v4.2.1:

HIGHLIGHTS:
- bugfixes & improvements

DETAILS:
- NRD: fixed regression previously introduced in projection matrix decomposition
- REBLUR: fixed touching of data outside of rectangle, set by Dynamic Resolution Scaling
- Cmake: "NRD_Shaders" target doesn't get created if "NRD_USE_PRECOMPILED_SHADERS" is not set
- updated MathLib

CMakeLists.txt

@@ -114,11 +114,6 @@ endif ()
 
 message ("NRD shaders output path: '${NRD_SHADERS_PATH}'")
 
-if (NRD_USE_PRECOMPILED_SHADERS)
-    target_include_directories (${PROJECT_NAME} PRIVATE "${NRD_SHADERS_PATH}")
-    target_compile_definitions (${PROJECT_NAME} PRIVATE NRD_USE_PRECOMPILED_SHADERS)
-endif ()
-
 target_include_directories (${PROJECT_NAME} PUBLIC "Include")
 target_include_directories (${PROJECT_NAME} PRIVATE "External")
 target_compile_definitions (${PROJECT_NAME} PRIVATE ${COMPILE_DEFINITIONS})
@@ -138,76 +133,85 @@ if (NOT TARGET ShaderMake)
 endif ()
 
 # Shaders
-file (GLOB_RECURSE SHADERS "Shaders/*.hlsl" "Shaders/*.hlsli" "External/MathLib/*.hlsli")
-set_source_files_properties (${SHADERS} PROPERTIES VS_TOOL_OVERRIDE "None")
+if (NRD_USE_PRECOMPILED_SHADERS)
+    target_include_directories (${PROJECT_NAME} PRIVATE "${NRD_SHADERS_PATH}")
+    target_compile_definitions (${PROJECT_NAME} PRIVATE NRD_USE_PRECOMPILED_SHADERS)
 
