Created by: wangkuiyi

If we represent tensors in the same memory footprint, is it correct that we can call cuDNN, MKL, and Neon to operate tensors on various devices?