I'm not completely sure this is correct. Also, assuming that it is correct, we should update utils.compute_elementwise_output_strides -- other operators run into the same problem.

This is the function we use to compute strides when the input is not "non_overlapping_and_dense": https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/TensorIterator.cpp#L1276.

If we think it is exactly input.contiguous.stride(), then we should add some unittests to check that

EDIT: .contiguous calls exactly that function so it is actually the correct thing. So my ask is that we should update utils.compute_elementwise_output_strides to call

if torch._prims_common.is_non_overlapping_and_dense(input):
    strides = input.stride()
else:
    strides = input.contiguous().stride()

@zou3519 Thank you for checking the correct behavior. I updated utils.compute_elementwise_output_strides to call

if torch._prims_common.is_non_overlapping_and_dense(input):
    strides = input.stride()
else:
    strides = input.contiguous().stride()

exactly when 1 tensor is given.
For the case of multiple tensors, it was not clear how to adapt this without more substantial changes. There is quite a bit of logic in compute_elementwise_output_logical_to_physical_perm. Let me know if you want me to update the behavior there as well.

No, let's fix the one tensor case first.

After we get this merged, I can look into the multiple tensor case if you'd like.

@laithsakka @bobrenjc93 any dynamic shapes issues around implementing this like this? Otherwise, if the tests pass, I will assume we can ship this.

I took a look at failing tests, still figuring it out. It is an issue that was introduced by the change with inductor.

def test_clone_not_memory_dense():
    def foo():
        x = torch.randn(10, 8).t()[::2, ::2]
        y = x.clone()
        return y
    y = foo()
    assert y.stride() == (1, 4)
    print("uncompiled")
    y = torch.compile(foo, backend="eager")()
    print("eager")
    assert y.stride() == (1, 4)
    y = torch.compile(foo, backend="aot_eager")()
    print("aot_eager")
    assert y.stride() == (1, 4)
    y = torch.compile(foo, backend="inductor")()
    print("inductor")
    print(y.stride())
    assert y.stride() == (1, 4)

we prob should rename def is_non_overlapping_and_dense to
is_non_overlapping_and_dense_or_false. its fine its reasonable definition of unbacked semantics its probably would be good to ensure inductor behave similarly with unbacked inputs.

@laithsakka Thank you for the note, these commits were merged on PR, #163017. However that PR did not fully capture all the correct behavior. When the future work is implemented, I will make a note of this preserved naming scheme.