cc @vadimkantorov . This should satisfy your request, right? Let me know if it is not the case.

It seems so.

I'm a surprised this is satisfied because @colesbury suggested a new function reshape would be solving this.

Will there be an easy way to check if two shapes are compatible? So that I can call contiguous() if I can't get away with view

I can now see there's isViewable and that view throws if it returns False. Should this function be exposed to Python as well? So that the user doesn't have to rely on exceptions for control flow

Will this PR enable also that example torch.Tensor(1,1).expand(1,2).view(-1)?

@vadimkantorov : Yes it enables that example. Thanks for bringing that up! It helped me identify and fix a bug with zero-strided tensors in my previous code :). I just added tests for that case.

Hmmm about reshape, I didn't notice that comment before. It seems that it should do same thing as the view in this PR. If we decided to use reshape name, I can update this PR.

The original example in #3653 and torch.Tensor(1,1).expand(1,2).view(-1) are different in that the original example doesn't change numel() while the second is supposed to?

Please check with @colesbury for this design and semantics :)

They both don't change numel. In the second example, we have

view:       2
size:    1, 2
stride:  1, 0

The "contiguous" chunk here is all dimensions, i.e., [1, 2]. "numel" in this chunk is exactly the same as "numel" of the view chunk [2].

Nice!

We should still add reshape(), which is like view() but will make a copy if isViewable() is False.

If you decide to implement reshape() later, I think you can move both reshape() and view() to NativeFunctions in ATen (example). The advantage of that is you only need on implementation of stride/size calculations for all types.