Benchmark Cython vs Numba
A work-in-progress assessment of Cython vs Numba with Numpy as a baseline for various array manipulation. So far
- Pairwise distances (
pairwise_distances.py): a list of 3D coordinates as input, the list of all distances between any two points as output; - Laplacian filter (
laplacian.py): a 512x512 image as input, compute the Laplacian, filter the resulting image with a threshold.
Running the benchmarks is as simple as running the scripts with python. There are some prerequisites of course: the easiest is to install them using conda and the exported environment.yml:
% conda env create -f environment.yml
This will create an environment named numerics: activate it before running the benchmarks.
Conclusions so far.
- On MacOS Sierra with all packages installed with conda, with Westmere X5690
- Pairwise distances: Cython and Numba are both about 5 times faster than Numpy
- Laplacian: Numba is 2-3 times faster than Cython which is more than twice faster than Numpy