NRD updated to v4.9.2:

- utilized "NRIResourceAllocator" extension for ~ALL allocations in the app
- PT: removed incorrectly traced glass in PSR space
- PT: glass tracing uses Bayer dithering to reduce overall noise
- Linux: minor fix for DLSS in Cmake
- returned back NORMAL mode instead of OCCLUSION (regression)
- updated deps
- updated tests

CMakeLists.txt

@@ -142,7 +142,7 @@ if (WIN32)
 else ()
     find_library (NGX_D_LIB NAMES libnvsdk_ngx.a PATHS "External/NGX/lib/Linux_x86_64")
     find_library (NGX_LIB NAMES libnvsdk_ngx.a PATHS "External/NGX/lib/Linux_x86_64")
-    find_file (DLSS_SR_DLL NAMES libnvidia-ngx-dlss.so.3.7.0 PATHS "External/NGX/lib/Linux_x86_64/rel")
+    find_file (DLSS_SR_DLL NAMES libnvidia-ngx-dlss.so.3.7.10 PATHS "External/NGX/lib/Linux_x86_64/rel")
 endif ()
 
 if (NGX_D_LIB AND NGX_LIB AND DLSS_SR_DLL)

Shaders/Composition.cs.hlsl

@@ -202,7 +202,7 @@ void main( int2 pixelPos : SV_DispatchThreadId )
     float3 diffDemod = ( 1.0 - Fenv ) * albedo * 0.99 + 0.01;
     float3 specDemod = Fenv * 0.99 + 0.01;
 
-    if( normMaterialID == MATERIAL_ID_HAIR / 3.0 && NRD_NORMAL_ENCODING == NRD_NORMAL_ENCODING_R10G10B10A2_UNORM )
+    if( normMaterialID == MATERIAL_ID_HAIR / MATERIAL_NORM )
     {
         diffDemod = 1.0;
         specDemod = 1.0;

Shaders/DlssBefore.cs.hlsl

@@ -10,7 +10,7 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 #include "Include/Shared.hlsli"
 
-NRI_RESOURCE( RWTexture2D<float>, gOut_ViewZ, u, 0, 1 );
+NRI_RESOURCE( RWTexture2D<float>, gInOut_ViewZ, u, 0, 1 );
 NRI_RESOURCE( RWTexture2D<float3>, gInOut_Mv, u, 1, 1 );
 
 [numthreads( 16, 16, 1 )]
@@ -22,14 +22,14 @@ void main( uint2 pixelPos : SV_DispatchThreadId )
     if( pixelUv.x > 1.0 || pixelUv.y > 1.0 )
         return;
 
-    float viewZ = gOut_ViewZ[ pixelPos ];
+    float viewZ = gInOut_ViewZ[ pixelPos ];
     float3 Xv = Geometry::ReconstructViewPosition( pixelUv, gCameraFrustum, viewZ, gOrthoMode );
 
     // Recalculate viewZ to depth ( needed for SR )
     if( gSR )
     {
         float4 clipPos = Geometry::ProjectiveTransform( gViewToClip, Xv );
-        gOut_ViewZ[ pixelPos ] = clipPos.z / clipPos.w;
+        gInOut_ViewZ[ pixelPos ] = clipPos.z / clipPos.w;
     }
 
     // Patch MV, because 2D MVs needed

Shaders/Include/Shared.hlsli

@@ -21,7 +21,7 @@ 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                            OCCLUSION // NORMAL, SH, OCCLUSION, DIRECTIONAL_OCCLUSION
+#define NRD_MODE                            NORMAL // NORMAL, SH, OCCLUSION, DIRECTIONAL_OCCLUSION
 #define SIGMA_TRANSLUCENT                   1
 
 // Default = 1
@@ -94,6 +94,7 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 #define MATERIAL_ID_METAL                   1
 #define MATERIAL_ID_PSR                     2
 #define MATERIAL_ID_HAIR                    3
+#define MATERIAL_NORM                       3.0
 
 // Mip mode
 #define MIP_VISIBILITY                      0 // for visibility: emission, shadow and alpha mask
@@ -506,8 +507,6 @@ 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/Taa.cs.hlsl

@@ -10,9 +10,10 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 #include "Include/Shared.hlsli"
 
