Writing Custom Array Containers

NumPy's dispatch mechanism, is the recommended way to write custom N-dimensional array containers that are compatible with the NumPy API.

This provides custom implementations of NumPy functionality.

Applications include dask arrays, and N-dimensional array distributed across multiple nodes, and cupy arrays. Cupy arrays are an N-dimensional array on a GPU.

NumPy allows a class to indicate that it would like to handle computations in a custom defined way through the interfaces arrayufun and array_function.

The first array_ufunc_ covers Universal functions (ufunc), a class of functions that include, for example, numpy.multiply and numpy.sin

The _array_ufunc receives:

• ufunc, a function like numpy.multiply
• method, a string, differentiating between numpy.multiply and other numpy multiply
• inputs, which could be a mixture of different types