面向边缘智能计算的异构并行计算平台综述

doi:10.3778/j.issn.1002-8331.2204-0187

摘要/Abstract

摘要： 边缘智能计算对硬件资源的需求复杂多元，传统计算平台难以为继，异构并行计算平台成为边缘智能算法落地的关键途径之一。以深度学习算法和边缘计算为牵引，对异构并行计算平台展开研究。一方面，阐述了传统计算平台适配实现边缘智能计算的优缺点，指出边缘端应用场景中传统计算平台算力与功耗矛盾突出等局限性，并以指令模型、通讯机制和存储体系三个关键技术为线索梳理技术发展脉络。另一方面，从运算速度、功耗等角度重点对比分析了近年来典型异构平台较新的代表性产品，然后针对不同应用场景和约束条件给出了异构平台的选择建议：优先选择CPU+X组合的异构平台。功耗要求严格约束下的应用建议优先选择CPU+FPGA组合；功能迭代更新快的场景建议优先选择CPU+GPU组合；算法成熟且对实时性和功耗均具有高要求的应用优先选择ASIC计算平台。提出了异构并行计算平台在指令模型统一、通讯机制轻量化、存储体系灵活性以及开发生态完备化四个方面的问题与挑战，期望能为该领域研究人员带来一定的启发。

关键词: 异构并行架构, 边缘计算, 智能计算, 深度学习, 嵌入式设备, 硬件加速

Abstract: Edge intelligent computing necessitates a high demand for hardware resources, which traditional computing platforms are hard to meet. As a result, heterogeneous parallel computing platforms have emerged as one of the most important means of implementing edge intelligent algorithms. This paper investigates heterogeneous parallel computing platforms with a focus on deep learning and edge computing. On the one hand, the advantages and disadvantages of the traditional computing platforms in edge intelligent computing are elaborated, and the limitations of traditional computing platforms in edge application scenarios, such as the prominent contradiction between computing power and power consumption, are pointed out. And the development of technology is combed based on the instruction model, communication mechanism, and storage system, respectively. On the other hand, the typical products in recent years of heterogeneous platforms are compared in terms of computational power, power consumption, and other factors. Then the selection suggestions for different conditions and constraints are given. It is advised to select the heterogeneous platform of the CPU+X combination. For applications with strict power consumption requirements, it is recommended to choose the heterogeneous platform of the CPU+FPGA combination. For applications with fast iteration and updates, it is recommended to choose the heterogeneous platform of the CPU+GPU combination. The ASIC computing platform is suitable for applications with mature algorithms and high requirements for real-time performance and power consumption. Finally, the challenges of instruction model unification, communication mechanism lightweight, storage system flexibility, and development environment completion are discussed, with the expectation of bringing some inspiration for researchers in this field.

Key words: heterogeneous parallel platform, edge computing, intelligent computing, deep learning, embedded device, hardware acceleration

万朵, 胡谋法, 肖山竹, 张焱. 面向边缘智能计算的异构并行计算平台综述[J]. 计算机工程与应用, 2023, 59(1): 15-25.

WAN Duo, HU Moufa, XIAO Shanzhu, ZHANG Yan. Survey on Heterogeneous Parallel Computing Platform for Edge Intelligent Computing[J]. Computer Engineering and Applications, 2023, 59(1): 15-25.

