NRD v3.4.0:

HIGHLIGHTS:

- NRD: anisotropic curvature estimation
- REBLUR/RELAX: improved specular reprojection on curved surfaces
- REBLUR: added fast history
- REBLUR: introduced denoisers working in spherical harmonic (SH) space
- REBLUR: reduced memory usage
- REBLUR: temporal stabilization pass can be turned off eliminating associated overhead
- REBLUR: specular tracking improvements
- REBLUR: better history reconstruction in disoccluded regions
- RELAX: improved performance
- RELAX: improved anti-firefly filtering
- RELAX: improved history confidence usage

DETAILS:

- NRD: anisotropic curvature estimation
- NRD: exposed helper "GetMethodString"
- NRD: gDebug is now available in Common.hlsli
- NRD: updated README and UPDATE

- REBLUR: added SH denoisers
- REBLUR: switched to YCoCg
- REBLUR: added fast history
- REBLUR: rewritten HistoryFix (no mips anymore)
- REBLUR: reduced memory usage
- REBLUR: improved error compensation and reduced bias
- REBLUR: fix for checkerboard resolve in advanced prepass mode
- REBLUR: removed virtual history amount correction which is not needed anymore
- REBLUR: reduced bias in pre-pass
- REBLUR: fixed laggy specular reprojection on bumpy surfaces
- REBLUR: specular confidence doesn't affect hit distance weight
- REBLUR: specular prepass setting should not affect diffuse only denoiser (and vice versa)
- REBLUR: improved reprojection on curved surfaces
- REBLUR: simplified temporal stabilization params
- REBLUR: fixed wrong output in DIFFUSE_DIRECTIONAL_OCCLUSION
- REBLUR: simplified roughness weight in temporal accumulation pass
- REBLUR: removed gFramerateScale based acceleration
- REBLUR: tuned "GetSpecAccumSpeed" (consistent parallax usage)
- REBLUR: tuned prepass parameters
- REBLUR: fixed & optimized hitT for specular tracking
- REBLUR: fixed (corrected) curvature if its sign is oscillating
- REBLUR: fixed (returned) virtual motion based normal weight in temporal accumulation
- REBLUR: added prev-prev test for roughness
- REBLUR: improved firefly suppression
- REBLUR: removed clamping to noisy input for specular
- REBLUR: dropped unnecessary output for sky pixels
- REBLUR: tweaked surface based motion accumulation
- REBLUR: hitT check for virtual motion in temporal accumulation replaced with parallax check
- REBLUR: improved prev-prev test
- REBLUR: prev-prev test checks surface similarity
- REBLUR: performance optimizations
- REBLUR: tweaked hit distance weight
- REBLUR: fixed regressions
- REBLUR: improved virtual motion roughness confidence in temporal accumulation
- REBLUR: maximized code reuse
- REBLUR: refactoring

- RELAX: removed geometry weights in firefly filter
- RELAX: switched from disocclusion threshold to spatial filtering threshold in disocclusion fix
- RELAX: added prev-prev test for roughness
- RELAX: improved confidence buffers handling
- RELAX: performance optimizations

- NRD integration: now computes memory usage

Author: dzhdanNV

External/MathLib

@@ -11,18 +11,15 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 /*
 CREDITS:
     Developed by:
-        Library, REBLUR and SIGMA denoisers:
-            Dmitry Zhdan ([email protected])
-
-        ReLAX denoiser:
-            Tim Cheblokov ([email protected])
-            Pawel Kozlowski ([email protected])
+        Dmitry Zhdan ([email protected])
+        Tim Cheblokov ([email protected])
 
     Special thanks:
-        Evgeny Makarov (NVIDIA) - denoising ideas
-        Ivan Fedorov (NVIDIA) - interface
-        Ivan Povarov (NVIDIA) - QA, integrations and feedback
-        Oles Shyshkovtsov (4A GAMES) - initial idea of recurrent blurring
+        Pawel Kozlowski (NVIDIA)
+        Evgeny Makarov (NVIDIA)
+        Ivan Fedorov (NVIDIA)
+        Ivan Povarov (NVIDIA)
+        Oles Shyshkovtsov (4A GAMES) for initial idea of recurrent blurring
 */
 
