search

rapidsai / cuspatial

CUDA-accelerated GIS and spatiotemporal algorithms
https://docs.rapids.ai/api/cuspatial/stable/
Apache License 2.0
612 stars 154 forks source link

Introduce data structure to simplify multigeometry API #703

Open isVoid opened 2 years ago

isVoid commented 2 years ago

[UPDATE] 04/10/2023

Recent development added multipoint_range, multilinestring_range, multipolygon_range, which are flexible views over geometry arrays. cuspatial's API should be refactored using these data structures. For refactor demos. see: https://github.com/rapidsai/cuspatial/pull/979 (polygons argument of quadtree_point_in_polygon) and https://github.com/rapidsai/cuspatial/blob/branch-23.06/cpp/include/cuspatial/experimental/linestring_distance.cuh

Original Post

As demonstrated in https://github.com/rapidsai/cuspatial/pull/677, the geometry input of the header only API can quickly get out of control for complex geometry. We should try to simplify the APIs to improve developer experience.

Device view over the physical memory of nested types

The data structure should be a view over the physical memory layouts, thus it should be cheap to construct the data structure on host, passing the structure to device and invoke the method on device.

Unlike cudf::column_view and cudf::column_device_view, the data structure is completely templated (not type erased), and always assumes the view are device views.

Assumes GeoArrow memory layout (#649)

The structures holds a view to offset arrays and point arrays. Offset arrays are assumed to always conform to arrow's offset array layout, which is specified in (https://arrow.apache.org/docs/format/Columnar.html#variable-size-binary-layout).

Accessors

Due to the complex nature of geometries, developers may want to support different patterns of access from the data structure.

Element-wise accessors (top-down access)

The simplest of all is element-wise accessor. A kernel is launched on a per-geometry level, accessors like .element(idx) should return a geometry object from the array.

Component-wise accessors (bottom-up access)

Other parallel patterns may require a thread to work on one component from the array, such as paralleled on a point of a multilinestring_array. In this case bottom-up traversal utilities should be supported, and should be easily implemented by a binary search in the offset array.

- [ ] https://github.com/rapidsai/cuspatial/issues/704
- [ ] https://github.com/rapidsai/cuspatial/issues/705
- [ ] https://github.com/rapidsai/cuspatial/issues/706
- [ ] https://github.com/rapidsai/cuspatial/issues/707
- [ ] https://github.com/rapidsai/cuspatial/issues/708
- [x] Simplify header-only pairwise_point_in_polygon API
isVoid commented 1 year ago

Since the related PR lists haven't all completed. Reopening this issue.