NRD updated to v4.9.1:

- OCCLUSION: better settings
- OCCLUSION: added blue noise usage to compensate the lack of temporal stabilization
- UI: various improvements
- reworked post processing pipeline
- quality of life improvements
- fixed several minor bugs
- updated deps

Shaders.cfg

@@ -7,8 +7,8 @@ SharcClear.cs.hlsl -T cs
 SharcHashCopy.cs.hlsl -T cs
 SharcResolve.cs.hlsl -T cs
 SharcUpdate.cs.hlsl -T cs
-Temporal.cs.hlsl -T cs
+Taa.cs.hlsl -T cs
 TraceOpaque.cs.hlsl -T cs
 TraceTransparent.cs.hlsl -T cs
-Upsample.cs.hlsl -T cs
-UpsampleNis.cs.hlsl -T cs
+Final.cs.hlsl -T cs
+Nis.cs.hlsl -T cs

Shaders/DlssAfter.cs.hlsl

@@ -23,18 +23,10 @@ void main( uint2 pixelPos : SV_DispatchThreadId )
 
     float3 color = gOut_Image[ pixelPos ];
 
-    // Tonemap
+    bool isSrgb = gIsSrgb && ( gOnScreen == SHOW_FINAL || gOnScreen == SHOW_BASE_COLOR );
     color = ApplyTonemap( color );
-
-    // Dithering
-    Rng::Hash::Initialize( pixelPos, gFrameIndex );
-
-    float rnd = Rng::Hash::GetFloat( );
-    color += ( rnd * 2.0 - 1.0 ) / 127.0;
-
-    // Conversion
-    if( gIsSrgb && ( gOnScreen == SHOW_FINAL || gOnScreen == SHOW_BASE_COLOR ) )
-        color = Color::ToSrgb( color );
+    if( isSrgb )
+        color = Color::ToSrgb( saturate( color ) );
 
     // Output
     gOut_Image[ pixelPos ] = color;

Shaders/DlssBefore.cs.hlsl

@@ -10,11 +10,8 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 #include "Include/Shared.hlsli"
 
-NRI_RESOURCE( Texture2D<float4>, gIn_ComposedLighting_ViewZ, t, 0, 1 );
-
 NRI_RESOURCE( RWTexture2D<float>, gOut_ViewZ, u, 0, 1 );
 NRI_RESOURCE( RWTexture2D<float3>, gInOut_Mv, u, 1, 1 );
-NRI_RESOURCE( RWTexture2D<float3>, gOut_FinalImage, u, 2, 1 );
 
 [numthreads( 16, 16, 1 )]
 void main( uint2 pixelPos : SV_DispatchThreadId )
@@ -36,19 +33,12 @@ void main( uint2 pixelPos : SV_DispatchThreadId )
     }
 
     // Patch MV, because 2D MVs needed
-    float3 mv = gInOut_Mv[ pixelPos ];
     if( gIsWorldSpaceMotionEnabled )
     {
+        float3 mv = gInOut_Mv[ pixelPos ];
         float3 Xprev = Geometry::AffineTransform( gViewToWorld, Xv ) + mv;
         float2 pixelUvPrev = Geometry::GetScreenUv( gWorldToClipPrev, Xprev );
         mv.xy = ( pixelUvPrev - pixelUv ) * gRenderSize;
         gInOut_Mv[ pixelPos ] = mv;
     }
-
-    // Apply exposure
-    float3 Lsum = gIn_ComposedLighting_ViewZ[ pixelPos ].xyz;
-    Lsum = ApplyExposure( Lsum );
-
-    // Output
-    gOut_FinalImage[ pixelPos ] = Lsum;
 }

Shaders/Upsample.cs.hlsl → Shaders/Final.cs.hlsl

@@ -10,10 +10,11 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 #include "Include/Shared.hlsli"
 
