Improved SSAO performance with Blue Noise

RELEASE-NOTES.md

@@ -79,6 +79,8 @@ The main goal is that the new content looks the same in RBDOOM-3-BFG as in Blend
 
 * Improved Shadow Mapping performance by reducing the number of taps from 12 to 6 and keeping a good quality using dithering the result with Blue Noise magic by Alan Wolfe
 
+* Improved Screen Space Ambient Occlusion performance by enhancing the quality with Blue Noise and skipping the expensive extra bilateral filtering pass
+
 * Updated idRenderLog to support RenderDoc and Nvidia's Nsight and only issue OpenGL or Vulkan debug commands if the debug extensions are detected. Reference: https://devblogs.nvidia.com/best-practices-gpu-performance-events/
 
 * Artistic Style C++ beautifier configuration has slightly changed to work closer to Clang Format's behaviour

base/renderprogs/AmbientOcclusion_AO.ps.hlsl

@@ -94,6 +94,7 @@ const float projScale = 500.0;
 // *INDENT-OFF*
 uniform sampler2D samp0 : register( s0 ); // view normal/roughness
 uniform sampler2D samp1 : register( s1 ); // view depth
+uniform sampler2D samp2	: register( s2 ); // blue noise 256
 
 #define CS_Z_buffer		samp1
 
@@ -109,7 +110,22 @@ struct PS_OUT
 // *INDENT-ON*
 
 
+float BlueNoise( float2 n, float x )
+{
+	float noise = tex2D( samp2, ( n.xy / 256.0 ) ).r;
 
+#if TEMPORALLY_VARY_TAPS
+	noise = fract( noise + 0.61803398875 * rpJitterTexOffset.z * x );
+#else
+	noise = fract( noise );
+#endif
+
+	noise = RemapNoiseTriErp( noise );
+
+	//noise = noise * 2.0 - 1.0;
+
+	return noise;
+}
 
 /** Used for packing Z into the GB channels */
 // float CSZToKey( float z )
@@ -390,13 +406,20 @@ void main( PS_IN fragment, out PS_OUT result )
 	}
 #endif
 
+#if 1
+	float randomPatternRotationAngle = BlueNoise( ssP.xy, 10.0 ) * 10.0;
+	//float randomPatternRotationAngle = InterleavedGradientNoise( ssP.xy ) * 10.0;
+#else
+
 	// Hash function used in the HPG12 AlchemyAO paper
 	float randomPatternRotationAngle = float( ( ( 3 * ssP.x ) ^ ( ssP.y + ssP.x * ssP.y ) )
 #if TEMPORALLY_VARY_TAPS
-									   + rpJitterTexOffset.x
+									   + rpJitterTexOffset.z
 #endif
 											) * 10.0;
 
+#endif
+
 	// Choose the screen-space sample radius
 	// proportional to the projected area of the sphere
 	float ssDiskRadius = -projScale * radius / C.z;

base/renderprogs/interactionSM.ps.hlsl

@@ -303,7 +303,7 @@ void main( PS_IN fragment, out PS_OUT result )
 	float shadow = 0.0;
 
 	// RB: casting a float to int and using it as index can really kill the performance ...
-	float numSamples = 6.0; //int(rpScreenCorrectionFactor.w);
+	float numSamples = 6.0;
 	float stepSize = 1.0 / numSamples;
 
 	//float4 jitterTC = ( fragment.position * rpScreenCorrectionFactor ) + rpJitterTexOffset;

neo/renderer/OpenGL/RenderDebug_GL.cpp

@@ -2697,7 +2697,7 @@ void idRenderBackend::DBG_TestImage()
 	}
 
 	// Set State
-	GL_State( GLS_DEPTHFUNC_ALWAYS | GLS_SRCBLEND_ONE | GLS_DSTBLEND_ZERO );
+	GL_State( GLS_SRCBLEND_ONE | GLS_DSTBLEND_ZERO | GLS_DEPTHMASK | GLS_DEPTHFUNC_ALWAYS | GLS_CULL_TWOSIDED );
 
 	// Set Parms
 	float texS[4] = { 1.0f, 0.0f, 0.0f, 0.0f };
@@ -2708,6 +2708,7 @@ void idRenderBackend::DBG_TestImage()
 	float texGenEnabled[4] = { 0, 0, 0, 0 };
 	renderProgManager.SetRenderParm( RENDERPARM_TEXGEN_0_ENABLED, texGenEnabled );
 
