[MLIR][TORCH] Add E2E support for aten.as_strided op

[vivekkhandelwal1]

This commit adds the e2e support for the aten.as_strided op by decomposing it into a series of other torch operations.

Fixes #4191.

The failing tests for Tosa config are tracked by #4272.

[vivekkhandelwal1]

The CI will fail for the nightly run. That's because of a different issue related to the PyTorch integration. I will fix that in a separate patch.

[vivekkhandelwal1]

@sjarus @justin-ngo-arm Some of the tests are failing for Tosa config since I have added a decomp for the as_strided op.

[justin-ngo-arm]

@vivekkhandelwal1 Can you let me know which tests are failing for TOSA? Do you want us to fix those for you, or will you fix them/add them to XFail?

[vivekkhandelwal1]

@vivekkhandelwal1 Can you let me know which tests are failing for TOSA? Do you want us to fix those for you, or will you fix them/add them to XFail?

@justin-ngo-arm In this patch, I have added a decomposition for the as_strided op. As a result, the tests that were passing earlier for the Tosa config are now failing because, instead of being lowered in the TorchToTosa pipeline, the op is being decomposed and cannot be converted to Tosa. Hence, the tests that passed when you added the lowering for as_strided, some of them are now failing.

In total, 11 tests are failing; they are as follows:

ChunkListUnpackDynamic_Module_basic
ChunkListUnpackUnevenDynamic_Module_basic
ChunkListUnpackUneven_Module_basic
ChunkListUnpack_Module_basic
NativeGroupNormModule_basic
SplitTensorGetItem_Module_basic
SplitTensorLastSmallerModule_basic
SplitTensorListUnpackModule_basic
SplitTensorNegativeDimModule_basic
SplitWithSizesListUnpackModule_basic
SplitWithSizes_Module_basic

I don't know how much time it would take to fix these tests, and that has to be done in a separate patch. So, if it's not blocking you, then can I just xfail them in this patch? And, you may raise a separate PR with the fix where the tests can be re-enabled.

[justin-ngo-arm]

@vivekkhandelwal1 Marking them with XFail works for me. Can you do me a favor by creating an issue with a list of those failed tests and assign them to me please? That will serve as a better reminder for me to track this, and I will fix them when I have time to do so. Thanks a lot!

[vivekkhandelwal1]

@vivekkhandelwal1 Marking them with XFail works for me. Can you do me a favor by creating an issue with a list of those failed tests and assign them to me please? That will serve as a better reminder for me to track this, and I will fix them when I have time to do so. Thanks a lot!

Sure, I will do that.

[vivekkhandelwal1]

In this patch, I have added a decomposition for the as_strided op. As a result, the tests that were passing earlier for the Tosa config are now failing because, instead of being lowered in the TorchToTosa pipeline, the op is being decomposed and cannot be converted to Tosa. Hence, the tests that passed when you added the lowering for as_strided, some of them are now failing.

In total, 11 tests are failing; they are as follows:

ChunkListUnpackDynamic_Module_basic
ChunkListUnpackUnevenDynamic_Module_basic
ChunkListUnpackUneven_Module_basic
ChunkListUnpack_Module_basic
NativeGroupNormModule_basic
SplitTensorGetItem_Module_basic
SplitTensorLastSmallerModule_basic
SplitTensorListUnpackModule_basic
SplitTensorNegativeDimModule_basic
SplitWithSizesListUnpackModule_basic
SplitWithSizes_Module_basic

Hi @justin-ngo-arm, I have created an issue here #4272.

[justin-ngo-arm]

Thank you @vivekkhandelwal1! I really appreciate it.

[praveen-g-ctt]

LGTM

[sahas3]

LGTM.

A clarification: Is there a general guidelines for adding tests? Are e2e tests sufficient or LIT tests are required as well?

Thanks!

[sjarus]

Looks like @justin-ngo-arm has this covered @vivekkhandelwal1 . I tried to add him on the issue, but for some reason his ID doesn't show up as someone who could be added as reviewer on the issue.