With this proposal, the definition of Reshape would be sufficiently restrictive that I think

• the documentation should be extended to specify in more detail what "the new shape is not compatible with the old shape" means
• the functionality should be hidden behind a TryReshape method.

Also, note that the bit where numpy and the current TensorSpan deviate in #106539 is just a.Reshape(a.Lengths), on which numpy returns a whereas the proposal here would throw an exception, which it should probably not (although maybe there is a reasonable definition of Reshape that doesn't allow it and I just don't see it, see the first bullet point).

@soerenwolfers I am about to push a change that will let a.Reshape(a.Lengths) always work.

We will always have some difference between numpy because numpy will allocate new memory behind the scenes if it needs to and we cannot allocate new memory for the TensorSpan. It should work just fine for Tensor though as most operations copy for that in comparison to the TensorSpan where its usually not a copy, so we shouldn't get into the memory situations where a reshape wouldn't work correctly.

I agree a TryReshape would be a good thing to have for TensorSpan but not for Tensor. @tannergooding any thoughts on any of this?

/azp run runtime-dev-innerloop

/backport to release/9.0

Started backporting to release/9.0: https://github.com/dotnet/runtime/actions/runs/11052885500

@microsoft-github-policy-service rerun

@dotnet-policy-service rerun