 #pragma once
@@ -31,9 +28,9 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 #include <cstddef>
 
 #define NRD_VERSION_MAJOR 3
-#define NRD_VERSION_MINOR 3
-#define NRD_VERSION_BUILD 1
-#define NRD_VERSION_DATE "16 June 2022"
+#define NRD_VERSION_MINOR 4
+#define NRD_VERSION_BUILD 0
+#define NRD_VERSION_DATE "14 July 2022"
 
 #if defined(_MSC_VER)
     #define NRD_CALL __fastcall
@@ -62,5 +59,8 @@ namespace nrd
     NRD_API Result NRD_CALL SetMethodSettings(Denoiser& denoiser, Method method, const void* methodSettings);
     NRD_API Result NRD_CALL GetComputeDispatches(Denoiser& denoiser, const CommonSettings& commonSettings, const DispatchDesc*& dispatchDescs, uint32_t& dispatchDescNum);
     NRD_API void NRD_CALL DestroyDenoiser(Denoiser& denoiser);
+
+    // Helpers
     NRD_API const char* GetResourceTypeString(ResourceType resourceType);
+    NRD_API const char* GetMethodString(Method method);
 }

Include/NRD.h

@@ -11,7 +11,7 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 #pragma once
 
 #define NRD_DESCS_VERSION_MAJOR 3
-#define NRD_DESCS_VERSION_MINOR 3
+#define NRD_DESCS_VERSION_MINOR 4
 
 static_assert (NRD_VERSION_MAJOR == NRD_DESCS_VERSION_MAJOR && NRD_VERSION_MINOR == NRD_DESCS_VERSION_MINOR, "Please, update all NRD SDK files");
 
@@ -44,6 +44,11 @@ namespace nrd
         // OUTPUTS - OUT_DIFF_HITDIST
         REBLUR_DIFFUSE_OCCLUSION,
 
+        // INPUTS - IN_MV, IN_NORMAL_ROUGHNESS, IN_VIEWZ, IN_DIFF_SH0, IN_DIFF_SH1
+        // OPTIONAL INPUTS - IN_DIFF_CONFIDENCE
+        // OUTPUTS - OUT_DIFF_SH0, OUT_DIFF_SH1
+        REBLUR_DIFFUSE_SH,
+
         // INPUTS - IN_MV, IN_NORMAL_ROUGHNESS, IN_VIEWZ, IN_SPEC_RADIANCE_HITDIST,
         // OPTIONAL INPUTS - IN_SPEC_DIRECTION_PDF, IN_SPEC_CONFIDENCE
         // OUTPUTS - OUT_SPEC_RADIANCE_HITDIST
@@ -53,6 +58,11 @@ namespace nrd
         // OUTPUTS - OUT_SPEC_HITDIST
         REBLUR_SPECULAR_OCCLUSION,
 
+        // INPUTS - IN_MV, IN_NORMAL_ROUGHNESS, IN_VIEWZ, IN_SPEC_SH0, IN_SPEC_SH1
+        // OPTIONAL INPUTS - IN_SPEC_CONFIDENCE
+        // OUTPUTS - OUT_SPEC_SH0, OUT_SPEC_SH1
+        REBLUR_SPECULAR_SH,
+
         // INPUTS - IN_MV, IN_NORMAL_ROUGHNESS, IN_VIEWZ, IN_DIFF_RADIANCE_HITDIST, IN_SPEC_RADIANCE_HITDIST,
         // OPTIONAL INPUTS - IN_DIFF_DIRECTION_PDF, IN_SPEC_DIRECTION_PDF, IN_DIFF_CONFIDENCE,  IN_SPEC_CONFIDENCE
         // OUTPUTS - OUT_DIFF_RADIANCE_HITDIST, OUT_SPEC_RADIANCE_HITDIST
@@ -62,6 +72,11 @@ namespace nrd
         // OUTPUTS - OUT_DIFF_HITDIST, OUT_SPEC_HITDIST
         REBLUR_DIFFUSE_SPECULAR_OCCLUSION,
 