参考文献

[1] ZAHRAN M.Heterogeneous computing：hardware and software perspectives[M].New York：Association for Computing Machinery，2019：15-18.
[2] MOLCHANOVP，TYREE S，KARRAS T，et al.Pruning convolutional neural networks for resource inference[J].arXiv：1611.06440，2016.
[3] IANDOLA F N，HAN S，MOSKEWICZ M W，et al.SqueezeNet：AlexNet-level accuracy with 50x fewer parameters and <0.5?MB model size[J].arXiv：1602.07360，2016.
[4] LIU Z，LI J G，SHEN Z Q，et al.Learning efficient convolutional networks through network slimming[C]//Proceedings of the IEEE International Conference on Computer Vision，2017：2755-2763.
[5] 李欣瑶，刘飞阳，李鹏.嵌入式智能计算加速技术综述[C]//中国航空科学技术大会论文集，2019：996-1004.
LI X Y，LIU F Y，LI P.Survey of embedded intelligent computing acceleration technology[C]//Proceedings of the China Aeronautical Science and Technology Conference，2019：996-1004.
[6] 陈桂林，马胜，郭阳.硬件加速神经网络综述[J].计算机研究与发展，2019，56（2）：240-253.
CHEN G L，MA S，GUO Y.Survey on accelerating neural network with hardware[J].Journal of Computer Research and Development，2019，56（2）：240-253.
[7] 张政馗，庞为光，谢文静，等.面向实时应用的深度学习研究综述[J].软件学报，2020，31（9）：2654?2677.
ZHANG Z D，PANG W G，XIE W J，et al.Deep learning for real-time applications：A survey[J].Journal of Software，2020，31（9）：2654-2677.
[8] 卢冶，陈瑶，李涛，等.面向边缘计算的嵌入式FPGA卷积神经网络构建方法[J].计算机研究与发展，2018，55（3）：551-562.
LU Y，CHEN Y，LI T，et al.Convolutional neural network construction method for embedded FPGAs oriented edge computing[J].Journal of Computer Research and Development，2018，55（3）：551-562.
[9] 张孟逸.基于FPGA的卷积神经网络并行加速器设计[D].哈尔滨：哈尔滨理工大学，2019.
ZHANG M Y.Parallel accelerator design for convolutional neutral networks based on FPGA[D].Harbin：Harbin University of Science and Technology，2019.
[10] 中国信息通信研究院.先进计算发展研究报告[EB/OL].（2018-12-19）[2022-03-22].http：//www.caict.ac.cn/kxyj/qwfb/bps/201812/P020181218519867225833.pdf.
CAICT.Research report on the development of advanced computing[EB/OL].（2018-12-19）[202-03-22].http：//www.caict.ac.cn/kxyj/qwfb/bps/201812/P020181218519867225833.pdf.
[11] VENKAT A，TULLSEN D M.Harnessing ISA diversity：Design of a heterogeneous-ISA chip multiprocessor[C]//proceedings of the ACM/IEEE 41st International Symposium on Computer Architecture，2014：121-132.
[12] 邢景.基于ARM NEON目标检测网络算法的加速技术研究[D].哈尔滨：哈尔滨工业大学，2020.
XING J.Research on object detection network algorithm accelerating technology based on ARM NEON[D].Harbin：Harbin Institute of Technology，2020.
[13] 刘湘，魏鑫，陈禾，等.应用于ARM-NEON的图像二维卷积高效实现方法[C]//第十二届全国信号和智能信息处理与应用学术会议论文集，2018：1-6.
LIU X，WEI X，CHEN H，et al.An efficient implementation for image two-dimensional convolution on ARM-NEON[C]//Proceedings of the 12nd National Conference on Signal and Intelligent Information Processing and Applications，2018：1-6.
[14] CADENCE.Enabling embedded vision neural network DSPs[EB/OL].（2021-06-12）[2022-03-22].https：//www.cadence.com/content/dam/cadence-www/global/en_US/documents/tools/ip/tensilica-ip/tip-vision-c5.pdf.
[15] YANG C，CHEN S M，ZHANG J，et al.A novel DSP architecture for scientific computing and deep learning[J].IEEE Access，2019，7：36413-36425.
[16] LINDHOLM E，NICKOLLS J，OBERMAN S，et al.NVIDIA tesla：A unified graphics and computing architecture[J].IEEE Micro，2008，28（2）：39-55.
[17] 李宏亮，郑方，郝子宇，等.面向智能计算的国产众核处理器架构研究[J].中国科学：信息科学，2019，49（3）：247-255.
LI H L，ZHENG F，HAO Z Y，et al.Research on homegrown manycore architecture for intelligent computing[J].SCIENTIA SINICA Informationis，2019，49（3）：247-255.
