Hmm if new shapes are as same as the old shapes, ideep::tensor::reshape will not create a copy and thus y will share the same memory as x. This could be messy since then there will be two aten opaque tensors sharing the same underlying ideep tensor.

@bddppq We wanted to keep the semantics of reshape as much as possible here, i.e. do shallow copy on plain-format MKL-DNN tensor (layout of which is contiguous) and do deep copy otherwise (see https://github.com/intel/ideep/blob/d7304e0345c3f0647cb020a71682d680b9f0424a/include/ideep/tensor.hpp#L1100). For the shallow copy case, ideep::tensor y{x} would make y copy the metadata of x while share its underlying storage. This is the same semantics as the shallow copy of a native CPU tensor. So this should work fine here.

@jgong5 It's fine to follow the semantics of cpu tensor to return the input tensor as output if the shapes are the same, but it's not fine to let two different aten opaque tensors to hold the same underlying ideep tensor.

@bddppq But y and x are two different ideep::tensor object, right? They just share the same underlying storage while hold their own metadata. It is similar to how two CPU tensors share the same underlying storage while have their own metatdata. Make sense?

@bddppq @XiaobingSuper OK. Sounds like we should call y.set_descriptor(x.get_descriptor()) explicitly after ideep::tensor y{x} in order to guarantee y uses a new MKL-DNN descriptor while sharing the underlying buffer of x. We should have wrapped an explicit shallow copy API in ideep to do such a thing actually.

@jgong5 @XiaobingSuper I would say just add the following lines at the beginning of this function :-)

if (self.sizes() == inferred_size) {
  return self;
}

Yes, I will change it. Thanks!

@bddppq, @jgong5 , changed it, thanks!

@bddppq y.reshape would create a new descriptor by calling set_descriptor() so additional y.set_descriptor(x.get_descriptor()) is not necessary here. With the recent change of @XiaobingSuper , the behavior of mkldnn_reshape looks like below, mimicking the semantics of reshape of native CPU tensor:

1. If the requested shape is the same as self, return self directly.
2. If the format of MKL-DNN tensor is public (or plain), the storage is contiguous, the returned MKL-DNN tensor shares the storage of self and copies all metadata of self with the new shape.
3. For other situations, return a deep-copied tensor of self with the new shape.

@jgong5 @XiaobingSuper This sounds good to me.

Why is this part of the public API? If you want to have helper functions prepend an underscore to their name!