Background

In #739 I introduce boundary exclusion to the is_point_in_polygon kernel that is used both by the original point_in_polygon as well as quadtree_point_in_polygon.

This addition adds 6 multiplies and 15 adds to the kernel per iteration. Because we added python benchmarks in 22.10 I was able to benchmark the new feature branch against the 22.12 branch.

The benchmark runs each point_in_polygon algorithm on 50m points and 30 polygons taken from the naturalearth_lowres feature database that is available in GeoPandas. You can run the benchmark with

cd cuspatial/python/cuspatial/benchmarks
pytest api/bench_api.py::bench_point_in_polygon
pytest api/bench_api.py::bench_quadtree_point_in_polygon

The default benchmark only tests 10m points, not 50m. You'll have to modify the tests to change to 50m. Finally, to run the benchmark using the new is_point_in_polygon kernel requires that you checkout my PR branch and rebuild cuspatial cpp and python.

Benchmark result

On my machine with a GV100 the boundary-exclusion branch is 30% faster, for some reason, than the original, taking 871msec on 50m points with boundary-exclusion and 1161msec without the new changes. Quadtree point in polygon is 10% slower with boundary exclusion at 36msec and 32msec with the original.