Operator-level Reconstruction Method of YOLOv11 for Ascend NPU

doi:10.19678/j.issn.1000-3428.0260413

Computer Engineering ›› 2026, Vol. 52 ›› Issue (9): 365-375. doi: 10.19678/j.issn.1000-3428.0260413

• High-Performance Computing and Big Data • Previous Articles Next Articles

Operator-level Reconstruction Method of YOLOv11 for Ascend NPU

XIE Jingyuan¹, ZHA Kaiwen², LIU Pengju³, TIAN Chunwei⁴^,*()

1. School of Software, Northwestern Polytechnical University, Xi'an 710072, Shaanxi, China
2. Shenzhen University WeBank Institute of FinTech, Shenzhen 518060, Guangdong, China
3. Young Scientists Innovation Center, Zhengzhou Advanced Research Institute, Harbin Institute of Technology, Zhengzhou 450000, Henan, China
4. School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001, Heilongjiang, China

Received:2026-03-31 Revised:2026-05-27 Online:2026-09-15 Published:2026-07-22
Contact: TIAN Chunwei

面向昇腾NPU的YOLOv11算子级重构方法

谢静远¹, 查凯文², 刘鹏举³, 田春伟⁴^,*()

1. 西北工业大学软件学院, 陕西西安 710072
2. 深圳大学微众金融科技学院, 广东深圳 518060
3. 哈尔滨工业大学郑州高等研究院青年科学家创新中心, 河南郑州 450000
4. 哈尔滨工业大学计算机科学与技术学院, 黑龙江哈尔滨 150001

通讯作者: 田春伟
作者简介:
谢静远, 男, 硕士研究生, 主研方向为图像识别、算子优化
查凯文, 本科生
刘鹏举, 助理研究员
田春伟(通信作者), 教授
基金资助:
中国人工智能学会-昇腾CANN学术基金; 国家自然科学基金面上项目(62576123); 黑龙江省自然科学基金青年项目(B类)(YQ2025F003); 河南省科技攻关计划(262102211043)

Abstract

Abstract:

This paper studies operator-level reconstruction for the structural mismatch between YOLOv11 and Ascend Neural Processing Unit (NPU). The Spatial Pyramid Pooling-Fast (SPPF), C3K2, and C2PSA modules are optimized without changing network semantics or model scale. Three Ascend C operators are designed: the SPPF operator uses on-chip data loop and halo cache to reduce redundant global-memory traffic in multi-stage pooling; the C3K2 operator integrates multi-core task assignment and multi-queue asynchronous pipelining to reduce fine-grained kernel launch overhead; and the C2PSA operator reconstructs attention communication through a parallel reduction-broadcast primitive. On an Ascend 910B NPU, the complete reconstruction reduces the training time per epoch by 23.2% and improves the training throughput by 27.6% on the COCO dataset. The results show that matching Ascend on-chip memory, asynchronous queues, and multi-core synchronization mechanisms improves the training execution efficiency of key YOLOv11 modules and keeps inference performance stable. It can provide verifiable operator mapping schemes for the deployment of complex object detection networks on the Ascend platform.

Key words: YOLO optimization, Ascend Neural Processing Unit (NPU), operator-level reconstruction, Ascend C, framework adaptation

摘要：

针对YOLOv11的快速空间金字塔池化(SPPF)、C3K2与C2PSA模块在昇腾神经网络处理器(NPU)上存在的冗余访存、细粒度核函数启动和多分支同步通信瓶颈, 本文研究一种不改变网络语义和参数规模的算子级重构方法。基于Ascend C设计3类硬件友好融合算子: SPPF采用片上数据闭环与边界缓存减少多级池化的全局内存往返; C3K2采用多核任务分配和多队列异步流水线降低细粒度算子串行开销; C2PSA采用并行归约-广播机制重构注意力分支间通信。在Ascend 910B NPU上, 完整重构后COCO数据集上单轮训练时间缩短23.2%, 训练吞吐量提升27.6%。实验结果表明, 所提方法通过适配昇腾片上存储、异步队列和多核同步机制, 提高了YOLOv11关键模块的训练执行效率, 并保持推理性能稳定, 可为复杂目标检测网络在昇腾平台上的部署提供可复核的算子映射方案。

关键词: YOLO优化, 昇腾神经网络处理器, 算子级重构, Ascend C, 框架适配

XIE Jingyuan, ZHA Kaiwen, LIU Pengju, TIAN Chunwei. Operator-level Reconstruction Method of YOLOv11 for Ascend NPU[J]. Computer Engineering, 2026, 52(9): 365-375.

