Complete comparison of overlapped CUDA target preparation¶
Not adopted. Runtime sources are restored to the previous revision. The candidate and its 51 passing device/unit tests are retained for a controlled repeat, rather than enabled as a demonstrated whole-calibration improvement.
| Scope | Sequential preparation | Overlapped preparation | Observed ratio |
|---|---|---|---|
| Complete worker | 97.706 s | 86.766 s | 1.126x |
| Unchanged four CPU fits, including covariance | 16.384 s | 18.928 s | 0.866x |
| Unchanged CUDA normal preparation | 7.259 s | 5.895 s | 1.231x |
| Pair | Sequential complete worker | Overlapped complete worker | Observed ratio |
|---|---|---|---|
| 1 (sequential first) | 72.916 s | 76.616 s | 0.952x |
| 2 (overlap first) | 122.497 s | 96.917 s | 1.264x |
The paired comparisons change direction: one is slower and the other faster. Unchanged CPU fit and GPU normal times also vary substantially. The median ratio is descriptive; these runs establish no stable end-to-end speed gain. Adoption is deferred pending a repeat under less host contention.
Two alternating fresh-process pairs replay the public GLIM Livox bag's first six seconds with four map rounds. This is a descriptive observation on an unreserved host. One worker prepares the local map, captured native tree and CUDA target while the caller deskews and downsamples the source. Registration waits for the same prepared target, with all original kernels and arithmetic. One preparation is outstanding at most, and result snapshots close the pool. The worker uses the caller's device and a reusable private stream, synchronized before publishing the target. Rotation callbacks remain on the caller thread.
The previous revision is 88e4a3a1db0e67dc538e3faf73c9a12b570ab3de and the current revision
is 7fe26771f6fb2a8017be54b9053b112fcaf2efc0. The input bag SHA-256 is
f57a2e85339316d4235ab693ab2876aac5e113625c7f8b0f7873b62f92e09643. No raw scans or compiler caches are redistributed.
Each worker uses one BLAS thread and a separate initially empty CuPy compilation cache. Whole timing includes startup/imports, decoding, every odometry pass, target/source work and worker shutdown, association, fits/covariance, noise updates, evidence, hashing, saves and source verification. It includes first GPU initialization and compilation. The unchanged dense policy uses 0.03 m voxels, three local-map scans, an opt-in 1 nm coordinate grid and motion grids of 1 micrometre translation and 0.1 microradian rotation-vector components. Those grids remain disabled by default. CPU odometry remains the default.
All complete physical/evidence fields, initial estimates, execution identities and GPU work counters match exactly, excluding only provenance timestamps, commands and source revisions. Every full fit record matches exactly excluding wall time, including covariance, trajectory, calibration, costs, dimensions, status and iterations. Both pairs satisfy the stronger exact-output protocol. The final real-data verdict remains INCONCLUSIVE; there is no independent extrinsic truth. Return budgets, neighborhoods, convergence and controls are unchanged. No public option, dependency or result-schema change is introduced.
Unchanged fit and normal timings expose host variation, and the complete-worker ratio is specific to these runs. Ratios from distinct scopes or cohorts are not multiplied. Overlap can increase simultaneous CPU working memory; the number of preparations is bounded, rather than process RSS.
The index binds all four full results/measurements/logs, both source capsules,
launcher, schema, publisher and an existing Windows test-failure reproduction.
prototype-validation additionally binds the candidate's successful 51-test
device/unit run, exact test files, validation launcher, source manifest and log.
The archive audit checks these bindings without importing an accelerator.
test_default_cache_dir_respects_xdg expects HOME to redirect Path.home();
it fails with both current and frozen baseline sources locally. The baseline
receipt binds that source manifest, validation log and reproduction launcher.
The test and production cache-path policy are unchanged.
Audit without a GPU or bag:
python -m tools.audit_bspline_lidar_imu_benchmark docs/assets/overlapped_target_preparation