEnd.colorMask[3]);
 		backEnd.depthFill = qfalse;
 
-		// If we're using multisampling, resolve the depth first
 		if (tr.msaaResolveFbo)
 		{
+			// If we're using multisampling, resolve the depth first
 			FBO_FastBlit(tr.renderFbo, NULL, tr.msaaResolveFbo, NULL, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
 		}
+		else if (tr.renderFbo == NULL)
+		{
+			// If we're rendering directly to the screen, copy the depth to a texture
+			GL_BindToTMU(tr.renderDepthImage, 0);
+			qglCopyTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH_COMPONENT24, 0, 0, glConfig.vidWidth, glConfig.vidHeight, 0);
+		}
 
 		if (r_ssao->integer)
-		{			
-			vec2_t srcTexScale;
-			vec4_t color;
-			vec4_t quadVerts[4];
-			vec2_t texCoords[4];
-			vec2_t invTexRes;
-			
-			matrix_t idmatrix;
-
-			srcTexScale[0] = srcTexScale[1] = 1.0f;
-			color[0] = color[1] = color[2] = color[3] = 1.0f;
-
-			FBO_Bind(tr.hdrDepthFbo);
-
-			qglViewport(0, 0, tr.hdrDepthFbo->width, tr.hdrDepthFbo->height);
-			qglScissor(0, 0, tr.hdrDepthFbo->width, tr.hdrDepthFbo->height);
-
-			Matrix16Identity(idmatrix);
-
-			VectorSet4(quadVerts[0], -1,  1, 0, 1);
-			VectorSet4(quadVerts[1],  1,  1, 0, 1);
-			VectorSet4(quadVerts[2],  1, -1, 0, 1);
-			VectorSet4(quadVerts[3], -1, -1, 0, 1);
-
-			texCoords[0][0] = 0; texCoords[0][1] = 1;
-			texCoords[1][0] = 1; texCoords[1][1] = 1;
-			texCoords[2][0] = 1; texCoords[2][1] = 0;
-			texCoords[3][0] = 0; texCoords[3][1] = 0;
-
-			invTexRes[0] = 0.0f;
-			invTexRes[1] = 0.0f;
-
-			GL_State( GLS_DEPTHTEST_DISABLE );
-
-			GLSL_BindProgram(&tr.textureColorShader);
-			
-			GL_BindToTMU(tr.renderDepthImage, TB_COLORMAP);
-
-			GLSL_SetUniformMatrix16(&tr.textureColorShader, TEXTURECOLOR_UNIFORM_MODELVIEWPROJECTIONMATRIX, idmatrix);
-			GLSL_SetUniformVec4(&tr.textureColorShader, TEXTURECOLOR_UNIFORM_COLOR, color);
-			GLSL_SetUniformVec2(&tr.textureColorShader, TEXTURECOLOR_UNIFORM_INVTEXRES, invTexRes);
-			GLSL_SetUniformVec2(&tr.textureColorShader, TEXTURECOLOR_UNIFORM_AUTOEXPOSUREMINMAX, tr.refdef.autoExposureMinMax);
-			GLSL_SetUniformVec3(&tr.textureColorShader, TEXTURECOLOR_UNIFORM_TONEMINAVGMAXLINEAR, tr.refdef.toneMinAvgMaxLinear);
-
-			RB_InstantQuad2(quadVerts, texCoords); //, color, shaderProgram, invTexRes);
+		{
+			// need the depth in a texture we can do GL_LINEAR sampling on, so copy it to an HDR image
+			FBO_BlitFromTexture(tr.renderDepthImage, NULL, NULL, tr.hdrDepthFbo, NULL, NULL, NULL, 0);
 		}
 
 		if (backEnd.viewParms.flags & VPF_USESUNLIGHT)
@@ -1584,7 +1548,7 @@ const void *RB_ClearDepth(const void *data)
 
 	if (glRefConfig.framebufferObject)
 	{
-		if (backEnd.framePostProcessed)
+		if (!tr.renderFbo || backEnd.framePostProcessed)
 		{
 			FBO_Bind(tr.screenScratchFbo);
 		}
@@ -1599,7 +1563,7 @@ const void *RB_ClearDepth(const void *data)
 	// if we're doing MSAA, clear the depth texture for the resolve buffer
 	if (tr.msaaResolveFbo)
 	{
-		FBO_Bind(tr.screenScratchFbo);
+		FBO_Bind(tr.msaaResolveFbo);
 		qglClear(GL_DEPTH_BUFFER_BIT);
 	}
 
