vibespatial.io.osm_pbf_memory

Physical-work admission for bounded PBF relation batches.

Ordinary relations are partitioned by expanded coordinates and member graph bytes. A single relation that exceeds the workspace budget uses CUDA managed scratch: kernels still execute on the GPU and the CUDA driver can page its working set into host backing memory. No geometry is evaluated on the host.

Functions

relation_workspace(work_bytes, budget, *[, kind])
concatenate_results(results)	Assemble the final native columns with an explicit paging admission.

Module Contents

vibespatial.io.osm_pbf_memory.relation_workspace(work_bytes, budget, *, kind='relation')

vibespatial.io.osm_pbf_memory.concatenate_results(results)

Assemble the final native columns with an explicit paging admission.

Input geometry buffers are retained through a native composition. libcudf writes final attribute columns once; when that transient copy exceeds the remaining device envelope, its allocation uses CUDA managed backing.