OpenGL2: Reimplement soft overbright to avoid a framebuffer blit.

8 files changed, 70 insertions, 79 deletions

diff --git a/src/renderergl2/tr_backend.c b/src/renderergl2/tr_backend.c
index 559243ca..b9cbedd9 100644
--- a/src/renderergl2/tr_backend.c
+++ b/src/renderergl2/tr_backend.c
@@ -450,7 +450,7 @@ void RB_BeginDrawingView (void) {
 		{
 			if (!tr.renderFbo || (backEnd.framePostProcessed && (backEnd.refdef.rdflags & RDF_NOWORLDMODEL)))
 			{
-				FBO_Bind(tr.screenScratchFbo);
+				FBO_Bind(NULL);
 			}
 			else
 			{
@@ -860,7 +860,7 @@ void RE_StretchRaw (int x, int y, int w, int h, int cols, int rows, const byte *
 	{
 		if (!tr.renderFbo || backEnd.framePostProcessed)
 		{
-			FBO_Bind(tr.screenScratchFbo);
+			FBO_Bind(NULL);
 		}
 		else
 		{
@@ -947,7 +947,7 @@ const void *RB_StretchPic ( const void *data ) {
 	{
 		if (!tr.renderFbo || backEnd.framePostProcessed)
 		{
-			FBO_Bind(tr.screenScratchFbo);
+			FBO_Bind(NULL);
 		}
 		else
 		{
@@ -1444,7 +1444,7 @@ const void *RB_ClearDepth(const void *data)
 	{
 		if (!tr.renderFbo || backEnd.framePostProcessed)
 		{
-			FBO_Bind(tr.screenScratchFbo);
+			FBO_Bind(NULL);
 		}
 		else
 		{
@@ -1513,32 +1513,13 @@ const void	*RB_SwapBuffers( const void *data ) {
 			{
 				// 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);
+				FBO_FastBlit(tr.msaaResolveFbo, NULL, NULL, 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);
+				FBO_FastBlit(tr.renderFbo, NULL, NULL, NULL, GL_COLOR_BUFFER_BIT, GL_NEAREST);
 			}
 		}
-
-		if (tr.screenScratchFbo)
-		{
-			vec4_t color;
-
-			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 ) {
@@ -1655,11 +1636,11 @@ const void *RB_PostProcess(const void *data)
 		if (r_hdr->integer && (r_toneMap->integer || r_forceToneMap->integer))
 		{
 			autoExposure = r_autoExposure->integer || r_forceAutoExposure->integer;
-			RB_ToneMap(srcFbo, srcBox, tr.screenScratchFbo, dstBox, autoExposure);
+			RB_ToneMap(srcFbo, srcBox, NULL, dstBox, autoExposure);
 		}
 		else if (r_cameraExposure->value == 0.0f)
 		{
-			FBO_FastBlit(srcFbo, srcBox, tr.screenScratchFbo, dstBox, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+			FBO_FastBlit(srcFbo, srcBox, NULL, dstBox, GL_COLOR_BUFFER_BIT, GL_NEAREST);
 		}
 		else
 		{
@@ -1670,15 +1651,15 @@ const void *RB_PostProcess(const void *data)
 			color[2] = pow(2, r_cameraExposure->value); //exp2(r_cameraExposure->value);
 			color[3] = 1.0f;
 
-			FBO_Blit(srcFbo, srcBox, NULL, tr.screenScratchFbo, dstBox, NULL, color, 0);
+			FBO_Blit(srcFbo, srcBox, NULL, NULL, dstBox, NULL, color, 0);
 		}
 	}
 
 	if (r_drawSunRays->integer)
-		RB_SunRays(tr.screenScratchFbo, srcBox, tr.screenScratchFbo, dstBox);
+		RB_SunRays(NULL, srcBox, NULL, dstBox);
 
 	if (1)
-		RB_BokehBlur(tr.screenScratchFbo, srcBox, tr.screenScratchFbo, dstBox, backEnd.refdef.blurFactor);
+		RB_BokehBlur(NULL, srcBox, NULL, dstBox, backEnd.refdef.blurFactor);
 	else
 		RB_GaussianBlur(backEnd.refdef.blurFactor);
 
