[TorchToLinalg] Direct lowering from Torch to Linalg for torch.aten.convolution_backward

[a-sidorova]

Description:

• Added the direct lowering pass for torch.aten.convolution_backward from Torch to Linalg. Enabled this pass by default. The pass generates linalg.generic ops instead of linalg.conv_<> for better lowering.
• Removed the previous pass DecomposeAtenConvolutionBackwardOp from Torch/Transforms/DecomposeComplexOps.cpp.
• Created new lit tests for backward convolution in the separate file convolution_backward.mlir. Also added more test cases for better test coverage.
• Added new e2e tests for backward convolution for better test coverage.

Issue:

• Improve lowering of backward convolution to linalg #4355

[a-sidorova]

@zjgarvey hey! May I ask you to take a look when you're available? Thank you in advance for the review.

[zjgarvey]

Nice! This is an excellent start.

We need to keep the existing decomposition for other backends. I have a few other comments for you to look at, but that's the biggest blocker right now.

[a-sidorova]

@zjgarvey thank you for review! I have applied your comments in the latest commit. Could you please take a look at the changes one more time?

[zjgarvey]

A few more comments before we merge. Sorry if this took a while to get back to.

[a-sidorova]

@zjgarvey thank you for the one more review! I applied your comments - may I ask you to take a look again?

[zjgarvey]

I think this is good with the caveat that you revert the commit changing OpTy::create -> rewriter.create as rewriter.create is indeed deprecated now. It's not a huge deal since we will need to go through the whole codebase and update this eventually, but would be good to do this before merging.

Additionally, feel free to restore the llvm_unreachable if you think it's appropriate.

[a-sidorova]

I think this is good with the caveat that you revert the commit changing OpTy::create -> rewriter.create as rewriter.create is indeed deprecated now. It's not a huge deal since we will need to go through the whole codebase and update this eventually, but would be good to do this before merging.

Additionally, feel free to restore the llvm_unreachable if you think it's appropriate.

Thank you for the review! I reverted the both commits - May I ask you to relaunch CI please? It should be passed now.

[zjgarvey]

https://github.com/llvm/torch-mlir/actions/runs/20088891516/job/57707427422?pr=4384#step:8:4063

I think this is a quick thing to fix.

[a-sidorova]

https://github.com/llvm/torch-mlir/actions/runs/20088891516/job/57707427422?pr=4384#step:8:4063

I think this is a quick thing to fix.

My bad - missed this one error within other previous errors. Fixed in 24b0df0. Could you please relaunch one more time again please? 😃

[sjain-stanford]

Could you please relaunch one more time again please? 😃

I just did, but we need you to get triage access so you can trigger builds. I'll ask around.

[zjgarvey]

Ill add her to triage. Should have done that a while ago

[a-sidorova]

Ill add her to triage. Should have done that a while ago

@zjgarvey thanks! I pushed the commit with updates in xfail sets - CI has been automatically launched.

[zjgarvey]

There appears to be a hang on the fx importer tests. If you know of any that would cause a hang, you might need to add them to the crashing set instead of xfail.

[a-sidorova]

@zjgarvey I corrected xfail and crashed tests for configs in e2e test in 3d1fa85 - dilated bwd convolution was stuck somewhere. But now CI has been successfully passed and we can merge this PR to master