@@ -1648,47 +1612,38 @@ const void	*RB_SwapBuffers( const void *data ) {
 
 	if (glRefConfig.framebufferObject)
 	{
-		// copy final image to screen
-		vec4_t color;
-
-		if (backEnd.framePostProcessed)
-		{
-			// frame was postprocessed into screen fbo, copy from there
-		}
-		else if (!glRefConfig.framebuffer_srgb)
+		if (!backEnd.framePostProcessed)
 		{
-			// Copy render to screenscratch, possibly resolving MSAA
-			FBO_FastBlit(tr.renderFbo, NULL, tr.screenScratchFbo, NULL, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+			if (tr.msaaResolveFbo && r_hdr->integer)
+			{
+				// Resolving an RGB16F MSAA FBO to the screen messes with the brightness, so resolve to an RGB16F FBO first
+				FBO_FastBlit(tr.renderFbo, NULL, tr.msaaResolveFbo, NULL, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+				FBO_FastBlit(tr.msaaResolveFbo, NULL, tr.screenScratchFbo, NULL, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+			}
+			else if (tr.renderFbo)
+			{
+				FBO_FastBlit(tr.renderFbo, NULL, tr.screenScratchFbo, NULL, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+			}
 		}
-		else
+
+		if (tr.screenScratchFbo)
 		{
-			FBO_t *srcFbo = tr.renderFbo;
+			vec4_t color;
 
-			if (tr.msaaResolveFbo)
-			{
-				// Resolve the MSAA before copying
-				FBO_FastBlit(srcFbo, NULL, tr.msaaResolveFbo, NULL, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+			color[0] =
+			color[1] =
+			color[2] = pow(2, tr.overbrightBits); //exp2(tr.overbrightBits);
+			color[3] = 1.0f;
 
-				srcFbo = tr.msaaResolveFbo;
-			}
+			// turn off colormask when copying final image
+			if (backEnd.colorMask[0] || backEnd.colorMask[1] || backEnd.colorMask[2] || backEnd.colorMask[3])
+				qglColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
+				
+			FBO_Blit(tr.screenScratchFbo, NULL, NULL, NULL, NULL, NULL, color, 0);
 
-			// need to copy from resolve to screenscratch to fix gamma
-			FBO_Blit(srcFbo, NULL, NULL, tr.screenScratchFbo, NULL, NULL, NULL, 0);
+			if (backEnd.colorMask[0] || backEnd.colorMask[1] || backEnd.colorMask[2] || backEnd.colorMask[3])
+				qglColorMask(!backEnd.colorMask[0], !backEnd.colorMask[1], !backEnd.colorMask[2], !backEnd.colorMask[3]);
 		}
-		
-		color[0] =
-		color[1] =
-		color[2] = pow(2, tr.overbrightBits); //exp2(tr.overbrightBits);
-		color[3] = 1.0f;
-
-		// turn off colormask when copying final image
-		if (backEnd.colorMask[0] || backEnd.colorMask[1] || backEnd.colorMask[2] || backEnd.colorMask[3])
-			qglColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
-			
-		FBO_Blit(tr.screenScratchFbo, NULL, NULL, NULL, NULL, NULL, color, 0);
-
-		if (backEnd.colorMask[0] || backEnd.colorMask[1] || backEnd.colorMask[2] || backEnd.colorMask[3])
-			qglColorMask(!backEnd.colorMask[0], !backEnd.colorMask[1], !backEnd.colorMask[2], !backEnd.colorMask[3]);
 	}
 
 	if ( !glState.finishCalled ) {
@@ -1750,11 +1705,9 @@ const void *RB_PostProcess(const void *data)
 	FBO_t *srcFbo;
 	qboolean autoExposure;
 
-	if (!glRefConfig.framebufferObject)
+	if (!glRefConfig.framebufferObject || !r_postProcess->integer)
 	{
 		// do nothing
-		backEnd.framePostProcessed = qtrue;
-
 		return (const void *)(cmd + 1);
 	}
 
@@ -1766,36 +1719,37 @@ const void *RB_PostProcess(const void *data)
 		srcFbo = tr.msaaResolveFbo;
 	}
 