+        // INPUTS - IN_MV, IN_NORMAL_ROUGHNESS, IN_VIEWZ, IN_DIFF_SH0, IN_DIFF_SH1, IN_SPEC_SH0, IN_SPEC_SH1
+        // OPTIONAL INPUTS - IN_DIFF_CONFIDENCE,  IN_SPEC_CONFIDENCE
+        // OUTPUTS - OUT_DIFF_SH0, OUT_DIFF_SH1, OUT_SPEC_SH0, OUT_SPEC_SH1
+        REBLUR_DIFFUSE_SPECULAR_SH,
+
         // INPUTS - IN_MV, IN_NORMAL_ROUGHNESS, IN_VIEWZ, IN_DIFF_DIRECTION_HITDIST,
         // OPTIONAL INPUTS - IN_DIFF_DIRECTION_PDF, IN_DIFF_CONFIDENCE
         // OUTPUTS - OUT_DIFF_DIRECTION_HITDIST
@@ -122,53 +137,64 @@ namespace nrd
     enum class ResourceType : uint32_t
     {
         //=============================================================================================================================
-        // INPUTS
+        // COMMON INPUTS
         //=============================================================================================================================
 
         // 3D world space motion (RGBA16f+) or 2D screen space motion (RG16f+), MVs must be non-jittered, MV = previous - current
         IN_MV,
 
-        // Data must match encoding / decoding in "NRD_FrontEnd_PackNormalAndRoughness" / "NRD_FrontEnd_UnpackNormalAndRoughness" (RGBA8+ or R10G10B10A2+)
+        // Data must match encoding in "NRD_FrontEnd_PackNormalAndRoughness" and "NRD_FrontEnd_UnpackNormalAndRoughness" (RGBA8+)
         IN_NORMAL_ROUGHNESS,
 
         // Linear view depth for primary rays (R16f+)
         IN_VIEWZ,
 
-        // RELAX: Data must be packed using "RELAX_FrontEnd_PackRadianceAndHitDist"
-        // REBLUR: Data must be packed using "REBLUR_FrontEnd_PackRadianceAndHitDist" (and "REBLUR_FrontEnd_GetNormHitDist" for hit distance) (RGBA16f+)
-        // Noisy and weighted sum of hit distances along a path (using NRD_GetCorrectedHitDist), excluding primary hit distance.
-        // REBLUR works with normalized hit distances. It's recommended to use REBLUR_FrontEnd_GetNormHitDist for hit distance normalization.
-        // Normalization parameters should be passed into NRD as HitDistanceParameters for diffuse and specular separately for internal
-        // hit distance denormalization.
+        //=============================================================================================================================
+        // INPUTS
+        //=============================================================================================================================
+
+        // Noisy radiance and hit distance (RGBA16f+)
+        //      REBLUR: use "REBLUR_FrontEnd_PackRadianceAndNormHitDist" for encoding
+        //      RELAX: use "RELAX_FrontEnd_PackRadianceAndHitDist" for encoding
         IN_DIFF_RADIANCE_HITDIST,
         IN_SPEC_RADIANCE_HITDIST,
 
-        // REBLUR: Data must be packed using "REBLUR_FrontEnd_GetNormHitDist" (R8+)
+        // Noisy hit distance (R8+)
+        //      REBLUR: use "REBLUR_FrontEnd_GetNormHitDist" for encoding
         IN_DIFF_HITDIST,
         IN_SPEC_HITDIST,
 
-        // REBLUR: Data must be packed using "REBLUR_FrontEnd_PackDirectionAndHitDist" (RGBA8+)
+        // Noisy bent normal and normalized hit distance (RGBA8+)
+        //      REBLUR: use "REBLUR_FrontEnd_PackDirectionAndNormHitDist" for encoding
         IN_DIFF_DIRECTION_HITDIST,
 