谢静远, 查凯文, 刘鹏举, 田春伟. 面向昇腾NPU的YOLOv11算子级重构方法[J]. 计算机工程, 2026, 52(9): 365-375.

/ Recommend / Download Citations

URL: https://www.ecice06.com/EN/10.19678/j.issn.1000-3428.0260413

https://www.ecice06.com/EN/Y2026/V52/I9/365

Figures/Tables 11

Fig.1 Overall framework of YOLO-Ascend operator-level reconstruction for the Ascend NPU

Fig.2 Visual comparison of detection results of different methods on the PASCAL VOC dataset

Fig.3 Visual comparison of detection results of different methods on the COCO dataset

Fig.4 Visual comparison of feature responses from different methods on the CIFAR-100 dataset

References 37

1	贾晓芬, 吴雪茹, 赵佰亭. 绝缘子自爆缺陷的轻量化检测网络DE-YOLO. 电子测量与仪器学报, 2023, 37 (5): 28- 35.
	JIA X F , WU X R , ZHAO B T . Lightweight detection network for insulator self-detonation defect DE-YOLO. Journal of Electronic Measurement and Instrumentation, 2023, 37 (5): 28- 35.
2	TIAN C W , LIU K , ZHANG B , et al. A dynamic transformer network for vehicle detection. IEEE Transactions on Consumer Electronics, 2025, 71 (2): 2387- 2394. doi: 10.1109/TCE.2025.3565318
3	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, USA: IEEE Press, 2016: 779-788.
4	Ultralytics. Ultralytics YOLO11[EB/OL]. (2024-09-30)[2026-03-28]. https://github.com/ultralytics/ultralytics.
5	田春伟, 宋明键, 左旺孟, 等. 卷积神经网络在图像超分辨上的应用. 智能系统学报, 2025, 20 (3): 719- 749.
	TIAN C W , SONG M J , ZUO W M , et al. Application of convolutional neural networks in image super-resolution. CAAI Transactions on Intelligent Systems, 2025, 20 (3): 719- 749.
6	LIAO H, TU J J, XIA J, et al. Ascend: a scalable and unified architecture for ubiquitous deep neural network computing: industry track paper[C]//Proceedings of the IEEE International Symposium on High-Performance Computer Architecture (HPCA). Seoul, Republic of Korea: IEEE Press, 2021: 789-801.
7	WONG A, FAMUORI M, SHAFIEE M J, et al. YOLO nano: a highly compact you only look once convolutional neural network for object detection[C]//Proceedings of the 15th Workshop on Energy Efficient Machine Learning and Cognitive Computing-NeurIPS Edition (EMC2-NIPS). Vancouver, Canada: IEEE Press, 2019: 22-25.
8	SHEN X, DONG P Y, LU L, et al. Agile-quant: activation-guided quantization for faster inference of LLMs on the edge[C]//Proceedings of the AAAI Conference on Artificial Intelligence. [S. l.]: AAAI Press, 2024: 18944-18951.
9	NIU W, GUAN J X, WANG Y Z, et al. DNNFusion: accelerating deep neural networks execution with advanced operator fusion[C]//Proceedings of the 42nd ACM SIGPLAN International Conference on Programming Language Design and Implementation. New York, USA: ACM Press, 2021: 883-898.
10	包振山, 郭俊南, 张文博, 等. UltraAcc: 基于FPGA流水架构的低功耗高性能CNN加速器定制设计. 计算机学报, 2023, 46 (6): 1139- 1155.
	BAO Z S , GUO J N , ZHANG W B , et al. UltraAcc: a customized low power and high performance CNN accelerator with dataflow on FPGAs. Chinese Journal of Computers, 2023, 46 (6): 1139- 1155.
11	PASZKE A, GROSS S, MASSA F, et al. PyTorch: an imperative style, high-performance deep learning library[C]// Advances in Neural Information Processing Systems 32. Vancouver, Canada: Curran Associates, Inc., 2019: 8024-8035.
12	EVERINGHAM M , VAN GOOL L , WILLIAMS C K I , et al. The Pascal Visual Object Classes (VOC) challenge. International Journal of Computer Vision, 2010, 88 (2): 303- 338. doi: 10.1007/s11263-009-0275-4
