Original Version

This version works fine for CPU, MIC and GPU platforms.

1. On CPU, it supports all the convolution processes.
2. On MIC, it could use the same function as CPU. Besides, it also incorporates libxsmm in the forward phase.
3. On GPU, it adopts Cudnn and Cublas. Specifically, we optimize GEMM and Convolution (in progress) on Kepler GPU.