Added overload methods for Numpy's ufunc.reduce

[kc611]

Reference an existing issue

Resolves #6313 #2751

This PR aims to implement the numpy.ufunc.METHOD as a callable inside njit-ed functions so that we can do stuff like:

import numba as nb
import numpy as np

@nb.njit
def add_twice(x, y):
    return 2*x + 2*y

custom_vectorize = nb.vectorize([], identity=None, target="cpu")
vec_add = custom_vectorize(add_twice)

@nb.njit
def test_fn(x, axis):
    return vec_add.reduce(x, axis) 

print(test_fn(np.ones((10, 20)), 0))

[kc611]

The current issue I was facing is that this call cannot detect the implemented dufunc_reduce_direct method in C code. Is it a type issue ?

[gmarkall]

You need to add the symbol - try using llvmlite.binding.add_symbol('dufunc_reduce_direct', <addr of dufunc_reduce_direct>)

[kc611]

I'm also not too familiar with this C function declarations of Numba so I might be doing something wrong here.
(Looking at the failing build)

[gmarkall]

I don't think you need to wrap the return type in NUMBA_EXPORT_FUNC (at least for now) - can you try without it?

[gmarkall]

Also, I think this should be returning a PyObject* since, that's what a PyCFunctionWithKeywords returns: https://docs.python.org/3/c-api/structures.html#c.PyCFunctionWithKeywords