Skip to content

Tile large adaptive steering grids without changing cutting resolution - #372

Merged
10-X-eng merged 2 commits into
mainfrom
feat/193-active-front
Nov 21, 2026
Merged

10-X-eng merged 2 commits into
mainfrom
feat/193-active-front

Conversation

@10-X-eng

@10-X-eng 10-X-eng commented Nov 21, 2026 •

Copy link
Copy Markdown
Contributor

Closes #193.

A 2 mm end mill clearing a 400 × 400 mm blank at 0.25 mm radial engagement needs an 8,168 × 8,168 steering grid. Both AdaptivePocket and adaptive Roughing refuse that job on the base commit before the first cut.

Large layers now store cleared heights by tile, read neighbouring material across tile boundaries and track the best enclosed entry per tile. The candidate order, 0.05 mm fine lattice, stroke proofs and stock-generation checks are shared with the existing planner. The planning-memory check still applies after bounding the operation. Cancellation during entry classification requires the classification to finish on retry.

The end-to-end regressions machine a connected rim around a square boss from solid stock through both public operations. They check decoded posted motion, stock contact, removal on all four sides, the floor, the retained boss and the proven engagement angle. The tile regression compares candidates with whole-grid Euclidean distance transforms, including after a cut and rollback.

Linux release, 400 × 400 mm solid stock, 2 mm cutter and 0.25 mm radial engagement. The declared angular limit is 41.409622°. Both checked posts passed with zero stock-contact violations or uncertain motions. The independent stock checks cover the floor, all four cut sides and the retained boss.

Operation Motions Strokes proven / certified Maximum proven engagement Peak RSS
Adaptive pocket 707 683 / 683 41.408653° 1.074 GiB
Adaptive roughing 698 677 / 677 41.408653° 1.074 GiB

All 64 workloads completed on both commits with identical recorded toolpaths, material, cycle times and engagement results. Release runs used the same Linux host and three repeats per workload. To isolate the engine from Python and CAD startup, the four adaptive benchmark manifests were also replayed against the two release binaries in alternating order, pinned to the same CPU. The native columns are medians after warm-up.

Existing benchmark Full run seconds, base → PR Native CPU seconds, base → PR Native peak MiB, base → PR
Adaptive pocket 0.578 → 0.670 0.522 → 0.502 18.2 → 17.1
Pocket with corner cap 0.348 → 0.290 0.279 → 0.259 17.8 → 15.7
Adaptive roughing 15.331 → 13.522 10.475 → 10.556 52.4 → 44.1
Large adaptive roughing 16.545 → 14.808 12.390 → 12.279 163.8 → 120.2

The short pocket's full-run slowdown did not repeat in either pinned comparison. Individual timings varied with host load. The four adaptive workloads together took 32.80 → 29.29 seconds in the full benchmark. All four executed fewer native instructions and used less native peak memory in the direct comparisons.

Linux workspace tests passed (458 unit, 666 integration, one doctest, five ignored). The 30 mm dense controls and 400 mm tiled jobs passed through the release build. Linux and Windows CI passed, including the installed Windows wheel, LinuxCNC interpreter, minimum Rust version and benchmark expectations.

@10-X-eng
10-X-eng marked this pull request as ready for review November 21, 2026 18:13
@10-X-eng
10-X-eng added this pull request to the merge queue Nov 21, 2026
Merged via the queue into main with commit b636cc4 Nov 21, 2026
9 checks passed
@10-X-eng
10-X-eng deleted the feat/193-active-front branch November 21, 2026 18:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

engagement_front: multi-resolution steering grid for small cutters on large stock

1 participant