LAUSR

Abstract Compared with fast convolution methods such as im2col and the fast Fourier transform, Winograd-based convolution, which has been widely applied to accelerate convolutional neural networks (CNNs), can provide high performance with smaller filters. Although there are several reported studies on the algorithmic optimization of CNNs, most of them are targeted at hardware architectures. The existing implementations of the Winograd method perform well below what one would expect, due to the fact that the tile size of Winograd-based convolution is usually empirical and the features of each convolution layer are ignored. This study aims to fill this gap and focuses on the efficient implementation of Winograd-based convolution in the CNN model. Specifically, we discuss the causes of poor performance, calculate the coefficient of computation complexity model and demonstrate a speedup in the inference process using an elaborate tile-fusion method, which derives the optimal tile size for each convolution layer in a CNN model. Compared with the representative existing implementations of CuDNN with a 4 × 4 tile, Arm Compute Library with a 6 × 6 tile, and NNPACK with an 8 × 8 tile, the results show significant performance improvements on of up to 1.89 × , 1.29 × and 1.17 × , respectively.