一种CNN网络卷积并行计算方法

摘要

卷积神经网络(CNN)在图像分类、自然语言处理中广泛应用.区别于其他神经网络模型的最主要特点是卷积神经网络中使用的卷积操作.提出一种将卷积运算拆分为向量运算的方法,使得卷积网络部署在嵌入式平台时可以利用DSP、 ARM NEON等硬件加速器进行加速,其简单的取址方式增加了cache命中率,大大提升了卷积运算的效率.此外,该方法也可以作为ASIC设计专用的卷积神经网络加速器.%Convolutional neural network (CNN) offers many innovative applications in image classification and nature language processing.CNN using convolution operation is distinct from other NN.A new method transforming convolution calculation into vector operation is proposed.The proposed method optimizes the convolution efficiency by some common hardware accelerator and improves cache hit rate.A hardware accelerator for CNN is also designed by the method.