Created by: LihangLiu

以下是我在fluid1.0上跑sequence_pad的结果，cpu和cuda跑出来的output不一样，如下 case1: 一致

Input: 
x is 1-level lod tensor
x.lod = [[0, 8,]]
x.dims = [8, 1]

Output:
CPU version: y.dims = [1, 8, 1]
CUDA version: y.dims = [8, 1]

case2: 不一致

Input: 
x is 1-level lod tensor
x.lod = [[0, 8, 12]]
x.dims = [12, 1]

Output:
CPU version: y.dims = [2, 8, 1]
CUDA version: y.dims = [2, 8, 1]

当x有多个sequence的时候，结果是一样的。当x实际只有一个sequence的时候，结果是不一样的。