alexeybelkov / parinterp

Python binding of 2D delaunay triangulation and parallel linear interpolation

parinterp

Python library for parallel linear 2D interpolation, it consists of two main parts:

• Delaunay Triangulation from delaunator-cpp lib, still sequential part
• Linear 2D interpolation via barycentric coordinates, parallel part

Being almost fully written in C++ and binded via pybind11, library achieved very high perfomance

Installation

git clone https://github.com/alexeybelkov/parinterp
cd parinterp
pip install -r requirements.txt
./install.sh

Usage

import numpy as np
from parinterp import Linear2DInterpolator
n, m = 1024, 2
points = np.random.randint(low=0, high=1024, size=(n, m))
points = np.unique(points, axis=0)
x_points = points[: n // 2]
values = np.random.uniform(low=0.0, high=1.0, size=(len(x_points),))
interp_points = points[n // 2:]
n_jobs = -1 # will be equal to num of CPU cores
interpolator = Linear2DInterpolator(x_points, values, n_jobs)
# Also you can pass values to __call__ and rewrite the ones that were passed to __init__
interp_values = interpolator(interp_points, values + 1.0, fill_value=0.0)

TODO

• Currently, there are a lot of presumably unnecessary reallocations, try to find a way to remove them
• Using Python's scipy KDTree is more like a crutch, find fast C++ one-nearest-neighbour algorithm implementation
• Find a way to build a fast bijection $f : S \rightarrow {0, 1, ..., n}$, where $S \subset \mathbf{N}$
• Maybe consider this Parallel Nearest Neighbors in Low Dimensions with Batch Updates
• Compare perfomance with and without safety

Usefull links

Point location algorithm was highly inspired by paper Fast randomized point location without preprocessing in two- and three-dimensional Delaunay triangulations