-        // (Optional) Data must be packed using "NRD_FrontEnd_PackDirectionAndPdf" (RGBA8+)
-        // These inputs are needed only if PrePassMode::ADVANCED is used. The data can be averaged or weighted in case of many RPP. Encoding / decoding
-        // can be changed in "NRD_FrontEnd_PackDirectionAndPdf" and "NRD_FrontEnd_UnpackDirectionAndPdf" functions.
+        // Noisy SH data (2x RGBA16f+)
+        //      REBLUR: use "REBLUR_FrontEnd_PackSh" for encoding
+        IN_DIFF_SH0,
+        IN_DIFF_SH1,
+        IN_SPEC_SH0,
+        IN_SPEC_SH1,
+
+        // (Optional) Ray direction and sample PDF (RGBA8+)
+        // These inputs are needed only for PrePassMode::ADVANCED
+        //      REBLUR: use "NRD_FrontEnd_PackDirectionAndPdf" for encoding
         IN_DIFF_DIRECTION_PDF,
         IN_SPEC_DIRECTION_PDF,
 
         // (Optional) User-provided history confidence in range 0-1, i.e. antilag (R8+)
-        // Used only if "CommonSettings::isHistoryConfidenceInputsAvailable = true".
+        // Used only if "CommonSettings::isHistoryConfidenceInputsAvailable = true"
         IN_DIFF_CONFIDENCE,
         IN_SPEC_CONFIDENCE,
 
-        // SIGMA: Data must be packed using "SIGMA_FrontEnd_PackShadow" (3/5-args) (RG16+/RGBA8+)
+        // Noisy shadow data and optional translucency (RG16f+ and RGBA8+ for optional translucency)
+        //      SIGMA: use "SIGMA_FrontEnd_PackShadow" for encoding
         IN_SHADOWDATA,
         IN_SHADOW_TRANSLUCENCY,
 
-        // Any signal
+        // Noisy signal (R8+)
         IN_RADIANCE,
 
-        // Primary and secondary world space positions
+        // Primary and secondary world space positions (RGBA16f+)
         IN_DELTA_PRIMARY_POS,
         IN_DELTA_SECONDARY_POS,
 
@@ -178,37 +204,38 @@ namespace nrd
 
         // IMPORTANT: These textures can potentially be used as history buffers!
 
-        // REBLUR: Data must be unpacked using "REBLUR_BackEnd_UnpackRadianceAndHitDist"
-        //   .xyz - radiance, .w - normalized hit distance (RGBA16f+)
-        // RELAX: Data must be unpacked using "RELAX_BackEnd_UnpackRadianceAndHitDist"
-        //   .xyz - radiance (R11G11B10f+)
+        // Denoised radiance and hit distance
+        //      REBLUR: use "REBLUR_BackEnd_UnpackRadianceAndNormHitDist" for decoding (RGBA16f+)
+        //      RELAX: use "RELAX_BackEnd_UnpackRadiance" for decoding (R11G11B10f+)
         OUT_DIFF_RADIANCE_HITDIST,
         OUT_SPEC_RADIANCE_HITDIST,
 
-        // .x - normalized hit distance (R8+)
+        // Denoised SH data
+        //      REBLUR: use "REBLUR_BackEnd_UnpackSh" for decoding (2x RGBA16f+)
+        OUT_DIFF_SH0,
+        OUT_DIFF_SH1,
+        OUT_SPEC_SH0,
+        OUT_SPEC_SH1,
+
+        // Denoised normalized hit distance (R8+)
         OUT_DIFF_HITDIST,
         OUT_SPEC_HITDIST,
 
-        // .xyz - direction, .w - normalized hit distance (RGBA8+)
+        // Denoised bent normal and normalized hit distance (RGBA8+)
+        //      REBLUR: use "REBLUR_BackEnd_UnpackDirectionAndNormHitDist" for decoding
         OUT_DIFF_DIRECTION_HITDIST,
 
-        // SIGMA: Data must be unpacked using "SIGMA_BackEnd_UnpackShadow"
-        //   .x - shadow, .yzw - translucency (RGBA8+)
-        //   .x - shadow (R8+)
-        //   Usage:
-        //      shadowData = SIGMA_BackEnd_UnpackShadow( shadowData );
-        //      float3 finalShadowCommon = lerp( shadowData.yzw, 1.0, shadowData.x ); // or
-        //      float3 finalShadowExotic = shadowData.yzw * shadowData.x; // or
-        //      float3 finalShadowMoreExotic = shadowData.yzw;
+        // Denoised shadow and optional transcluceny (R8+ or RGBA8+)
+        //      SIGMA: use "SIGMA_BackEnd_UnpackShadow" for decoding
         OUT_SHADOW_TRANSLUCENCY,
 
