Add dim order variant clone operator

[Gasoonjia]

Currently clone operator is removed in the export pipeline without tracing its (possible) memory format / dim order changes within the copy and may lead to wrong memory format / dim order changes in the final exported model.

One of the essential features to solve the issue is creating dim order variant clone operator to ensure the correct dim order propagation in export graph. With the dim order variant clone op, we can make the TestRemoveCloneOpsTransform remove clone operator with dim order aware of.

cc. @digantdesai

Alternatives

No response

Additional context

No response

RFC (Optional)

No response

cc @JacobSzwejbka @angelayi

[digantdesai]

so,
(1) create clone_dim_order op
(2) update the MemoryFormatOps pass to replace clone with this op
(3) update TestRemoveCloneOpsTransform pass to be dim_order aware (not memory_format)

[keyprocedure]

I'd like to work on this! I have a few questions:

1. After registering clone_dim_order in dim_order_ops_registery.py, will MemoryFormatOpsPass automatically replace aten.clone with clone_dim_order, or are additional changes required to ensure that the mapping occurs?

2. Once _clone_dim_order.out is declared in edge_dialect_aten_op.yaml, where should the corresponding kernel implementation be placed?

3. If clone_dim_order changes the dim order of its input, is it treated as a permanent op, or does its removal from the graph before final export need to be implemented in a pass?

4. I wasn’t able to find TestRemoveCloneOpsTransform in the codebase. Is this referring to a test that should be added in test_memory_format_ops_pass.py to validate the removal of clone_dim_order?

[Gasoonjia]

@keyprocedure Thanks for taking this!

1. In the AOT side, yes, the replacement will be happen. We still need to impl runtime operators.
2. for its aten version, please put it here https://github.com/pytorch/executorch/tree/main/kernels/aten/cpu; for its portable version, please here https://github.com/pytorch/executorch/tree/main/kernels/portable/cpu
3. if it will change dim order, it should not be removed for dim order correctness.
4. i think it should be RemoveCloneOpsTransform pass:

   executorch/backends/transforms/remove_clone_ops.py

   Line 28 in c310efe

   class RemoveCloneOpsTransform(ExportPass):

   @digantdesai can you confirm it?