A Convolutional Computing Design Using Pulsating Arrays
Youyao Liu, Qifei Shi, Haihao Wang, Xin Liu · 2023
The training and reasoning of convolutional neural networks (CNN) require a large amount of computation, among which the most core operation is convolutional operation. In order to accelerate convolutional operation, many methods for accelerating convolution have emerged in recent years, including sliding window method, Winograd algorithm, etc. Although the sliding window method and Winograd algorithm can improve the computational efficiency of convolution to a certain extent, they also have some drawbacks. The sliding window method requires a large amount of data reuse and transmission, thus requiring high bandwidth and low latency memory and data paths, which is also the main challenge of its hardware implementation; The Winograd algorithm requires preprocessing of convolutional kernels and input data, which increases the computational and storage burden in the early stage. At the same time, for convolutional kernels of different sizes, matrix transformations need to be regenerated, which increases the difficulty of algorithm implementation. Based on this, this article adopts a pulsating array convolution calculation method based on fixed weights.