-        // .xyz - accumulated signal
+        // Denoised signal
         OUT_RADIANCE,
 
-        // .xy - 2D screen space virtual motion (RG16f+), MV = previous - current
+        // 2D screen space specular motion (RG16f+), MV = previous - current
         OUT_REFLECTION_MV,
 
-        // .xy - 2D screen space motion (RG16f+), MV = previous - current
+        // 2D screen space refraction motion (RG16f+), MV = previous - current
         OUT_DELTA_MV,
 
         //=============================================================================================================================

Include/NRDDescs.h

@@ -11,7 +11,7 @@ license agreement from NVIDIA CORPORATION is strictly prohibited.
 #pragma once
 
 #define NRD_SETTINGS_VERSION_MAJOR 3
-#define NRD_SETTINGS_VERSION_MINOR 3
+#define NRD_SETTINGS_VERSION_MINOR 4
 
 static_assert (NRD_VERSION_MAJOR == NRD_SETTINGS_VERSION_MAJOR && NRD_VERSION_MINOR == NRD_SETTINGS_VERSION_MINOR, "Please, update all NRD SDK files");
 
@@ -109,7 +109,7 @@ namespace nrd
         // (units) > 0 - use TLAS or tracing range (max value = NRD_FP16_MAX / NRD_FP16_VIEWZ_SCALE - 1 = 524031)
         float denoisingRange = 500000.0f;
 
-        // (normalized %)
+        // (normalized %) - if relative distance difference is greater than threshold, history gets reset (0.25-1.5% works well)
         float disocclusionThreshold = 0.005f;
 
         // [0; 1] - enables "noisy input / denoised output" comparison
@@ -191,7 +191,7 @@ namespace nrd
         float sensitivityToDarkness = 0.0f; // IMPORTANT: 0 is a bad default
 
         // Ideally, must be enabled, but since "sensitivityToDarkness" requires fine tuning from the app side it is disabled by default
-        bool enable = false;
+        bool enable = false; // IMPORTANT: doesn't affect "occlusion" denoisers
     };
 
     struct AntilagHitDistanceSettings
@@ -220,7 +220,10 @@ namespace nrd
         AntilagHitDistanceSettings antilagHitDistanceSettings = {};
 
         // [0; REBLUR_MAX_HISTORY_FRAME_NUM] - maximum number of linearly accumulated frames (= FPS * "time of accumulation")
-        uint32_t maxAccumulatedFrameNum = 31;
+        uint32_t maxAccumulatedFrameNum = 30;
+
+        // [0; REBLUR_MAX_HISTORY_FRAME_NUM] - maximum number of linearly accumulated frames in fast history
+        uint32_t maxFastAccumulatedFrameNum = 6;
 
         // (pixels) - pre-accumulation spatial reuse pass blur radius (0 = disabled, must be used in case of probabilistic sampling)
         float diffusePrepassBlurRadius = 30.0f;
@@ -253,9 +256,6 @@ namespace nrd
         // (normalized %) - represents maximum allowed deviation from local tangent plane
         float planeDistanceSensitivity = 0.005f;
 
-        // (normalized %) - adds a portion of input to the output of spatial passes
-        float inputMix = 0.0f;
-
         // If not OFF and used for DIFFUSE_SPECULAR, defines diffuse orientation, specular orientation is the opposite
         CheckerboardMode checkerboardMode = CheckerboardMode::OFF;
 
@@ -301,12 +301,12 @@ namespace nrd
         float specularPrepassBlurRadius = 50.0f;
 
         // [0; RELAX_MAX_HISTORY_FRAME_NUM] - maximum number of linearly accumulated frames ( = FPS * "time of accumulation")
-        uint32_t diffuseMaxAccumulatedFrameNum = 31;
-        uint32_t specularMaxAccumulatedFrameNum = 31;
+        uint32_t diffuseMaxAccumulatedFrameNum = 30;
+        uint32_t specularMaxAccumulatedFrameNum = 30;
 
