Created by: tonyyang-svail

During the benchmarking of https://github.com/dzhwinter/benchmark/blob/master/fluid/mnist.py, I found conv_2d with cuDNN is slow. It turns out the element_wise_add element_wise_add_grad bias takes about 80% of the time while the actual cudnn_conv only takes 10%.

@dzhwinter This looks pretty bad to me. Please confirm if you've seen the similar results during your benchmarking? I am using nvprof and NVIDIA Visual Profiler.

I am wondering if we can do one of the following.

1. See if can improve the elementwise_add, but I am not sure how hard it is in Eigen. @reyoung
2. Combine conv and add bias into one operator, like CudnnConvBaseLayer.cpp in v2.

We need to improve this due to the importance of this operator in vision task.

Issue related: https://github.com/PaddlePaddle/Paddle/issues/7862.