计算机科学

计算机科学 ›› 2026, Vol. 53 ›› Issue (6): 193-202.doi: 10.11896/jsjkx.251000093

• 高性能计算 • 上一篇    下一篇

面向寒武纪MLU加速卡的高性能图像预处理算子研究

李菲1, 刘松1, 郭松坚1, 刘嘉政1, 张颖1, 洪龙伟2, 张博宣2   

  1. 1 西安交通大学计算机科学与技术学院 西安 710049
    2 西安交通大学软件学院 西安 710049
  • 收稿日期:2025-10-23 修回日期:2025-12-22 出版日期:2026-06-15 发布日期:2026-06-09
  • 通讯作者: 刘松(liusong@xjtu.edu.cn)
  • 作者简介:(lifei@stu.xjtu.edu.cn)
  • 基金资助:
    国家重点研发计划 (2022YFB4501604)

High-performance Image Preprocessing Operators for Cambricon MLU Accelerator Card

LI Fei1, LIU Song1, GUO Songjian1, LIU Jiazheng1, ZHANG Ying1, HONG Longwei2, ZHANG Boxuan2   

  1. 1 School of Computer Science and Technology,Xi'an Jiaotong University,Xi'an 710049,China
    2 School of Software Engineering,Xi'an Jiaotong University,Xi'an 710049,China
  • Received:2025-10-23 Revised:2025-12-22 Published:2026-06-15 Online:2026-06-09
  • About author:LI Fei,born in 1999,Ph.D candidate,is a member of CCF(No.L1866G).His main research interests include high performance computing and parallel algorithm.
    LIU Song,born in 1987,Ph.D,associate professor,Ph.D supervisor,is a member of CCF(No.A0055M).His main research interests include parallel computing and code optimization.
  • Supported by:
    National Key R & D Program of China(2022YFB4501604).

PDF (PC)

65

摘要： 图像预处理是机器学习任务中的关键环节,其计算效率直接影响着模型训练和推理性能。传统CPU计算难以满足高分辨率图像和大规模数据集的实时性处理需求,而神经网络处理单元(NPU)凭借高性能成为加速图像预处理的理想平台。然而,图像预处理算子的多样性和访存密集特性与NPU的矩阵运算优化模式不完全匹配,给NPU的适配带来极大的挑战。为此,基于飞腾CPU和寒武纪MLU异构计算平台,提出了高效的图像预处理算子加速方法。通过深入分析MLU的多核并行架构和存储体系,重新设计了算子的计算逻辑和任务划分,结合多核并行调度、向量化计算、三级存储结构以及双缓冲机制等优化策略,显著提升了算子的计算性能,展示了NPU在图像预处理加速中的潜力。实验结果表明,优化后的10种常见的图像预处理算子相较于PaddlePaddle原生算子性能提升了33.22%~234.46%。在深度学习和传统的机器学习任务中,所提算法分别实现了12.47%与48.63%的端到端性能提升。

关键词: 图像预处理, 寒武纪MLU, 高性能计算, 并行计算, 性能优化

Abstract: Image preprocessing is a critical component in machine learning tasks,and its computational efficiency directly impacts the performance of model training and inference.Traditional CPU computations struggle to meet the real-time processing demands of high-resolution images and large-scale datasets,whereas Neural Processing Units(NPUs),with their high performance,emerge as ideal platforms for accelerating image preprocessing.However,the diversity and memory access intensiveness of image preprocessing operations do not fully align with the matrix operation optimization mode of NPUs,posing significant challenges for their adaptation.Based on the Phytium CPU and Cambricon MLU heterogeneous computing platform,this paper proposes an efficient acceleration method for image preprocessing operators.By thoroughly analyzing the multi-core parallel architecture and sto-rage system of the MLU,the computational logic and task partitioning of the operators are redesigned.Combined with optimization strategies such as multi-core parallel scheduling,vectorized computation,three-level storage structure,and double buffering mechanism,the computational performance of the operators is significantly enhanced,demonstrating the potential of NPUs in accelerating image preprocessing.Experimental results show that the optimized ten common image preprocessing operators achieve a performance improvement of 33.22% to 234.46% compared to the native operators in PaddlePaddle.In deep learning and traditional machine learning tasks,end-to-end performance improvements of 12.47% and 48.63% are achieved,respectively.