-if (WIN32)
-    add_custom_target (${PROJECT_NAME}_Shaders ALL
-        COMMAND ShaderMake --useAPI
-            --header ${NRD_SHADER_BINARIES} --flatten --stripReflection --compiler "${DXC_PATH}"
-            --sourceDir "Shaders/Source"
-            --allResourcesBound
-            -p DXIL --WX
-            -c Shaders.cfg
-            -o "${NRD_SHADERS_PATH}"
-            -I "External/MathLib"
-            -I "Shaders/Include"
-            -I "Shaders/Resources"
-            -D NRD_NORMAL_ENCODING=${NRD_NORMAL_ENCODING}
-            -D NRD_ROUGHNESS_ENCODING=${NRD_ROUGHNESS_ENCODING}
-            -D NRD_INTERNAL
-        COMMAND ShaderMake --useAPI
-            --header ${NRD_SHADER_BINARIES} --flatten --stripReflection --compiler "${DXC_SPIRV_PATH}"
-            --sourceDir "Shaders/Source"
-            --allResourcesBound
-            -p SPIRV --WX
-            -c Shaders.cfg
-            -o "${NRD_SHADERS_PATH}"
-            -I "External/MathLib"
-            -I "Shaders/Include"
-            -I "Shaders/Resources"
-            -D NRD_NORMAL_ENCODING=${NRD_NORMAL_ENCODING}
-            -D NRD_ROUGHNESS_ENCODING=${NRD_ROUGHNESS_ENCODING}
-            -D NRD_INTERNAL
-        COMMAND ShaderMake --useAPI
-            --header ${NRD_SHADER_BINARIES} --flatten --stripReflection --compiler "${FXC_PATH}"
-            --sourceDir "Shaders/Source"
-            --allResourcesBound
-            -p DXBC --WX
-            -c Shaders.cfg
-            -o "${NRD_SHADERS_PATH}"
-            -I "External/MathLib"
-            -I "Shaders/Include"
-            -I "Shaders/Resources"
-            -D NRD_NORMAL_ENCODING=${NRD_NORMAL_ENCODING}
-            -D NRD_ROUGHNESS_ENCODING=${NRD_ROUGHNESS_ENCODING}
-            -D NRD_INTERNAL
-        DEPENDS ShaderMake
-        WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
-        VERBATIM
-        SOURCES ${SHADERS}
-    )
-else ()
-    add_custom_target (${PROJECT_NAME}_Shaders ALL
-        COMMAND ShaderMake
-            --header ${NRD_SHADER_BINARIES} --flatten --stripReflection --compiler "${DXC_SPIRV_PATH}"
-            --sourceDir "Shaders/Source"
-            --allResourcesBound
-            -p SPIRV --WX
-            -c Shaders.cfg
-            -o "${NRD_SHADERS_PATH}"
-            -I "External/MathLib"
-            -I "Shaders/Include"
-            -I "Shaders/Resources"
-            -D NRD_NORMAL_ENCODING=${NRD_NORMAL_ENCODING}
-            -D NRD_ROUGHNESS_ENCODING=${NRD_ROUGHNESS_ENCODING}
-            -D NRD_INTERNAL
-        DEPENDS ShaderMake
-        WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
-        VERBATIM
-        SOURCES ${SHADERS}
-    )
-endif ()
+    file (GLOB_RECURSE SHADERS "Shaders/*.hlsl" "Shaders/*.hlsli" "External/MathLib/*.hlsli")
+    set_source_files_properties (${SHADERS} PROPERTIES VS_TOOL_OVERRIDE "None")
+
+    if (WIN32)
+        add_custom_target (${PROJECT_NAME}_Shaders ALL
+            COMMAND ShaderMake --useAPI
+                --header ${NRD_SHADER_BINARIES} --flatten --stripReflection --compiler "${DXC_PATH}"
+                --sourceDir "Shaders/Source"
+                --allResourcesBound
+                -p DXIL --WX
+                -c Shaders.cfg
+                -o "${NRD_SHADERS_PATH}"
+                -I "External/MathLib"
+                -I "Shaders/Include"
+                -I "Shaders/Resources"
+                -D NRD_NORMAL_ENCODING=${NRD_NORMAL_ENCODING}
+                -D NRD_ROUGHNESS_ENCODING=${NRD_ROUGHNESS_ENCODING}
+                -D NRD_INTERNAL
+            COMMAND ShaderMake --useAPI
+                --header ${NRD_SHADER_BINARIES} --flatten --stripReflection --compiler "${DXC_SPIRV_PATH}"
+                --sourceDir "Shaders/Source"
+                --allResourcesBound
+                --sRegShift 100
+                --tRegShift 200
+                --bRegShift 300
+                --uRegShift 400
+                -p SPIRV --WX
+                -c Shaders.cfg
+                -o "${NRD_SHADERS_PATH}"
+                -I "External/MathLib"
+                -I "Shaders/Include"
+                -I "Shaders/Resources"
+                -D NRD_NORMAL_ENCODING=${NRD_NORMAL_ENCODING}
+                -D NRD_ROUGHNESS_ENCODING=${NRD_ROUGHNESS_ENCODING}
+                -D NRD_INTERNAL
+            COMMAND ShaderMake --useAPI
+                --header ${NRD_SHADER_BINARIES} --flatten --stripReflection --compiler "${FXC_PATH}"
+                --sourceDir "Shaders/Source"
+                --allResourcesBound
+                -p DXBC --WX
+                -c Shaders.cfg
+                -o "${NRD_SHADERS_PATH}"
+                -I "External/MathLib"
+                -I "Shaders/Include"
+                -I "Shaders/Resources"
+                -D NRD_NORMAL_ENCODING=${NRD_NORMAL_ENCODING}
+                -D NRD_ROUGHNESS_ENCODING=${NRD_ROUGHNESS_ENCODING}
+                -D NRD_INTERNAL
+            DEPENDS ShaderMake
+            WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
+            VERBATIM
+            SOURCES ${SHADERS}
+        )
+    else ()
+        add_custom_target (${PROJECT_NAME}_Shaders ALL
+            COMMAND ShaderMake
+                --header ${NRD_SHADER_BINARIES} --flatten --stripReflection --compiler "${DXC_SPIRV_PATH}"
+                --sourceDir "Shaders/Source"
+                --allResourcesBound
+                -p SPIRV --WX
+                -c Shaders.cfg
+                -o "${NRD_SHADERS_PATH}"
+                -I "External/MathLib"
+                -I "Shaders/Include"
+                -I "Shaders/Resources"
+                -D NRD_NORMAL_ENCODING=${NRD_NORMAL_ENCODING}
+                -D NRD_ROUGHNESS_ENCODING=${NRD_ROUGHNESS_ENCODING}
+                -D NRD_INTERNAL
+            DEPENDS ShaderMake
+            WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
+            VERBATIM
+            SOURCES ${SHADERS}
+        )
+    endif ()
 
