Created by: qingqing01

1. It's no problem with relative precision 1e-3 when testing several times on my local machine.
2. But the testing failed with 1e-2 in the TeamCity. And the output size is 4096, only one value's relative precision is over 1e-2.

[00:15:50]	I8018 16:15:50.212294 30060 test_NetworkCompare.cpp:172] ------------------------------ Check Network Output ------------------------------
[00:15:50]	I8018 16:15:50.212296 30060 test_NetworkCompare.cpp:176] OUTPUT VALUE: 0
[00:15:50]	I8018 16:15:50.212419 30060 test_NetworkCompare.cpp:164] Row: 8, network A output : -1.13249e-05 network B output : -1.14441e-05
[00:15:50]	/paddle/paddle/gserver/tests/test_NetworkCompare.cpp:168: Failure
[00:15:50]	Expected: 0
[00:15:50]	To be equal to: nNum
[00:15:50]	Which is: 1

https://paddleci.ngrok.io/viewLog.html?buildId=6071&buildTypeId=Paddle_GpuUnittest&tab=buildLog&_focus=4229