v2.2.0

2.2.0

Adds DXR Clusters support through the preview D3D12 Agility SDK. The D3D12 back-end can
build its cluster acceleration structures through the native DXR cluster API as well as
NVAPI, and shaders read cluster data through the Shader Model 6.10 ClusterID() and
TriangleObjectPositions() intrinsics. See
DXR Clusters.

DXR Clusters

• Built by default; CMake fetches the 1.722.0-preview Agility SDK from NuGet and the
  matching v1.10.2605.37 preview DXC. Running it needs NVIDIA driver 621.01 or later and
  Windows Developer Mode, and the sample falls back to NVAPI otherwise.
  -DNVRHI_D3D12_WITH_DXRCLUSTER=OFF builds against the retail Agility SDK instead
• A DXR cluster build keeps the NVAPI path: by default it uses DXR clusters when the
  device supports them and NVAPI otherwise. --cluster-api dxr|nvapi forces one, and the
  window title and log name the API in use. The same build still runs on Vulkan with -vk
• Known issue: on the 1.722 runtime, the NVAPI cluster intrinsics return invalid data
  under D3D12 GPU-based validation (-gd). With --cluster-api nvapi -gd the path
  tracing shader uses SM 6.10 DXR cluster intrinsics instead when Windows Developer Mode
  is on

Platform

• Build on Linux with the Vulkan back-end; DLSS is Windows-only, since Streamline ships
  Windows binaries only
• Mark the wave-dependent compute shaders as 32-lane waves, so D3D12 runs on any device
  that supports 32-lane waves
• Disable use of the NVAPI shader extensions in the path tracer on non-NVIDIA GPUs
• Update Donut and NVRHI

Subdivision Surfaces

• Instance subdivision meshes: an OBJ referenced by several .scene.json nodes loads
  once, and each node can offset its animation; adds the amy_instanced scene
• Adapt the fine and coarse tessellation rates to the tessellation budgets on the GPU,
  stepping down when cluster, CLAS or vertex memory load passes 85% and recovering below
  60%; on by default, --no-adaptive-tess-rate turns it off. The Adaptive readout names
  the limiting budget
• Hold the last keyframe at the end of an animation loop instead of blending it into
  frame 0

Cluster LOD

• Pack per-geometry LOD metadata into shared buffers and build it in parallel at scene
  load; Zorah's metadata build drops from 2465 to 406 ms
• Import a glTF referenced by several .scene.json nodes once
• Color the adaptive error readout by streaming load, matching the tessellation readout

UI and Profiling

• Improve tracking of VRAM usage in the VRAM Budget window: it counts occupied VRAM
  rather than packed sizes, and now tracks AS build scratch, DLSS and DLSS-RR, BLAS and
  TLAS, envmap sampling tables and subdivision surfaces
• Add a Shot List window to jump the camera between saved positions, save the current
  camera and run captures without exiting
• Edit the camera string and the field of view in the Camera section

Bug Fixes

• Vulkan screenshots were sRGB-encoded twice
• Applying a texture setting in the VRAM Budget window reset the Cluster LOD streaming
  budgets
• Vulkan overflowed its per-instance tessellation constants past ~5 subdivision instances
• Loading a scene without subdivision surfaces kept the previous scene's tessellation
  stats