Key words: Image preprocessing, Cambrian MLU, High performance computing, Parallel computing, Performance optimization

中图分类号: 

  • TP338.6
[1]SHORTEN C,KHOSHGOFTAAR T M.A survey on image data augmentation for deep learning[J].Journal of Big Data,2019,6(1):1-48.
[2]MINH T N,SINN M,LAM H T,et al.Automated image data preprocessing with deep reinforcement learning[J].arXiv:1806.05886,2018.
[3]GYAWALI D.Comparative analysis ofcpu and gpu profiling for deep learning models[J].arXiv:2309.02521,2023.
[4]XIAO H,SUN L P,LI C L,et al.Histogram statistical image enhancement parallel algorithm for GPU[J].Journal of Frontiers of Computer Science & Technology,2022,16(10):2273-2285.
[5]LIU B,ZHOU H,BIAN C J,et al.Target detection systembased on lightweight Yolov5 algorithm based on aerospace-grade NPU[J].Chinese Journal of Space Science,2025,45(4):1-11.
[6]WENG X,IVANOVIC B,WANG Y,et al.Para-drive:Paralle-lized architecture for real-time autonomous driving[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2024:15449-15458.
[7]DE SILVA U,FERNANDO L,LIK B L P,et al.Large language models for video surveillance applications[C]//2024 IEEE Region 10 Conference(TENCON).IEEE,2024:563-566.
[8]CHEN T,DU Z,SUN N,et al.Diannao:A small-footprint high-throughput accelerator for ubiquitous machine-learning[J].ACM SIGARCH Computer Architecture News,2014,42(1):269-284.
[9]LU W Z,ZHANG F,HE Y X,et al.Performance evaluation and optimization of Huawei Ascend neural network accelerator[J].Chinese Journal of Computers,2022,45(8):1618-1637.
[10]ZHANG S,DU Z,ZHANG L,et al.Cambricon-X:An accelerator for sparse neural networks[C]//2016 49th Annual IEEE/ACM International Symposium on Microarchitecture(MICRO).IEEE,2016:1-12.
[11]JIAO Y,HAN L,LONG X.Hanguang 800 NPU-the ultimate AI inference solution for data centers[C]//2020 IEEE Hot Chips 32 Symposium(HCS).IEEE Computer Society,2020:1-29.
[12]LI S.China's largest AI computing chip is launched:A look at Suiyuan Technology's Suisi chip and Yunsui accelerator card [J].Microcomputer,2021(24):89-93.
[13]LEE K J.Architecture of neural processing unit for deep neural networks[M]//Advances in Computers.Elsevier,2021:217-245.
[14]CHEN J,BAI G,LIANG S,et al.Automatic image cropping:A computational complexity study[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.2016:507-515.
[15]ZHANG Y,GUO R Q,SHENG Y Y.CV-CUDA High-Perfor-mance Image Processing Acceleration Library [EB/OL].(2022-11-22)[2025-06-03].https://developer.nvidia.com/zh-cn/blog/cv-cuda-high-performance-image-processing.
[16]MA Y J,YU D H,WU T,et al.PaddlePaddle:An open source deep learning platform derived from industrial practice [J].Frontiers of Data and Computing Development,2019,1(5):105-115.
[17]LIAO H,TU J,XIA J,et al.Ascend:a scalable and unified architecture for ubiquitous deep neural network computing:Industry track paper[C]//2021 IEEE International Symposium on High-Performance Computer Architecture(HPCA).IEEE,2021:789-801.
[18]LIU S,DU Z,TAO J,et al.Cambricon:An instruction set architecture for neural networks[J].ACM SIGARCH Computer Architecture News,2016,44(3):393-405.
[19]GUO H,ZHAO Y,LI Z,et al.Cambricon-u:A systolic random increment memory architecture for unary computing[C]//Proceedings of the 56th Annual IEEE/ACM International Sympo-sium on Microarchitecture.2023:424-437.
[20]HAO Y,ZHAO Y,LIU C,et al.Cambricon-p:A bitflow architecture for arbitrary precision computing[C]//2022 55th IEEE/ACM International Symposium on Microarchitecture(MICRO).IEEE,2022:57-72.
[21]SONG X,WEN Y,HU X,et al.Cambricon-r:A fully fused accelerator for real-time learning of neural scene representation[C]//Proceedings of the 56th Annual IEEE/ACM International Symposium on Microarchitecture.2023:1305-1318.
[22]ZHAO Y,LIU C,DU Z,et al.Cambricon-Q:A hybrid architecture for efficient training[C]//2021 ACM/IEEE 48th Annual International Symposium on Computer Architecture(ISCA).IEEE,2021:706-719.
[23]JOUPPI N P,YOUNG C,PATIL N,et al.In-datacenter performance analysis of a tensor processing unit[C]//Proceedings of the 44th Annual International Symposium on Computer Architecture.2017:1-12.
