CUDA spline plane fitting

The optional CUDA plane backend removes the largest point-dependent work from the joint B-spline fitter: per-return pose evaluation, point-to-plane residuals, analytic control derivatives and the control normal matrix/gradient. It fuses these operations and directly reduces the seven-block control band, avoiding the large plane Jacobian during LM iterations. The final Jacobian is assembled for the existing marginal covariance computation.

Each normal entry has one owner, with a fixed row reduction order and no floating-point atomic additions. Kernels use FP64 and disable fused multiply-add. Referenced control-edge logarithms keep the native CPU near-pi guard. Immutable measurements and their time schedule stay on device for one fit; changing the trajectory invalidates residual/derivative reuse. A conservative 1 GiB payload check precedes device uploads. This is not a process RSS or CUDA allocator bound.

IMU residuals/derivatives, clock-offset central differences and their segment selection, LM damping/acceptance/stopping, the solve, marginal covariance, holdout and signed controls keep their CPU implementation. Return budgets, measurement weights and numerical convergence thresholds are unchanged. NumPy remains the default and does not import CuPy.

Complete public-data comparison

Scope	NumPy planes	CUDA planes	Observed ratio
Complete worker	65.884 s	61.146 s	1.077x
Four fits, including covariance	10.239 s	3.506 s	2.921x

Two alternating fresh-process pairs replay six seconds of the public GLIM Livox bag with four map rounds and one BLAS thread. Each worker has a separate empty CuPy compiler cache; first initialization/compilation are included. Both sides use identical CUDA scan odometry and the same dense odometry settings. Host background load is unreserved. These are observations for this workload, with different scopes for fit and whole-worker ratios.

The bound archive retains every full result, full fit, measurement/log, source capsule/manifest, launcher/schema and source-bound correctness run. The CPU-only audit checks the actual adapter source identity and work against each round's plane count. Fit status, iterations and dimensions must match exactly; all calibration and trajectory fields and covariance are compared at explicit roundoff tolerances. The existing physical/evidence checks apply to the complete pipeline. This path does not claim bit equality. The clip has no independent extrinsic truth and the final policy verdict remains INCONCLUSIVE.

Usage

Install the optional CUDA dependency and explicitly select the plane backend:

pip install 'calibrex[cuda]'
calibrex imu-lidar joint path/to/public.bag \
  --lidar-topic /livox/points --imu-topic /livox/imu --acceleration-unit g \
  --dataset-family glim_versatile --dataset-license 'upstream AIST GLIM dataset terms' \
  --odometry-backend cuda --fit-plane-backend cuda --output calibration.json

The Python fitter accepts BSplineLidarImuFitOptions(plane_backend="cuda"). Its result reports the executed kernel identity and actual work. Bag artifacts also retain every round's identity/counters in options.fit_execution when CUDA planes are requested. Existing artifact schemas are unchanged.

python -m tools.benchmark_bspline_lidar_imu --comparison cuda-cuda \
  --baseline-commit c3845d1eb42326e6da11a235b5e6239c33448160 \
  --current-fit-plane-backend cuda --max-seconds 6 --max-rounds 4 --repeats 2 \
  --odometry-voxel-size-m .03 --odometry-local-map-scans 3 \
  --odometry-geometry-quantum-m 1e-9 \
  --odometry-motion-translation-quantum-m 1e-6 \
  --odometry-motion-rotation-quantum-rad 1e-7 \
  --output outputs/glim-cuda-spline-planes-pairs
python -m tools.audit_bspline_lidar_imu_benchmark docs/assets/cuda_spline_planes

Use committed runtime sources and a new output directory. Keep source, HEAD, comparison launcher and schema fixed until all workers finish.