Reduce by key

[syurkevi]

Implements one dimensional reduce by key for all backends.

[umar456]

I only did a quick run through this PR. I didn't look into the logic just yet. There is a lot of stuff here. There are a couple of suggestions I have made. Your tests are only testing your float code paths. They should be testing all types. There is also a lot of repetition there. I think things can be improved. It is exciting that we will finally have a reduce by key in ArrayFire. Awesome work.

[9prady9]

@syurkevi Is this ready for review ? Has umar's feedback been addressed ?

[syurkevi]

@syurkevi Is this ready for review ? Has umar's feedback been addressed ?

All outstanding feedback has been addressed, however we were also considering handling the multidimensional cases. If the multidimensional case needs to be added then it isn't ready yet.

[mlloreda]

@syurkevi this doesn't build for me unless I add #include "spdlog/sinks/stdout_sinks.h" to Logger.cpp.

[9prady9]

That's probably so because this branch is quite outdated from master.

Rebased latest master and minor repairs

[9prady9]

Looks good. I need to do another pass on the kernels of each backend.

[mark-poscablo]

forge's commit SHA is older than master's in this PR. Should that be updated?

[WilliamTambellini]

Hi @mark-poscablo
Could you tell if that PR would allow to do a "ragged max" :
input
1 2
3 4
5 6
keys
3 2
expected output
5 4

• the max of the first col is 5 because we are considering the 3 elements of the first column
• the max of the 2nd col is supposed to be 6 but that max reduction returns 4 because the algo only has to look at the 2 first elements of the 2nd column

Tks
W.

[mark-poscablo]

Hi @WilliamTambellini , we should probably ask the author @syurkevi, but I think this PR does not address that use case. Reduce by key would do something like the following:

Using maxByKey:

Input : 1 2 3 4 5 6 7 8 9
Keys  : 1 1 1 2 2 2 3 3 3
Output: 3 6 9

It essentially does the reduction operation on groups, where elements that have the same keys are in the same group.

[9prady9]

I have reviewed api level (c/cpp/unified) source code and CPU backend implementation. I am still going through CUDA backend. Will submit another set of comments once that is done.

[9prady9]

Is there an order to the keys that will be present in the output keys for all the newly added *ByKey functions ? We need to mention this somewhere in the documentation whatever the expectation should be.