基于 Zynq 的卷积神经网络加速器设计

针对卷积神经网络(convolutional neural network, CNN)嵌入式部署资源开销大、运行速度慢等问题，文章提出一种以Tiny-YOLOv3作为算法模型的CNN硬件加速器。首先，基于Tiny-YOLOv3网络各层的特性和要求设计CNN加速器实现方案，将权重系数按位分割，面向单bit权重设计卷积加速器，通过逐位实施达到处理速度和识别率的高效平衡；然后，采用查表选择法实现卷积算子的乘加运算，设计一款 $ 6\times3\times16 $的三维加速器计算阵列，可单周期完成288个卷积窗口计算；最后，在Xilinx Zynq UltraScale+MPSoC系列芯片上对设计的CNN加速器进行性能测试。实验结果表明，该CNN加速器在200 MHz频率下具有518.4 GOPS的算力，比现有的解决方案性能提高了约63%。

In view of the problems of high cost and slow running of embedded deployment of convolutional neural network(CNN), a CNN hardware accelerator using Tiny-YOLOv3 as algorithm model is presented. According to the characteristics and requirements of each layer of Tiny-YOLOv3 network, a CNN accelerator implementation scheme is proposed, which divides the weight coefficients bit by bit, designs the convolution accelerator for single bit weight, and achieves a high-efficiency compromise between processing speed and recognition efficiency through bit-by-bit implementation. The multiplication and addition of convolution operator is implemented by a table-looking selection method. A $ 6 \times 3 \times 16 $ 3D accelerator calculation array is designed, and 288 convolution windows can be calculated in a single cycle. Finally, the performance of CNN accelerator is tested on Xilinx Zynq UltraScale+MPSoC series chips. The results show that the designed CNN accelerator has a computational power of 518.4 GOPS at 200 MHz, achieving a performance improvement of about 63% compared to existing solutions.