Created by: xiaolil1

We met in-place issue when enable MKLDNN Reshape OP. For the case: no-mkldnn op ---> mkldnn op ( Reshape or other in-place supported Ops).

We investigated this issue and found the "data transform" will generate new tensor and change the input tensor of Reshape Op, while the output not, which will break the in-place between input and output.

To solve this issue, we developed two solutions, and submitted two PRs. First one is #16914 , there will be overhead when "in-place" occurs. Second one is #16915 , overhead only happens on "in-place" occurs in "data transform", but will add 5 parameters for "TransformData".

We have verified correctness of these two solutions and will send out performance result later.