         // [0; RELAX_MAX_HISTORY_FRAME_NUM] - maximum number of linearly accumulated frames in fast history
-        uint32_t diffuseMaxFastAccumulatedFrameNum = 8;
-        uint32_t specularMaxFastAccumulatedFrameNum = 8;
+        uint32_t diffuseMaxFastAccumulatedFrameNum = 6;
+        uint32_t specularMaxFastAccumulatedFrameNum = 6;
 
         // A-trous edge stopping Luminance sensitivity
         float diffusePhiLuminance = 2.0f;
@@ -350,6 +350,11 @@ namespace nrd
         // (normalized %) - A-trous edge stopping depth threshold
         float depthThreshold = 0.01f;
 
+        // Confidence inputs can affect spatial blurs, relaxing some weights in areas with low confidence
+        float confidenceDrivenRelaxationMultiplier = 0.0f;
+        float confidenceDrivenLuminanceEdgeStoppingRelaxation = 0.0f;
+        float confidenceDrivenNormalEdgeStoppingRelaxation = 0.0f;
+
         // How much we relax roughness based rejection in areas where specular reprojection is low
         float luminanceEdgeStoppingRelaxation = 0.5f;
         float normalEdgeStoppingRelaxation = 0.3f;
@@ -384,8 +389,8 @@ namespace nrd
     {
         float prepassBlurRadius = 0.0f;
 
-        uint32_t diffuseMaxAccumulatedFrameNum = 31;
-        uint32_t diffuseMaxFastAccumulatedFrameNum = 8;
+        uint32_t diffuseMaxAccumulatedFrameNum = 30;
+        uint32_t diffuseMaxFastAccumulatedFrameNum = 6;
 
         float diffusePhiLuminance = 2.0f;
         float diffuseLobeAngleFraction = 0.5f;
@@ -401,6 +406,10 @@ namespace nrd
         float minLuminanceWeight = 0.0f;
         float depthThreshold = 0.01f;
 
+        float confidenceDrivenRelaxationMultiplier = 0.0f;
+        float confidenceDrivenLuminanceEdgeStoppingRelaxation = 0.0f;
+        float confidenceDrivenNormalEdgeStoppingRelaxation = 0.0f;
+
         CheckerboardMode checkerboardMode = CheckerboardMode::OFF;
         HitDistanceReconstructionMode hitDistanceReconstructionMode = HitDistanceReconstructionMode::OFF;
 
@@ -415,8 +424,8 @@ namespace nrd
     {
         float prepassBlurRadius = 50.0f;
 
-        uint32_t specularMaxAccumulatedFrameNum = 31;
-        uint32_t specularMaxFastAccumulatedFrameNum = 8;
+        uint32_t specularMaxAccumulatedFrameNum = 30;
+        uint32_t specularMaxFastAccumulatedFrameNum = 6;
 
         float specularPhiLuminance = 1.0f;
         float diffuseLobeAngleFraction = 0.5f;
@@ -437,6 +446,10 @@ namespace nrd
         float minLuminanceWeight = 0.0f;
         float depthThreshold = 0.01f;
 
+        float confidenceDrivenRelaxationMultiplier = 0.0f;
+        float confidenceDrivenLuminanceEdgeStoppingRelaxation = 0.0f;
+        float confidenceDrivenNormalEdgeStoppingRelaxation = 0.0f;
+
         float luminanceEdgeStoppingRelaxation = 0.5f;
         float normalEdgeStoppingRelaxation = 0.3f;
         float roughnessEdgeStoppingRelaxation = 0.3f;
@@ -456,7 +469,7 @@ namespace nrd
     struct ReferenceSettings
     {
         // (>= 0) - maximum number of linearly accumulated frames ( = FPS * "time of accumulation")
-        uint32_t maxAccumulatedFrameNum = 3600;
+        uint32_t maxAccumulatedFrameNum = 1024;
     };
 
     // SPECULAR_REFLECTION_MV