Public GLIM joint B-spline CPU acceleration

The complete native bag calibration runs 2.554x faster on this public 12-second clip. Its four complete spline fits, including marginal covariance, run 6.320x faster. Three alternating fresh-process pairs use one BLAS thread and the same four map rounds and 30-iteration fit caps. GPU is not required.

Scope	Prior median	Analytic median	Speedup
Complete worker	325.065 s	127.257 s	2.554x
Four fits, including covariance	251.629 s	39.818 s	6.320x

Trial	Variant	Whole seconds	Fit seconds
0	prior	343.659968	259.109449
0	current	132.007329	50.824268
1	current	118.642378	39.817663
1	prior	325.065390	251.628897
2	prior	306.357344	231.575880
2	current	127.257188	39.379822

Method and source binding

• Prior: b89e50bbd9e48a05a2acb6010ac46e50ece9220f.
• Current measured source: 4f3325740f9c821f1d198315f4534ae37482d181.
• Original AIST GLIM livox.bag SHA-256: f57a2e85339316d4235ab693ab2876aac5e113625c7f8b0f7873b62f92e09643.
• PointCloud2 /livox/points, point-time t; IMU /livox/imu, acceleration in g.
• The first 12 seconds yield a roughly 11.5-second reliable odometry segment and 118 spline controls; the full 61-second recording is not timed here.
• Whole time runs from process launch through bag reading, angular-rate initialization, gyro-deskewed odometry, map association and all four fits, covariance, noise update, holdout and 32 signed controls, result saving, input hashing and verified imports. Source preparation in the parent is outside the timed worker. Fit time includes each fit's covariance.

Every worker has its complete schema-valid result.json, full per-round covariance/trajectory/calibration in measurement.json, and original log. Both exact frozen source trees are in tar archives; their raw-byte manifests and each worker's actual module-import digests are retained. The executed benchmark launcher and its tool-only schema are frozen here. publication-index.json binds every published file except itself. The raw recording is not redistributed.

Result agreement and limits

All three quality comparisons pass. Policy, block/control, separability and convergence decisions agree; parameter estimates, std, last-round changes, noise, control chi-square and correlations meet the fixed protocol tolerances. For the first pair, rotation differs by 1.6e-09 deg, lever arm by 1.43e-10 m and clock offset by 5.33e-13 s. Whitened holdout RMS differs by at most 1.63e-12. The audit additionally checks full per-round covariance and trajectory arrays and costs.

All six workers report INCONCLUSIVE. This capture and its estimated plane map do not provide independent absolute extrinsic truth. Formal covariance holds the plane map fixed. This is a performance/regression comparison, not an accuracy or observability claim, and does not predict all recordings' speed.

The exploratory profiles under ignored outputs/ were used to identify the rotation finite differences. Their initial-odometry timing changed substantially between runs, so their total-time ratio is not used in these results. The single formal cohort above is the source of every reported speed ratio.

Audit and reproduce

python -m tools.audit_bspline_lidar_imu_benchmark
python -m tools.benchmark_bspline_lidar_imu \
  --baseline-commit b89e50bbd9e48a05a2acb6010ac46e50ece9220f \
  --bag data/public/glim_versatile/livox.bag \
  --max-seconds 12 --max-rounds 4 --repeats 3 \
  --output outputs/glim-joint-bspline-comparison

The comparison output must not exist and runtime source changes must be committed. The archive audit needs the ordinary CPU dependencies only; it does not open a bag or import CUDA. See the benchmark explanation.