At last, pytorch "@" operators can be called from fortran so that they can call dgemm()... Hahaha.