So, the behavior now is basically "Only use the input's memory format to determine strides when we have to resize the output and aren't given explicit strides", right? Posting a snippet below to confirm what I'm talking about.

input = torch.ones((2, 2, 2, 2), memory_format=torch.channels_last)
output = torch.ones((2, 2, 2, 2))
# output doesn't get resized, so we don't clobber the output's strides/memory_format.
torch._C._nn.upsample_nearest2d(input, (2,2), out=output).is_contiguous(memory_format=torch.channels_last)
# returns False

output = torch.ones(0)
# output gets resized, so we DO restride according to the input's memory format
torch._C._nn.upsample_nearest2d(input, (2,2), out=output).is_contiguous(memory_format=torch.channels_last)
# returns True

That all seems reasonable, since:

• I can imagine if a user explicitly specifies an output with the correct size, they probably don't want it to get re-strided?
• We also don't allow you to pass in an output tensor with non-zero size if it doesn't match the expected output size. (Well, the behavior is deprecated anyway). So really your only options are to pass in a zero-sized tensor, indicating that you want whatever properties the input tensor has, or you specify your own output tensor with the right size.

Just calling it out explicitly, since I think the behavior of the upsample kernels before porting them to structured was that they unconditionally copied the suggested memory format of the input (here).

Expected out behavior is described here https://github.com/pytorch/pytorch/wiki/Developer-FAQ#how-does-out-work-in-pytorch, it is what you describe. Thanks for calling it out explicitly, and you are correct that previous behavior of upsample didn't conform to those rules.

probably these tests belong in test_nn, but no need to change this PR.