-NRI_RESOURCE( Texture2D<float3>, gIn_Mv, t, 0, 1 );
-NRI_RESOURCE( Texture2D<float4>, gIn_ComposedLighting_ViewZ, t, 1, 1 );
-NRI_RESOURCE( Texture2D<float4>, gIn_History, t, 2, 1 );
+NRI_RESOURCE( Texture2D<float>, gIn_ViewZ, t, 0, 1 );
+NRI_RESOURCE( Texture2D<float3>, gIn_Mv, t, 1, 1 );
+NRI_RESOURCE( Texture2D<float3>, gIn_Composed, t, 2, 1 );
+NRI_RESOURCE( Texture2D<float4>, gIn_History, t, 3, 1 );
 
 NRI_RESOURCE( RWTexture2D<float4>, gOut_Result, u, 0, 1 );
 
@@ -41,9 +42,9 @@ void Preload( uint2 sharedPos, int2 globalPos )
 {
     globalPos = clamp( globalPos, 0, gRectSize - 1.0 );
 
-    float4 color_viewZ = gIn_ComposedLighting_ViewZ[ globalPos ];
-    color_viewZ.xyz = ApplyTonemap( color_viewZ.xyz );
-    color_viewZ.w = abs( color_viewZ.w ) * Math::Sign( gNearZ ) / FP16_VIEWZ_SCALE;
+    float4 color_viewZ;
+    color_viewZ.xyz = ApplyTonemap( gIn_Composed[ globalPos ] );
+    color_viewZ.w = gIn_ViewZ[ globalPos ];
 
     s_Data[ sharedPos.y ][ sharedPos.x ] = color_viewZ;
 }

Shaders/TraceOpaque.cs.hlsl

@@ -57,6 +57,10 @@ float2 GetBlueNoise( uint2 pixelPos, bool isCheckerboard, uint seed = 0 )
     float2 dither = ( float2( d & 3, d >> 2 ) + 0.5 ) * ( 1.0 / 4.0 );
     blue += ( dither.xyxy - 0.5 ) * ( 1.0 / 256.0 );
 
+    // Don't use blue noise in these cases
+    if( gDenoiserType == DENOISER_REFERENCE || gRR )
+        blue.xy = Rng::Hash::GetFloat2( );
+
     return saturate( blue.xy );
 }
 
@@ -273,7 +277,7 @@ TraceOpaqueResult TraceOpaque( TraceOpaqueDesc desc )
 
                     bool isSharcAllowed = geometryProps.tmin > voxelSize; // voxel angular size is acceptable
                     isSharcAllowed = isSharcAllowed && Rng::Hash::GetFloat( ) > Lcached.w; // probabilistically estimate the need
-                    isSharcAllowed = isSharcAllowed && desc.bounceNum == 0; // for PSR allowed only for the last bounce
+                    isSharcAllowed = isSharcAllowed && desc.bounceNum == 0; // allow only for the last bounce for PSR
 
                     float3 sharcRadiance = 0;
                     if( isSharcAllowed && SharcGetCachedRadiance( sharcState, sharcHitData, sharcRadiance, false ) )
@@ -299,21 +303,22 @@ TraceOpaqueResult TraceOpaque( TraceOpaqueDesc desc )
             psrLsum = Lcached.xyz;
 
             // Update materials and INF emission
+            float3 psrNormal = float3( 0, 0, 1 );
             if( !geometryProps.IsSky( ) )
             {
-                float3 psrNormal = Geometry::RotateVectorInverse( mirrorMatrix, materialProps.N );
+                psrNormal = Geometry::RotateVectorInverse( mirrorMatrix, materialProps.N );
 
                 gOut_BaseColor_Metalness[ desc.pixelPos ] = float4( Color::ToSrgb( materialProps.baseColor ), materialProps.metalness );
-
-                // IMPORTANT: use another set of materialID to avoid potential leaking, because PSR uses own de-modulation scheme
-                gOut_Normal_Roughness[ desc.pixelPos ] = NRD_FrontEnd_PackNormalAndRoughness( psrNormal, materialProps.roughness, MATERIAL_ID_PSR );
             }
             else
             {
                 // Composition pass doesn't apply "psrThroughput" to INF pixels
                 gOut_DirectEmission[ desc.pixelPos ] = psrLsum * psrThroughput;
             }
 
