Skip to content

Latest commit

 

History

History
86 lines (70 loc) · 4.88 KB

spacepoint-parser.md

File metadata and controls

86 lines (70 loc) · 4.88 KB

Space-point extraction

larsoft2supera::spacepoint extracts native recob::SpacePoint and optionally associated recob::Hit records into Atomic's standard-library-only supera/data/SpacePoint.h. Neither the parser nor its records depend on LArCV. This stage performs no cuts, voxel selection, truth matching, or output writing. The shared selection stage now connects these collections to EventRecord and the LArCV2 writer; see the output review. Raw cloud parsing remains independently usable.

Explicit input interpretation

convert_spacepoints(points, profile) retains collection order, positions in cm, chi-square, and optional native IDs. The required profile is either:

  • Covariance: preserves all six covariance entries, in native xx, xy, yy, xz, yz, zz order, in square cm.
  • IcarusCluster3D: interprets ErrXYZ()[1] as charge and [3] as charge asymmetry. The remaining slots are producer placeholders, not covariance. Charge units remain explicitly source-defined. This profile is for the audited ICARUS cluster3D producer, not every producer in ICARUS.

Floating-point invalidity and field-specific native sentinels are normalized at this boundary using the shared validity contract. Optional columns distinguish absent information from present-but-empty products.

Optional hit associations

read_spacepoints(event, point_tag, profile) reads points only. Passing a SpacePointHitInput additionally requests one association tag and an explicit list of allowed hit product tags. The Event-like template is tested with art::Event; gallery integration has not been validated.

The reader requests Assns<SpacePoint, Hit>; art supplies the partner of native Assns<Hit, SpacePoint> products. A table must refer to exactly the requested point product. Mixed point-product tables are unsupported and rejected. Missing requested products, null endpoints, out-of-range keys, unexpected hit products, and duplicate configured product identities fail before returning a record. Points-only mode never loads hits or associations.

The shared collection owns each referenced hit once per source product/key per read call. Flat links reference these compact hit collections, with offsets marking each point's span. Repeated links and serialized within-point order are preserved. Original hit keys, event-scoped product identity, configured source tag, and full native product size survive compaction. There are no art pointers or handles in the result. Sharing across separate point collection reads is not yet implemented.

All native hit scalar fields are retained, including wire address, channel, view/signal codes, timing, amplitude, summed ADC, integral, errors, fit quality, multiplicity, local index, and degrees of freedom. Source enum codes retain their native meaning; no detector semantics are inferred from them.

Validation and performance

tools/integration.sh runs native converter tests, installed-consumer validation, and tests/spacepoint_integration.py against the checksum-pinned cached sample. The latter checks every point and the full association graph, with native fixtures for shared hits, duplicates, multiple products, empty events, and eight malformed/missing-input cases. tools/spacepoint-checks.sh adds converter coverage and address/undefined-behavior sanitizer checks using genuine native classes.

For three independent timing repetitions, run spacepoint_integration.py with --repeats 3 after the normal native setup and an optimized build. Its CSV files separate point access, conversion, hit access, association access, and link/hit conversion. Optional timing performs no per-object clock reads. Validation is outside timed regions; timing repetitions disable validation. Product access includes framework loading/decompression and depends on cache state.

owned_capacity_bytes counts owned vector payload capacity, excluding strings, transient indexing arrays, framework products, and allocator overhead. Use the MemoryTracker report for whole-process peak memory. Historical whole-Supera timings in performance.md include much more work and cannot serve as a direct parser speedup denominator.

Normalization reports aggregate the read call. Hit examples identify the first point referencing that hit, and each distinct referenced hit contributes once. They do not claim to be separate per-hit-product reports. Converter sanitizer coverage does not instrument external art libraries or the templated event reader; the latter is exercised by native framework fixtures and full-sample validation.

The local native check completed with 58/58 converter lines and 2/2 functions covered (45/70 gcov branches, including compiler-generated exception paths). Reports live in results/spacepoint-parser-checks; the final checksum-pinned framework validation is in results/spacepoint-parser-final. The shared model also passes its dependency-free CTest and clang-tidy checks.