vibespatial.spatial.point_grid_index¶

Reusable dense-cell index for conservative point-tree candidates.

The carrier stores each point exactly once. Predicate queries may consume its cell-aligned bbox superset because an exact predicate immediately refines the relation. Bbox-only queries continue to use the exact Morton path.

Classes¶

PreparedPointGridIndex

Device-resident point rows grouped into a fixed square grid.

Functions¶

`point_grid_index_kernels`()
`point_grid_preflight`(...)	Inspect grid shape and bytes without building or querying the provider.
`prepare_point_grid_index`(...)	Build or return the NativeSpatialIndex-owned dense point grid.
`point_grid_relation_superset_query`(native_index, ...)	Materialize one admitted full relation candidate superset.
`point_grid_superset_query`(...)	Return cell-conservative pairs for immediate exact refinement.
`point_grid_query_row_partitions`(...)	Plan query-row slices whose point-grid supersets fit a pair budget.
`point_grid_candidate_not_in_other_superset`(...)	Mark candidates absent from an aligned point grid's prior superset.

Module Contents¶

class vibespatial.spatial.point_grid_index.PreparedPointGridIndex¶

Device-resident point rows grouped into a fixed square grid.

grid_size: int¶

cache_key: object¶

precision_plan: object¶

readiness: object¶

xmin: float¶

ymin: float¶

xmax: float¶

ymax: float¶

sorted_tree_rows: object¶

cell_counts: object¶

cell_offsets: object¶

integral_counts: object¶

property device_bytes: int¶

vibespatial.spatial.point_grid_index.point_grid_index_kernels()¶

vibespatial.spatial.point_grid_index.point_grid_preflight(native_index, *, query_count: int, pair_budget: int, force_eligible: bool = False) → tuple[vibespatial.spatial.point_partition.PointPartitionPreflight | None, vibespatial.spatial.point_partition.PointPartitionDecline | None]¶: Inspect grid shape and bytes without building or querying the provider.

vibespatial.spatial.point_grid_index.prepare_point_grid_index(native_index, *, query_count: int = 0, pair_budget: int = 0, force_eligible: bool = False, admission=None) → tuple[PreparedPointGridIndex | None, vibespatial.spatial.point_partition.PointPartitionDecline | None]¶: Build or return the NativeSpatialIndex-owned dense point grid.

vibespatial.spatial.point_grid_index.point_grid_relation_superset_query(native_index, query_bounds)¶

Materialize one admitted full relation candidate superset.

Relation-producing public APIs are pair-shaped by contract. They retain a single named allocation fence for the exact output capacity, then use the same provenance-bound guarded scatter as bounded reducers.

vibespatial.spatial.point_grid_index.point_grid_superset_query(native_index, query_slice: vibespatial.spatial.point_partition.PointPartitionQuerySlice) → vibespatial.spatial.query_types._DeviceCandidates | None¶: Return cell-conservative pairs for immediate exact refinement.

vibespatial.spatial.point_grid_index.point_grid_query_row_partitions(native_index, query_bounds, *, pair_budget: int, force_eligible: bool = False, admission=None) → tuple[vibespatial.spatial.point_partition.PointPartitionQueryPlan | None, vibespatial.spatial.point_partition.PointPartitionDecline | None]¶

Plan query-row slices whose point-grid supersets fit a pair budget.

Fixed-size row blocks are reduced on-device and cross once as compact planning metadata. Greedy block slices use exact count sums as capacities, so every admitted tile fits pair_budget without a per-tile allocation fence, padded launch, or full row-count export. Oversized blocks are refined down to single-row metadata; consumers then scan those rows in bounded dense tree-row tiles.

vibespatial.spatial.point_grid_index.point_grid_candidate_not_in_other_superset(native_index, prior_plan: vibespatial.spatial.point_partition.PointPartitionQueryPlan, candidate_query_rows, candidate_tree_rows)¶

Mark candidates absent from an aligned point grid’s prior superset.

This is an internal paired-reduction primitive. Candidate rows refer to two aligned point columns, so candidate_tree_rows can address the geometry owned by flat_index directly. Oversized query rows are marked already seen because their prior reduction scanned every aligned tree row.