@@ -1686,27 +1667,27 @@ const void *RB_PostProcess(const void *data)
 	{
 		vec4i_t dstBox;
 		VectorSet4(dstBox, 0, 0, 128, 128);
-		FBO_BlitFromTexture(tr.sunShadowDepthImage[0], NULL, NULL, tr.screenScratchFbo, dstBox, NULL, NULL, 0);
+		FBO_BlitFromTexture(tr.sunShadowDepthImage[0], NULL, NULL, NULL, dstBox, NULL, NULL, 0);
 		VectorSet4(dstBox, 128, 0, 128, 128);
-		FBO_BlitFromTexture(tr.sunShadowDepthImage[1], NULL, NULL, tr.screenScratchFbo, dstBox, NULL, NULL, 0);
+		FBO_BlitFromTexture(tr.sunShadowDepthImage[1], NULL, NULL, NULL, dstBox, NULL, NULL, 0);
 		VectorSet4(dstBox, 256, 0, 128, 128);
-		FBO_BlitFromTexture(tr.sunShadowDepthImage[2], NULL, NULL, tr.screenScratchFbo, dstBox, NULL, NULL, 0);
+		FBO_BlitFromTexture(tr.sunShadowDepthImage[2], NULL, NULL, NULL, dstBox, NULL, NULL, 0);
 	}
 
 	if (0)
 	{
 		vec4i_t dstBox;
 		VectorSet4(dstBox, 256, glConfig.vidHeight - 256, 256, 256);
-		FBO_BlitFromTexture(tr.renderDepthImage, NULL, NULL, tr.screenScratchFbo, dstBox, NULL, NULL, 0);
+		FBO_BlitFromTexture(tr.renderDepthImage, NULL, NULL, NULL, dstBox, NULL, NULL, 0);
 		VectorSet4(dstBox, 512, glConfig.vidHeight - 256, 256, 256);
-		FBO_BlitFromTexture(tr.screenShadowImage, NULL, NULL, tr.screenScratchFbo, dstBox, NULL, NULL, 0);
+		FBO_BlitFromTexture(tr.screenShadowImage, NULL, NULL, NULL, dstBox, NULL, NULL, 0);
 	}
 
 	if (0)
 	{
 		vec4i_t dstBox;
 		VectorSet4(dstBox, 256, glConfig.vidHeight - 256, 256, 256);
-		FBO_BlitFromTexture(tr.sunRaysImage, NULL, NULL, tr.screenScratchFbo, dstBox, NULL, NULL, 0);
+		FBO_BlitFromTexture(tr.sunRaysImage, NULL, NULL, NULL, dstBox, NULL, NULL, 0);
 	}
 
 #if 0
@@ -1718,8 +1699,8 @@ const void *RB_PostProcess(const void *data)
 		if (cubemapIndex)
 		{
 			VectorSet4(dstBox, 0, glConfig.vidHeight - 256, 256, 256);
-			//FBO_BlitFromTexture(tr.renderCubeImage, NULL, NULL, tr.screenScratchFbo, dstBox, &tr.testcubeShader, NULL, 0);
-			FBO_BlitFromTexture(tr.cubemaps[cubemapIndex - 1], NULL, NULL, tr.screenScratchFbo, dstBox, &tr.testcubeShader, NULL, 0);
+			//FBO_BlitFromTexture(tr.renderCubeImage, NULL, NULL, NULL, dstBox, &tr.testcubeShader, NULL, 0);
+			FBO_BlitFromTexture(tr.cubemaps[cubemapIndex - 1], NULL, NULL, NULL, dstBox, &tr.testcubeShader, NULL, 0);
 		}
 	}
 #endif
diff --git a/src/renderergl2/tr_cmds.c b/src/renderergl2/tr_cmds.c
index e6741ee3..540b2a1d 100644
--- a/src/renderergl2/tr_cmds.c
+++ b/src/renderergl2/tr_cmds.c
@@ -528,12 +528,6 @@ void RE_BeginFrame( stereoFrame_t stereoFrame ) {
 						qglClear(GL_COLOR_BUFFER_BIT);
 					}
 
