NRD SAMPLE

All-in-one repository including all relevant pieces to see NRD (NVIDIA Real-time Denoisers) in action. The sample is cross-platform, it's based on NRI (NVIDIA Rendering Interface) to bring cross-GraphicsAPI support.

NRD sample is a land for high performance path tracing for games. Some features to highlight:

• minimalistic path tracer utilizing Trace Ray Inline
• HALF resolution (checkerboard), FULL resolution and FULL resolution tracing with PROBABILISTIC diffuse / specular selection at the primary hit
• NRD denoising, including occlusion-only and spherical harmonics / gaussian modes
• overhead-free multi-bounce propagation (even in case of a single bounce) based on reusing the previously denoised frame
• reference accumulation
• several rays per pixel and bounces
• realistic glass with multi-bounce reflections and refractions
• physically based ambient estimation using RT
• mip level calculation
• curvature estimation

A NOTE ABOUT THE TRACER

The path tracer in the sample has been designed to respect performance. Instead of using a commonly used solution, which in general looks like:

// Resources
ByteAddressBuffer g_BindlessBuffers[];
Texture2D g_BindlessTextures[];
StructuredBuffer<InstanceData> g_InstanceData;
StructuredBuffer<GeometryData> g_GeometryData;
StructuredBuffer<MaterialData> g_MaterialData;

// Geometry fetching
uint instanceIndex = rayQuery.InstanceIndex();
uint geometryIndex = rayQuery.GeometryIndex();
uint primitiveIndex = rayQuery.PrimitiveIndex();

InstanceData instanceData = g_InstanceData[ instanceIndex ];
GeometryData geometryData = g_GeometryData[ instanceData.geometryBaseIndex + geometryIndex ];

ByteAddressBuffer indexBuffer = g_BindlessBuffers[ NonUniformResourceIndex( geometryData.indexBufferIndex ) ];
ByteAddressBuffer vertexBuffer = g_BindlessBuffers[ NonUniformResourceIndex( geometryData.vertexBufferIndex ) ];

uint3 indices = indexBuffer.Load3( geometryData.indexOffset + primitiveIndex * INDEX_STRIDE );

float3 p0 = DecodePosition( vertexBuffer.Load3( geometryData.vertexOffset + indices[0] * VERTEX_STRIDE ) );
float3 p1 = DecodePosition( vertexBuffer.Load3( geometryData.vertexOffset + indices[1] * VERTEX_STRIDE ) );
float3 p2 = DecodePosition( vertexBuffer.Load3( geometryData.vertexOffset + indices[2] * VERTEX_STRIDE ) );
float3 p = Interpolate( p0, p1, p2, barycentrics );

float3 n0 = DecodeNormal( vertexBuffer.Load3( geometryData.vertexOffset + offset1 + indices[0] * VERTEX_STRIDE ) );
float3 n1 = DecodeNormal( vertexBuffer.Load3( geometryData.vertexOffset + offset1 + indices[1] * VERTEX_STRIDE ) );
float3 n2 = DecodeNormal( vertexBuffer.Load3( geometryData.vertexOffset + offset1 + indices[2] * VERTEX_STRIDE ) );
float3 n = Interpolate( n0, n1, n2, barycentrics );
n = Rotate( instanceData.transform );

float2 uv0 = DecodeUv( vertexBuffer.Load2( geometryData.vertexOffset + offset2 + indices[0] * VERTEX_STRIDE ) );
float2 uv1 = DecodeUv( vertexBuffer.Load2( geometryData.vertexOffset + offset2 + indices[1] * VERTEX_STRIDE ) );
float2 uv2 = DecodeUv( vertexBuffer.Load2( geometryData.vertexOffset + offset2 + indices[2] * VERTEX_STRIDE ) );
float2 uv = Interpolate( uv0, uv1, uv2, barycentrics );

// Material fetching
MaterialData materialData = g_MaterialData[ geometryData.materialIndex ];

Texture2D texture1 = g_BindlessTextures[ NonUniformResourceIndex( materialData.textureIndex1 ) ];
float4 data1 = texture1.SampleLevel( ... );

Texture2D texture2 = g_BindlessTextures[ NonUniformResourceIndex( materialData.textureIndex2 ) ];
float4 data2 = texture2.SampleLevel( ... );

Texture2D texture3 = g_BindlessTextures[ NonUniformResourceIndex( materialData.textureIndex3 ) ];
float4 data1 = texture3.SampleLevel( ... );

To get a vertex data we need:

• fetch the vertex data per element through 4 indirections
• vertex position is interpolated despite that it's already in BVH
• in our case to fetch all vertex data we need to do 14 HLSL fetches

The path tracer uses the following scheme:

