bvh-build-traversallisted
Install: claude install-skill rustcircle/c101
# BVH build and traversal
Rule zero: report SAH cost, box and triangle tests per ray and Mrays/s on a named machine before and
after every builder change, and assert identical hits (count and sum of t) across builders. Tested
harness and recorded runs: [references/bvh-benchmark.md](references/bvh-benchmark.md).
## 1. SAH cost model and quality metrics
```text
split cost C = C_trav + (A_L/A) N_L C_isect + (A_R/A) N_R C_isect leaf cost C_isect N
tree cost C_tree = Σ_inner (A_n/A_root) C_trav + Σ_leaf (A_l/A_root) N_l C_isect
```
- A is surface area. C_trav/C_isect ≈ 1 for scalar binary nodes (the benchmark uses 1/1), lower with
SIMD triangle tests, higher for wide nodes; state the constants next to every reported cost.
- Leaf rule: stop when C_isect N ≤ the best split cost or N ≤ 1-2, but force a split above a cap (16
here) so a degenerate SAH estimate cannot create huge leaves.
- Report: SAH cost, end-point overlap (EPO, Aila-Karras-Laine 2013, which predicts traversal cost better
than SAH when boxes overlap), box and triangle tests per ray for primary and diffuse rays, Mrays/s,
build ms, nodes and bytes, max depth, and a node-visit heatmap per builder (PGM or EXR).
## 2. Builder decision table
| Need | Builder | Notes |
|---|---|---|
| production CPU tracing | Embree 4 (`RTC_BUILD_QUALITY_HIGH` uses spatial splits) or tinybvh | do not ship a hand-rolled builder |
| static scene, offline quality | binned SAH with 16-32 bins | O(n log n); the baseline fo