A Tile-based Feature map Compression Method for Multi-stride Winograd Algorithm

There are three commonly used convolution acceleration algorithms: Winograd, FFT, and FFA. However, these methods face challenges when accelerating multi-step convolution in hardware. In this paper, a Tile-based Feature map Compression Method (TFCM) was proposed to accelerate both convolutions with filter (3,2) (i.e., filter size is 3 and stride is 2) and (3,1) by using Winograd. Furthermore, the feasibility of TFCM was also discussed based on some popular CNNs. In addition, a Winograd-based Acceleration Unit (WAU) was designed to implement TFCM. The evaluation results show that our proposed methods can achieve 45.93%~75.28% multiplication savings when applied to VGG16, MobileNetV1. Moreover, our accelerator can accomplish 1.71 TOPS while deploying VGG16, achieving a 1.27× to 2× improvement in DSP efficiency per MHz compared with the state-of-the-art accelerators.