// Resources
StructuredBuffer<InstanceData>, g_InstanceData;
StructuredBuffer<PrimitiveData> g_PrimitiveData;
Texture2D g_BindlessTextures[];

// Geometry fetching
uint instanceIndex = rayQuery.InstanceIndex();
uint geometryIndex = rayQuery.GeometryIndex();
uint primitiveIndex = rayQuery.PrimitiveIndex();

InstanceData instanceData = g_InstanceData[ instanceIndex + geometryIndex ];
PrimitiveData primitiveData = g_PrimitiveData[ primitiveIndex ];

float3x3 mObjectToWorld = (float3x3)rayQuery.ObjectToWorld3x4();
if( instanceData.isStatic )
    mObjectToWorld = (float3x3)instanceData.mWorldToWorldPrev;

float3 p = rayOrigin + rayDirection * rayQuery.RayT();

float3 n0 = DecodeNormal( primitiveData.n0 );
float3 n1 = DecodeNormal( primitiveData.n1 );
float3 n2 = DecodeNormal( primitiveData.n2 );
float3 n = Interpolate( n0, n1, n2, barycentrics );
n = Rotate( mObjectToWorld );

float2 uv0 = DecodeUv( primitiveData.uv0 );
float2 uv1 = DecodeUv( primitiveData.uv1 );
float2 uv2 = DecodeUv( primitiveData.uv2 );
float2 uv = Interpolate( uv0, uv1, uv2, barycentrics );

// Material fetching
Texture2D texture1 = g_BindlessTextures[ NonUniformResourceIndex( instanceData.textureBaseIndex ) ];
float4 data1 = texture1.SampleLevel( ... );

Texture2D texture2 = g_BindlessTextures[ NonUniformResourceIndex( instanceData.textureBaseIndex + 1 ) ];
float4 data2 = texture2.SampleLevel( ... );

Texture2D texture3 = g_BindlessTextures[ NonUniformResourceIndex( instanceData.textureBaseIndex + 2 ) ];
float4 data1 = texture3.SampleLevel( ... );

To get a vertex data we need:

• fetch the primitive data explicitly without indirections
• interpolate vertex elements using primitive data
• in our case to fetch all vertex data we need to do 2 HLSL fetches

This approach simplifies and accelerates ray tracing, but adds difficulties to BVH management. Deleting a BLAS adds a contiguous region of free elements in g_PrimitiveData, which needs to be tracked and potentially re-used in the future when a suitable object appears. If estimated geometry sizes are known, this memory-fragmentation-free approach is more than applicable.

BUILD INSTRUCTIONS

• Install Cmake 3.15+
• Install on
  • Windows: latest WindowsSDK and VulkanSDK
  • Linux (x86-64): latest VulkanSDK, libx11-dev, libxrandr-dev, libwayland-dev
  • Linux (aarch64): find a precompiled binary for DXC, libx11-dev, libxrandr-dev, libwayland-dev
• Build (variant 1) - using Git and CMake explicitly
  • Clone project and init submodules
  • Generate and build project using CMake
• Build (variant 2) - by running scripts:
  • Run 1-Deploy
  • Run 2-Build

CMAKE OPTIONS

• USE_MINIMAL_DATA=ON - download minimal resource package (90MB)
• DISABLE_SHADER_COMPILATION=ON - disable compilation of shaders (shaders can be built on other platform)
• DXC_CUSTOM_PATH=custom/path/to/dxc - custom path to DXC (will be used if VulkanSDK is not found)
• USE_DXC_FROM_PACKMAN_ON_AARCH64=OFF - use default path for DXC

HOW TO RUN

• Run 3-Run NRD sample script and answer the cmdline questions to set the runtime parameters
• If Smart Command Line Arguments extension for Visual Studio is installed, all command line arguments will be loaded into corresponding window
• The executables can be found in _Bin. The executable loads resources from _Data, therefore please run the samples with working directory set to the project root folder (needed pieces of the command line can be found in 3-Run NRD sample script)

REQUIREMENTS

Any ray tracing compatible GPU.

USAGE

• Right mouse button + W/S/A/D - move camera
• Mouse scroll - accelerate / decelerate
• F1 - toggle "gDebug" (can be useful for debugging and experiments)
• F2 - go to next test (only if TESTS section is unfolded)
• F3 - toggle emission
• Tab - UI toggle
• Space - animation toggle
• PgUp/PgDown - switch between denoisers

By default NRD is used in common mode. But it can also be used in occlusion-only (including directional) and SH (spherical harmonics) modes in the sample. To change the behavior NRD_MODE macro needs to be changed from NORMAL to OCCLUSION, SH or DIRECTIONAL_OCCLUSION in Shared.hlsli.

Notes:

• RELAX doesn't support AO / SO denoising. If RELAX is the current denoiser, ambient term will be flat, but energy correct.