The architecture round. Loop inverted from per-sample-per-op to per-op-per-block,
operators pruned per block by interval bounds, profiler moved out of the hot loop.
my build, 3 runs: 1.013 / 1.036 / 1.075 s (adb34df was 1.418 / 1.445 / 1.458)
mean 1.041 vs 1.440 = -28.5%, 1.39x
geometry SHA-256 b3e5f4c3..9b377 IDENTICAL to adb34df
pruning removed 20,999 / 31,320 operator-block visits = 67.1%
surviving ops per block: LOD0 6.63 LOD1 6.30 LOD2 5.23 LOD3 4.41 LOD4 3.89
(out of 22)
⚠️ I PREDICTED THE DISPATCH OVERHEAD WAS "THE 10x". IT WAS NOT.
864,688 virtual dispatches per chunk sounded decisive and is worth ~20-28%. Real, and
worth having, but the order of magnitude is elsewhere. That is my fourth wrong
performance prediction in this project, and the pattern is identical each time: I reason
from a call site's shape instead of measuring what the data actually costs.
⭐⭐ AND THE "5-20x GAME/EXPORT GAP" WAS MY ARITHMETIC ERROR. IT DOES NOT EXIST.
I told the owner the largest unexplained number in the project was that export does
22 ms/tile while the game does 115-500 ms. That 22 ms came from dividing 1.4 s of WALL
CLOCK by 64 tiles -- ignoring that 15 workers ran them in parallel. Per-job means on the
same build:
nonshared block export 447 ms/job
shared-grid export 261 ms/job
game LOD0 meshing 253 ms/tile p50
They are the same. The game is not anomalously slow; generation simply costs ~250 ms per
LOD0 tile. The shared grid's benefit is real but modest, and a neighbourhood cache is NOT
justified by these numbers -- I would have had Codex build one on the strength of a
division error.
GAME PATH, final build (841 profiles, 447 applies)
meshed p50/p95/max LOD0 253/521/697 ms LOD1 79/130/135 LOD2 19/42/51
LOD3 36/64/81 LOD4 93/238/271
LOD0 request-to-ready p50 47.7 ms p95 389 ms max 798 ms
worker queue p95 84 ms, generation p95 344 ms
at 6 m/s: p95 = 2.3 m of travel, max = 4.8 m
empty LOD0 tiles never enter the mesher; classifier p50 8.7 ms, p95 17.2 ms
apply is not the bottleneck: p50 0.107 ms, p95 0.161 ms, max 0.296 ms
⭐ THE BINDING CONSTRAINT IS COMPUTE, NOT POLICY:
logical cores 16, automatic core cap 14, effective generation cap 14, CAP HIT.
Apply limits never bind. We are saturating the machine, which means the remaining wins
must come from doing less work, not from scheduling it better.
NOT DONE, with reasons rather than excuses:
- Coarse-first LOD refinement: coarse tiles carry NO COLLISION, so streaming them first
would let the player fall through. Needs a conservative collision proxy or an L0
handoff first. Correct refusal.
- Heading prefetch: desired tiles are sorted by distance only (VoxelWorld.cpp:1748).
Velocity-weighted prediction with a bounded budget and expiry is the shape, not built.
A failed experiment, reported rather than buried: eagerly filling every classifier leaf
with EvalBlock regressed export to 3.2-3.3 s because it defeated the classifier's early
exit. Removed.
Build 43 actions, 0 errors. Peak RSS under 2.1 GB. Sources: Epic's Tasks System and
Level Streaming docs, Lorensen-Cline.
Cumulative on this region: 342.2 s -> 1.02 s.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012aheo2crrZmHuYtPpdY2Cb
Fr0zkaandClaude Opus 5
5ddd96664d