+#if 1
 	// not really necessary but just for clarity
 	const float screenWidth = 1.0f;
 	const float screenHeight = 1.0f;
@@ -2716,7 +2717,7 @@ void idRenderBackend::DBG_TestImage()
 
 	float scale[16] = { 0 };
 	scale[0] = w; // scale
-	scale[5] = h; // scale
+	scale[5] = -h; // scale
 	scale[12] = halfScreenWidth - ( halfScreenWidth * w ); // translate
 	scale[13] = halfScreenHeight - ( halfScreenHeight * h ); // translate
 	scale[10] = 1.0f;
@@ -2737,11 +2738,10 @@ void idRenderBackend::DBG_TestImage()
 	float projMatrixTranspose[16];
 	R_MatrixTranspose( finalOrtho, projMatrixTranspose );
 	renderProgManager.SetRenderParms( RENDERPARM_MVPMATRIX_X, projMatrixTranspose, 4 );
-
+#else
-//	glMatrixMode( GL_PROJECTION );
+	// draw texture over entire screen
-//	glLoadMatrixf( finalOrtho );
+	RB_SetMVP( renderMatrix_identity );
-//	glMatrixMode( GL_MODELVIEW );
+#endif
-//	glLoadIdentity();
 
 	// Set Color
 	GL_Color( 1, 1, 1, 1 );
@@ -2767,6 +2767,7 @@ void idRenderBackend::DBG_TestImage()
 
 	// Draw!
 	DrawElementsWithCounters( &testImageSurface );
+	//DrawElementsWithCounters( &unitSquareSurface );
 }
 
 // RB begin

neo/renderer/RenderBackend.cpp

@@ -2068,7 +2068,7 @@ void idRenderBackend::AmbientPass( const drawSurf_t* const* drawSurfs, int numDr
 	}
 	else
 	{
-		if( r_forceAmbient.GetFloat() <= 0 || r_skipAmbient.GetBool() )
+		if( r_forceAmbient.GetFloat() <= 0 )//|| r_skipAmbient.GetBool() )
 		{
 			// clear gbuffer
 			GL_Clear( true, false, false, 0, 0.0f, 0.0f, 0.0f, 1.0f, false );
@@ -4990,19 +4990,31 @@ void idRenderBackend::DrawScreenSpaceAmbientOcclusion( const viewDef_t* _viewDef
 #endif
 	SetVertexParms( RENDERPARM_MODELMATRIX_X, viewDef->unprojectionToCameraRenderMatrix[0], 4 );
 
+	const static int BLUENOISE_SIZE = 256;
+	const float jitterSampleScale = 1.0f;
+
+	float jitterTexScale[4];
+	jitterTexScale[0] = r_shadowMapJitterScale.GetFloat() * jitterSampleScale;	// TODO shadow buffer size fraction shadowMapSize / maxShadowMapSize
+	jitterTexScale[1] = r_shadowMapJitterScale.GetFloat() * jitterSampleScale;
+	jitterTexScale[2] = -r_shadowMapBiasScale.GetFloat();
+	jitterTexScale[3] = 0.0f;
+	SetFragmentParm( RENDERPARM_JITTERTEXSCALE, jitterTexScale ); // rpJitterTexScale
+
 	float jitterTexOffset[4];
+	jitterTexOffset[0] = 1.0f / BLUENOISE_SIZE;
+	jitterTexOffset[1] = 1.0f / BLUENOISE_SIZE;
+
 	if( r_shadowMapRandomizeJitter.GetBool() )
 	{
-		jitterTexOffset[0] = ( rand() & 255 ) / 255.0;
+		jitterTexOffset[2] = Sys_Milliseconds() / 1000.0f;
-		jitterTexOffset[1] = ( rand() & 255 ) / 255.0;
+		jitterTexOffset[3] = tr.frameCount % 64;
 	}
 	else
 	{
-		jitterTexOffset[0] = 0;
+		jitterTexOffset[2] = 0.0f;
-		jitterTexOffset[1] = 0;
+		jitterTexOffset[3] = 0.0f;
 	}
-	jitterTexOffset[2] = viewDef->renderView.time[0] * 0.001f;
+
-	jitterTexOffset[3] = 0.0f;
 	SetFragmentParm( RENDERPARM_JITTERTEXOFFSET, jitterTexOffset ); // rpJitterTexOffset
 
 	GL_SelectTexture( 0 );
@@ -5018,6 +5030,9 @@ void idRenderBackend::DrawScreenSpaceAmbientOcclusion( const viewDef_t* _viewDef
 		globalImages->currentDepthImage->Bind();
 	}
 
