Created by: tonyyang-svail

Profiler

Use either nvprof or paddle.v2.fluid.profiler

Benchmark:

1. resnet_cifar10: https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/tests/book/test_image_classification_train.py
2. RNN

Steps:

1. Update cuDNN, this includes:
   1. Update NVIDIA Container to CUDA 9 and cuDNN 7 (Current version is CUDA 8 and cuDNN 5)
   2. Link cuDNN into Paddle in cudnn.h
   3. Search over Paddle Operators, directly use CuDNN when possible
2. Reduce the number of kernel launch of the kernel is written through eigen