Performance & parallelism
This is stale as of the tier-1 parallelization work — see "Parallelism (per
This is stale as of the tier-1 parallelization work — see "Parallelism (per
converter)" above for the current, accurate picture. (Kept here,
corrected, since other sections still link to this heading.) Tile work — decode → SPZ/GLB encode
(gzip) → write — is embarrassingly parallel and independent per tile. All five operational
converters (rad, sog, lcc, copc, potree) now fan tiles out via parallelMap (bounded concurrency =
cores−1) + async gzip, verified byte/content-identical to the prior serial versions; i3s (draft)
remains serial. Memory is already bounded per-chunk (esp. SOG/LCC: decode one chunk → write its
tiles → release). The outstanding piece is tier 2 — a real worker_threads pool for true
multicore decode (WASM/CPU-bound: LAZ, WebP, brotli) rather than just bounded concurrency on the
main thread's event loop; see the per-converter table for where that headroom is biggest (SOG's WebP
decode, mainly).