You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
solidcore: variable-round skin refinement re-solves spans already within tolerance #15482
#15257 holds a variable round's skin to its balls between stations. Its refinement loop re-solves every span's middle section in every round, including spans whose skin was already measured within tolerance in the round before. On the variable-round test that refines most, this more than doubles the time.
Where (solidcore/crates/solidcore/src/blend/fillet_intersector.rs, #15257's branch):
flush_curve, lines 1982-2011: each round rebuilds the skin (build_flush_geometry), calls skin_splits on all stations, merges the splits into the stations, re-equalises and loops again (at most SKIN_ROUNDS = 6 rounds, at most 64 times the stations).
skin_splits, line 2154: solve_mid_point (a Newton solve of the section, upstream step 5a) runs for every span on every round. The rest of the measurement is cheap by comparison: a projection onto the skin, then 9 evaluations against the section's envelope cone.
A span's middle section depends only on its two stations. Stations that stay put keep the same section, but it is solved again each round. Only the skin changes between rounds: the global C2 interpolation moves it everywhere when a station is added, so every span does still need re-measuring.
Timing (pinned to P-cores, one test at a time, #15257 against its base #15411):
fillet_variable_radius_closes_onto_vertex::taper_closing_to_a_point_rounds_the_wedge: 4.35 s -> 11.31 s (2.6x). The radius runs to zero, so the skin refines most there.
Other variable-round tests are at 1.0x-1.14x (for example a_record_ordered_variable_round_keeps_the_near_end_and_tie_rule 12.62 -> 14.44 s). The sum over 122 variable-round tests in a parallel gate run is 1.37x.
Proposal (performance only, results unchanged): keep each span's solved middle section across rounds, keyed by its two stations, and re-solve only the sections of spans that a new station split. Every span would still be re-measured against the rebuilt skin. Since the same sections are measured, the stations and the result are bit-identical, which a goldens run can confirm.
Correctness comes first: #15257 lands as it is, and this is its follow-up.
Written by Claude Code (Anthropic) on behalf of @brett19.
Activity