-NRI_RESOURCE( Texture2D<float4>, gIn_Image, t, 0, 1 );
-NRI_RESOURCE( Texture2D<float4>, gIn_Validation, t, 1, 1 );
+NRI_RESOURCE( Texture2D<float4>, gIn_PostAA, t, 0, 1 );
+NRI_RESOURCE( Texture2D<float4>, gIn_PreAA, t, 1, 1 );
+NRI_RESOURCE( Texture2D<float4>, gIn_Validation, t, 2, 1 );
 
-NRI_RESOURCE( RWTexture2D<float3>, gOut_Image, u, 0, 1 );
+NRI_RESOURCE( RWTexture2D<float3>, gOut_Final, u, 0, 1 );
 
 [numthreads( 16, 16, 1 )]
 void main( uint2 pixelPos : SV_DispatchThreadId )
@@ -24,23 +25,26 @@ void main( uint2 pixelPos : SV_DispatchThreadId )
     if( pixelUv.x > 1.0 || pixelUv.y > 1.0 )
         return;
 
-    // DLSS upscales render resolution to output resolution
-    float2 rectSize = gRectSize;
-    float2 invRenderSize = gInvRenderSize;
-    if( gSR || gRR )
-    {
-        rectSize = gOutputSize;
-        invRenderSize = gInvOutputSize;
-    }
-
     // Upsampling
-    float2 pixelPosScaled = pixelUv * rectSize;
-    float3 upsampled = BicubicFilterNoCorners( gIn_Image, gLinearSampler, pixelPosScaled, invRenderSize, 0.66 ).xyz;
+    float3 upsampled = BicubicFilterNoCorners( gIn_PostAA, gLinearSampler, pixelUv * gOutputSize, gInvOutputSize, 0.66 ).xyz;
+
+    // Noisy input
+    float3 input = gIn_PreAA.SampleLevel( gNearestSampler, pixelUv * gRectSize * gInvRenderSize, 0 ).xyz;
+
+    bool isSrgb = gIsSrgb && ( gOnScreen == SHOW_FINAL || gOnScreen == SHOW_BASE_COLOR );
+    input = ApplyTonemap( input );
+    if( isSrgb )
+        input = Color::ToSrgb( saturate( input ) );
 
     // Split screen - noisy input / denoised output
-    float3 input = gIn_Image.SampleLevel( gNearestSampler, pixelPosScaled * invRenderSize, 0 ).xyz;
     float3 result = pixelUv.x < gSeparator ? input : upsampled;
 
+    // Dithering
+    Rng::Hash::Initialize( pixelPos, gFrameIndex );
+
+    float rnd = Rng::Hash::GetFloat( );
+    result += ( rnd - 0.5 ) / ( isSrgb ? 256.0 : 1024.0 );
+
     // Split screen - vertical line
     float verticalLine = saturate( 1.0 - abs( pixelUv.x - gSeparator ) * gWindowSize.x / 3.5 );
     verticalLine = saturate( verticalLine / 0.5 );
@@ -52,10 +56,10 @@ void main( uint2 pixelPos : SV_DispatchThreadId )
     // Validation layer
     if( gValidation )
     {
-        float4 validation = gIn_Validation.SampleLevel( gLinearSampler, pixelUv, 0 );
+        float4 validation = gIn_Validation.SampleLevel( gNearestSampler, pixelUv, 0 );
         result = lerp( result, validation.xyz, validation.w );
     }
 
     // Output
-    gOut_Image[ pixelPos ] = result;
+    gOut_Final[ pixelPos ] = result;
 }

Shaders/Include/RaytracingShared.hlsli

@@ -579,10 +579,6 @@ float ReprojectIrradiance(
     float weight = 1.0;
     float2 pixelUv = float2( pixelPos + 0.5 ) * gInvRectSize;
 
