Embree comparison, adopt its ideas (occlusion/watertight/instancing/SBVH/BVH4/masks/motion/curves), and extract the crust-rt kernel crate

[doubleailes]

Starts with a deep analysis of Embree 4.4.x vs crust's intersection layer (docs/embree_comparison.md), implements all seven of the doc's "adopt natively" recommendations in safe Rust, then completes the "openqmc move": the whole intersection layer extracted into a dedicated crust-rt crate behind an Embree-shaped, ID-based API.

The comparison

docs/embree_comparison.md analyzes Embree's geometry types, BVH builders, query API, and ISA/SYCL support against crust's intersection layer, documents what crust covers that Embree doesn't (the whole renderer above the query line), and ranks which Embree ideas were worth adopting. Its head-to-head table and addenda reflect the final state of this branch.

Adopted, one commit each

1. Dedicated occlusion query — early-exit, unordered traversal for NEE shadow rays (the rtcIntersect/rtcOccluded split).
2. Watertight triangles — Woop et al. 2013 replaces epsilon-based Möller–Trumbore: dominant-axis shear, 2D signed edge functions with an f64 fallback on exact-zero ties, division-free t range test. Pinned by 10k-sample shared-edge/shared-vertex pinhole tests.
3. Instancing — local-space mesh BVHs shared across prims with identical points/topology/material (content hash + memoized material Arcs); rays transform into local space, normals map back by inverse-transpose.
4. Parallel deterministic BVH build — subtrees above 4096 references build via rayon::join; split decisions are input-only, so the tree is byte-identical run-to-run (pinned by a build-twice test).
5. SBVH spatial splits — reference-based build (Stich et al. 2009): binned object SAH plus an α-gated binned spatial split with entry/exit counting; straddling references are clipped (exact Sutherland–Hodgman for triangles) and duplicated across children.
6. BVH4 — the binary tree collapses into 4-wide SoA nodes; traversal tests four child slabs per visit with Vec4 min/max (clamped reciprocals keep zero-direction lanes NaN-free), near-to-far lane ordering for closest-hit, early exit for occlusion.
7. Ray masks, motion blur, curves — crust:rayMask visibility categories (camera/shadow/indirect); crust:motion:translate transform motion blur (QMC shutter time on primary rays, inherited along the path; instances lerp their placement per ray with a conservative union-of-endpoints bbox); RoundCurveSegment sphere-swept cones with UsdGeomBasisCurves import (linear direct, cubic bezier/bspline/catmullRom flattened).

The crust-rt kernel crate

The intersection layer is now its own crate (the way openqmc-rs factored out sampling), deliberately Embree-shaped so Embree bindings could slot in behind the same seam:

• Geometry variants (triangle meshes with optional per-vertex shading normals, spheres, round curves, single-level instances with motion) attach to a SceneBuilder with per-geometry masks; commit() builds; Scene::intersect/Scene::occluded mirror rtcIntersect1/rtcOccluded1.
• Hits are plain Copy RayHits carrying geom_id/prim_id — the kernel never sees materials. Depends only on glam + rayon; 27 unit tests + doctest.
• crust-core consumes it through rt_world.rs: WorldBuilder::attach(geometry, material) -> geom_id and a committed World that resolves hits back to (HitRecord, &dyn Material, geom_id, prim_id). Light attribution moved from Arc-address identity to geom_id (LightList::find_by_geom) — cleaner and faster than pointer comparison. The old Hittable/HittableList/Bvh/primitive types are gone from crust-core.

Measured performance

Stress scene from scripts/gen_stress_scene.py (383k world-baked triangles: 36 identical 9.8k-tri spheres, a 30k thin-shard field, an occluder slab for shadow rays), 640×360, median of 3 runs:

build	spp=1 (parse+build+1 pass)	spp=16	render cost / spp
pre-adoption (021d8dd)	2.40 s	26.91 s	1.63 s
+ occlusion query, watertight (b4a48ab)	1.91 s	20.39 s	1.23 s
+ instancing, parallel SBVH (3736eb3)	1.02 s	8.33 s	0.49 s
+ BVH4 (3671527)	0.89 s	6.60 s	0.38 s
+ crust-rt extraction (HEAD)	—	6.46 s	—

4.1× end-to-end at 16 spp (4.3× on pure per-sample render cost, 2.7× on startup), with the final image matching the pre-adoption render to a mean abs pixel difference of 1.5e-5, and the kernel extraction adding zero overhead (pixel-identical renders, unchanged wall time). Peak RSS rose from ~78 MB to ~96 MB on this scene (build-transient allocations). On the small bundled samples (shading-bound) wall time is unchanged.

Verification

• 125 tests green across the workspace (27 kernel + 76 core + 10 USD integration + others), including new integration tests over two new sample scenes (samples/curves.usda, samples/motionblur.usda) asserting exact strand-hit distances, shutter-time hits/misses, and mask filtering.
• Cornellbox and the stress scene render pixel-identical before/after the kernel extraction; curves/motionblur renders visually confirmed (streaks, camera-invisible shadow card, hair strands).
• Zero compiler warnings; benches compile.

Known caveats (documented in CLAUDE.md)

Motion blur is transform-only with linear matrix lerp (no deformation/quaternion motion); cubic curves are flattened rather than intersected exactly; mesh sharing requires identical geometry and material — UsdGeomPointInstancer / instanceable arcs are not consumed; emissive curves/instances are not light-list entries. world.count() now counts geometries (a rect light's two triangles are one mesh geometry).

🤖 Generated with Claude Code

https://claude.ai/code/session_011XTp9J7yinpSvfjX5tmh4F

[qodo-code-review]

Qodo reviews are paused for this user.

Troubleshooting steps vary by plan Learn more →

On a Teams plan?
Reviews resume once this user has a paid seat and their Git account is linked in Qodo.
Link Git account →

Using GitHub Enterprise Server, GitLab Self-Managed, or Bitbucket Data Center?
These require an Enterprise plan - Contact us
Contact us →