-	if (r_postProcess->integer && r_ssao->integer)
+	if (r_ssao->integer)
 	{
-		vec4i_t dstBox;
-		VectorSet4(dstBox, 0, 0, srcFbo->width, srcFbo->height);
-		FBO_BlitFromTexture(tr.screenSsaoImage, NULL, NULL, srcFbo, dstBox, NULL, NULL, GLS_SRCBLEND_DST_COLOR | GLS_DSTBLEND_ZERO);
+		FBO_BlitFromTexture(tr.screenSsaoImage, NULL, NULL, srcFbo, NULL, NULL, NULL, GLS_SRCBLEND_DST_COLOR | GLS_DSTBLEND_ZERO);
 	}
 
-	if (r_postProcess->integer && (r_toneMap->integer || r_forceToneMap->integer))
-	{
-		autoExposure = r_autoExposure->integer || r_forceAutoExposure;
-		RB_ToneMap(srcFbo, autoExposure);
-	}
-	else if (!glRefConfig.framebuffer_srgb && r_cameraExposure->value == 0.0f)
-	{
-		FBO_FastBlit(srcFbo, NULL, tr.screenScratchFbo, NULL, GL_COLOR_BUFFER_BIT, GL_NEAREST);
-	}
-	else
+	if (srcFbo)
 	{
-		vec4_t color;
+		if (r_hdr->integer && (r_toneMap->integer || r_forceToneMap->integer))
+		{
+			autoExposure = r_autoExposure->integer || r_forceAutoExposure->integer;
+			RB_ToneMap(srcFbo, autoExposure);
+		}
+		else if (!glRefConfig.framebuffer_srgb && r_cameraExposure->value == 0.0f)
+		{
+			FBO_FastBlit(srcFbo, NULL, tr.screenScratchFbo, NULL, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+		}
+		else
+		{
+			vec4_t color;
 
-		color[0] =
-		color[1] =
-		color[2] = pow(2, r_cameraExposure->value); //exp2(r_cameraExposure->value);
-		color[3] = 1.0f;
+			color[0] =
+			color[1] =
+			color[2] = pow(2, r_cameraExposure->value); //exp2(r_cameraExposure->value);
+			color[3] = 1.0f;
 
-		FBO_Blit(srcFbo, NULL, NULL, tr.screenScratchFbo, NULL, NULL, color, 0);
+			FBO_Blit(srcFbo, NULL, NULL, tr.screenScratchFbo, NULL, NULL, color, 0);
+		}
 	}
 
 #ifdef REACTION
-	if (r_postProcess->integer && glRefConfig.framebufferObject)
+	if (1)
 	{
 		RB_GodRays();
 
@@ -1820,16 +1774,16 @@ const void *RB_PostProcess(const void *data)
 	if (0)
 	{
 		vec4i_t dstBox;
-		VectorSet4(dstBox, 256, tr.screenScratchFbo->height - 256, 256, 256);
+		VectorSet4(dstBox, 256, glConfig.vidHeight - 256, 256, 256);
 		FBO_BlitFromTexture(tr.renderDepthImage, NULL, NULL, tr.screenScratchFbo, dstBox, NULL, NULL, 0);
-		VectorSet4(dstBox, 512, tr.screenScratchFbo->height - 256, 256, 256);
+		VectorSet4(dstBox, 512, glConfig.vidHeight - 256, 256, 256);
 		FBO_BlitFromTexture(tr.screenShadowImage, NULL, NULL, tr.screenScratchFbo, dstBox, NULL, NULL, 0);
 	}
 
 	if (0)
 	{
 		vec4i_t dstBox;
-		VectorSet4(dstBox, 256, tr.screenScratchFbo->height - 256, 256, 256);
+		VectorSet4(dstBox, 256, glConfig.vidHeight - 256, 256, 256);
 		FBO_BlitFromTexture(tr.renderImage, NULL, NULL, tr.screenScratchFbo, dstBox, NULL, NULL, 0);
 	}
 
diff --git a/src/rend2/tr_cmds.c b/src/rend2/tr_cmds.c
index 9945da3f..2f67d221 100644
--- a/src/rend2/tr_cmds.c
+++ b/src/rend2/tr_cmds.c
@@ -502,29 +502,21 @@ void RE_BeginFrame( stereoFrame_t stereoFrame ) {
 				if (glRefConfig.framebufferObject)
 				{
 					// clear all framebuffers
-					// FIXME: must be a better way to do this
-					int i;
-
-					for (i = 0; i < 3; i++)
+					if (tr.msaaResolveFbo)
 					{
-						if (i == 1 && !tr.msaaResolveFbo)
-							continue;
-
-						switch(i)
-						{
-							case 0:
-								FBO_Bind(tr.renderFbo);
-								break;
-
-							case 1:
-								FBO_Bind(tr.msaaResolveFbo);
-								break;
+						FBO_Bind(tr.msaaResolveFbo);
+						qglClear(GL_COLOR_BUFFER_BIT);
+					}
 
