Some dtypes not supported by numba

[aulemahal]

Describe the bug
Big-endian floats (and maybe others) are not supported by numba, which leads to TypingError on some sparse operations.

To Reproduce

import numpy as np
import sparse as sp

little = np.array([1, 1, 1, 1, 1, 1], dtype='<f8')
big = np.array([1, 1, 1, 1, 1, 1], dtype='>f8')

Anp = np.random.randint(100, size=(4, 6)) / 100
A = sp.COO.from_numpy(Anp)

A.dot(little) # works
A.dot(big) # fails

The other way around is the same;

Abig = sp.COO.from_numpy(Anp.astype('>f8'))

Abig.dot(little) # fails
Abig.dot(big) # fails

Expected behavior
I expected the operation to work in all cases.

Could on-the-fly byte swap be implemented? May be with a warning?

In this specific case, on-the-fly byte swapping is what numpy does:

r = Anp.dot(big)  
r.dtype # <f8, little-endian

For the case with Abig, where the sparse array itself is in a unsupported dtype, if not on-the-fly conversion is done, maybe there could be an issue raised, or at least a warning, on creation? To signify that the input array doesn't support the full features of sparse.

System

• OS and version: Ubuntu 18.04
• sparse version : 0.13.0
• NumPy version : 1.21.2
• Numba version : 0.53.1

Additional context

[hameerabbasi]

Hmm, this is an upstream Numba bug, as you already said. It's hard to detect the full scope of these kinds of things without try-catch.

[hameerabbasi]

On-the-fly swapping could be done, in theory. That's something I would consider cleaner.