[uncategorized_lowerings] Add lowering for torch.aten.round.decimals

[meshtag]

Implement missing lowering for the op in a similar fashion as done by torch inductor. Also fix data movement and reduce op variants patterns to correctly handle explicitly declared legal ops.

Inductor decomposition ref: https://github.com/pytorch/pytorch/blob/main/torch/_inductor/decomposition.py#L223.

[vivekkhandelwal1]

I am not able to understand why the lowering for the op torch.aten.round.decimals is added as a custom op lowering, why don't do it the right way?

Please follow the steps specified here https://github.com/llvm/torch-mlir/blob/main/docs/Torch-ops-E2E-implementation.md to add the op lowering.

[vivekkhandelwal1]

I am not able to understand why the lowering for the op torch.aten.round.decimals is added as a custom op lowering, why don't do it the right way?

Please follow the steps specified here https://github.com/llvm/torch-mlir/blob/main/docs/Torch-ops-E2E-implementation.md to add the op lowering.

Hi @meshtag, is this PR still of interest to you?

[meshtag]

Hi @meshtag, is this PR still of interest to you?

Hey @vivekkhandelwal1, looks like I missed this PR. Apologies for that.

I am not able to understand why the lowering for the op torch.aten.round.decimals is added as a custom op lowering, why don't do it the right way?

AFAICT, we have two options here:

1. Get this op directly in the mlir-world and then lower it from there.
2. Register this op for decomposition (it already exists in torch, as pointed above) and then decompose it at a higher level (ideally before the fx->mlir translation layer) and then directly handle the decomposed op (which is already supported).

This PR tries to go via path 1. I am not sure if we should discard path 2 though, can you please share your thoughts on this.

Please follow the steps specified here https://github.com/llvm/torch-mlir/blob/main/docs/Torch-ops-E2E-implementation.md to add the op lowering.

Sure, I can do this (if we want to go via this path). Thanks for pointing it out.

[vivekkhandelwal1]

I am not able to understand why the lowering for the op torch.aten.round.decimals is added as a custom op lowering, why don't do it the right way?

AFAICT, we have two options here:

1. Get this op directly in the mlir-world and then lower it from there.
2. Register this op for decomposition (it already exists in torch, as pointed above) and then decompose it at a higher level (ideally before the fx->mlir translation layer) and then directly handle the decomposed op (which is already supported).

This PR tries to go via path 1. I am not sure if we should discard path 2 though, can you please share your thoughts on this.

If you want to go thorough the path 1, then I don't think you have to do much. Just register this op in Torch-MLIR, and the same lowering which you have written can be used. Also, you would be able to test the correctness of your lowering e2e.