[18] KECKLER S W，DALLY W J，KHAILANY B，et al.GPUs and the future of parallel computing[J].IEEE Micro，2011，31（5）：7-17.
[19] LIU S L，DU Z D，TAO J H，et al.Cambricon：An instruction set architecture for neural networks[C]//Proceedings of the ACM/IEEE 43rd Annual International Symposium on Computer Architecture，2016：393-405.
[20] 韩栋，周聖元，支天，等.智能芯片的评述和展望[J].计算机研究与发展，2019，56（1）：7-22.
HAN D，ZHOU S Y，ZHI T，et al.A survey of artificial intelligence chip[J].Journal of Computer Research and Development，2019，56（1）：7-22.
[21] LI W C，YANG C J，FANG W C.A real-time emotion recognition system based on an AI system-on-chip design[C]//Proceedings of the International SoC Design Conference（ISOCC），2020：29-30.
[22] WANG X Y，MAGNO M，CAVIGELLI L，et al.FANN-on-MCU：An open-source toolkit for energy-efficient neural network Inference at the edge of the Internet of Things[J].IEEE Internet of Things Journal，2020，7（5）：4403-4417.
[23] DESOLI G，CHAWLA N，BOESCH T，et al.A 2.9TOPS/W deep convolutional neural network SoC in FD-SOI 28 nm for intelligent embedded systems[C]//Proceedings of the Solid-State Circuits Conference，2017：238-239.
[24] FOLEY D，DANSKIN J.Ultra-performance pascal GPU and NVLink interconnect[J].IEEE Micro，2017，37（2）：7-17.
[25] 卢丽强，郑思泽，肖倾城，等.面向卷积神经网络的FPGA设计[J].中国科学：信息科学，2019，49（3）：277-294.
LU L Q，ZHENG S Z，XIAO Q C，et al.Accelerating convolutional neural networks on FPGAs[J].SCIENTIA SINICA Informationis，2019，49（3）：277-294.
[26] 王明钊，程华，王宇泽，等.基于精度可变乘法器的脉动阵列[J].南京大学学报（自然科学），2020，56（6）：885-891.
WANG M Z，CHENG H，WANG Y Z，et al.Systolic array based on precision variable multiplier[J].Journal of Nanjing university（Natural Science），2020，56（6）：885-891.
[27] PARMAR Y，SRIDHARAN K.A high-performance VLSI architecture for a self-feedback convolutional neural network[J].IEEE Transactions on Circuits and Systems II：Express Briefs，2021，68（1）：456-460.
[28] MOOLCHANDANI D，KUMAR A，SARANGI S R.Accelerating CNN inference on ASICs：A survey[J].Journal of Systems Architecture，2021，113：101887.
[29] CHEN Y H，YANG T J，EMER J，et al.Eyeriss V2：A flexible accelerator for emerging deep neural networks on mobile devices[J].IEEE Journal on Emerging and Selected Topics in Circuits and Systems，2019，9（2）：292-308.
[30] 任小广.面向多级SPM存储的并行程序优化——任务与数据的关联调度[D].长沙：国防科学技术大学，2010.
REN X G.Multi-level SPM based parallel programme optimization—associated scheduling of task and data[D].Changsha：National University of Defense Technology，2010.
[31] WULF W A，MCKEE S A.Hitting the memory wall[J].ACM SIGARCH Computer Architecture News，1995，23（1）：20-24.
[32] CHEN T Sh，DU Z D，SUN N H，et al.DianNao：A small-footprint high-throughput accelerator for ubiquitous machine-learning[J].Architecture Support for Programming Languages and Operating Systems，2014，49（4）：269-284.
[33] GEE J D，SMITH A J.The performance impact of vector processor cashes[C]//Proceedings of the Twenty-Fifth Hawaii International Conference on System Sciences，1992：437-448.
[34] RYOO S，RODRIGUES C I，STONE S S，et al.Program optimization space pruning for a multithreaded gpu[C]//Proceedings of the 6th Annual IEEE/ACM International Symposium on Code Generation and Optimization，2008：195-204.
[35] 汪黎.面向软件管理片上存储器的编译优化技术研究[D].长沙：国防科学技术大学，2009.
WANG L.Compiler optimization for software-managed on-chip memory[D].Changsha：National University of Defense Technology，2009.
