Indexing not Stable

[Inzlinger]

Description
When using indices to acces array elements the order in the result is not stable for different splits.

To Reproduce

import heat as ht

a = ht.array([0,1], split = 0)

indices = ht.array([0, 1, 0, 1], split=0)

b = a[indices]
b.balance_()

print(b._DNDarray__array)

print(b)

With 1 process this gives [0, 1, 0, 1] as expected
Run this with n=2 and it changes the order to [0, 0, 1, 1].

Expected behavior
The same ordering in the result for all splits.

[ClaudiaComito]

Thanks @Inzlinger , I'm looking into getitem for other reasons (including #656) so I'll check this out as well if @coquelin77 agrees.

[coquelin77]

i am pretty sure that this is caused by the assumption if the index is a split DNDarray, it is assumed to be non-repeating. when the indices are gathered they are then sorted. this is to determine which process has the requisite data. this could also be done with a mask on the index instead. then the local key which is generated would use this mask based on where the mask == the local rank number. or something along these lines.

@ClaudiaComito I have no objections to you looking into this one. i dont think that I will get to this for a little while otherwise. when i have more time available i will look at this again.

[ClaudiaComito]

@Inzlinger solving this issues requires communication, i.e. a distributed getitem (more like a get_distributed_item?). This would be needed and awesome and is on our list. If you want to start, you're very welcome! If you want we can organize a chat about how to proceed.