-					if (tr.screenScratchFbo)
-					{
-						FBO_Bind(tr.screenScratchFbo);
-						qglClear(GL_COLOR_BUFFER_BIT);
-					}
-
 					FBO_Bind(NULL);
 				}
 
diff --git a/src/renderergl2/tr_fbo.c b/src/renderergl2/tr_fbo.c
index a1d1339b..d9fd917f 100644
--- a/src/renderergl2/tr_fbo.c
+++ b/src/renderergl2/tr_fbo.c
@@ -526,22 +526,6 @@ 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);
-		FBO_Bind(tr.screenScratchFbo);
-		
-		//FBO_CreateBuffer(tr.screenScratchFbo, format, 0, 0);
-		FBO_AttachTextureImage(tr.screenScratchImage, 0);
-
-		// FIXME: hack: share zbuffer between render fbo and pre-screen fbo
-		//FBO_CreateBuffer(tr.screenScratchFbo, GL_DEPTH_COMPONENT24_ARB, 0, 0);
-		R_AttachFBOTextureDepth(tr.renderDepthImage->texnum);
-
-		R_CheckFBO(tr.screenScratchFbo);
-	}
-
 	for (i = 0; i < 2; i++)
 	{
 		tr.quarterFbo[i] = FBO_Create(va("_quarter%d", i), tr.quarterImage[i]->width, tr.quarterImage[i]->height);
diff --git a/src/renderergl2/tr_image.c b/src/renderergl2/tr_image.c
index 2467aac8..2b5798ac 100644
--- a/src/renderergl2/tr_image.c
+++ b/src/renderergl2/tr_image.c
@@ -2932,9 +2932,6 @@ void R_CreateBuiltinImages( void ) {
 		if (r_drawSunRays->integer)
 			tr.sunRaysImage = R_CreateImage("*sunRays", NULL, width, height, IMGTYPE_COLORALPHA, IMGFLAG_NO_COMPRESSION | IMGFLAG_CLAMPTOEDGE, rgbFormat);
 
-		if (r_softOverbright->integer)
-			tr.screenScratchImage = R_CreateImage("*screenScratch", NULL, width, height, IMGTYPE_COLORALPHA, IMGFLAG_NO_COMPRESSION | IMGFLAG_CLAMPTOEDGE, rgbFormat);
-
 		if (glRefConfig.framebufferObject)
 		{
 			tr.renderDepthImage  = R_CreateImage("*renderdepth",  NULL, width, height, IMGTYPE_COLORALPHA, IMGFLAG_NO_COMPRESSION | IMGFLAG_CLAMPTOEDGE, GL_DEPTH_COMPONENT24_ARB);
diff --git a/src/renderergl2/tr_local.h b/src/renderergl2/tr_local.h
index da3edba2..6f56f139 100644
--- a/src/renderergl2/tr_local.h
+++ b/src/renderergl2/tr_local.h
@@ -1621,7 +1621,6 @@ typedef struct {
 	image_t					*renderDepthImage;
 	image_t					*pshadowMaps[MAX_DRAWN_PSHADOWS];
 	image_t					*textureScratchImage[2];
-	image_t					*screenScratchImage;
 	image_t                 *quarterImage[2];
 	image_t					*calcLevelsImage;
 	image_t					*targetLevelsImage;
@@ -1640,7 +1639,6 @@ typedef struct {
 	FBO_t					*depthFbo;
 	FBO_t					*pshadowFbos[MAX_DRAWN_PSHADOWS];
 	FBO_t					*textureScratchFbo[2];
-	FBO_t					*screenScratchFbo;
 	FBO_t                   *quarterFbo[2];
 	FBO_t					*calcLevelsFbo;
 	FBO_t					*targetLevelsFbo;
diff --git a/src/renderergl2/tr_postprocess.c b/src/renderergl2/tr_postprocess.c
index cab9d42b..dba19d4a 100644
--- a/src/renderergl2/tr_postprocess.c
+++ b/src/renderergl2/tr_postprocess.c
@@ -255,10 +255,20 @@ static void RB_RadialBlur(FBO_t *srcFbo, FBO_t *dstFbo, int passes, float stretc
 			float s1 = iscale + s0;
 			float t1 = iscale + t0;
 