-							case 2:
-								FBO_Bind(tr.screenScratchFbo);
-								break;
-						}
+					if (tr.renderFbo)
+					{
+						FBO_Bind(tr.renderFbo);
+						qglClear(GL_COLOR_BUFFER_BIT);
+					}
 
+					if (tr.screenScratchFbo)
+					{
+						FBO_Bind(tr.screenScratchFbo);
 						qglClear(GL_COLOR_BUFFER_BIT);
 					}
 
diff --git a/src/rend2/tr_fbo.c b/src/rend2/tr_fbo.c
index 04f9ff26..0e84d944 100644
--- a/src/rend2/tr_fbo.c
+++ b/src/rend2/tr_fbo.c
@@ -309,7 +309,7 @@ FBO_Bind
 */
 void FBO_Bind(FBO_t * fbo)
 {
-	if (fbo && glState.currentFBO == fbo)
+	if (glState.currentFBO == fbo)
 		return;
 		
 	if (r_logFile->integer)
@@ -405,6 +405,8 @@ void FBO_Init(void)
 		ri.Cvar_SetValue("r_ext_framebuffer_multisample", (float)multisample);
 	}
 	
+	// only create a render FBO if we need to resolve MSAA or do HDR
+	// otherwise just render straight to the screen (tr.renderFbo = NULL)
 	if (multisample && glRefConfig.framebufferMultisample)
 	{
 		tr.renderFbo = FBO_Create("_render", tr.renderDepthImage->width, tr.renderDepthImage->height);
@@ -427,7 +429,7 @@ void FBO_Init(void)
 
 		R_CheckFBO(tr.msaaResolveFbo);
 	}
-	else
+	else if (r_hdr->integer)
 	{
 		tr.renderFbo = FBO_Create("_render", tr.renderDepthImage->width, tr.renderDepthImage->height);
 		FBO_Bind(tr.renderFbo);
@@ -443,10 +445,13 @@ void FBO_Init(void)
 
 	// clear render buffer
 	// this fixes the corrupt screen bug with r_hdr 1 on older hardware
-	FBO_Bind(tr.renderFbo);
-	qglClearColor( 1, 0, 0.5, 1 );
-	qglClear( GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT );
-	FBO_Bind(NULL);
+	if (tr.renderFbo)
+	{
+		FBO_Bind(tr.renderFbo);
+		qglClearColor( 1, 0, 0.5, 1 );
+		qglClear( GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT );
+		FBO_Bind(NULL);
+	}
 
 #ifdef REACTION
 	{
@@ -525,6 +530,7 @@ void FBO_Init(void)
 		R_CheckFBO(tr.targetLevelsFbo);
 	}
 
+	if (r_softOverbright->integer)
 	{
 		//tr.screenScratchFbo = FBO_Create("_screenscratch", width, height);
 		tr.screenScratchFbo = FBO_Create("_screenscratch", tr.screenScratchImage->width, tr.screenScratchImage->height);
diff --git a/src/rend2/tr_image.c b/src/rend2/tr_image.c
index eaf133b4..c00bfbbc 100644
--- a/src/rend2/tr_image.c
+++ b/src/rend2/tr_image.c
@@ -2943,6 +2943,7 @@ void R_CreateBuiltinImages( void ) {
 		tr.godRaysImage = R_CreateImage("*godRays", NULL, width, height, IMGTYPE_COLORALPHA, IMGFLAG_NO_COMPRESSION | IMGFLAG_CLAMPTOEDGE, GL_RGBA8);
 #endif
 
+		if (r_softOverbright->integer)
 		{
 			int format;
 
@@ -3033,14 +3034,14 @@ void R_SetColorMappings( void ) {
 		tr.overbrightBits = 0;		// need hardware gamma for overbright
 	}
 
-	// never overbright in windowed mode
-	if ( 0 /* !glConfig.isFullscreen */ ) 
+	// never overbright in windowed mode without soft overbright
+	if ( !glConfig.isFullscreen && !r_softOverbright->integer ) 
 	{
 		tr.overbrightBits = 0;
 	}
 