-    // Ignore undenoised regions ( split screen mode is active )
-    weight *= float( pixelUv.x > gSeparator );
-    weight *= float( uv.x > gSeparator );
-
     // Relaxed checks for refractions
     float viewZ = abs( Geometry::AffineTransform( isPrevFrame ? gWorldToViewPrev : gWorldToView, geometryProps.X ).z );
     float err = ( viewZ - prevViewZ ) * Math::PositiveRcp( max( viewZ, prevViewZ ) );

Shaders/Include/Shared.hlsli

@@ -21,14 +21,15 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 // SH - SH (spherical harmonics or spherical gaussian) denoisers
 // OCCLUSION - OCCLUSION (ambient or specular occlusion only) denoisers
 // DIRECTIONAL_OCCLUSION - DIRECTIONAL_OCCLUSION (ambient occlusion in SH mode) denoisers
-#define NRD_MODE                            NORMAL // NORMAL, SH, OCCLUSION, DIRECTIONAL_OCCLUSION
+#define NRD_MODE                            OCCLUSION // NORMAL, SH, OCCLUSION, DIRECTIONAL_OCCLUSION
 #define SIGMA_TRANSLUCENT                   1
 
 // Default = 1
 #define USE_IMPORTANCE_SAMPLING             1
 #define USE_PSR                             1 // allow primary surface replacement
 #define USE_SHARC_DITHERING                 1 // must be in [0; 1] range
 #define USE_TRANSLUCENCY                    1 // translucent foliage
+#define USE_NIS                             1 // NIS filter (debug only)
 
 // Default = 0
 #define USE_SANITIZATION                    0 // NRD sample is NAN/INF free
@@ -129,10 +130,22 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 #define SHARC_NORMAL_DITHER                 0.003
 #define SHARC_POS_DITHER                    0.001
 
+#define BLUE_NOISE_SPATIAL_DIM              128 // see StaticTexture::ScramblingRanking
+#define BLUE_NOISE_TEMPORAL_DIM             4 // good values: 4-8 for shadows, 8-16 for occlusion, 8-32 for lighting
+
 #define MORPH_MAX_ACTIVE_TARGETS_NUM        8u
 #define MORPH_ELEMENTS_PER_ROW_NUM          4
 #define MORPH_ROWS_NUM                      ( MORPH_MAX_ACTIVE_TARGETS_NUM / MORPH_ELEMENTS_PER_ROW_NUM )
 
+#define NIS_SCALER                          1
+#define NIS_HDR_MODE                        0
+#define NIS_BLOCK_WIDTH                     32
+#define NIS_BLOCK_HEIGHT                    32
+#define NIS_THREAD_GROUP_SIZE               128
+#define NIS_USE_HALF_PRECISION              1
+#define NIS_HLSL                            1
+#define NIS_VIEWPORT_SUPPORT                0
+
 // Instance flags
 #define FLAG_FIRST_BIT                      25 // this + number of flags must be <= 32
 #define NON_FLAG_MASK                       ( ( 1 << FLAG_FIRST_BIT ) - 1 )
@@ -493,6 +506,8 @@ float GetCircleOfConfusion( float distance ) // diameter
 #define SKY_INTENSITY 1.0
 #define SUN_INTENSITY 10.0
 
+// TODO: add dither, use USE_FP11 = true
+
 float3 GetSunIntensity( float3 v )
 {
     float b = dot( v, gSunDirection.xyz );

Shaders/UpsampleNis.cs.hlsl → Shaders/Nis.cs.hlsl

@@ -54,23 +54,26 @@ NRI_RESOURCE( RWTexture2D<float4>, gOut_Image, u, 0, 1 );
 #pragma warning( disable : 3203 )
 
 // Main
-#define NIS_SCALER                  1
-#define NIS_HDR_MODE                0
-#define NIS_BLOCK_WIDTH             32
-#define NIS_BLOCK_HEIGHT            32
-#define NIS_THREAD_GROUP_SIZE       128
-#define NIS_USE_HALF_PRECISION      1
-#define NIS_HLSL                    1
-#define NIS_VIEWPORT_SUPPORT        0
-
 #if( NRI_SHADER_MODEL >= 62 )
     #define NIS_HLSL_6_2            1
 #endif
 
