Created by: wangkuiyi

It seems that caffe2::TensorShape and other tensor shape constraints are not supposed to be used in CUDA kernels. Does this mean that we can remove the CUDA code generation macros from dim.h?