Add more core NumPy Functions

[hameerabbasi]

• [x] broadcast_arrays/broadcast_to (Quansight-Labs/uarray#150)
• [x] concatenate/stack (Quansight-Labs/uarray#151)
• [x] errstate (#17)
• [x] pad (#12)
• [x] Set routines (Quansight-Labs/uarray#144)
  • [x] unique
  • [x] in1d
  • [x] intersect1d
  • [x] isin
  • [x] setdiff1d
  • [x] setxor1d
  • [x] union1d
• [x] Sorting routines
  • [x] sort (Quansight-Labs/uarray#145)
  • [x] lexsort (Quansight-Labs/uarray#145)
  • [x] argsort (Quansight-Labs/uarray#155)
  • [x] msort (Quansight-Labs/uarray#155)
  • [x] sort_complex (Quansight-Labs/uarray#155)
  • [x] partition (Quansight-Labs/uarray#155)
  • [x] argpartition (Quansight-Labs/uarray#155)
• [x] Searching
  • [x] argmax
  • [x] nanargmax (#10)
  • [x] argmin
  • [x] nanargmin (#10)
  • [x] argwhere (#10)
  • [x] nonzero (#10)
  • [x] flatnonzero (#10)
  • [x] where (#10)
  • [x] searchsorted (#12)
  • [x] extract (#12)
• [x] Counting: count_nonzero (#16)

[rgommers]

Maybe errstate & co, the set routines (unique in particular) and the sort routines? and np.pad?

[hameerabbasi]

Updated the list. 😄

[hameerabbasi]

Question: Do we want to provide a default implementation?

[rgommers]

Question: Do we want to provide a default implementation?

Depends on what that means I guess. I think it may be fine if

import unumpy as np
np.arange(5)

gives an error. However I think if you import one backend it should work with explicitly having to set the backend.

import unumpy as np
from unumpy.backends import XndBackend  # or some such line
np.arange(5)

Or do you mean the order in which backends are chosen?

[hameerabbasi]

Ah, I mean (to be specific): Do we want a default implementation for MultiMethods in terms of others, if it is reasonably efficient?

[rgommers]

That doesn't make it much clearer for me ..... Can you give a concrete example?

[hameerabbasi]

So, I currently dispatch sum to add.reduce with the appropriate arguments. Do we want to do something similar with some of these other functions?

[rgommers]

Hmm, I'd guess not. The signatures in numpy for those two things don't even match, and other libraries may not match numpy. Why not leave that up to the backend?

[hameerabbasi]

To make it easier for the backend... And make gradual implementation easier while still having the option for a more complex or efficient one.

[rgommers]

How many functions can this be done for? sum, min, max, what else?

[hameerabbasi]

A lot of them, actually... nan-reductions most of the time, usual reduction functions, stack/concatenate/block, argwhere/where/nonzero, etc.

To be clear, the backend can still override them. There will be a big bold disclaimer saying these can change.

[rgommers]

Hmm yeah okay, I can see how that saves doing the exact same thing in multiple backends. So yes with a clear statement in the docs that this is done, why not?

[hameerabbasi]

Broadcasting and concat/stack done.

[hameerabbasi]

Searching and pad done!

[hameerabbasi]

Closed this since everything here is done.