+	GL_SelectTexture( 2 );
+	globalImages->blueNoiseImage256->Bind();
+
 	DrawElementsWithCounters( &unitSquareSurface );
 
 	if( r_ssaoFiltering.GetBool() )

neo/renderer/RenderProgs_embedded.h

@@ -2798,6 +2798,7 @@ static const cgShaderDef_t cg_renderprogs[] =
 		"// *INDENT-OFF*\n"
 		"uniform sampler2D samp0 : register( s0 ); // view normal/roughness\n"
 		"uniform sampler2D samp1 : register( s1 ); // view depth\n"
+		"uniform sampler2D samp2	: register( s2 ); // blue noise 256\n"
 		"\n"
 		"#define CS_Z_buffer		samp1\n"
 		"\n"
@@ -2813,7 +2814,22 @@ static const cgShaderDef_t cg_renderprogs[] =
 		"// *INDENT-ON*\n"
 		"\n"
 		"\n"
+		"float BlueNoise( float2 n, float x )\n"
+		"{\n"
+		"	float noise = tex2D( samp2, ( n.xy / 256.0 ) ).r;\n"
 		"\n"
+		"#if TEMPORALLY_VARY_TAPS\n"
+		"	noise = fract( noise + 0.61803398875 * rpJitterTexOffset.z * x );\n"
+		"#else\n"
+		"	noise = fract( noise );\n"
+		"#endif\n"
+		"\n"
+		"	noise = RemapNoiseTriErp( noise );\n"
+		"\n"
+		"	//noise = noise * 2.0 - 1.0;\n"
+		"\n"
+		"	return noise;\n"
+		"}\n"
 		"\n"
 		"/** Used for packing Z into the GB channels */\n"
 		"// float CSZToKey( float z )\n"
@@ -3094,13 +3110,20 @@ static const cgShaderDef_t cg_renderprogs[] =
 		"	}\n"
 		"#endif\n"
 		"\n"
+		"#if 1\n"
+		"	float randomPatternRotationAngle = BlueNoise( ssP.xy, 10.0 ) * 10.0;\n"
+		"	//float randomPatternRotationAngle = InterleavedGradientNoise( ssP.xy ) * 10.0;\n"
+		"#else\n"
+		"\n"
 		"	// Hash function used in the HPG12 AlchemyAO paper\n"
 		"	float randomPatternRotationAngle = float( ( ( 3 * ssP.x ) ^ ( ssP.y + ssP.x * ssP.y ) )\n"
 		"#if TEMPORALLY_VARY_TAPS\n"
-		"									   + rpJitterTexOffset.x\n"
+		"									   + rpJitterTexOffset.z\n"
 		"#endif\n"
 		"											) * 10.0;\n"
 		"\n"
+		"#endif\n"
+		"\n"
 		"	// Choose the screen-space sample radius\n"
 		"	// proportional to the projected area of the sphere\n"
 		"	float ssDiskRadius = -projScale * radius / C.z;\n"
@@ -10068,7 +10091,7 @@ static const cgShaderDef_t cg_renderprogs[] =
 		"	float shadow = 0.0;\n"
 		"\n"
 		"	// RB: casting a float to int and using it as index can really kill the performance ...\n"
-		"	float numSamples = 6.0; //int(rpScreenCorrectionFactor.w);\n"
+		"	float numSamples = 6.0;\n"
 		"	float stepSize = 1.0 / numSamples;\n"
 		"\n"
 		"	//float4 jitterTC = ( fragment.position * rpScreenCorrectionFactor ) + rpJitterTexOffset;\n"

neo/renderer/RenderSystem_init.cpp

@@ -293,7 +293,7 @@ idCVar r_ssgiFiltering( "r_ssgiFiltering", "1", CVAR_RENDERER | CVAR_BOOL, "" );
 
 idCVar r_useSSAO( "r_useSSAO", "1", CVAR_RENDERER | CVAR_ARCHIVE | CVAR_BOOL, "use screen space ambient occlusion to darken corners" );
 idCVar r_ssaoDebug( "r_ssaoDebug", "0", CVAR_RENDERER | CVAR_INTEGER, "" );
-idCVar r_ssaoFiltering( "r_ssaoFiltering", "1", CVAR_RENDERER | CVAR_BOOL, "" );
+idCVar r_ssaoFiltering( "r_ssaoFiltering", "0", CVAR_RENDERER | CVAR_BOOL, "" );
 idCVar r_useHierarchicalDepthBuffer( "r_useHierarchicalDepthBuffer", "1", CVAR_RENDERER | CVAR_BOOL, "" );
 
 idCVar r_usePBR( "r_usePBR", "1", CVAR_RENDERER | CVAR_ARCHIVE | CVAR_BOOL, "use PBR and Image Based Lighting instead of old Quake 4 style ambient lighting" );