Need to add some new algorithm(like direct, winograd) for convolution calculation.

[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.

[hedaoyuan]

Convolution performance

Convolution layer performance in a ResNet model with 3x3 kernel. Test in a raspberry environment, and build with NNPACK. We can see that some layers with gemm algorithm performance better, while some layers with wt8x8 algorithm performance better.

[Xreki]

NNPACK is supported in v2 Paddle and new features won't be implemented based on v2, so close this issue.