+            // IMPORTANT: use another materialID to avoid potential leaking, because PSR uses own de-modulation scheme. Also needed for tracing of transparent stuff
+            gOut_Normal_Roughness[ desc.pixelPos ] = NRD_FrontEnd_PackNormalAndRoughness( psrNormal, materialProps.roughness, MATERIAL_ID_PSR );
+
             // Update viewZ
             float3 Xvirtual = desc.geometryProps.X - desc.geometryProps.V * accumulatedHitDist;
             viewZ = Geometry::AffineTransform( gWorldToView, Xvirtual ).z;
@@ -377,7 +382,7 @@ TraceOpaqueResult TraceOpaque( TraceOpaqueDesc desc )
 
                 // Clamp probability to a sane range to guarantee a sample in 3x3 ( or 5x5 ) area
                 float rnd = Rng::Hash::GetFloat( );
-                if( gTracingMode == RESOLUTION_FULL_PROBABILISTIC && bounceIndex == 1 )
+                if( gTracingMode == RESOLUTION_FULL_PROBABILISTIC && bounceIndex == 1 && !gRR )
                 {
                     diffuseProbability = float( diffuseProbability != 0.0 ) * clamp( diffuseProbability, gMinProbability, 1.0 - gMinProbability );
                     rnd = Sequence::Bayer4x4( desc.pixelPos, gFrameIndex ) + rnd / 16.0;
@@ -623,7 +628,7 @@ TraceOpaqueResult TraceOpaque( TraceOpaqueDesc desc )
                         float3 Xglobal = GetGlobalPos( geometryProps.X );
                         uint level = GetGridLevel( Xglobal, gridParameters );
                         float voxelSize = GetVoxelSize( level, gridParameters );
-                        float smc = GetSpecMagicCurve( accumulatedDiffuseLikeMotion ); // better than "materialProps.roughness"
+                        float smc = GetSpecMagicCurve( materialProps.roughness );
 
                         float3x3 mBasis = Geometry::GetBasis( geometryProps.N );
                         float2 rndScaled = ( Rng::Hash::GetFloat2( ) - 0.5 ) * voxelSize * USE_SHARC_DITHERING;
@@ -933,8 +938,6 @@ void main( uint2 pixelPos : SV_DispatchThreadId )
 
     // Sun shadow
     float2 rnd = GetBlueNoise( pixelPos, false );
-    if( gDenoiserType == DENOISER_REFERENCE )
-        rnd = Rng::Hash::GetFloat2( );
     rnd = ImportanceSampling::Cosine::GetRay( rnd ).xy;
     rnd *= gTanSunAngularRadius;
 

Shaders/TraceTransparent.cs.hlsl

@@ -16,11 +16,12 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 // Inputs
 NRI_RESOURCE( Texture2D<float>, gIn_ViewZ, t, 0, 1 );
-NRI_RESOURCE( Texture2D<float3>, gIn_ComposedDiff, t, 1, 1 );
-NRI_RESOURCE( Texture2D<float4>, gIn_ComposedSpec_ViewZ, t, 2, 1 );
+NRI_RESOURCE( Texture2D<float4>, gIn_Normal_Roughness, t, 1, 1 );
+NRI_RESOURCE( Texture2D<float3>, gIn_ComposedDiff, t, 2, 1 );
+NRI_RESOURCE( Texture2D<float4>, gIn_ComposedSpec_ViewZ, t, 3, 1 );
 
 // Outputs
-NRI_RESOURCE( RWTexture2D<float4>, gOut_Composed, u, 0, 1 );
+NRI_RESOURCE( RWTexture2D<float3>, gOut_Composed, u, 0, 1 );
 NRI_RESOURCE( RWTexture2D<float3>, gInOut_Mv, u, 1, 1 );
 
 //========================================================================================
@@ -66,7 +67,13 @@ float3 TraceTransparent( TraceTransparentDesc desc )
         if( bounce == 1 )
             pathThroughput *= isReflection ? F : 1.0 - F;
         else
-            isReflection = Rng::Hash::GetFloat( ) < F;
+        {
+            float rnd = Sequence::Bayer4x4( desc.pixelPos, gFrameIndex + bounce );
+            if( gDenoiserType == DENOISER_REFERENCE || gRR )
+                rnd = Rng::Hash::GetFloat( );
+
+            isReflection = rnd < F;
+        }
 
