NRD v3.4.2:

HIGHLIGHTS:
- bug fixes & improvements

DETAILS:
- REBLUR: fixed broken resource bindings in REBLUR_DIRECTIONAL_OCCLUSION denoiser
- REBLUR: tweaked plane distance sensitivity after history reset to minimize overblur
- REBLUR: tuned blur radius in spatial passes if number of accumulated frames is low (gives IQ and performance)
- REBLUR: switched anti-firefly to 3x3 instead of 5x5
- REBLUR: anti-firefly and fast history clamping improvements

Include/NRD.h

@@ -29,8 +29,8 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 #define NRD_VERSION_MAJOR 3
 #define NRD_VERSION_MINOR 4
-#define NRD_VERSION_BUILD 1
-#define NRD_VERSION_DATE "17 August 2022"
+#define NRD_VERSION_BUILD 2
+#define NRD_VERSION_DATE "18 August 2022"
 
 #if defined(_MSC_VER)
     #define NRD_CALL __fastcall

README.md

@@ -1,4 +1,4 @@
-# NVIDIA Real-time Denoisers v3.4.1 (NRD)
+# NVIDIA Real-time Denoisers v3.4.2 (NRD)
 
 ## SAMPLE APP
 

Resources/Version.h

@@ -23,7 +23,7 @@ Versioning rules:
 
 #define VERSION_MAJOR                   3
 #define VERSION_MINOR                   4
-#define VERSION_BUILD                   1
+#define VERSION_BUILD                   2
 #define VERSION_REVISION                0
 
 #define VERSION_STRING STR(VERSION_MAJOR.VERSION_MINOR.VERSION_BUILD.VERSION_REVISION)

Shaders/Include/Common.hlsli

@@ -346,9 +346,10 @@ float2 GetKernelSampleCoordinates( float4x4 mToClip, float3 offset, float3 X, fl
 
 // Weight parameters
 
-float2 GetGeometryWeightParams( float planeDistSensitivity, float frustumHeight, float3 Xv, float3 Nv, float scale = 1.0 )
+float2 GetGeometryWeightParams( float planeDistSensitivity, float frustumHeight, float3 Xv, float3 Nv, float nonLinearAccumSpeed )
 {
-    float a = scale / ( planeDistSensitivity * frustumHeight + 1e-6 );
+    float relaxation = lerp( 1.0, 0.25, nonLinearAccumSpeed );
+    float a = relaxation / ( planeDistSensitivity * frustumHeight + 1e-6 );
     float b = -dot( Nv, Xv ) * a;
 
     return float2( a, b );

Shaders/Include/REBLUR/REBLUR_Common.hlsli

@@ -103,10 +103,10 @@ float4 UnpackInternalData1( float4 p )
 float4 PackInternalData2( float fbits, float curvature, float virtualHistoryAmount, float hitDistScaleForTracking, float viewZ )
 {
     // BITS:
-    // 0        - free
+    // 0        - free // TODO: can be used to store "skip HistoryFix" bit
     // 1        - CatRom flag
     // 2,3,4,5  - occlusion 2x2
-    // 6        - free
+    // 6        - free // TODO: free!
     // 7        - curvature sign
 
     float pixelSize = PixelRadiusToWorld( gUnproject, gOrthoMode, 1.0, viewZ );
@@ -436,8 +436,11 @@ float GetBlurRadius(
     float r = lerp( gMinConvergedStateBaseRadiusScale, 1.0, nonLinearAccumSpeed );
 
     // Avoid over-blurring on contact ( tests 76, 95, 120 )
-    hitDistFactor = lerp( hitDistFactor, 1.0, nonLinearAccumSpeed );
-    r *= hitDistFactor; // TODO: reduce influence if reprojection confidence is low?
+    // TODO: reduce "hitDistFactor" influence if reprojection confidence is low?
+    // TODO: if luminance stoppers are used, blur radius should depend less on "hitDistFactor"
+    float relaxedHitDistFactor = lerp( 1.0, hitDistFactor, roughness );
+    hitDistFactor = lerp( hitDistFactor, relaxedHitDistFactor, nonLinearAccumSpeed );
+    r *= hitDistFactor;
 
     // IMPORTANT: do not apply "hitDistFactor" to radiusBias because it is responsible for error compensation
 
