Created by: kexinzhao

Currently, float16 op only supports conv2d, it is straightforward to extend it to support conv3d.