Created by: hedaoyuan

At present, there are only sgemm and CUDNN-based convolution implementations in Paddle. In a model training or prediction without a GPU, can only select convolution calculations based on sgemm. However, based on the sgemm convolution calculation, performance is not optimal in many scenarios. see here

We also encountered the problem of convolution computing performance when deploying Paddle into some product environments. At the same time, there are many excellent convolution implementation libraries, I think we can try to import it into Paddle to improve the Paddle convolution calculation performance.