@@ -451,7 +454,7 @@ float GetBlurRadius(
     // Disable spatial filtering if radius is 0
     r *= float( radius != 0 );
 
-    return r; // TODO: if luminance stoppers are used, blur radius should depend less on hitDistFactor
+    return r;
 }
 
 float GetBlurRadiusScaleBasingOnTrimming( float roughness, float3 trimmingParams )

Shaders/Include/REBLUR/REBLUR_Common_DiffuseSpatialFilter.hlsli

@@ -31,9 +31,11 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 #else
         float2 diffInternalData = float2( 1.0 / ( 1.0 + internalData1.x ), internalData1.x );
         float2 sum = 1.0;
+
         float boost = saturate( 1.0 - diffInternalData.y / REBLUR_FIXED_FRAME_NUM );
-        float radius = gBlurRadius;
-        radius *= ( 1.0 + 2.0 * boost ) / 3.0;
+        boost *= NoV;
+
+        float radius = gBlurRadius * ( 1.0 + 2.0 * boost ) / 3.0;
 #endif
 
     #if( REBLUR_SPATIAL_MODE == REBLUR_PRE_BLUR )
@@ -51,8 +53,6 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
     #endif
 
         // Blur radius
-        float3 Vv = GetViewVector( Xv, true );
-
         float frustumHeight = PixelRadiusToWorld( gUnproject, gOrthoMode, gRectSize.y, viewZ );
         float hitDistScale = _REBLUR_GetHitDistanceNormalization( viewZ, gHitDistParams, 1.0 );
         float hitDist = diff.w * hitDistScale;
@@ -70,7 +70,7 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 
         // Denoising
         float2x3 TvBv = GetKernelBasis( Nv, Nv, 1.0, worldBlurRadius ); // D = N
-        float2 geometryWeightParams = GetGeometryWeightParams( gPlaneDistSensitivity, frustumHeight, Xv, Nv, lerp( 1.0, REBLUR_PLANE_DIST_MIN_SENSITIVITY_SCALE, diffInternalData.x ) );
+        float2 geometryWeightParams = GetGeometryWeightParams( gPlaneDistSensitivity, frustumHeight, Xv, Nv, diffInternalData.x );
         float normalWeightParams = GetNormalWeightParams( diffInternalData.x, lobeAngleFractionScale );
         float2 hitDistanceWeightParams = GetHitDistanceWeightParams( diff.w, diffInternalData.x );
 

Shaders/Include/REBLUR/REBLUR_Common_SpecularSpatialFilter.hlsli

@@ -32,9 +32,12 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 #else
         float2 specInternalData = float2( 1.0 / ( 1.0 + internalData1.z ), internalData1.z );
         float2 sum = 1.0;
+
         float boost = saturate( 1.0 - specInternalData.y / REBLUR_FIXED_FRAME_NUM );
-        float radius = gBlurRadius;
-        radius *= ( 1.0 + 2.0 * boost * GetSpecMagicCurve2( roughness ) ) / 3.0;
+        boost *= NoV;
+        boost *= GetSpecMagicCurve2( roughness );
+
+        float radius = gBlurRadius * ( 1.0 + 2.0 * boost ) / 3.0;
 #endif
         radius *= GetBlurRadiusScaleBasingOnTrimming( roughness, gSpecLobeTrimmingParams.xyz );
 
@@ -57,7 +60,6 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
     #endif
 
         // Blur radius
-        float3 Vv = GetViewVector( Xv, true );
         float4 Dv = STL::ImportanceSampling::GetSpecularDominantDirection( Nv, Vv, roughness, STL_SPECULAR_DOMINANT_DIRECTION_G2 );
         float NoD = abs( dot( Nv, Dv.xyz ) );
 