[24]MARKIDIS S,DER CHIEN S W,LAURE E,et al.Nvidia tensor core programmability,performance & precision[C]//2018 IEEE International Parallel and Distributed Processing Sympo-sium Workshops(IPDPSW).IEEE,2018:522-531.
[25]HICKMANN B,CHEN J,ROTZIN M,et al.Intelnervana neural network processor-t(nnp-t)fused floating point many-term dot product[C]//2020 IEEE 27th Symposium on Computer Arithmetic(ARITH).IEEE,2020:133-136.
[26]BBI R,XU T,XU M,et al.Paddlepaddle:A production-oriented deep learning platform facilitating the competency of enterprises[C]//2022 IEEE 24th Int Conf on High Performance Computing & Communications;8th Int Conf on Data Science & Systems;20th Int Conf on Smart City;8th Int Conf on Dependability in Sensor,Cloud & Big Data Systems & Application(HPCC/DSS/SmartCity/DependSys).IEEE,2022:92-99.
[27]VASILE C E,ULMĂMEI A A,BÎRĂ C.Image ProcessingHardware Acceleration－A Review of Operations Involved and Current Hardware Approaches[J].Journal of Imaging,2024,10(12):298.
[28]YANG H Y,LI C M,WANG X P,et al.Image collaborativeparallel processing model in CPU/GPU heterogeneous environment[J].Integration Technology,2017,6(5):8-18.
[29]NAZ N,HASEEB MALIK A,KHURSHID A B,et al.Efficient processing of image processing applications on CPU/GPU[J].Mathematical Problems in Engineering,2020,2020(1):4839876.
[30]ALHUMAIDAN B,ALGHOFAILY S,AL QHAHTANI M,et al.Parallel image processing:Taking grayscale conversion using openmp as an example[J].Journal of Computer and Communications,2024,12(2):1-10.
[31]XIAO S Y,WANG L,DU Y,et al.OpenCL acceleration algorithm for image median filtering based on heterogeneousplatforms[J].Journal of Hebei University(Natural Science Edition),2024,44(1):92.
[32]MÁNDI Á,MÁTÉ J,RÓZSA D,et al.Hardware acceleratedimage processing on FPGA based PYNQ-Z2 board[J].CarpathianJournal of Electronic and Computer Engineering,2021,14(1):20-23.
[33]YUAN H,DING D,FAN Z,et al.A real-time image processing hardware acceleration method based onfpga[C]//2021 6th International Conference on Computational Intelligence and Applications(ICCIA).IEEE,2021:200-205.
[34]CHEN W,ZHANG C S,LIU S.An image processing acceleration method based on domestic accelerator card:CN 202410455829.4 [P].2024-07-05.
[35]LI Y.The Investigation of DeiT model Based on PaddlePaddleFramework on CIFAR-10 Dataset Image Classification[C]//2023 International Conference on Image,Algorithms and Artificial Intelligence(ICIAAI 2023).Atlantis Press,2023:1062-1067.
[36]CAMBRICON TECHNOLOGIES.CAMBRICON BANG C/C++Programming Guide [EB/OL].(2023-09-12)[2025-06-03].https://www.cambricon.com/docs/sdk_1.15.0/cntoolkit_3.7.2/programming_guide_1.7.0/hardware_implementation/index.html.
[37]ABOUELNAGA Y,ALI O S,RADY H,et al.Cifar-10:Knn-based ensemble of classifiers[C]//2016 International Confe-rence on Computational Science and Computational Intelligence(CSCI).IEEE,2016:1192-1195.
[38]MOKHAIRI M,ENGKU FADZLI HASAN S A,NURSHAZWANI K.Comparison of image classification techniques using CALTECH 101 dataset[J].Journal of Theoretical and Applied Information Technology,2015,71(1):79-86.
[39]BAIDU PADDLEPADDLE.PaddlePaddle Deep Learning Plat-form User Guide [EB/OL].[2025-7-18].https://www.paddlepaddle.org.cn/documentation/docs/zh/3.0-beta/guides/hardware_support/mlu/support_cn.html.
[40]HE K,ZHANG X,REN S,et al.Deep residual learning for image recognition[C]//Proceedings of theIEEE Conference on Computer Vision and Pattern Recognition.2016:770-778.
[41]RIGATTI S J.Random forest[J].Journal of Insurance Medi-cine,2017,47(1):31-39.
[42]STEINBACH M,TAN P N.kNN:k-nearest neighbors[M]//The Top Ten Algorithms in Data Mining.Chapman and Hall/CRC,2009:165-176.
[43]DE VILLE B.Decision trees[J].Wiley Interdisciplinary Re-views:Computational Statistics,2013,5(6):448-455.
Viewed
Full text


Abstract

Cited
  Shared   
  Discussed   
No Suggested Reading articles found!
渝ICP备16013121号-4
地址：重庆市两江新区洪湖西路18号    邮编：401121  
电话：023-63500828（编辑部） 023-67039612（会议/专题） 023-67039623（财务/发票）
E-mail：jsjkx12@163.com
本系统由北京玛格泰克科技发展有限公司设计开发