 #include "NVIDIAImageScaling/NIS/NIS_Scaler.h"
 
+#if( USE_NIS == 1 )
+
 [numthreads( NIS_THREAD_GROUP_SIZE, 1, 1 )]
 void main( uint2 blockId : SV_GroupID, uint threadId : SV_GroupThreadID )
 {
     NVScaler( blockId, threadId );
 }
+
+#else
+
+[numthreads( NIS_BLOCK_WIDTH, NIS_BLOCK_HEIGHT, 1 )]
+void main( int2 pixelPos : SV_DispatchThreadId )
+{
+    gOut_Image[ pixelPos ] = gIn_Image[ pixelPos ];
+}
+
+#endif

Shaders/Temporal.cs.hlsl → Shaders/Taa.cs.hlsl

@@ -42,7 +42,6 @@ void Preload( uint2 sharedPos, int2 globalPos )
     globalPos = clamp( globalPos, 0, gRectSize - 1.0 );
 
     float4 color_viewZ = gIn_ComposedLighting_ViewZ[ globalPos ];
-    color_viewZ.xyz = ApplyExposure( color_viewZ.xyz );
     color_viewZ.xyz = ApplyTonemap( color_viewZ.xyz );
     color_viewZ.w = abs( color_viewZ.w ) * Math::Sign( gNearZ ) / FP16_VIEWZ_SCALE;
 
@@ -128,6 +127,7 @@ void main( int2 threadPos : SV_GroupThreadId, int2 pixelPos : SV_DispatchThreadI
 
     history.xyz = max( history.xyz, 0.0 ); // yes, not "saturate"
 
+    // Remove transfer
     bool isSrgb = gIsSrgb && ( gOnScreen == SHOW_FINAL || gOnScreen == SHOW_BASE_COLOR );
     if( isSrgb )
         history.xyz = Color::FromSrgb( history.xyz );
@@ -138,7 +138,7 @@ void main( int2 threadPos : SV_GroupThreadId, int2 pixelPos : SV_DispatchThreadI
 
     // Disocclusion #1
     bool isInScreen = all( saturate( pixelUvPrev ) == pixelUvPrev );
-    mixRate = ( !isInScreen || pixelUv.x < gSeparator ) ? 1.0 : mixRate;
+    mixRate = !isInScreen ? 1.0 : mixRate;
 
     // Disocclusion #2
     float3 clampedHistory = Color::ClampAabb( m1, sigma, history.xyz );
@@ -158,24 +158,10 @@ void main( int2 threadPos : SV_GroupThreadId, int2 pixelPos : SV_DispatchThreadI
     // Final mix
     float3 result = lerp( clampedHistory, input, mixRate );
 
-    // Split screen - vertical line
-    float verticalLine = saturate( 1.0 - abs( pixelUv.x - gSeparator ) * gRectSize.x / 3.5 );
-    verticalLine = saturate( verticalLine / 0.5 );
-    verticalLine *= float( gSeparator != 0.0 );
-    verticalLine *= float( gRenderSize.x == gRectSize.x );
-
-    const float3 nvColor = float3( 118.0, 185.0, 0.0 ) / 255.0;
-    result = lerp( result, nvColor * verticalLine, verticalLine );
-
-    // Dithering
-    Rng::Hash::Initialize( pixelPos, gFrameIndex );
-
-    float rnd = Rng::Hash::GetFloat( );
-    result += ( rnd * 2.0 - 1.0 ) / 1023.0;
-
-    // Output
+    // Apply transfer
     if( isSrgb )
         result = Color::ToSrgb( result );
 
+    // Output
     gOut_Result[ pixelPos ] = float4( result, mixRate );
 }