Rejected conservative FP32 pruning for CUDA odometry

Not adopted. The integrated candidate preserves all outputs, but its final complete-worker ratio is only 1.038x. The existing asynchronous-normal runtime is retained. The experimental implementation and its tests are preserved in the source capsules and the measured revisions below.

The experimental CUDA scan adapter prunes spatial searches using outward FP32 intervals while retaining the original FP64 candidate distances, neighbor ordering, normals, residuals and systems. Both box/plane rejection and a candidate-point prefilter use lower distance bounds. A rejected branch cannot contain a closer candidate under the retained arithmetic.

Conversions enclose each coordinate or split plane using directed rounding. Subtraction, nonnegative squared products and sums round downward, with an additional FP32 margin before comparing against the FP64 candidate threshold. Nonfinite lower bounds disable rejection. CUDA documents these directed conversions and arithmetic intrinsics. The implementation does not copy external kernel code.

Preparation is limited to 32–65,536 target points with coordinate magnitude at most 1e6. A relative extent guard keeps tiny translated clouds on the native path. Each registration iteration also checks source/matrix magnitude. The native tree arrays and original FP64 kernels remain available for fallback. Intervals add bounded target storage inside the existing payload policy; this does not bound total allocator memory or process RSS.

A completion event orders cross-stream consumers and drains preparation on abandonment. The target also records completion of its initial capture, covering registration on another stream without normal priming. No CPU thread is added. The map-registration adapter keeps its existing implementation.

Complete public-data comparison

Integrated candidate	Previous whole median	Candidate whole median	Ratio
Repeated target extent scans	54.383 s	63.992 s	0.850x
Reuse existing KD-tree extrema	53.534 s	51.576 s	1.038x

The first candidate is revision 61921e57; the revised candidate is 2249a4f7. Both compare against bb131739, which already includes asynchronous normal preparation. The revised pair ratios are 1.018x and 1.060x. The extent reuse removes a duplicate point-cloud reduction; it does not establish a substantial whole-calibration improvement. Both complete cohorts are retained.

Two alternating fresh-process pairs replay six seconds of the public GLIM Livox bag with four map rounds, one BLAS thread, CUDA scan odometry and CUDA fitter planes. Each worker starts with a separate empty CuPy compiler cache. Whole time includes imports, initialization/compilation, every initial/gyro/feedback/final odometry pass, association, all fits/covariance, noise updates, holdout/controls, hashing, saves and frozen-import verification. No warmup is discarded.

The exact gate permits only the source-bound kernel identity, declared pruning tag/counters and run metadata to differ. Every calibration/evidence field, full per-round fit, trajectory, covariance, iteration and previous work counter must match exactly. Actual interval query rows and fused calls are checked against the total GPU work. CPU-only auditing verifies the source capsules, executed imports, kernel identities, full fits, quality and aggregate timings.

The bound archive retains all eight results and full fits, raw logs, source capsules/manifests, launcher/schema, source-bound regression run and schema comparison. Raw bag data and compiler caches are not redistributed. Tests cover neighbor membership and distance bits, ties, coordinate scales, translated clouds, payload boundaries, cross-stream reads, abandoned trees and native fallback.

Host load is unreserved. These ratios describe this dense six-second workload, not all point clouds, and are not multiplied by earlier cohorts. The clip has no independent extrinsic truth and its final verdict remains INCONCLUSIVE. CPU stays the default. The dense benchmark's coordinate/motion grids remain off by default, and all 130 public schema/ledger exports are unchanged.

Reproduce

python -m tools.benchmark_bspline_lidar_imu --comparison cuda-cuda \
  --baseline-commit bb131739c0d14d30fc0e1660ce4f2ee587d6d32e \
  --prior-fit-plane-backend cuda --current-fit-plane-backend cuda \
  --require-exact-spatial-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-interval-pruning-pairs
python -m tools.audit_bspline_lidar_imu_benchmark docs/assets/cuda_interval_pruning

Reproduction requires the measured candidate revision 2249a4f7 and its frozen launcher, since the interval path is not enabled in the retained runtime. Keep runtime source, HEAD, launcher and comparison schema fixed until every worker finishes. Use a new output directory. The source archive contains the candidate implementation without a public configuration or schema migration.