Actual-frame LiDAR ICP CUDA evaluation¶
These cohorts use distinct consecutive Velodyne frames from the public KITTI Raw dataset, sequence 2011_09_30_drive_0018_sync. No copied or synthetic target cloud is used. Source scans contain roughly 125,000 points. Official OXTS navigation and the daily IMU-to-Velodyne calibration provide an initialization, followed by a declared 3 cm/2 cm/1.5 cm translation and small rotation perturbation. The fixture retains exact raw-file, navigation, calibration and source hashes.
The primary comparison uses default 50-iteration ICP with complete spatial holdout, local rematching and six initial-value diagnostics. Breadth and small cohorts use a predefined 20-iteration cap. Both methods retain automatic bounded diagnostic parallelism; all workers set and verify one internal BLAS thread. The primary has three alternating pairs; other cohorts have one smoke pair.
| Cohort | Iteration cap | Native whole median | CUDA whole median | Ratio | Status |
|---|---|---|---|---|---|
| pilot | 20 | 63.468311 s | 29.026270 s | 2.187x | max_iterations |
| primary-115 | 50 | 55.983115 s | 25.303228 s | 2.212x | converged |
| breadth-230 | 20 | 29.587284 s | 15.634474 s | 1.892x | max_iterations |
| breadth-345 | 20 | 28.698942 s | 15.961634 s | 1.798x | max_iterations |
| small-115 | 20 | 2.798039 s | 3.527018 s | 0.793x | max_iterations |
The primary whole-worker ratio is 2.212x; complete-solver ratio is 2.390x. All primary solves converge. Complete native result bytes match exactly across CPU/CUDA and repetitions: pose, history, splits, correspondences, holdout, rematching diagnostics, six start trials and symmetry flags. The other cohorts also retain exact complete results within each cohort. All fourteen completed workers are saved, including the 20-iteration pilot that reached its cap.
| Cohort | Pair | Order | Native whole wall | CUDA whole wall |
|---|---|---|---|---|
| pilot | 1 | native, cuda | 63.468311 s | 29.026270 s |
| primary-115 | 1 | native, cuda | 55.983115 s | 25.303228 s |
| primary-115 | 2 | cuda, native | 55.384290 s | 25.332281 s |
| primary-115 | 3 | native, cuda | 57.209044 s | 25.228562 s |
| breadth-230 | 1 | native, cuda | 29.587284 s | 15.634474 s |
| breadth-345 | 1 | native, cuda | 28.698942 s | 15.961634 s |
| small-115 | 1 | native, cuda | 2.798039 s | 3.527018 s |
| Cohort | Initial holdout RMSE | Final holdout RMSE | Navigation translation difference |
|---|---|---|---|
| pilot | 0.153941 m | 0.142109 m | 0.032132 m |
| primary-115 | 0.153941 m | 0.142109 m | 0.032134 m |
| breadth-230 | 0.147279 m | 0.142171 m | 0.009685 m |
| breadth-345 | 0.180983 m | 0.176710 m | 0.027798 m |
| small-115 | 1.331548 m | 1.135372 m | 0.439500 m |
Holdout scores are the native fresh-target-rematching scores on the spatial source split. The navigation pose is used for initialization and consistency checking; these raw rotating scans are not deskewed. Navigation differences are not independently established calibration accuracy. This evaluates registration between time-separated frames of the same LiDAR.
The 768-point cohort launches zero GPU spatial queries and uses native fallback. Cold CUDA import/initialization makes its whole worker slower. Use the native solver for small clouds. Breadth smoke ratios are local observations and are not median-confidence estimates; their iteration caps remain explicitly reported.
Every CUDA worker starts with a fresh empty disk cache. Whole time includes startup, input verification, point preparation, solve, all diagnostics, source guards, initial holdout evaluation and saves. Solver time includes cold CUDA import, compilation, transfers and native diagnostics, but excludes input preparation and the additional initial-holdout evaluation. No cold or slow trial is discarded. OS caches persist. Cohort speed ratios are never pooled or multiplied.
The protocol was saved before formal execution. The primary uses source commit
0a30e10021124f52e7afcb44e5b5cf00ba8475c7. All measured files and fixtures remained
unchanged throughout each worker. Source snapshots preserve measured bytes and
normalized Git identity. Identical result bytes are stored once per digest,
with a reference from every worker. Only local path metadata in the acquisition
manifest was made portable. Raw pointcloud files remain external public inputs.
Reproduce with the existing official range downloader and optional CUDA extra:
python -m pip install -e ".[cuda]" threadpoolctl
python -m tools.download_kitti_raw_lidar_windows 2011_09_30_drive_0018_sync --center-frame-ids 115,230,345 --neighbor-radius 1 --output-root data/public/kitti_icp --manifest-output outputs/kitti-selection.json
# Put the official daily calib_imu_to_velo.txt in data/public/kitti_icp.
python -m tools.prepare_kitti_icp_pairs --data-root data/public/kitti_icp --calibration data/public/kitti_icp/calib_imu_to_velo.txt --source-frames 115 --max-iterations 50 --output outputs/kitti-pairs
python tools/benchmark_cuda_point_to_point_icp.py --pair-fixture outputs/kitti-pairs/0000000115-0000000116.json --data-root data/public/kitti_icp --output outputs/kitti-real-icp --repeats 3
python tools/audit_actual_icp_archive.py
The launcher applies its thread budget before Python/NumPy starts and records actual loaded BLAS thread counts. It does not change settings in an existing application process. The earlier failures with unbounded nested threading remain documented in the parent archive; their cause remains unresolved.