perf(generation): memoise room player-fit points -- 3.8x less worker CPU in game

The sampler (8e40bdd) showed ~3/4 of in-game worker CPU inside BuildChunkCache ->
VF_FindPlayerFitPointForRoom, recomputed for the same rooms on every per-worker cache rebuild.
The search reads only the strate params, the room site, the strate Z bounds and the warp seed,
so its answer is pure. It is now memoised process-wide.

- The key is exact, 97 words: every FStrateGenerationParams field via VF_STRATE_PARAM_FIELDS
  (83 scalar fields, audited: no pointers or TArrays), the room centre/radii/hash/bOrigin,
  both Z bounds, the warp seed and a key version. The hash only picks a shard; lookup compares
  every word. A miss is the worst case, and eviction just recomputes the same value.
- 4096 sets x 4 ways, a per-slot reader/writer state, a per-set lock only on miss (single
  flight), 7.1 MB fixed. The search algorithm is unchanged.
- voxel.PlayerFitMemo (CVar + command line), default on. The logs confirm =0 reaches the path.
- Diagnostics: the SdfCacheBuild count moves to BuildChunkCache entry, per-tile cache_builds in
  the tile profile, tile coordinates in the startup trace, and memo stats logged at EndPlay.

Game path, two runs each, sampler off (Codex; checked against its report):
                      worker CPU     LOD0 ready p50 / p95      tiles applied
  static  off -> on   183.7 -> 48.2 s  (3.81x)   109 / 178 -> 103 / 172 ms   841 = 841
  moving  off -> on   256.9 -> 124.7 s (2.06x)   104 / 192 ->  78 / 157 ms   2226 -> 2348
Hit rate 99.95%, 463 real searches in the moving run, 0 evictions.

Field and capability: per-tile triangle counts are identical on/off for 841/841 tiles
(669,834 triangles), the export OBJ SHA-256 is unchanged (b3e5f4c3..9b377), and player-fit
20,830 / walk-reachable 10,909 are unchanged.

Next, measured by this round: coarse tiles thrash the 4-chunk room-cache window. BuildChunkCache
calls per tile are 1.7 at LOD0, 18 at LOD2, 633 at LOD3 and 1,062 at LOD4.

Known gap, closed in the next round: nothing yet asserts that VF_STRATE_PARAM_FIELDS covers every
FStrateGenerationParams field. A field added to the struct but not to the list would be missing
from the memo key.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VZS94UH8bSsBeg5sZaxe2Y
This commit is contained in:
Fr0zkaandClaude Opus 5 committed 2026-09-13 04:48:44 +02:00
1 parent 8e40bdd92c
commit 2990009f3b
6 files changed
+515 -9

No files matched your search

@@ -2745,11 +2745,6 @@ float UVoxelGenerator::GetDensityWithParams(float WorldX, float WorldY, float Wo
if (bNeedRebuild)
{
if (VoxelDensityProfile::AreCountersEnabled())
{
VoxelDensityProfile::AddCounter(
VoxelDensityProfile::ECounter::SdfCacheBuild);
}
// The lowered evaluator uses the same window-invariance contract as the prepared
// room-graph source, but keeps a four-chunk XY window per worker. A tile worker can
// visit neighbouring chunk keys in an arbitrary order; the larger deterministic