 	// never overbright with tonemapping
-	if ( r_toneMap->integer )
+	if ( r_toneMap->integer && r_hdr->integer )
 	{
 		tr.overbrightBits = 0;
 	}
@@ -3077,8 +3078,11 @@ void R_SetColorMappings( void ) {
 
 	shift = tr.overbrightBits;
 
-	if (glRefConfig.framebufferObject)
+	// no shift with soft overbright
+	if (r_softOverbright->integer)
+	{
 		shift = 0;
+	}
 
 	for ( i = 0; i < 256; i++ ) {
 		if ( g == 1 ) {
diff --git a/src/rend2/tr_init.c b/src/rend2/tr_init.c
index 0f14c081..7b384106 100644
--- a/src/rend2/tr_init.c
+++ b/src/rend2/tr_init.c
@@ -107,6 +107,9 @@ cvar_t  *r_mergeMultidraws;
 cvar_t  *r_mergeLeafSurfaces;
 
 cvar_t  *r_cameraExposure;
+
+cvar_t  *r_softOverbright;
+
 cvar_t  *r_hdr;
 cvar_t  *r_postProcess;
 
@@ -1149,6 +1152,8 @@ void R_Register( void )
 	r_greyscale = ri.Cvar_Get("r_greyscale", "0", CVAR_ARCHIVE | CVAR_LATCH);
 	ri.Cvar_CheckRange(r_greyscale, 0, 1, qfalse);
 
+	r_softOverbright = ri.Cvar_Get( "r_softOverbright", "1", CVAR_ARCHIVE | CVAR_LATCH );
+
 	r_hdr = ri.Cvar_Get( "r_hdr", "1", CVAR_ARCHIVE | CVAR_LATCH );
 	r_postProcess = ri.Cvar_Get( "r_postProcess", "1", CVAR_ARCHIVE );
 
diff --git a/src/rend2/tr_local.h b/src/rend2/tr_local.h
index 84829dc6..77b11f90 100644
--- a/src/rend2/tr_local.h
+++ b/src/rend2/tr_local.h
@@ -2092,6 +2092,8 @@ extern	cvar_t	*r_anaglyphMode;
 extern  cvar_t  *r_mergeMultidraws;
 extern  cvar_t  *r_mergeLeafSurfaces;
 
+extern  cvar_t  *r_softOverbright;
+
 extern  cvar_t  *r_hdr;
 extern  cvar_t  *r_postProcess;
 
diff --git a/src/rend2/tr_postprocess.c b/src/rend2/tr_postprocess.c
index 0cb162b3..73fa7175 100644
--- a/src/rend2/tr_postprocess.c
+++ b/src/rend2/tr_postprocess.c
@@ -112,14 +112,16 @@ void RB_BokehBlur(float blur)
 		if (blur > 0.0f)
 		{
 			// create a quarter texture
-			FBO_Blit(tr.screenScratchFbo, NULL, NULL, tr.quarterFbo[0], NULL, NULL, NULL, 0);
+			//FBO_Blit(NULL, NULL, NULL, tr.quarterFbo[0], NULL, NULL, NULL, 0);
+			FBO_FastBlit(tr.screenScratchFbo, NULL, tr.quarterFbo[0], NULL, GL_COLOR_BUFFER_BIT, GL_LINEAR);
 		}
 
 #ifndef HQ_BLUR
 		if (blur > 1.0f)
 		{
 			// create a 1/16th texture
-			FBO_Blit(tr.quarterFbo[0], NULL, NULL, tr.textureScratchFbo[0], NULL, NULL, NULL, 0);
+			//FBO_Blit(tr.quarterFbo[0], NULL, NULL, tr.textureScratchFbo[0], NULL, NULL, NULL, 0);
+			FBO_FastBlit(tr.quarterFbo[0], NULL, tr.textureScratchFbo[0], NULL, GL_COLOR_BUFFER_BIT, GL_LINEAR);
 		}
 #endif
 
@@ -340,16 +342,15 @@ void RB_GodRays(void)
 		VectorSet4(color, mul, mul, mul, 1);
 
