vibespatial.kernels.core.wkb_decode

GPU WKB decode kernel pipeline (P1a).

Implements a 5-stage GPU pipeline that reads WKB byte streams directly on device and produces OwnedGeometryArray results without host round-trips.

Stages:

  1. Header scan kernel – 1 thread per WKB record, reads endian + type tag

  2. Family partition – CuPy boolean masks to bucket records by geometry type

  3. Size scan + offset computation – sizing kernels + CCCL exclusive_sum

  4. Family decode kernels – per-family coordinate extraction

  5. Assembly – build OwnedGeometryArray via io_pylibcudf helpers

ADR-0033 dispatch tiers:

Tier 1 (custom NVRTC) for geometry-specific decode Tier 3a (CCCL) for prefix-sum offset computation

Attributes

logger

Functions

decode_wkb_device_pipeline(...)

GPU WKB decode pipeline.

Module Contents

vibespatial.kernels.core.wkb_decode.logger
vibespatial.kernels.core.wkb_decode.decode_wkb_device_pipeline(payload_device, record_offsets_device, record_count: int) → vibespatial.geometry.owned.OwnedGeometryArray

GPU WKB decode pipeline.

Reads WKB byte streams directly on GPU using custom CUDA kernels. Returns an OwnedGeometryArray with device-resident geometry buffers.

Parameters

payload_devicedevice array (uint8)

Contiguous WKB byte payload on device.

record_offsets_devicedevice array (int32)

Byte offsets for each record (length = record_count + 1).

record_countint

Number of WKB records.

Returns

OwnedGeometryArray

Device-resident geometry array.