Repository navigation
macOS: ae3d's standing ragdolls cost about 2x per step after 1409bf2..89d8f79 #120
Description
Moving ae3d from aephysics 1409bf2 to 89d8f79 (main, with #101 to #112) fails ae3d's tests/test_motion_cost.ae on macOS CI. Windows is unchanged.
Test. 16 POWERED box-man ragdolls standing on one ground: the fixed step's cost per figure, best of 5 runs over 120 steps after 60. The bound is 80 µs.
| macOS CI (GitHub runner) | Windows local | |
|---|---|---|
| 1409bf2 | passes (< 80 µs) | 40.2 µs |
| 89d8f79 | 86.67 µs, twice, to 0.02% (1387.2 and 1387.5 µs a step); 4 figures 78.4 µs | 40.3 µs |
The two macOS readings agree to 0.02%. That steadiness looks systematic, not runner noise. Reverting only the submodule pin turned macOS green again (ae3d PR #644, runs 37538020309 failed and 3f7e696's passed).
Suspects, from the range: #106 "independent worlds stepped at once: worker slots of their own per world", #103's parallel-step race fixes (allocation counters, the stepping thread's index), and #107's joint layout. A per-step worker wake or slot cost on macOS's scheduler would fit a cost that grows with the figures but sits well above Windows.
To reproduce: in ae3d, ./build.sh tests/test_motion_cost.ae on macOS with deps/aephysics at each pin.
ae3d stays on 1409bf2 until this is understood.
Activity
Bisect #120: host scheduler, alternating rounds
nicolas-maman commentedon Oct 6, 2026
Bisected on a macOS arm64 runner (Apple M1, virtual, 3 cores), and I can't reproduce a regression in aephysics.
Repro without ae3d. 16 powered ragdolls standing on one ground: an align spring at the pelvis, motor anchors driven to a moving pose every step, 8 sub-steps, 3 workers. Best of 5 × 120 steps after 60, the empty world subtracted. That's the shape of
test_motion_cost, built directly on aephysics. I ran it both with the world's own scheduler and with one host scheduler lent to a fresh world on each run, as ae3d's engines do.Every merge in 1409bf2..89d8f79, one job (µs per figure at 16, 3 workers): 68, 57, 34, 29, 33, 50, 27, 28, 30, 31, 23, 69, 24, 28, 50, 48, 94, 47, 45, 35. No step, no trend. The old pin is the slowest reading.
The two pins alternated, three rounds, same job:
Best at each pin: 45.0 vs 45.0. A single best-of-5 reading moves by 2x on this runner at a fixed commit. ae3d's own test notes the same thing (#534: all three 16-figure runs at 87 µs where 4 and 32 read 43). 86.7 µs is that slow mode, and two readings agreeing to 0.02% would be two runs that landed in it.
On Windows the repro reads 22.5 µs at 1409bf2 and 21.5 µs at 89d8f79. Current main is the same or faster on both platforms.
Suggestion for ae3d: bump the pin again and rerun macOS a few times. If the test has to hold on this runner, a best over more runs or several rounds interleaved with the empty world would keep the slow mode out of a single verdict. Closing as not reproducible in aephysics. Reopen with a run that shows it at a fixed runner and I'll dig again.