Optimize LCC reader for large scenes (~5.6x faster)

[slimbuck]

Summary

Optimizes the LCC (XGrids) reader to handle large scenes efficiently. Tested on a 139M splat scene (9GB on disk), load time drops from 200s to 36s (~5.6x speedup) with significantly reduced peak memory usage.

Changes

• Eliminate per-splat allocations: decodeRotation no longer creates a temporary array per call; SH decoding no longer creates 15 Vec3 objects per splat. New decodeRotationInto writes directly to output arrays.
• Remove double-buffering: processUnit writes directly into the shared output arrays instead of allocating per-unit intermediate typed arrays and copying with .set().
• Concurrent I/O: Units are dispatched with a bounded worker pool instead of sequential await per unit. No measurable impact for local disk, but benefits network-based file systems (browser URL loading).
• Pre-combine LODs: All selected LODs are decoded directly into a single pre-allocated DataTable, eliminating the expensive post-read combine() step which was allocating ~35GB of intermediate buffers and copies for large scenes.
• Replace DataView with typed array views: The hot decode loop uses Float32Array, Uint16Array, and Uint8Array views over the input buffer instead of DataView.get*() calls, avoiding per-access bounds checks and endianness handling overhead.

Performance

Benchmarked on a 139,722,883 splat LCC scene (Grow-House.lcc):

Version	Time	Speedup
Before	200s	1x
+ No allocations, no double-buffering	81s	2.5x
+ Pre-combined LODs (no combine step)	39s	5.1x
+ Typed array views	36s	5.6x

Memory

• Eliminated ~35GB of transient allocations from the combine() step (per-unit intermediate arrays + final column merge)
• Per-unit temporary Float32Array allocations removed (previously allocated and copied per unit, now writes directly to global arrays)
• Billions of per-splat temporary object allocations eliminated (quaternion arrays, Vec3 instances), reducing GC pressure

[Copilot]

Pull request overview

Optimizes the XGrids LCC reader hot path to significantly improve load time and memory usage on very large scenes by decoding directly into preallocated output buffers, using typed-array views for faster access, and concurrently dispatching unit reads.

Changes:

• Decode rotations/SH coefficients without per-splat temporary allocations (write directly into output arrays).
• Remove per-unit intermediate buffers by writing unit data straight into shared global arrays.
• Add bounded-concurrency unit decoding and pre-combine selected LODs into a single DataTable (plus environment table).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.