Resident odometry in LiDAR map calibration: rejected¶
Connecting the resident geometry pipeline to the reference odometry of LiDAR–LiDAR map calibration did not improve complete execution time over the existing CUDA backend. The candidate is not adopted. Production source remains the runtime already covered by PR #270.
| Triad | CPU (s) | Existing CUDA (s) | Resident candidate (s) | CUDA / candidate |
|---|---|---|---|---|
| 1 | 47.083 | 18.070 | 18.761 | 0.963x |
| 2 | 47.667 | 16.747 | 18.836 | 0.889x |
| 3 | 54.222 | 16.308 | 17.457 | 0.934x |
| Median time | 47.667 | 16.747 | 18.761 | 0.893x |
All three candidate workers were slower than their CUDA counterparts. The ratio of median times means a 12.0% increase in complete process time. The candidate is 2.541x faster than CPU in this cohort, but the existing CUDA implementation is faster still. That CPU ratio does not establish an incremental benefit from the candidate and is not an adoption claim.
Workload and scope¶
Three fresh cyclic triads replay six seconds of the spent public GLIM Livox bag. The same point-cloud topic supplies reference and target scans. The target receives fixed Gaussian measurement noise (0.01 m, seed 20261018), avoiding a zero-residual copied-sensor fit. This is an execution control with no independent dual-LiDAR accuracy evidence.
Reference odometry uses 0.03 m voxels, three local-map scans, a 1e-9 m geometry grid, a 1e-6 m translation grid and a 1e-7 rad rotation grid in every arm. The reference pass is performed once. Map fitting stays on CPU in all arms, including the training fit, four block jackknife fits, full covariance, held-out scoring and all six axis controls. Eleven target samples span four training blocks and two held-out blocks; all 60 reference scans are retained. Sampling and filtering follow the same declared options in every worker.
Whole-process time includes startup, imports, fresh compilation, both bag streams, target-noise generation, reference odometry, map construction, calibration/evidence, full saves, input/source verification and hashes of prepared sample arrays. Every process has an independent initially empty CuPy compiler cache and one BLAS thread. The host was unreserved. There is no retained-cache, longer-recording or portable-device speed claim.
The predeclared speed gates require a median CPU/candidate ratio of at least 2 and a median CUDA/candidate ratio above 1. The second gate fails. No timing worker is discarded, and no favorable replacement cohort is substituted.
Quality and source retention¶
All 30 CPU/GPU full-fit comparisons retain the original map numerical
tolerances, including covariance: abs 1e-9 / rel 1e-6 generally and abs 1e-13 /
rel 1e-5 for covariance. Status, iteration and correspondence counts, policy,
degrees of freedom and control decisions match exactly. Existing/candidate
CUDA artifacts, all 15 paired full fits and every prepared pose/map/target
array digest match all float bits after only declared execution metadata and
run provenance are normalized. The copied-sensor policy status is pass;
that does not provide independent accuracy evidence for a two-LiDAR rig.
The candidate passed 100 relevant host/device tests, strict mypy for Linux and Windows over 396 modules, and Ruff. Its two new test files exercise option validation, CLI provenance and invalid-grid errors, synthetic room calibration, complete sample/evidence parity, both deskew modes and stream restoration on decoding failure.
The bound archive retains all nine workers, every complete fit, the original launcher/protocol and the candidate qualification. Its complete prior source is the already published source capsule; a small candidate delta replaces exactly three modules. A complete manifest checks all 396 reconstructed modules, preserving source identity while avoiding duplicate unchanged source archives. The experimental source is retained for inspection and is absent from the production changes.
The CPU-only auditor checks input/source/import/kernel/cache/thread bindings, full numerical and evidence parity, sample digests and all original timings. It independently recomputes the failed adoption gate. Tests reject altered payloads, a rebound adoption claim, forged kernel identity and a rebound source delta.
python -m tools.audit_lidar_map_resident_screen \
docs/assets/lidar_map_resident_screen --check