@@ -84,7 +86,7 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 
         // Denoising
         float2x3 TvBv = GetKernelBasis( Dv.xyz, Nv, NoD, worldBlurRadius, roughness, specInternalData.x );
-        float2 geometryWeightParams = GetGeometryWeightParams( gPlaneDistSensitivity, frustumHeight, Xv, Nv, lerp( 1.0, REBLUR_PLANE_DIST_MIN_SENSITIVITY_SCALE, specInternalData.x ) );
+        float2 geometryWeightParams = GetGeometryWeightParams( gPlaneDistSensitivity, frustumHeight, Xv, Nv, specInternalData.x );
         float normalWeightParams = GetNormalWeightParams( specInternalData.x, lobeAngleFractionScale, roughness );
         float2 hitDistanceWeightParams = GetHitDistanceWeightParams( spec.w, specInternalData.x, roughness );
         float2 roughnessWeightParams = GetRoughnessWeightParams( roughness, roughnessFractionScale );
@@ -98,7 +100,6 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
         float3 L = dirPdf.xyz;
         float3 V = STL::Geometry::RotateVector( gViewToWorld, Vv );
         float3 H = normalize( L + V );
-        float NoV = abs( dot( N, V ) );
         float NoL = saturate( dot( N, L ) );
         float NoH = saturate( dot( N, H ) );
         float VoH = saturate( dot( V, H ) );

Shaders/Include/REBLUR/REBLUR_Config.hlsli

@@ -18,9 +18,9 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 #define REBLUR_USE_HISTORY_FIX                                  1
 #define REBLUR_USE_YCOCG                                        1
 #define REBLUR_USE_FAST_HISTORY                                 1
-#define REBLUR_USE_5X5_ANTI_FIREFLY                             1 // can be 0 for clean signals
 
 // Switches ( default 0 )
+#define REBLUR_USE_5X5_ANTI_FIREFLY                             0 // can be 1 for dirty signals, but will be less useful
 #define REBLUR_USE_SCREEN_SPACE_SAMPLING                        0
 #define REBLUR_USE_ANTILAG_NOT_INVOKING_HISTORY_FIX             0
 #define REBLUR_USE_ACCUM_SPEED_NONLINEAR_INTERPOLATION          0
@@ -70,10 +70,9 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 #define REBLUR_POST_BLUR_FRACTION_SCALE                         0.5
 
 #define REBLUR_VIRTUAL_MOTION_NORMAL_WEIGHT_ITERATION_NUM       2
-#define REBLUR_COLOR_CLAMPING_SIGMA_SCALE                       2.0
+#define REBLUR_COLOR_CLAMPING_SIGMA_SCALE                       1.5 // TODO: was 2.0, but we can use even 1.0 because the fast history is noisy, while the main history is denoised
 #define REBLUR_SPEC_ACCUM_BASE_POWER                            ( 0.4 + 0.2 * exp2( -gFramerateScale ) ) // bigger values = more aggressive rejection
 #define REBLUR_SPEC_ACCUM_CURVE                                 ( 1.0 - exp2( -gFramerateScale ) ) // smaller values = more aggressive rejection
-#define REBLUR_PLANE_DIST_MIN_SENSITIVITY_SCALE                 0.05
 #define REBLUR_TS_SIGMA_AMPLITUDE                               ( 3.0 * gFramerateScale )
 #define REBLUR_TS_ACCUM_TIME                                    ( gFramerateScale * 30.0 * 0.5 ) // = FPS * seconds
 #define REBLUR_PARALLAX_SCALE                                   ( 2.0 * gFramerateScale ) // TODO: is it possible to use 1 with tweaks in other parameters?

Shaders/Include/REBLUR/REBLUR_DiffuseSpecular_Blur.hlsli

@@ -54,6 +54,9 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
     float3 Xv = STL::Geometry::ReconstructViewPosition( pixelUv, gFrustum, viewZ, gOrthoMode );
     float4 rotator = GetBlurKernelRotation( REBLUR_BLUR_ROTATOR_MODE, pixelPos, gRotator, gFrameIndex );
 
+    float3 Vv = GetViewVector( Xv, true );
+    float NoV = abs( dot( Nv, Vv ) );
+
     // Spatial filtering
     #define REBLUR_SPATIAL_MODE REBLUR_BLUR
 

Shaders/Include/REBLUR/REBLUR_DiffuseSpecular_HistoryFix.hlsli

@@ -95,14 +95,16 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
 
     #ifndef REBLUR_PERFORMANCE_MODE
         #ifdef REBLUR_DIFFUSE
-            float diffM1 = s_DiffLuma[ smemPos.y ][ smemPos.x ];
+            float diffCenter = s_DiffLuma[ smemPos.y ][ smemPos.x ];
+            float diffM1 = diffCenter;
             float diffM2 = diffM1 * diffM1;
             float diffMax = -NRD_INF;
             float diffMin = NRD_INF;
         #endif
 
