fix pixel_shuffle return empty

[zpcore]

Fixes #5886. Basically lowered the op using the XlaOps instead of going through pytorch decomposition.

Removed ExpectCounterChanged("xla::permute_copy", cpp_test::GetIgnoredCounters()); because we don't call xla::permute_copy in the lowering.

[JackCaoG]

Do you know if pixel_shuffle a view op?

[zpcore]

Do you know if pixel_shuffle a view op?

This is how pytorch do the composition: https://github.com/pytorch/pytorch/blob/58047205ed098c04ec045e66fc39dcc70b60600b/torch/_refs/nn/functional/__init__.py#L1183-L1190.

XlaOp:Reshape served the functionality of view op here. That's why we don't see view in our lowering.

[JackCaoG]

I guess my question is, if you do something like

input = torch.randn(1, 9, 4, 4)
output = torch.nn.functional.pixel_shuffle(input, 3)
input[0][0][0][0] = 1

I would expect output to also change since it is a view op. Do we still have that property with this lowering?

[zpcore]

No, it's not. Just tested on both CPU and XLA device, the input and output are in different shapes.

[JackCaoG]

ok.. yea input and output will be different shape, but if on CPU you modify the source tensor, output tensor won't get modified, it is a not a view op

[zpcore]

Oh I see, I should modify the contents to test the view. Thanks!