 		// first, downsample the framebuffer
-		VectorSet4(srcBox, 0, 0, tr.godRaysFbo->width, tr.godRaysFbo->height);
-		VectorSet4(dstBox, 0, 0, tr.quarterFbo[0]->width, tr.quarterFbo[0]->height);
-		FBO_Blit(tr.godRaysFbo, srcBox, texScale, tr.quarterFbo[0], dstBox, &tr.textureColorShader, color, 0);
-		
 		if (colorize)
 		{
-			VectorSet4(srcBox, 0, 0, tr.screenScratchFbo->width, tr.screenScratchFbo->height);
-			FBO_Blit(tr.screenScratchFbo, srcBox, texScale, tr.quarterFbo[0], dstBox, &tr.textureColorShader, color, 
-			         GLS_SRCBLEND_DST_COLOR | GLS_DSTBLEND_ZERO);
-		}	
+			FBO_FastBlit(tr.screenScratchFbo, NULL, tr.quarterFbo[0], NULL, GL_COLOR_BUFFER_BIT, GL_LINEAR);
+			FBO_Blit(tr.godRaysFbo, NULL, NULL, tr.quarterFbo[0], NULL, NULL, color, GLS_SRCBLEND_DST_COLOR | GLS_DSTBLEND_ZERO);
+		}
+		else
+		{
+			FBO_FastBlit(tr.godRaysFbo, NULL, tr.quarterFbo[0], NULL, GL_COLOR_BUFFER_BIT, GL_LINEAR);
+		}
 	}
 
 	// radial blur passes, ping-ponging between the two quarter-size buffers
@@ -375,7 +376,7 @@ void RB_GodRays(void)
 		VectorSet4(color, mul, mul, mul, 1);
 
 		VectorSet4(srcBox, 0, 0, tr.quarterFbo[0]->width, tr.quarterFbo[0]->height);
-		VectorSet4(dstBox, 0, 0, tr.screenScratchFbo->width, tr.screenScratchFbo->height);
+		VectorSet4(dstBox, 0, 0, glConfig.vidWidth, glConfig.vidHeight);
 		FBO_Blit(tr.quarterFbo[0], srcBox, texScale, tr.screenScratchFbo, dstBox, &tr.textureColorShader, color, GLS_SRCBLEND_ONE | GLS_DSTBLEND_ONE);
 	}
 }
@@ -462,13 +463,8 @@ void RB_GaussianBlur(float blur)
 		VectorSet4(color, 1, 1, 1, 1);
 
 		// first, downsample the framebuffer
-		VectorSet4(srcBox, 0, 0, tr.screenScratchFbo->width, tr.screenScratchFbo->height);
-		VectorSet4(dstBox, 0, 0, tr.quarterFbo[0]->width, tr.quarterFbo[0]->height);
-		FBO_Blit(tr.screenScratchFbo, srcBox, texScale, tr.quarterFbo[0], dstBox, &tr.textureColorShader, color, 0);
-
-		VectorSet4(srcBox, 0, 0, tr.quarterFbo[0]->width, tr.quarterFbo[0]->height);
-		VectorSet4(dstBox, 0, 0, tr.textureScratchFbo[0]->width, tr.textureScratchFbo[0]->height);
-		FBO_Blit(tr.quarterFbo[0], srcBox, texScale, tr.textureScratchFbo[0], dstBox, &tr.textureColorShader, color, 0);
+		FBO_FastBlit(tr.screenScratchFbo, NULL, tr.quarterFbo[0], NULL, GL_COLOR_BUFFER_BIT, GL_LINEAR);
+		FBO_FastBlit(tr.quarterFbo[0], NULL, tr.textureScratchFbo[0], NULL, GL_COLOR_BUFFER_BIT, GL_LINEAR);
 
 		// set the alpha channel
 		VectorSet4(srcBox, 0, 0, tr.whiteImage->width, tr.whiteImage->height);
@@ -483,7 +479,7 @@ void RB_GaussianBlur(float blur)
 
 		// finally, merge back to framebuffer
 		VectorSet4(srcBox, 0, 0, tr.textureScratchFbo[0]->width, tr.textureScratchFbo[0]->height);
-		VectorSet4(dstBox, 0, 0, tr.screenScratchFbo->width,     tr.screenScratchFbo->height);
+		VectorSet4(dstBox, 0, 0, glConfig.vidWidth,              glConfig.vidHeight);
 		color[3] = factor;
 		FBO_Blit(tr.textureScratchFbo[0], srcBox, texScale, tr.screenScratchFbo, dstBox, &tr.textureColorShader, color, GLS_SRCBLEND_SRC_ALPHA | GLS_DSTBLEND_ONE_MINUS_SRC_ALPHA);
 	}