-set_property (TARGET ${PROJECT_NAME}_Shaders PROPERTY FOLDER ${PROJECT_NAME})
-add_dependencies (${PROJECT_NAME} ${PROJECT_NAME}_Shaders)
+    set_property (TARGET ${PROJECT_NAME}_Shaders PROPERTY FOLDER ${PROJECT_NAME})
+    add_dependencies (${PROJECT_NAME} ${PROJECT_NAME}_Shaders)
+endif ()

Include/NRD.h

@@ -29,8 +29,8 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 
 #define NRD_VERSION_MAJOR 4
 #define NRD_VERSION_MINOR 2
-#define NRD_VERSION_BUILD 0
-#define NRD_VERSION_DATE "23 May 2023"
+#define NRD_VERSION_BUILD 1
+#define NRD_VERSION_DATE "25 May 2023"
 
 #if defined(_MSC_VER)
     #define NRD_CALL __fastcall

README.md

@@ -1,4 +1,4 @@
-# NVIDIA REAL-TIME DENOISERS v4.2.0 (NRD)
+# NVIDIA REAL-TIME DENOISERS v4.2.1 (NRD)
 
 [![Build NRD SDK](https://github.com/NVIDIAGameWorks/RayTracingDenoiser/actions/workflows/build.yml/badge.svg)](https://github.com/NVIDIAGameWorks/RayTracingDenoiser/actions/workflows/build.yml)
 

Resources/Version.h

@@ -23,6 +23,6 @@ Versioning rules:
 
 #define VERSION_MAJOR                   4
 #define VERSION_MINOR                   2
-#define VERSION_BUILD                   0
+#define VERSION_BUILD                   1
 
 #define VERSION_STRING STR(VERSION_MAJOR.VERSION_MINOR.VERSION_BUILD encoding=NRD_NORMAL_ENCODING.NRD_ROUGHNESS_ENCODING)

Shaders/Include/Common.hlsli

@@ -440,6 +440,7 @@ float GetGaussianWeight( float r )
 
 // Upsampling
 
+// TODO: used only for history resampling. This is why "gRectSizePrev" is hardcoded
 #define _BicubicFilterNoCornersWithFallbackToBilinearFilterWithCustomWeights_Init \
     /* Catmul-Rom with 12 taps ( excluding corners ) */ \
     float2 centerPos = floor( samplePos - 0.5 ) + 0.5; \
@@ -463,11 +464,12 @@ float GetGaussianWeight( float r )
     w4 = useBicubic ? w4 : 0.0; \
     float sum = dot( w, 1.0 ) + w4; \
     /* Texture coordinates */ \
-    float2 uv0 = centerPos + ( useBicubic ? float2( tc.x, -1.0 ) : float2( 0, 0 ) ); \
-    float2 uv1 = centerPos + ( useBicubic ? float2( -1.0, tc.y ) : float2( 1, 0 ) ); \
-    float2 uv2 = centerPos + ( useBicubic ? float2( tc.x, tc.y ) : float2( 0, 1 ) ); \
-    float2 uv3 = centerPos + ( useBicubic ? float2( 2.0, tc.y ) : float2( 1, 1 ) ); \
-    float2 uv4 = centerPos + ( useBicubic ? float2( tc.x, 2.0 ) : f ); // can be used to get a free bilinear sample after some massaging
+    float2 uv0 = min( centerPos + ( useBicubic ? float2( tc.x, -1.0 ) : float2( 0, 0 ) ), gRectSizePrev - 1.0 ); \
+    float2 uv1 = min( centerPos + ( useBicubic ? float2( -1.0, tc.y ) : float2( 1, 0 ) ), gRectSizePrev - 1.0 ); \
+    float2 uv2 = min( centerPos + ( useBicubic ? float2( tc.x, tc.y ) : float2( 0, 1 ) ), gRectSizePrev - 1.0 ); \
+    float2 uv3 = min( centerPos + ( useBicubic ? float2( 2.0, tc.y ) : float2( 1, 1 ) ), gRectSizePrev - 1.0 ); \
+    float2 uv4 = min( centerPos + ( useBicubic ? float2( tc.x, 2.0 ) : f ), gRectSizePrev - 1.0 ); \
+    float4 bilinearTaps
 
