Created by: jczaja

PaddlePaddle often applies reshape to Tensors for example in Tensor::Reshape or in some ops that do falttening. MKL-DNN code then rearannge data to NCHW layout and continue operation in that not very efficent memory arrangement. MKLDNN 1.0 introduce reshape of memory objects (as well as subview) that probably can be used to remove conversions to NCHW arrangement (and perhaps allow to implement broadcasting with subview)