Created by: wangkuiyi

The implementation contains three parts:

1. VisitPlace (35 lines) https://github.com/PaddlePaddle/Paddle/blob/037ce12ee4f3167fc31f2fdd21e5390b42c806b8/paddle/fluid/platform/place.h#L97-L132

2. The CPU implementation of TensorContainsNaN (~100 lines)

   https://github.com/PaddlePaddle/Paddle/blob/7b54f16855ca3aaef6ff1cdb9bf7a4d4f1d8b6c1/paddle/fluid/framework/tensor_util.cc#L139-L241

3. The CUDA implementation (~100 lines)

https://github.com/PaddlePaddle/Paddle/blob/8370fa82789bc4f05a3239700681818b0ff224ae/paddle/fluid/framework/tensor_util.cu#L19-L117

There uses about 235 lines of C++ code to test if a tensor contains NaN or INF!