 /*
 IMPORTANT:
@@ -488,10 +490,12 @@ IMPORTANT:
 
 #define _BilinearFilterWithCustomWeights_Color( color, tex ) \
     /* Sampling */ \
-    color = tex.SampleLevel( gNearestClamp, centerPos * invTextureSize, 0 ) * bilinearCustomWeights.x; \
-    color += tex.SampleLevel( gNearestClamp, centerPos * invTextureSize, 0, int2( 1, 0 ) ) * bilinearCustomWeights.y; \
-    color += tex.SampleLevel( gNearestClamp, centerPos * invTextureSize, 0, int2( 0, 1 ) ) * bilinearCustomWeights.z; \
-    color += tex.SampleLevel( gNearestClamp, centerPos * invTextureSize, 0, int2( 1, 1 ) ) * bilinearCustomWeights.w; \
+    bilinearTaps = centerPos.xyxy + float4( 0.0, 0.0, 1.0, 1.0 ); \
+    bilinearTaps.zw = min( bilinearTaps.zw, gRectSizePrev - 1.0 ); \
+    color = tex.SampleLevel( gNearestClamp, bilinearTaps.xy * invTextureSize, 0 ) * bilinearCustomWeights.x; \
+    color += tex.SampleLevel( gNearestClamp, bilinearTaps.zy * invTextureSize, 0 ) * bilinearCustomWeights.y; \
+    color += tex.SampleLevel( gNearestClamp, bilinearTaps.xw * invTextureSize, 0 ) * bilinearCustomWeights.z; \
+    color += tex.SampleLevel( gNearestClamp, bilinearTaps.zw * invTextureSize, 0 ) * bilinearCustomWeights.w; \
     /* Normalize similarly to "STL::Filtering::ApplyBilinearCustomWeights()" */ \
     sum = dot( bilinearCustomWeights, 1.0 ); \
     color = sum < 0.0001 ? 0 : color * rcp( sum );

Shaders/Include/REBLUR/REBLUR_Common_DiffuseSpatialFilter.hlsli

@@ -145,9 +145,9 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
             w *= GetCombinedWeight( geometryWeightParams, Nv, Xvs, normalWeightParams, N, Ns );
             w *= lerp( minHitDistWeight, 1.0, GetHitDistanceWeight( hitDistanceWeightParams, ExtractHitDist( s ) ) );
 
-            // Get rid of potentially bad values outside of the screen ( important for checkerboard )
-            w = IsInScreen( uv ) ? w : 0.0;
-            s = w ? s : 0.0;
+            // Get rid of potentially bad values outside of the screen
+            w = ( IsInScreen( uv ) && !isnan( w ) ) ? w : 0.0;
+            s = w != 0.0 ? s : 0.0;
 
             // Accumulate
             sum += w;

Shaders/Include/REBLUR/REBLUR_Common_SpecularSpatialFilter.hlsli

@@ -184,9 +184,9 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
             #endif
         #endif
 
-            // Get rid of potentially bad values outside of the screen ( important for checkerboard )
-            w = IsInScreen( uv ) ? w : 0.0;
-            s = w ? s : 0.0;
+            // Get rid of potentially bad values outside of the screen
+            w = ( IsInScreen( uv ) && !isnan( w ) ) ? w : 0.0;
+            s = w != 0.0 ? s : 0.0;
 
