CPU map normals and parallel queries¶
LiDAR–LiDAR map registration can compute native map normals when a fit first
reaches a point and retain them for later iterations, jackknife fits, held-out
scoring and controls. MapSampleOptions(query_workers=8) also runs independent
SciPy query rows on eight CPU workers. Both correspondence searches and normal
neighborhoods use this budget. The default remains one worker.
Two separate development cohorts compare the existing one-worker CPU path with reached-point normals and eight query workers. Complete fresh-process ratios are 1.987x and 2.138x. All six paired eight-worker processes are faster, but these unreserved-host observations do not establish a portable minimum of 2x. Bag decoding and reference odometry are outside this workload. A later whole-bag GLIM comparison includes those stages and observes 1.249x at 0.2 m map voxels and 1.687x at 0.03 m map voxels.
Complete measurements¶
Each worker reads all original points from nine spent public KITTI scans, constructs six local-map samples, runs the training fit and four block jackknife fits, computes covariance and held-out evidence, runs all six axis controls, validates and saves the complete artifact, and verifies input and executed-source hashes. Every process has one BLAS thread. Startup, imports, reads, preparation, all evaluation work, verification and saves enter wall time.
| First cohort | Previous CPU, 1 worker (s) | Reached normals, 1 worker (s) | Reached normals, 8 workers (s) |
|---|---|---|---|
| Triad 1 | 29.589 | 22.552 | 13.914 |
| Triad 2 | 28.931 | 24.062 | 16.403 |
| Triad 3 | 30.331 | 27.784 | 14.889 |
| Median | 29.589 | 24.062 | 14.889 |
The one-worker ratio is 1.230x; the eight-worker ratio is 1.987x. The initial 2x advancement gate is missed. Its original comparator also flagged different artifact creation times. Original records, errors and the negative advancement decision remain unchanged. The independent audit normalizes that one additional run-metadata field and verifies that all physical and evidence bits match.
| Follow-up cohort | Previous CPU, 1 worker (s) | Reached normals, 8 workers (s) | Query-certificate candidate, 8 workers (s) |
|---|---|---|---|
| Triad 1 | 41.999 | 19.387 | 23.074 |
| Triad 2 | 41.458 | 19.978 | 23.417 |
| Triad 3 | 34.002 | 15.395 | 23.653 |
| Median | 41.458 | 19.387 | 23.417 |
The retained middle variant is 2.138x faster than the previous CPU path in this cohort. The certificate experiment tries to retain nearest hits while triangle bounds exclude every competing point. It reduces searches but adds enough memory traffic and checks to make complete execution 20.8% slower than reached normals alone (0.828x). It fails its original gates and is excluded from the implementation. No unfavorable observation is discarded or replaced.
Keeping reached normals and explicit query budgets is a post-hoc engineering choice, informed by both cohorts and bit parity. It is not a claim that either frozen screen passed every advancement gate. The two cohorts have different instrumentation and host conditions; their times and ratios are not pooled or multiplied. No complete two-LiDAR bag speedup is established by these records.
Geometry, scope and retained work¶
The workload uses KITTI 2011_09_30_drive_0018_sync, frames 114–116, 229–231
and 344–346. Six samples reuse three raw windows under generated reference
poses. Targets copy central scans with fixed 0.01 m Gaussian noise, seed
20261012. Four sample blocks train and two hold out. These overlapping copied
samples are an execution control, with no independent dual-LiDAR accuracy or
temporal-validation evidence. Raw windows are not motion compensated.
Normal computations fall from 2,286,890 to 1,247,324 rows, 45.5% fewer. The retained change still searches every correspondence query: 9,782,926 rows per complete evaluation. It retains neighborhood size, native tie ordering, covariance arithmetic, NumPy eigenspaces, planarity without an eigenvalue floor, distance gates, Huber loss, all iteration limits and every evidence call.
All 18 workers retain all 90 complete fits. The independent audit compares 60 paired fits, full evaluation records, complete covariance and saved artifacts bit for bit after only declared run metadata and query-budget options are normalized. Decisions, counts, iterations, policies and all controls match. Prepared array and split hashes also match. These are execution-agreement checks, not independent accuracy or SOTA evidence.
Each sample owns its normal cache. A normal is computed once, with one extra
completion byte per map point. Explicit MapSample.prepare() still materializes
all normals, and caller-prepared normals retain their captured policy. No CUDA
or other optional dependency is imported by the default CPU path. The final
numerical implementation is the measured 09f27193c9418ec9b6584caea459661e047fe875;
a follow-up only records nondefault query budgets in bag-artifact options.
The rejected certificate runtime is retained as source evidence only.
Use and audit¶
calibrex lidar-lidar ros2 path/to/bag \
--reference-topic /reference/points --target-topic /target/points \
--initial initial.yaml --registration-query-workers 8 \
--dataset-family your_dataset --dataset-license "your dataset terms" \
--output outputs/lidar-pair.yaml
python -m tools.audit_cpu_map_query_cache docs/assets/cpu_map_query_cache --check
The CLI records the explicit budget in command provenance; the bag API records nondefault budgets in artifact options. Select a budget for the available CPU; small queries can spend more time starting threads. This option changes execution scheduling and introduces no result-schema change.
The bound archive retains both cohorts, all original errors and decisions, all complete worker artifacts and fits, launchers, input/source bindings, source deltas and the CPU-only audit. The shared prior source capsule is already published in the preceding resident-scan validation archive. No raw scans or fitted coordinate arrays are redistributed.