         #ifdef REBLUR_SPECULAR
-            float specM1 = s_SpecLuma[ smemPos.y ][ smemPos.x ];
+            float specCenter = s_SpecLuma[ smemPos.y ][ smemPos.x ];
+            float specM1 = specCenter;
             float specM2 = specM1 * specM1;
             float specMax = -NRD_INF;
             float specMin = NRD_INF;
@@ -153,19 +155,23 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
         #ifdef REBLUR_DIFFUSE
             float diffLuma = GetLuma( diff );
 
+            // Anti-firefly
+            float diffLumaClamped1 = clamp( diffLuma, diffMin, diffMax );
+            diffLuma = gAntiFirefly ? diffLumaClamped1 : diffLuma;
+
             // Fast history
             #if( REBLUR_USE_FAST_HISTORY == 1 && !defined( REBLUR_OCCLUSION ) )
                 diffM1 *= invSum;
                 diffM2 *= invSum;
                 float diffSigma = GetStdDev( diffM1, diffM2 ) * REBLUR_COLOR_CLAMPING_SIGMA_SCALE;
 
-                float diffLumaClamped1 = STL::Color::Clamp( diffM1, diffSigma, diffLuma );
-                diffLuma = lerp( diffLumaClamped1, diffLuma, 1.0 / ( 1.0 + useFastHistory * frameNumUnclamped.x ) );
-            #endif
+                // Seems that extending clamping range by the center helps to minimize potential bias
+                diffMin = min( diffM1 - diffSigma, diffCenter );
+                diffMax = max( diffM1 + diffSigma, diffCenter );
 
-            // Anti-firefly
-            float diffLumaClamped2 = clamp( diffLuma, diffMin, diffMax );
-            diffLuma = lerp( diffLumaClamped2, diffLuma, 1.0 / ( 1.0 + gAntiFirefly * frameNumUnclamped.x ) );
+                float diffLumaClamped2 = clamp( diffLuma, diffMin, diffMax );
+                diffLuma = lerp( diffLumaClamped2, diffLuma, 1.0 / ( 1.0 + useFastHistory * frameNumUnclamped.x ) );
+            #endif
 
             // Change luma
             diff = ChangeLuma( diff, diffLuma );
@@ -178,19 +184,23 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
         #ifdef REBLUR_SPECULAR
             float specLuma = GetLuma( spec );
 
+            // Anti-firefly
+            float specLumaClamped1 = clamp( specLuma, specMin, specMax );
+            specLuma = gAntiFirefly ? specLumaClamped1 : specLuma;
+
             // Fast history
             #if( REBLUR_USE_FAST_HISTORY == 1 && !defined( REBLUR_OCCLUSION ) )
                 specM1 *= invSum;
                 specM2 *= invSum;
                 float specSigma = GetStdDev( specM1, specM2 ) * REBLUR_COLOR_CLAMPING_SIGMA_SCALE;
 
-                float specLumaClamped1 = STL::Color::Clamp( specM1, specSigma, specLuma );
-                specLuma = lerp( specLumaClamped1, specLuma, 1.0 / ( 1.0 + useFastHistory * frameNumUnclamped.y ) );
-            #endif
+                // Seems that extending clamping range by the center helps to minimize potential bias
+                specMin = min( specM1 - specSigma, specCenter );
+                specMax = max( specM1 + specSigma, specCenter );
 
-            // Anti-firefly
-            float specLumaClamped2 = clamp( specLuma, specMin, specMax );
-            specLuma = lerp( specLumaClamped2, specLuma, 1.0 / ( 1.0 + gAntiFirefly * frameNumUnclamped.y ) );
+                float specLumaClamped2 = clamp( specLuma, specMin, specMax );
+                specLuma = lerp( specLumaClamped2, specLuma, 1.0 / ( 1.0 + useFastHistory * frameNumUnclamped.y ) );
+            #endif
 
             // Change luma
             spec = ChangeLuma( spec, specLuma );
@@ -271,7 +281,7 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
 
