Asynchronous CUDA normal preparation

CUDA scan odometry submits a bounded dense target's normal queries and FP64 PCA before CPU source deskew and voxel preprocessing. A pinned host buffer and completion event let that GPU work run during the CPU work. Registration waits for completion and applies the existing native tie, eigengap and classification repairs before reading any normals. No additional CPU thread is created.

The target owns every queued device buffer, pinned allocation and event until completion. Abandoning a target drains its queued writes. Failed repairs invalidate affected rows so a retry cannot consume partially prepared normals. Submission is limited to one 65,536-row chunk and a source-size hint of at least one tenth of the target; sparse hints and larger maps retain demand-driven preparation. Heavily cropped sources can still overestimate that hint.

Normal kernels, neighbor budgets, native repair rules, registration stopping, motion/geometry grids, fitting and evidence evaluation are unchanged. CUDA execution reports include normal_schedule=async-v1 and the actual gpu_async_normal_batches counter. CPU remains the default.

Complete public-data comparison

Scope	Synchronous preparation	Asynchronous preparation	Observed ratio
Complete worker	57.791 s	51.622 s	1.120x

Two alternating fresh-process pairs replay the first six seconds of the public GLIM Livox bag with four map rounds and one BLAS thread. Both use CUDA scan odometry and CUDA fitter planes. Every worker has its own initially empty CuPy compiler cache. Whole time includes imports, decoding, initialization and first compilation, every odometry pass, association, fitting/covariance, noise updates, holdout, signed controls, hashing, saves and frozen-import verification.

The exact comparison permits only the declared scheduling identity/counter and run metadata to differ. Every calibration/evidence field, full per-round fit, trajectory, covariance, iteration count and pre-existing work counter must match exactly. This verifies preservation of the previous output; the clip has no independent extrinsic truth and its final verdict remains INCONCLUSIVE.

Host load is unreserved. These observations apply to this dense six-second workload. They are not multiplied by ratios from other cohorts. The legacy normal-preparation timer measures host time in submission/completion calls, including repairs; it excludes CPU work between the two calls and does not time overlapped GPU work. Its ratio is not a normal-kernel arithmetic speedup.

The bound archive retains all four full results, per-round fits, logs, source capsules/manifests, launcher/schema, correctness receipt and schema-export comparison. The audit runs without CUDA. Raw bag data and compiler caches are not redistributed.

Reproduce

python -m tools.benchmark_bspline_lidar_imu --comparison cuda-cuda \
  --baseline-commit 105e1031206ba6c551d08a1b07742b0307e5f523 \
  --prior-fit-plane-backend cuda --current-fit-plane-backend cuda \
  --require-exact-preparation-results --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-async-normal-overlap-pairs
python -m tools.audit_bspline_lidar_imu_benchmark docs/assets/cuda_normal_overlap

Use committed runtime sources and a new output directory. Keep runtime source, HEAD, launcher and comparison schema fixed until all workers finish. The dense settings use three local scans, 3 cm voxels, a 1 nm coordinate grid and returned motion grids of 1 micrometre translation / 0.1 microradian rotation components. The grids remain off by default. The existing CUDA odometry option enables bounded preparation automatically; no public config or result schema changes.