Fused CUDA map registration

The CUDA LiDAR–LiDAR map backend combines point projection and exact two-neighbor search in one kernel, gathers source-ordered point-to-plane residuals in another, and forms weighted analytic Jacobians in a third. Both variants retain the previously adopted normal cache. This comparison isolates the registration dispatch change and its compilation cost.

Compiler-cache policy	Previous whole (s)	Current whole (s)	Ratio	Previous API (s)	Current API (s)	Ratio
Empty per worker, 3 pairs	19.202	13.253	1.449x	12.901	7.183	1.796x
Retained per method, 4 pairs	10.049	11.393	0.882x	4.249	3.921	1.084x

Cold mode gives every process a separate initially empty compiler cache. Retained mode gives each method an independent empty cache, then retains it for that method's subsequent fresh processes. Every process enters the medians, including the first compilation; there is no excluded warmup. Pair order alternates. The initial-compilation and retained-cache results have different scopes and must not be combined or multiplied with earlier normal-cache ratios.

The final cold-cache cohort has three positive pair ratios (1.510x, 1.449x and 1.298x). The retained-cache cohort's complete-process median is 13.4% slower, with pair ratios 1.576x, 0.759x, 1.118x and 1.049x. Its API median improves by 1.084x, but the complete fresh-process observations establish no end-to-end improvement with an existing cache. Startup/source verification and unchanged CPU work remain in those timings; shared-host variation is not isolated from implementation effects. Every observation remains available.

The generated compiler-cache program count falls from 47 to 21 on this workload. The cold-start improvement comes mainly from compiling and launching fewer small GPU operations. Once the compilation cache exists, the gain is substantially smaller. These measurements describe an unreserved host and do not establish portable speed guarantees or steady-state throughput for other map sizes.

Unchanged calibration work

Nine original KITTI raw scans from 2011_09_30_drive_0018_sync, frames 114–116, 229–231 and 344–346, form three-scan maps. The central scans provide copied targets with fixed 0.01 m Gaussian noise, seed 20261012. Six generated pose samples contain 2,286,890 map rows and 762,118 target rows; four blocks train and two hold out. Every original point, the training fit, four block jackknife fits, covariance, held-out scoring and all six known-bad controls remain in the API.

This copied-sensor execution control has no independent two-LiDAR extrinsic truth or independent temporal validation. The raw map windows are not motion compensated. Bag decoding and reference odometry are outside this workload.

All complete evaluation fields and saved schema-valid artifacts are compared under the existing map tolerances: absolute 1e-9 / relative 1e-6 for floating-point fields and absolute 1e-13 / relative 1e-5 for covariance. Counts, iterations, statuses, policy, estimated degrees of freedom and control decisions match exactly. Prepared arrays, raw scan hashes and train/holdout identities match. Both CUDA identities are independently checked against their frozen query, normal and registration kernel bytes. This is numerical tolerance parity, with no claim of bitwise floating-point equality.

All previously recorded work counters are identical in each pair. Both variants query 11,030,250 rows and compute 1,247,324 normals, with 7,691 native normal repairs and 78,474 native query repairs. The current implementation additionally reports 77 projection/query launches over 9,782,926 rows, 77 residual-gather launches over 5,678,656 rows and 59 weighted-Jacobian launches over 4,372,180 rows. The auditor checks these counters against the complete original query work.

Arithmetic, layout and memory

Projection, nearest-neighbor search, residuals and Jacobian arithmetic remain FP64, with fused multiply-add disabled in these explicit kernels. Existing GPU matrix products reduce the same weighted Jacobians into the normal equations. Six-dimensional updates, covariance and evidence policy retain their native definitions. Correspondence gates, tie guards, planarity, Huber loss and convergence limits are unchanged. Ambiguous ties and gate boundaries retain the original native queries and transforms.

One byte flag per source point reports validity and uncertainty. The host compacts these flags in increasing source order, avoiding a separate dynamic GPU nonzero/count operation. Normal selection reuses the existing bounded cache. Explicit FP64 row-major matrix snapshots support ordinary, column-major and strided caller arrays, including float32 input matrices. No new input-layout restriction is introduced.

Retained device input bytes remain 183,963,548 on this workload, under the existing 1 GiB bound. Scratch and allocator pools are additional. No new dependency, public calibration option, result schema, point cap or approximation is added. CPU remains the default backend.

Reproduction and independent audit

The baseline is 22c19ebc75ec61715c83a2e77e9c3e8eea34dffc. The measured current revision is 8fcc8d63052908e8fb06c0153a2e975acfb0dc9f. Obtain the original scans under the official KITTI terms, install calibrex[cuda], check out that current revision and run:

python -m tools.benchmark_cuda_map_preparation \
  --sequence path/to/2011_09_30_drive_0018_sync \
  --baseline-commit 22c19ebc75ec61715c83a2e77e9c3e8eea34dffc \
  --half-window 1 --repetitions 3 --output-dir outputs/map-registration-cold

Use a new output directory. For retained-cache measurements, use --compiler-cache retained --repetitions 4 and a separate output directory. All workers use one BLAS thread and verify their actual imports against frozen complete runtime manifests. Complete wall time includes startup, imports, input reads/preparation, GPU initialization/compilation, all evaluation work, full result saves, hashes and worker source verification. Parent guards and aggregation are outside the timed process.

The archives retain complete observations/results, both source capsules, manifests, exact launchers/schema, cache-file hashes and publication byte bindings. Original operator paths remain in provenance. Raw scans and compiled binaries are external. Recompute the evidence without a GPU or raw scans:

python -m tools.audit_cuda_map_preparation docs/assets/cuda_map_registration_fusion/cold
python -m tools.audit_cuda_map_preparation docs/assets/cuda_map_registration_fusion/retained

See the full cold records and retained-cache records.

Final local validation: 83 relevant unit/device/benchmark/archive tests pass, including FP32/FP64 matrices in C, column-major and strided layouts, native weighted systems, exact tie/gate selection and empty targets. Full-repository Ruff, strict mypy over 376 source files and strict MkDocs pass. All 130 schema and ledger exports are byte-identical to the measured baseline. CPU-only mutation tests reject changed fused kernel identities, projection/residual/Jacobian work and original query counters even after outer file hashes are rebound.