Writing Custom Array Containers

NumPy's dispatch protocols are the standard route for custom N-dimensional containers that interoperate with its API.

They let a container provide its own implementation of NumPy operations.

Dask arrays distribute N-dimensional work across chunks or nodes; CuPy arrays execute N-dimensional operations on GPUs.

A class opts into custom dispatch through __array_ufunc__ and __array_function__.

__array_ufunc__ handles universal functions such as numpy.multiply and numpy.sin.

__array_ufunc__ receives:

• ufunc, such as numpy.multiply
• method, identifying the requested ufunc method
• inputs, which may mix several types