Enhance the CPU Power of PaddlePaddle

[luotao1]

PaddlePaddle should enhance its CPU Power due to two reasons:

• For company, the number of CPU Cloud Clusters they used are much larger and bigger than that of GPU.
• For person, most laptops they used only have CPU processor.
• Some workloads, such as RNN, sparse update etc, are more suitable for CPU.

Thus, combined with PaddlePaddle's 10 aspects in 2018, several things related with CPU Power should be considered:

1. Fluid framework (Discussed with @jacquesqiao ): Add auto data format transform mechanism for different library and hardware. Add new fluid op with MKL-DNN/MKL library.
2. Multi-device Multi-thread (Discussed with @jacquesqiao ): Use the full capacity of CPU when multi-threading on both MKL and OpenBLAS library.
3. Distribution (Discussed with @typhoonzero ): Improve distributed training with LARS. Improve distributed training with LARS #6811
4. Benchmarking (Discussed with @dzhwinter ): CPU benchmarking is a part of whole benchmarking. Need to provide both training and inference benchmarking among different Math libraries, and compared with other DL frameworks.
5. Documents: Enhance the documents and select some models to tell users: 1) when and how to choose the most efficient Math Library, such as the usage range of MKL-DNN library; 2) how to use the full capacity of CPU.
6. Training (Discussed with @wanghaoshuang ): If our documents and models are more perfect, it would be much easier for user training.
7. NLP support (Discussed with @lcy-seso ): Enhance the CPU power of some NLP ops (mainly RNN/LSTM/GRU) based on specific workload.
8. Speech support (Discussed with @kuke ): Enhance the CPU power of some Speech ops (mainly RNN/CNN) based on specific workload.
9. Image support (Discussed with @qingqing01 ): Enhance the CPU power of some Image ops (mainly CNN and Detection) based on specific workload.
10. Inference (Discussed with @Xreki ): Enhance the CPU power of some Inference ops' forward computation based on specific workload. Easy to deploy. No need to change or configure the inference API to use MKL-DNN related operators.

Hope any suggestions!