| name | world42-cbt |
| description | World42's CBT/OCBT terrain-LOD knowledge and workflow. Every planet runs the single integrated, validated GPU pool-CBT (OCBT) terrain engine. Use when requests mention Concurrent Binary Trees, OCBT, large_cbt, bisectors, triangle-area/screen-space LOD, or the GPU topology engine in paths under src/systems/lod/, src/app/setup_lod_and_shadows.ts, src/game_world/stellar_system/, and camera-space contracts. |
| paper | Concurrent Binary Trees for Large-Scale Game Components — HPG 2024 — https://arxiv.org/abs/2407.02215 |
| reference_impl | https://github.com/AnisB/large_cbt (DX12/HLSL reference — see references/10_large_cbt_gpu_reference.md for WebGPU mapping) |
World42 CBT / OCBT terrain engine
The terrain-LOD knowledge base for World42's Concurrent Binary Tree path. There is one
terrain algorithm — the GPU pool-CBT (OCBT) engine — and every planet runs it. There
is no per-planet algorithm selection and no CDLOD anymore.
Current state (what this skill now covers)
This started as an MVP ("one planet CDLOD, another CBT, CBT as a CPU stub"). It has since
grown well past that — CDLOD has been fully removed, and the CBT path is now the sole
terrain backend: a fully integrated GPU pool-CBT (OCBT) engine
(Benyoub & Dupuy, HPG 2024): the CBT is a fixed-capacity memory-pool allocator and the
triangulation is stored as explicit bisectors, so per-frame cost and memory are
decoupled from subdivision depth.
Status of the OCBT engine:
- Concurrent GPU topology (split + merge) ported and proven watertight against the CPU
oracle (20/20 GPU-vs-oracle cross-check).
- Render path live behind
?terrain=terrain: df64 camera-relative decode, frustum culling, draw
compaction, indirect dispatch of the work-list passes.
- Crack-free render to ~level 27 at the surface; usable-depth ceiling ~45 (emulated df64),
structural hard cap 60 (u64 heap id).
- Remaining work is performance (kill the
O(capacity) floor), precision (push
past ~45), and productionization (promote from dev-only) — see
references/15_ocbt_future_work.md.
The cpu and gpu-implicit backends have been removed; terrainType is now the single value
'gpu-terrain' in terrain_scheduler.ts, and the GPU OCBT engine
(src/systems/lod/terrain/gpu/) is the only runtime path. The CPU mirror still exists, but
only as the test oracle, not as a runtime backend.
Guardrails
- Keep coordinate-space boundaries explicit (WorldDouble, render-space, planet-local).
- All planets run the single GPU OCBT engine — there is no per-planet algorithm choice.
- Validation = geometry, not labels: GPU and CPU oracle use different heap-id conventions,
so compare decoded vertices / neighbor reciprocity / watertightness, never raw slot ids.
- Keep rollout low-risk with feature flags and defaults (
?terrain=terrain dev override;
terrainType is the single value 'gpu-terrain').
Terrain pipeline (single engine, all planets)
- There is no per-planet algorithm selection: every planet builds the GPU OCBT pipeline in
scene setup (
setup_lod_and_shadows.ts, stellar_catalog_loader.ts).
- Subdivision is still governed by the screen-space-error / projected-area budget classifier
(conservative per-frame budget) — that classifier drives split/merge, not a backend choice.
Validate
./.claude/skills/world42-cbt/scripts/check-cbt-mvp-integration.ps1
npm run pw:validate
Plus: npm test (pool allocator, u64/df64 emulation, LEB decode, full CPU topology oracle +
adversarial stress suite) and the dev cross-check page ocbt-topo-test.html (GPU vs oracle,
direct + indirect paths, 20/20).
Resources
references/08_world42_mapping.md — map CBT concepts to current World42 modules.references/09_checklist.md — final review checklist.references/10_large_cbt_gpu_reference.md — authoritative GPU reference: full kernel
pipeline, memory layout, bitfield ops, HLSL→WGSL translation, priority roadmap.references/11_cpu_optimization_results.md — CPU optimization harness + measured results.references/12_offthread_cbt_worker.md — off-thread CBT worker notes.references/13_gpu_cbt_webgpu.md — implicit GPU CBT (WebGPU) path notes.references/14_ocbt_gpu_topology_design.md — as-built OCBT design: buffer/binding
table, per-frame dispatch order, correctness/WebGPU risks, convention resolutions.references/15_ocbt_future_work.md — OCBT improvement roadmap: ranked backlog
(performance / render robustness / productionization / precision), critique triage, and a
suggested order of attack.references/mvp-implementation-plan.md — original MVP sequence (historical).scripts/check-cbt-mvp-integration.ps1 — inspect routing and coordinate-space touch points.
History (CPU-stub MVP — superseded)
The original MVP delivered per-planet CDLOD/CBT routing with a CPU-stub CBT (correct
bisection + screen-space-area metric + neighbor conformity + typed-array pool, but no GPU
kernels or indirect draw). That stage is captured in references/mvp-implementation-plan.md
and references/11_cpu_optimization_results.md. It has been superseded by the GPU OCBT engine
above; the CPU mirror now serves as the test oracle, not the runtime path.
