Created by: kexinzhao

fix #9627 (closed)

Tensor core example: https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/

Enabling tensor core for float16 cudnn conv has been tested to provide significant speedup.