基于动态自适应计算引擎的MobileNetV3网络加速器设计  

Design of MobileNetV3 network accelerator based on dynamic adaptive computing engine

作　　者：项浩斌 杨瑞敏[1] 吴文涛 李春雷[1] 董燕[1,2] Xiang Haobin;Yang Ruimin;Wu Wentao;Li Chunlei;Dong Yan(School of Information and Communication Engineering,Zhongyuan University of Technology,Zhengzhou 450007,China;School of Automation Engineering,University of Electronic Science and Technology of China,Chengdu 610000,China)

机构地区：[1]中原工学院信息与通信工程学院,河南郑州450007 [2]电子科技大学自动化工程学院,四川成都610000

出　　处：《电子技术应用》2025年第1期8-17,共10页Application of Electronic Technique

摘　　要：现有面向高效轻量化MobileNetV3网络的加速方法通常采用高度定制的计算引擎进行模型计算,从而限制了加速器的可扩展性使其仅适用于小型网络或资源丰富的硬件平台。针对此问题,提出了基于动态自适应计算引擎的MobileNetV3网络加速器。首先,设计了局部感知区域卷积的流水线推理架构实现特征、权重的高度并行处理和缓冲调度。其次,提出全局自适应的点卷积方法优化点卷积,并结合空间探索获得最优的参数配置以实现最大计算并行性。此外,加速器可以根据模型参数变化动态配置以适应不同场景。实验结果显示加速器推理速度为8 F/s,是现有方法速度的2.7倍。Existing acceleration methods for efficient and lightweight MobileNetV3 networks usually use highly customized com‐puting engines for model calculations,which limits the scalability of the accelerator and makes it only applicable to small net‐works or resource-rich hardware platforms.To address this problem,this paper proposes a MobileNetV3 network accelerator based on a dynamic adaptive computing engine.Firstly,a pipeline inference architecture of local perception area convolution is designed to achieve highly parallel processing and buffer scheduling of features and weights.Secondly,a global adaptive point convolution method is proposed to optimize point convolution and combine spatial exploration to obtain the optimal parameter configuration to achieve maximum computational parallelism.In addition,the accelerator can be dynamically configured accord‐ing to model parameter changes to adapt to different scenarios.Experimental results show that the accelerator's inference speed is 8 F/s,which is 2.7 times as fast as existing methods.

关 键 词：卷积神经网络 并行计算 动态自适应 边缘设备 硬件加速 

分 类 号：TN791[电子电信—电路与系统] TP183[自动化与计算机技术—控制理论与控制工程]