         #if( REBLUR_SPATIAL_MODE == REBLUR_PRE_BLUR )
             // Min hit distance for tracking

Shaders/Include/REBLUR/REBLUR_DiffuseSpecular_Blur.hlsli

@@ -16,7 +16,7 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
 
     // Tile-based early out
     float isSky = gIn_Tiles[ pixelPos >> 4 ];
-    if( isSky != 0.0 )
+    if( isSky != 0.0 || pixelPos.x >= gRectSize.x || pixelPos.y >= gRectSize.y )
         return;
 
     // Early out

Shaders/Include/REBLUR/REBLUR_DiffuseSpecular_CopyStabilizedHistory.hlsli

@@ -13,7 +13,7 @@ NRD_EXPORT void NRD_CS_MAIN( uint2 pixelPos : SV_DispatchThreadId )
 {
     // Tile-based early out
     float isSky = gIn_Tiles[ pixelPos >> 4 ];
-    if( isSky != 0.0 )
+    if( isSky != 0.0 && !gIsRectChanged )
         return;
 
     // TODO: is it possible to introduce "CopyResource" in NRD API?

Shaders/Include/REBLUR/REBLUR_DiffuseSpecular_HistoryFix.hlsli

@@ -41,7 +41,7 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
     PRELOAD_INTO_SMEM_WITH_TILE_CHECK;
 
     // Tile-based early out
-    if( isSky != 0.0 )
+    if( isSky != 0.0 || pixelPos.x >= gRectSize.x || pixelPos.y >= gRectSize.y )
         return;
 
     // Early out
@@ -149,19 +149,23 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
                     float angle = STL::Math::AcosApprox( cosa );
 
                     // Accumulate
-                    float wd = IsInScreen( uv );
-                    wd *= CompareMaterials( materialID, materialIDs, gDiffMaterialMask );
-                    wd *= _ComputeWeight( NoX, diffGeometryWeightParams.x, diffGeometryWeightParams.y );
-                    wd *= _ComputeExponentialWeight( angle, diffNormalWeightParam, 0.0 );
+                    float w = IsInScreen( uv );
+                    w *= CompareMaterials( materialID, materialIDs, gDiffMaterialMask );
+                    w *= _ComputeWeight( NoX, diffGeometryWeightParams.x, diffGeometryWeightParams.y );
+                    w *= _ComputeExponentialWeight( angle, diffNormalWeightParam, 0.0 );
 
-                    REBLUR_TYPE d = gIn_Diff.SampleLevel( gNearestClamp, uvScaled, 0 );
+                    REBLUR_TYPE s = gIn_Diff.SampleLevel( gNearestClamp, uvScaled, 0 );
 
-                    diff += d * wd;
-                    sumd += wd;
+                    // Get rid of potentially bad values outside of the screen
+                    w = IsInScreen( uv ) ? w : 0.0; // no "!isnan" because "s" is not used for "w" calculations
+                    s = w != 0.0 ? s : 0.0;
+
+                    diff += s * w;
+                    sumd += w;
 
                     #ifdef REBLUR_SH
-                        float4 dh = gIn_DiffSh.SampleLevel( gNearestClamp, uvScaled, 0 );
-                        diffSh += dh * wd;
+                        float4 sh = gIn_DiffSh.SampleLevel( gNearestClamp, uvScaled, 0 );
+                        diffSh += sh * w;
                     #endif
                 }
             }
@@ -329,25 +333,29 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
                     float angle = STL::Math::AcosApprox( cosa );
 