         #ifdef REBLUR_DIFFUSE
             float diffNonLinearAccumSpeed = 1.0 / ( 1.0 + frameNumUnclamped.x );
-            float2 diffGeometryWeightParams = GetGeometryWeightParams( gPlaneDistSensitivity, frustumHeight, Xv, Nv, lerp( 1.0, REBLUR_PLANE_DIST_MIN_SENSITIVITY_SCALE, diffNonLinearAccumSpeed ) );
+            float2 diffGeometryWeightParams = GetGeometryWeightParams( gPlaneDistSensitivity, frustumHeight, Xv, Nv, diffNonLinearAccumSpeed );
             float diffNormalWeightParam = GetNormalWeightParams( diffNonLinearAccumSpeed, 1.0 );
         #endif
 
@@ -295,7 +305,7 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
             #endif
             float specNormalWeightParam = rcp( angle * specNonLinearAccumSpeed + NRD_NORMAL_ENCODING_ERROR );
 
-            float2 specGeometryWeightParams = GetGeometryWeightParams( gPlaneDistSensitivity, frustumHeight, Xv, Nv, lerp( 1.0, REBLUR_PLANE_DIST_MIN_SENSITIVITY_SCALE, specNonLinearAccumSpeed ) );
+            float2 specGeometryWeightParams = GetGeometryWeightParams( gPlaneDistSensitivity, frustumHeight, Xv, Nv, specNonLinearAccumSpeed );
             float2 specRoughnessWeightParams = GetRoughnessWeightParams( roughness, 1.0 );
             float2 specHitDistanceWeightParams = GetHitDistanceWeightParams( ExtractHitDist( spec ), specNonLinearAccumSpeed, roughness );
 

Shaders/Include/REBLUR/REBLUR_DiffuseSpecular_PostBlur.hlsli

@@ -59,6 +59,9 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
     float3 Xv = STL::Geometry::ReconstructViewPosition( pixelUv, gFrustum, viewZ, gOrthoMode );
     float4 rotator = GetBlurKernelRotation( REBLUR_POST_BLUR_ROTATOR_MODE, pixelPos, gRotator, gFrameIndex );
 
+    float3 Vv = GetViewVector( Xv, true );
+    float NoV = abs( dot( Nv, Vv ) );
+
     // Spatial filtering
     #define REBLUR_SPATIAL_MODE REBLUR_POST_BLUR
 

Shaders/Include/REBLUR/REBLUR_DiffuseSpecular_PrePass.hlsli

@@ -46,6 +46,9 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
     float3 Xv = STL::Geometry::ReconstructViewPosition( pixelUv, gFrustum, viewZ, gOrthoMode );
     float4 rotator = GetBlurKernelRotation( REBLUR_PRE_BLUR_ROTATOR_MODE, pixelPos, gRotator, gFrameIndex );
 
+    float3 Vv = GetViewVector( Xv, true );
+    float NoV = abs( dot( Nv, Vv ) );
+
     #define REBLUR_SPATIAL_MODE REBLUR_PRE_BLUR
 
     #ifdef REBLUR_DIFFUSE

Shaders/Include/SIGMA/SIGMA_Config.hlsli

@@ -11,7 +11,6 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 #define SIGMA_POISSON_SAMPLE_NUM                        8
 #define SIGMA_POISSON_SAMPLES                           g_Poisson8
 #define SIGMA_MAX_PIXEL_RADIUS                          32.0
-#define SIGMA_PLANE_DISTANCE_SCALE                      0.25
 #define SIGMA_MIN_HIT_DISTANCE_OUTPUT                   0.0001
 #define SIGMA_PENUMBRA_FIX_BLUR_RADIUS_ADDON            5.0
 #define SIGMA_MAX_SIGMA_SCALE                           3.5

Shaders/Include/SIGMA/SIGMA_Shadow_Blur.hlsli

@@ -152,7 +152,7 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
     sum = 1.0;
 
     float frustumHeight = PixelRadiusToWorld( gUnproject, gOrthoMode, gRectSize.y, viewZ );
-    float2 geometryWeightParams = GetGeometryWeightParams( gPlaneDistSensitivity, frustumHeight, Xv, Nv, SIGMA_PLANE_DISTANCE_SCALE );
+    float2 geometryWeightParams = GetGeometryWeightParams( gPlaneDistSensitivity, frustumHeight, Xv, Nv, 1.0 );
 
     [unroll]
     for( uint i = 0; i < SIGMA_POISSON_SAMPLE_NUM; i++ )