-			srcBox[0] = s0 * srcFbo->width;
-			srcBox[1] = t0 * srcFbo->height;
-			srcBox[2] = (s1 - s0) * srcFbo->width;
-			srcBox[3] = (t1 - t0) * srcFbo->height;
+			if (srcFbo)
+			{
+				srcBox[0] = s0 * srcFbo->width;
+				srcBox[1] = t0 * srcFbo->height;
+				srcBox[2] = (s1 - s0) * srcFbo->width;
+				srcBox[3] = (t1 - t0) * srcFbo->height;
+			}
+			else
+			{
+				srcBox[0] = s0 * glConfig.vidWidth;
+				srcBox[1] = t0 * glConfig.vidHeight;
+				srcBox[2] = (s1 - s0) * glConfig.vidWidth;
+				srcBox[3] = (t1 - t0) * glConfig.vidHeight;
+			}
 			
 			FBO_Blit(srcFbo, srcBox, texScale, dstFbo, dstBox, &tr.textureColorShader, color, GLS_SRCBLEND_ONE | GLS_DSTBLEND_ONE );
 
@@ -351,10 +361,20 @@ void RB_SunRays(FBO_t *srcFbo, vec4i_t srcBox, FBO_t *dstFbo, vec4i_t dstBox)
 
 		VectorSet4(color, mul, mul, mul, 1);
 
-		rayBox[0] = srcBox[0] * tr.sunRaysFbo->width  / srcFbo->width;
-		rayBox[1] = srcBox[1] * tr.sunRaysFbo->height / srcFbo->height;
-		rayBox[2] = srcBox[2] * tr.sunRaysFbo->width  / srcFbo->width;
-		rayBox[3] = srcBox[3] * tr.sunRaysFbo->height / srcFbo->height;
+		if (srcFbo)
+		{
+			rayBox[0] = srcBox[0] * tr.sunRaysFbo->width  / srcFbo->width;
+			rayBox[1] = srcBox[1] * tr.sunRaysFbo->height / srcFbo->height;
+			rayBox[2] = srcBox[2] * tr.sunRaysFbo->width  / srcFbo->width;
+			rayBox[3] = srcBox[3] * tr.sunRaysFbo->height / srcFbo->height;
+		}
+		else
+		{
+			rayBox[0] = srcBox[0] * tr.sunRaysFbo->width  / glConfig.vidWidth;
+			rayBox[1] = srcBox[1] * tr.sunRaysFbo->height / glConfig.vidHeight;
+			rayBox[2] = srcBox[2] * tr.sunRaysFbo->width  / glConfig.vidWidth;
+			rayBox[3] = srcBox[3] * tr.sunRaysFbo->height / glConfig.vidHeight;
+		}
 
 		quarterBox[0] = 0;
 		quarterBox[1] = tr.quarterFbo[0]->height;
@@ -480,7 +500,7 @@ void RB_GaussianBlur(float blur)
 		VectorSet4(color, 1, 1, 1, 1);
 
