Cluster LOD 1a: ingest -- clusterize + DAG build, attribute/weight-aware, disk cache #744

New issue

Closed

opened 2026-10-13 01:57:39 +00:00 by sky_line · 2 comments

sky_line commented

2026-10-13 01:57:39 +00:00

Owner

Part of #642. Wave 1. Depends on the wave-0 contract.

Split LOD0 of each mesh face into clusters; build the DAG (group neighbours, simplify with the group border locked, re-split) on worker threads, with monotonic errors and nested bounds.
Clusters never cross material / face boundaries; weld-boundary vertices (#641) and layer-contact regions (#611) locked.
Attribute- and skin-weight-aware simplification (see #618); per-cluster joint-stretch bound for animated error.
Pack into fixed-size pages, coarse to fine; on-disk cache keyed by mesh asset + face + builder version, with a byte cap and LRU (pattern of #274).
Budgeted so a crowded arrival never stalls the frame; stats for build time, queue, cache hit rate and size.

Rules for this work (all sub-tickets):

Behind RenderVKClusterLOD (default off). With it off, the existing rigged LOD controller (#561 / #613) and every current path are untouched.
Newest path first, chosen by extension presence; every slower path is an additive fallback, forceable on the newest card via RenderVKClusterForceTier.
Failures never crash: a missing page / failed allocation falls back to the coarser parent cut; no unbounded growth (fixed pools, byte budgets, live/cap/peak stats).

Part of #642. **Wave 1.** Depends on the wave-0 contract. - Split LOD0 of each mesh face into clusters; build the DAG (group neighbours, simplify with the group border locked, re-split) on worker threads, with monotonic errors and nested bounds. - Clusters never cross material / face boundaries; weld-boundary vertices (#641) and layer-contact regions (#611) locked. - Attribute- and skin-weight-aware simplification (see #618); per-cluster joint-stretch bound for animated error. - Pack into fixed-size pages, coarse to fine; on-disk cache keyed by mesh asset + face + builder version, with a byte cap and LRU (pattern of #274). - Budgeted so a crowded arrival never stalls the frame; stats for build time, queue, cache hit rate and size. **Rules for this work (all sub-tickets):** - Behind `RenderVKClusterLOD` (default off). With it off, the existing rigged LOD controller (#561 / #613) and every current path are untouched. - Newest path first, chosen by extension presence; every slower path is an additive fallback, forceable on the newest card via `RenderVKClusterForceTier`. - Failures never crash: a missing page / failed allocation falls back to the coarser parent cut; no unbounded growth (fixed pools, byte budgets, live/cap/peak stats).

sky_line added the

labels

2026-10-13 01:57:39 +00:00

sky_line added this to the Slipstream project

2026-10-13 01:57:39 +00:00

sky_line referenced this issue