[36] 张骏.便笺存储嵌入式系统中多层存储上的数据分配算法研究[D].长沙：湖南大学，2013.
ZHANG J.The research on data placement for multiple types of memory units in embedded system with scratch-pad memories[D].Changsha：Hunan University，2011.
[37] 刘文志，陈轶，吴长江.OpenCL异构并行计算：原理、机制与优化实践[M].北京：机械工业出版社，2015：1-3.
LIU W Z，CHEN Y，WU C J.OpenCL heterogeneous parallel computing：from principle to practice[M].Beijing：China Machine Press，2015：1-3.
[38] 周楠，胡娟，胡海明.多核处理器发展趋势及关键技术[J].计算机工程与设计，2018，39（2）：393-399.
ZHOU N，HU J，HU H M.Development trend and key techniques of multi?core processor[J].Computer Engineering and Design，2018，39（2）：393-399.
[39] 徐洋.图像去雾算法研究及其异构并行实现[D].哈尔滨：哈尔滨工业大学，2018.
XU Y.Research on image dehazing algorithm and heterogeneous parallel implementation[D].Harbin：Harbin Institute of Technology，2018.
[40] 陈书明，李振涛，万江华，等.“银河飞腾”高性能数字信号处理器研究进展[J].计算机研究与发展，2006，43（6）：993-1000.
CHEN S M，LI Z T，WAN J H，et al.Research and development of high performance YHFT digital signal processor[J].Journal of Computer Research and Development，2006，43（6）：993-1000.
[41] 杜承垚，袁景凌，陈旻骋，等.GPU加速与L-ORB特征提取的全景视频实时拼接[J].计算机研究与发展，2017，54（6）：1316-1325.
DU C Y，YUAN J L，CHEN M C，et al.Real-time panoramic video stitching based on GPU acceleration using local ORB feature extraction[J].Journal of Computer Research and Development，2017，54（6）：1316-1325.
[42] BAO C，XIE T，FENG W B，et al.A power-efficient optimizing framework FPGA accelerator based on Winograd for YOLO[J].IEEE Access，2020，8：94307-94317.
[43] DU Z D，FASTHUBER R，CHEN T S，et al.ShiDianNao：Shifting vision processing closer to the sensor[C]//Proceedings of the ACM/IEEE 42nd Annual International Symposium on Computer Architecture，2015：92-104.
[44] FANG S X，TIAN L，WANG J B，et al.Real-time object detection and semantic segmentation hardware system with deep learning networks[C]//Proceedings of the International Conference on Field-Programmable Technology（FPT），2018：389-392.
[45] 崔洲涓，安军社，陈长龙，等.基于PYNQ框架的深度卷积特征异构跟踪系统[J].计算机工程与应用，2021，57（4）：120-126.
CUI Z J，AN J H，CHEN C L，et al.Deep convolutional features heterogeneous tracking system based on PYNQ framework[J].Computer Engineering and Applications，2021，57（4）：120-126.
[46] 吴艳霞，梁楷，刘颖，等.深度学习FPGA加速器的进展与趋势[J].计算机学报，2019，42（11）：2461-2480.
WU Y X，LIANG K，LIU Y，et al.The progress and trends of FPGA-based accelerators in deep learning[J].Chinese Journal of Computers，2019，42（11）：2461-2480.
[47] WEI XUN D T，LIM Y L，SRIGRAROM S.Drone detection using YOLOv3 with transfer learning on NVIDIA Jetson TX2[C]//Proceedings of the 2nd International Symposium on Instrumentation，Control，Artificial Intelligence，and Robotics（ICA-SYMP），2021：1-6.
[48] MITTAL S.A Survey on optimized implementation of deep learning models on the NVIDIA Jetson platform[J].Journal of Systems Architecture，2019，97：428-442.
[49] OTHMAN N A，AYDIN I.A new deep learning application based on movidius NCS for embedded object detection and recognition[C]//Proceedings of the 2nd International Symposium on Multidisciplinary Studies and Innovative Technologies（ISMSIT），2018：1-5.
[50] 王云富.基于深度学习的交通标志检测研究与应用[D].大连：大连理工大学，2021.
WANG Y F.Research and application of traffic sign detection based on deep learning[D].Dalian：Dalian University of Technology，2021.
[51] 杨文静.基于嵌入式的水下鱼类识别技术研究[D].上海：上海海洋大学，2021.
YANG W J.Research on underwater fish recognition technology based on embedded[D].Shanghai：Shanghai Ocean University，2021.