                     // Accumulate
-                    float ws = IsInScreen( uv );
-                    ws *= CompareMaterials( materialID, materialIDs, gSpecMaterialMask );
-                    ws *= _ComputeWeight( NoX, specGeometryWeightParams.x, specGeometryWeightParams.y );
-                    ws *= _ComputeExponentialWeight( angle, specNormalWeightParam, 0.0 );
-                    ws *= _ComputeExponentialWeight( Ns.w * Ns.w, specRoughnessWeightParamsSq.x, specRoughnessWeightParamsSq.y );
+                    float w = IsInScreen( uv );
+                    w *= CompareMaterials( materialID, materialIDs, gSpecMaterialMask );
+                    w *= _ComputeWeight( NoX, specGeometryWeightParams.x, specGeometryWeightParams.y );
+                    w *= _ComputeExponentialWeight( angle, specNormalWeightParam, 0.0 );
+                    w *= _ComputeExponentialWeight( Ns.w * Ns.w, specRoughnessWeightParamsSq.x, specRoughnessWeightParamsSq.y );
 
                     REBLUR_TYPE s = gIn_Spec.SampleLevel( gNearestClamp, uvScaled, 0 );
 
                     // TODO: ideally "diffuseness at hit" needed...
                     // TODO: for roughness closer to REBLUR_HISTORY_FIX_BUMPED_ROUGHNESS "saturate( hitDistNormAtCenter - ExtractHitDist( s ) )" could be used.
                     // It allows bleeding of background to foreground, but not vice versa ( doesn't suit for 0 roughness )
-                    ws *= saturate( 1.0 - hitDistWeightScale * abs( ExtractHitDist( s ) - hitDistNormAtCenter ) / ( max( ExtractHitDist( s ), hitDistNormAtCenter ) + NRD_EPS ) );
+                    w *= saturate( 1.0 - hitDistWeightScale * abs( ExtractHitDist( s ) - hitDistNormAtCenter ) / ( max( ExtractHitDist( s ), hitDistNormAtCenter ) + NRD_EPS ) );
+
+                    // Get rid of potentially bad values outside of the screen
+                    w = ( IsInScreen( uv ) && !isnan( w ) ) ? w : 0.0;
+                    s = w != 0.0 ? s : 0.0;
 
-                    spec += s * ws;
-                    sums += ws;
+                    spec += s * w;
+                    sums += w;
 
                     #ifdef REBLUR_SH
                         float4 sh = gIn_SpecSh.SampleLevel( gNearestClamp, uvScaled, 0 );
-                        specSh += sh * ws;
+                        specSh += sh * w;
                     #endif
                 }
             }

Shaders/Include/REBLUR/REBLUR_DiffuseSpecular_HitDistReconstruction.hlsli

@@ -52,7 +52,7 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
     PRELOAD_INTO_SMEM_WITH_TILE_CHECK;
 
     // Tile-based early out
-    if( isSky != 0.0 )
+    if( isSky != 0.0 || pixelPos.x >= gRectSize.x || pixelPos.y >= gRectSize.y )
         return;
 
     // Early out

Shaders/Include/REBLUR/REBLUR_DiffuseSpecular_PostBlur.hlsli

@@ -16,7 +16,7 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
 
     // Tile-based early out
     float isSky = gIn_Tiles[ pixelPos >> 4 ];
-    if( isSky != 0.0 )
+    if( isSky != 0.0 || pixelPos.x >= gRectSize.x || pixelPos.y >= gRectSize.y )
     {
         // ~0 normal is needed to allow bilinear filter in TA ( 0 can't be used due to "division by zero" in "UnpackNormalRoughness" )
         gOut_Normal_Roughness[ pixelPos ] = PackNormalRoughness( 1.0 / 255.0 );
@@ -26,7 +26,7 @@ NRD_EXPORT void NRD_CS_MAIN( int2 threadPos : SV_GroupThreadId, int2 pixelPos :
 
     // Early out
     float viewZ = UnpackViewZ( gIn_ViewZ[ pixelPos ] );
-    if( viewZ > gDenoisingRange )
+    if( viewZ > gDenoisingRange || pixelPos.x >= gRectSize.x || pixelPos.y >= gRectSize.y )
     {
         // ~0 normal is needed to allow bilinear filter in TA ( 0 can't be used due to "division by zero" in "UnpackNormalRoughness" )
         gOut_Normal_Roughness[ pixelPos ] = PackNormalRoughness( 1.0 / 255.0 );