Support for array types other than NumPy

grlee77 · November 18, 2021, 10:54pm

I created a repo that has a demo of dispatching to different backends with a uarray-based approach

grlee77/uskimage-demo/blob/main/README.md

## General Overview

This repository is a proof-of-concept demo for different concrete backend
implementations for [scikit-image](https://scikit-image.org/). It is not
intended for use by downstream projects. In practice, such backends would
ideally be maintained by their upstream projects
([cuCIM](https://github.com/rapidsai/cucim), etc.) so that they can be tested,
ensuring that they stay in sync.

This respository is compatible with a specific [uarray](https://uarray.org/en/latest/)
development [branch of scikit-image](https://github.com/grlee77/scikit-image/tree/uarray)
which implements `uarray` multimethods for the majority of the scikit-image API.
Most of the multimethods were autogenerated using a script provided in the
tools directory here. That branch also differs from scikit-image main in that some
currently deprecated arguments have already been removed. Specifically,
`multichannel` was removed in favor of `channel_axis` and `selem` was
removed in favor of `footprint`.

## Installation

This file has been truncated. show original

Benchmark results and some discussion is given the README there. I mostly autogenerated the uarray multimethods on the scikit-image side using a script provided in the repository above, but it does create an additional layer of indirection that can make things a little harder for new contributors to understand. The Dask backend there is a more user-friendly way of using apply_parallel where a user would just have to register a dask_backend and never have to call apply_parallel (or the equivalent dask.array.map_blocks) directly.

We should make some corresponding demos of possibilities with the array_namespace approach that will become possible with NumPy 1.22. That approach is relatively simple to understand, but doesn’t help us with all of our functions that currently use Cython code.