Advanced indexing

[ipdemes]

Adding support for advanced indexing

[manopapad]

Note that, when deciding where to put the extra dimensions in a partial-basic, partial-advanced expression, NumPy considers constants as advanced arrays, so if it sees something like x[:,ind,:,1] it considers it two advanced arrays separated by :, and puts the added dimensions in the beginning:

import numpy as np
import cunumeric as cn
def foo(lib):
    x = lib.ones((3,4,5,6))
    ind = lib.ones((2,2), dtype=int)
    return x[:,ind,:,1].shape
print(foo(np))
print(foo(cn))

prints:

(2, 2, 3, 5)
(3, 2, 2, 5)

[ipdemes]

@manopapad : thank you for finding this bug. I will push the fix shortly

[ipdemes]

On the most recent checkout the index_routines.py test is failing on my local Mac, with:

I have fixed it by basically removing this check. In some cases, when index array is promoted at the front of the array, Legate will decide not to partition it and, instead, broadcast it to all index points. This doesn't change the logic in the task and doesn't effect correctness
@manopapad

[manopapad]

This is ready to merge. Great work!

[ipdemes]

This is ready to merge. Great work!

Thank you for your help!