임베디드 CPU에서 Winograd 컨볼루션 알고리즘에 기반한 Depthwise 컨볼루션의 구현

Abstract

The depthwise convolution layer is often used in lightweight convolutional neural networks (CNNs). In this study, the Winograd convolution algorithm is used to carry out depthwise convolution, and the proposed method is implemented using NEON instructions on an ARM Cortex-A53 CPU. The performance of the proposed method is evaluated with respect to two tensor storing orders (NCHW and NHWC) and various chunk sizes. In addition, we evaluate end-to-end inference latency of MobileNet-V2 using our method. The results show our method achieves about 2x speedup against the conventional convolution lowering method.