         // Compute ray
         float3 ray = reflect( -geometryProps.V, geometryProps.N );
@@ -184,8 +191,8 @@ void main( int2 pixelPos : SV_DispatchThreadId )
 
     // Composed without glass
     float3 diff = gIn_ComposedDiff[ pixelPos ];
-    float4 spec = gIn_ComposedSpec_ViewZ[ pixelPos ];
-    float3 Lsum = diff + spec.xyz * float( gOnScreen == SHOW_FINAL );
+    float3 spec = gIn_ComposedSpec_ViewZ[ pixelPos ].xyz;
+    float3 Lsum = diff + spec * float( gOnScreen == SHOW_FINAL );
 
     // Primary ray for transparent geometry only
     float3 cameraRayOrigin = ( float3 )0;
@@ -198,8 +205,12 @@ void main( int2 pixelPos : SV_DispatchThreadId )
 
     GeometryProps geometryPropsT = CastRay( cameraRayOrigin, cameraRayDirection, 0.0, tmin0, GetConeAngleFromRoughness( 0.0, 0.0 ), gWorldTlas, gTransparent == 0.0 ? 0 : GEOMETRY_ONLY_TRANSPARENT, 0 );
 
+    // Material ID
+    float normMaterialID;
+    float4 normalAndRoughness = NRD_FrontEnd_UnpackNormalAndRoughness( gIn_Normal_Roughness[ pixelPos ], normMaterialID );
+
     // Trace delta events
-    if( !geometryPropsT.IsSky( ) && geometryPropsT.tmin < tmin0 )
+    if( !geometryPropsT.IsSky( ) && geometryPropsT.tmin < tmin0 && normMaterialID != MATERIAL_ID_PSR / MATERIAL_NORM )
     {
         // Patch motion vectors replacing MV for the background with MV for the closest glass layer.
         // IMPORTANT: surface-based motion can be used only if the object is curved.
@@ -230,7 +241,5 @@ void main( int2 pixelPos : SV_DispatchThreadId )
     Lsum = ApplyExposure( Lsum );
 
     // Output
-    float z = spec.w;
-
-    gOut_Composed[ pixelPos ] = float4( Lsum, z );
+    gOut_Composed[ pixelPos ] = Lsum;
 }

Source/DLSS/DLSSIntegration.hpp

@@ -3,7 +3,7 @@
 #include <assert.h> // assert
 #include <stdio.h> // printf
 
-static_assert(NRI_VERSION_MAJOR >= 1 && NRI_VERSION_MINOR >= 140, "Unsupported NRI version!");
+static_assert(NRI_VERSION_MAJOR >= 1 && NRI_VERSION_MINOR >= 141, "Unsupported NRI version!");
 
 // An ugly temp workaround until DLSS fix the problem
 #ifndef _WIN32
@@ -141,8 +141,8 @@ bool DlssIntegration::InitializeLibrary(nri::Device& device, const char* appData
     else if (deviceDesc.graphicsAPI == nri::GraphicsAPI::VK)
     {
         VkDevice vkDevice = (VkDevice)NRI.GetDeviceNativeObject(*m_Device);
-        VkPhysicalDevice vkPhysicalDevice = (VkPhysicalDevice)NRI.GetVkPhysicalDevice(*m_Device);
-        VkInstance vkInstance = (VkInstance)NRI.GetVkInstance(*m_Device);
+        VkPhysicalDevice vkPhysicalDevice = (VkPhysicalDevice)NRI.GetPhysicalDeviceVK(*m_Device);
+        VkInstance vkInstance = (VkInstance)NRI.GetInstanceVK(*m_Device);
         result = NVSDK_NGX_VULKAN_Init(m_ApplicationId, path, vkInstance, vkPhysicalDevice, vkDevice);
         if (NVSDK_NGX_SUCCEED(result))
             result = NVSDK_NGX_VULKAN_GetCapabilityParameters(&m_NgxParameters);