Rejected conservative spatial interval candidate¶
Not adopted. Output equality does not establish a useful whole-worker gain.
| Scope | Previous CUDA | Interval candidate | Ratio |
|---|---|---|---|
| Complete worker | 54.383 s | 63.992 s | 0.850x |
| Repeat | Variant | Whole seconds |
|---|---|---|
| 0 | prior | 55.160675 |
| 0 | current | 65.272271 |
| 1 | current | 62.711611 |
| 1 | prior | 53.605573 |
This first version performs repeated CPU target-extent reductions. It is retained as measured, before reusing the existing cKDTree extrema.
Two alternating fresh-process pairs replay six seconds of the public GLIM Livox bag with four map rounds, one BLAS thread, CUDA scan odometry and CUDA fitter planes. Each worker has a separate initially empty CuPy compiler cache. Whole time includes imports, setup/compilation, every odometry pass, association, fits/covariance, noise updates, holdout/controls, hashes, saves and import checks. No warmup or worker is discarded. Host load is unreserved.
Baseline bb131739c0d14d30fc0e1660ce4f2ee587d6d32e; candidate 61921e5785876f9ee47ff030f034f11a357f25c2.
Bag SHA-256 f57a2e85339316d4235ab693ab2876aac5e113625c7f8b0f7873b62f92e09643. The unchanged dense policy is 3 cm voxels,
three local scans, a 1 nm coordinate grid and returned-motion grids of
1 micrometre / 0.1 microradian. CPU and disabled grids remain the defaults.
The candidate encloses boxes, split planes and points in outward FP32 intervals, using downward-rounded lower distance bounds to reject work. Actual candidate distances, tie ordering, normals, residuals/systems and native repairs stay FP64. Native arrays/kernels remain for bounded-path fallbacks; events order reads and drain abandoned writes. No extra CPU thread is introduced.
Every calibration/evidence field, complete fit, trajectory, covariance, iteration and previous work counter matches exactly. The exact gate allows only declared source-bound kernel identities/pruning counters and run metadata to differ. Actual interval work is positive in every measured pass and bounded by total GPU work. The clip has no independent extrinsic truth and its verdict remains INCONCLUSIVE. Ratios are not multiplied across cohorts.
All full results/fits/logs, deterministic source capsules/manifests, launcher, schema and publisher are SHA-256 bound. The audit runs without CUDA. Raw bag data and compiler caches are not redistributed. The retained runtime keeps the previous asynchronous-normal implementation; these interval kernels are experimental.