CUDA mutual point-to-point ICP controls¶
The optional adapter replaces forward and reciprocal nearest-neighbor searches with FP64 CUDA proposals over owned immutable spatial trees. It reuses the raw source tree by moving queries into its frame. Original NumPy world-frame distances and source-index tie ordering determine the result. Ambiguous rows are repaired natively; small, non-finite or extreme-coordinate searches fall back to the native evaluator. Pose updates, SVDs, reductions, spatial holdout and all six multi-start diagnostics remain native. The two captured trees are shared across starts; CPU execution imports no CuPy.
Primary observations use the final measured source 85a3f0d037dbbc29d3ecbd8a50ae721ada34f8bc.
They use public KITTI Raw scan 0000000115 (127,158 points), sequence
2011_09_30_drive_0018_sync. Its raw SHA-256 is
213948bd8c54c925aca7eb9f3b3c36e95e586b52ee8d9d64056c4fc4aeb12b2d.
The target is a deterministically transformed copy with 2 mm Gaussian noise;
seed, transform, native options and exact inputs are bound in every record.
This establishes execution equivalence on public data, not accuracy between
independently recorded sensors. No personal rosbag is needed.
| Scope | Native median | CUDA median | Ratio of medians |
|---|---|---|---|
| Whole worker | 46.777627 s | 24.407067 s | 1.917x |
| Complete solver | 43.813992 s | 21.320884 s | 2.055x |
| Pair | Order | Native whole wall | CUDA whole wall |
|---|---|---|---|
| 1 | native, cuda | 46.777627 s | 24.407067 s |
| 2 | cuda, native | 51.313975 s | 27.782913 s |
| 3 | native, cuda | 44.984903 s | 22.316691 s |
Whole worker includes interpreter startup, file loading, point preparation, full solve, diagnostics, source/input guards, hashes and result saves. Solver time includes optional CUDA import, initial upload, tree construction, cold kernel compilation and all calibration diagnostics; it excludes input preparation and result serialization. Every worker has its own initially empty CUDA disk-cache directory. OS caches persist. All three alternating pairs, cold trials and slower observations are included. No cohort ratios are pooled or multiplied; these are local observations rather than portable confidence estimates.
All final native calibration result bytes match between CPU and CUDA, including pose, complete history, correspondence count, source split IDs, holdout, rematching curvature, six multi-start trials and symmetry diagnostics. Both methods retain default bounded automatic diagnostic parallelism. BLAS/OpenMP internal threads are set to one in both fresh worker environments; every final record verifies the loaded BLAS runtime thread count. This avoids nested oversubscription.
Two earlier runs without that thread budget ended in access violations: one CUDA worker and one native CPU worker before any CUDA execution. Their logs and measured sources are retained; no completed calibration model exists for those failed workers. A successful unbounded GPU diagnostic and a sequential comparison are also retained, separately from the primary result. The root cause of the unbounded-thread failures remains unresolved. The sequential comparison does not establish a gain against the existing parallel CPU path.
The bounded precursor and pre-guard comparisons are retained in full. The final source adds a tested native fallback for extreme world coordinates; the final primary cohort was rerun afterwards. The earlier raw-source search probe remains explicitly kernel-only; its speed is not a solver speed claim.
The archive retains 22 complete worker results and 2 failed workers.
Identical complete JSON result bytes are stored once per digest, with an
explicit reference from every worker record. No fields are removed for parity.
Adapter provenance is retained separately in each CUDA measurement.
Source snapshots retain exact measured bytes and normalized Git source
identity. Only private paths in failed logs were replaced by portable markers;
original log digests are retained. Captured sources, measurements, logs and result bytes are digest-bound in
publication-index.json.
Validation: 198 distinct targeted tests across the final change and fallback checks, mypy on 357 source modules, and 126 byte-identical exported schemas. Real-device cases cover duplicates, near ties, wide coordinate scales, large translations, immutable captures, small/empty fallback and complete native calibration outcomes with sequential and parallel starts.
Reproduce after installing the optional CUDA extra and benchmark inspection helper:
python -m pip install -e ".[cuda]" threadpoolctl
python tools/benchmark_cuda_point_to_point_icp.py --input path/to/0000000115.bin --output outputs/icp-control --repeats 3
python tools/audit_cuda_point_icp_archive.py
The benchmark sets the worker BLAS thread environment before Python starts. The adapter does not modify process-global BLAS settings. For direct use, set OPENBLAS_NUM_THREADS, OMP_NUM_THREADS and MKL_NUM_THREADS before starting Python when enabling parallel diagnostic starts.
The portable publisher splits its privacy-pattern literal without changing behavior. Original and portable publisher digests are retained.