一种2D权值固定数据流架构的研究

Research on a 2D Weight Stationary Dataflow Architecture

随着人工智能算法的发展,卷积神经网络(CNN)在图像、音频等方面的应用越来越广泛,CNN算法的计算量也越来越大.权值固定数据流(WS)将权值固定在寄存器中,是一种最大化利用卷积重用和filter重用的数据流.不过当前的权值固定数据流结构存在建立流水线时间过长的问题.本文研究了一种去除PE(Process Element)行之间的FIFO,用加法器连接PE行的2D权值固定数据流结构.这种2D权值固定的数据流结构计算AlexNet时减少了近2.7倍建立流水线时间,并且能够灵活地调整卷积步长.

With the development of artificial intelligence algorithms,convolutional neural networks(CNN)is more and more widely used in image,audio and other aspects,and the amount of calculation of CNN algorithms is also increasing.Weight stationary(WS)is a dataflow that maximizes the use of convolutional reuse and filter reuse by fixing weights in registers.However,the current WS dataflow structure has the problem that the pipeline filling time is lager.This paper studies a 2D-WS dataflow structure that uses PE adders to remove the FIFO between PE lines.This kind of 2D-WS dataflow structure reduces the pipeline filling time by nearly 2.7 times when calculating AlexNet,and can flexibly adjust the stride size.