Implement tensor redistribution for binary operations

[Markus-Goetz]

After PR #165 is merged, implement binary operators for two tensor with set, but different split axis (e.g. split=0 and split=1).

Think of a rule to decide which tensor should be resplit. Proposal (might require modification):

1. attempt to redistribute tensors with fewer elements
2. should both have the same amount of elements, redistribute tensor that are split along a minor axis, i.e. split >= 1.
3. should both have the same number of elements and both have a minor axis, redistribute the one with larger minor axis

---

EDITED July 31st 2023 by @ClaudiaComito Here's the current status of binary operations heat.core._operations.__binary_op():

• [x] ensure that operand types are promoted correctly
• [x] ensure that binary ops are possible between operands split along the same axis but with misaligned lshape_maps (e.g. as a result of indexing)
• [ ] ensure that operands are on the same MPI communicator
• [ ] ensure that binary ops are possible between operands split along different axes

---

Reviewed within #1109

[Cdebus]

@Markus-Goetz I will try out these rules. However, what happens in 2) if none of the tensors is split among a minor axis?

[Markus-Goetz]

In this case you may simply add them locally as already implemented and working

[Cdebus]

No, I mean if they both have a split > 1, but along different axes... or is that unrealistic?

[Markus-Goetz]

That would be rule 3.) if both have a minor axis, than the one with the larger minor axis will be redistributed

[Markus-Goetz]

Couple more things that came to my mind that have to be considered:

• __binary_op should ensure that the comms of the two operands match. We cannot perform the operation if the data is distributed across different comms. Exception: the split of one of the operands is None. Local data can always be used

• Ensure that the operands types are properly promoted. See also types.promote_types.

[Cdebus]

Why was that closed? Issue is not fixed, PR #268 bumped due to issues with MPI communication and resplit

[coquelin77]

My apologies. i think I saw that the PR was closed (at the time it was a month after the bump, and i think that you had closed the PR). I am not sure but I think that I had thought that the PR was merged and then it was a stay issue which was not closed at the time of merging. Sorry @Cdebus

[Cdebus]

No worries ^^ I was just surprised, because fixing the _binary_op was my original intention but didn't work due to resplit due to allgather, and that's how I ended up with that mess

[ClaudiaComito]

Status? @coquelin77 @Markus-Goetz can we break this into smaller issues?

[ClaudiaComito]

Alright, here's the current status of binary operations heat.core._operations.__binary_op():

• [x] ensure that operand types are promoted correctly
• [x] ensure that binary ops are possible between operands split along the same axis but with misaligned lshape_maps (e.g. as a result of indexing)
• [ ] ensure that operands are on the same MPI communicator
• [ ] ensure that binary ops are possible between operands split along different axes

---

Reviewed within #1109

[github-actions[bot]]

Branch 167-Implement_tensor_redistribution_for_binary_operations created!

[mrfh92]

Stil active and ongoing work.

(Reviewed within #1109 )