 		// first, downsample the framebuffer
-		FBO_FastBlit(tr.screenScratchFbo, NULL, tr.quarterFbo[0], NULL, GL_COLOR_BUFFER_BIT, GL_LINEAR);
+		FBO_FastBlit(NULL, 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
@@ -498,6 +518,6 @@ void RB_GaussianBlur(float blur)
 		VectorSet4(srcBox, 0, 0, tr.textureScratchFbo[0]->width, tr.textureScratchFbo[0]->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);
+		FBO_Blit(tr.textureScratchFbo[0], srcBox, texScale, NULL, dstBox, &tr.textureColorShader, color, GLS_SRCBLEND_SRC_ALPHA | GLS_DSTBLEND_ONE_MINUS_SRC_ALPHA);
 	}
 }
diff --git a/src/renderergl2/tr_shade.c b/src/renderergl2/tr_shade.c
index d6837b2a..a23ae119 100644
--- a/src/renderergl2/tr_shade.c
+++ b/src/renderergl2/tr_shade.c
@@ -444,7 +444,7 @@ static void ProjectDlightTexture( void ) {
 }
 
 
-static void ComputeShaderColors( shaderStage_t *pStage, vec4_t baseColor, vec4_t vertColor )
+static void ComputeShaderColors( shaderStage_t *pStage, vec4_t baseColor, vec4_t vertColor, int blend )
 {
 	baseColor[0] = 
 	baseColor[1] =
@@ -602,6 +602,23 @@ static void ComputeShaderColors( shaderStage_t *pStage, vec4_t baseColor, vec4_t
 			break;
 	}
 
+	// multiply color by overbrightbits if this isn't a blend
+	if (r_softOverbright->integer && tr.overbrightBits 
+	 && !((blend & GLS_SRCBLEND_BITS) == GLS_SRCBLEND_DST_COLOR)
+	 && !((blend & GLS_SRCBLEND_BITS) == GLS_SRCBLEND_ONE_MINUS_DST_COLOR)
+	 && !((blend & GLS_DSTBLEND_BITS) == GLS_DSTBLEND_SRC_COLOR)
+	 && !((blend & GLS_DSTBLEND_BITS) == GLS_DSTBLEND_ONE_MINUS_SRC_COLOR))
+	{
+		float scale = 1 << tr.overbrightBits;
+
+		baseColor[0] *= scale;
+		baseColor[1] *= scale;
+		baseColor[2] *= scale;
+		vertColor[0] *= scale;
+		vertColor[1] *= scale;
+		vertColor[2] *= scale;
+	}
+
 	// FIXME: find some way to implement this.
 #if 0
 	// if in greyscale rendering mode turn all color values into greyscale.
@@ -770,7 +787,7 @@ static void ForwardDlight( void ) {
 			vec4_t baseColor;
 			vec4_t vertColor;
 
-			ComputeShaderColors(pStage, baseColor, vertColor);
+			ComputeShaderColors(pStage, baseColor, vertColor, GLS_SRCBLEND_ONE | GLS_DSTBLEND_ONE);
 
 			GLSL_SetUniformVec4(sp, UNIFORM_BASECOLOR, baseColor);
 			GLSL_SetUniformVec4(sp, UNIFORM_VERTCOLOR, vertColor);
@@ -1142,7 +1159,7 @@ static void RB_IterateStagesGeneric( shaderCommands_t *input )
 			vec4_t baseColor;
 			vec4_t vertColor;
 
-			ComputeShaderColors(pStage, baseColor, vertColor);
+			ComputeShaderColors(pStage, baseColor, vertColor, pStage->stateBits);
 
 			if ((backEnd.refdef.colorScale != 1.0f) && !(backEnd.refdef.rdflags & RDF_NOWORLDMODEL))
 			{
diff --git a/src/renderergl2/tr_sky.c b/src/renderergl2/tr_sky.c
index 7c15f732..cb3b9434 100644
--- a/src/renderergl2/tr_sky.c
+++ b/src/renderergl2/tr_sky.c
@@ -450,7 +450,7 @@ static void DrawSkySide( struct image_s *image, const int mins[2], const int max
 		
 		color[0] = 
 		color[1] = 
-		color[2] = tr.identityLight * backEnd.refdef.colorScale;
+		color[2] = (r_softOverbright->integer ? 1.0 : tr.identityLight) * backEnd.refdef.colorScale;
 		color[3] = 1.0f;
 		GLSL_SetUniformVec4(sp, UNIFORM_BASECOLOR, color);