13	LIN T Y, MAIRE M, BELONGIE S, et al. Microsoft COCO: common objects in context[C]//Proceedings of European Conference on Computer Vision. Berlin, Germany: Springer, 2014: 740-755.
14	KRIZHEVSKY A, HINTON G. Learning multiple layers of features from tiny images[EB/OL]. [2026-03-28]. https://www.semanticscholar.org/paper/Learning-Multiple-Layers-of-Features-from-Tiny-Krizhevsky/5d90f06bb70a0a3dced62413346235c02b1aa086.
15	LOSHCHILOV I, HUTTER F. Decoupled weight decay regularization[C]//Proceedings of the 7th International Conference on Learning Representations. New Orleans, USA: OpenReview. net, 2019: 1-10.
16	LIU K H, WANG M R, LIU T J, et al. YOLOv8-SwinT: a real time object detector via fusion of YOLOv8 and Swin Transformer[C]//Proceedings of IEEE ICCE-TW 2026. Kaohsiung, China: IEEE Press, 2026: 141-142.
17	LEI X C, WU S Q, WU W L, et al. MambaNeXt-YOLO: a hybrid state space model for real-time object detection[EB/OL]. [2026-03-28]. https://arxiv.org/abs/2506.03654.
18	QU Z , QU H M , YIN H N , et al. A method of object detection network with progressive feature fusion and reverse attention for traffic scenes. Signal, Image and Video Processing, 2025, 19 (11): 942. doi: 10.1007/s11760-025-04524-7
19	WANG C Y, YEH I H, MARK LIAO H Y. YOLOv9: learning what you want to learn using programmable gradient information[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Berlin, Germany: Springer, 2025: 1-21. .
20	CHEN H, CHEN K, DING G G, et al. YOLOv10: real-time end-to-end object detection[C]//Advances in Neural Information Processing Systems 37. Vancouver, Canada: Neural Information Processing Systems Foundation, Inc. (NeurIPS), 2024: 107984-108011.
21	BIAN Z T , YAO B , LI Q . SPAFPN: wear a multi-scale feature fusion scarf around neck for real-time object detector. Information Fusion, 2025, 119, 103034. doi: 10.1016/j.inffus.2025.103034
22	LIU Z , WU J H , CAI Y F , et al. Dual-stage feature specialization network for robust visual object detection in autonomous vehicles. Scientific Reports, 2025, 15, 15501. doi: 10.1038/s41598-025-99363-4
23	WANG X, JIN Y, CHEN L, et al. Dynamic graph induced contour-aware heat conduction network for event-based object detection[EB/OL]. [2026-03-28]. https://arxiv.org/abs/2505.12908.
24	JAFARI A A, ANBARJAFARI G. GEYOLO-AHC: a hybrid graph-enhanced framework with adaptive heat conduction for scalable, real-time object detection[EB/OL]. [2026-03-28]. https://www.techrxiv.org/doi/full/10.36227/techrxiv.175393417.70423928/v1.
25	REN S Q , HE K M , GIRSHICK R , et al. Faster R-CNN: towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39 (6): 1137- 1149. doi: 10.1109/TPAMI.2016.2577031
26	DAI X Y, CHEN Y P, XIAO B, et al. Dynamic head: unifying object detection heads with attentions[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville, USA: IEEE Press, 2021: 7369-7378.
27	CHEN K, PANG J M, WANG J Q, et al. Hybrid task cascade for instance segmentation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, USA: IEEE Press, 2019: 4969-4978.
28	MENG D P, CHEN X K, FAN Z J, et al. Conditional DETR for fast training convergence[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Montreal, Canada: IEEE Press, 2021: 3631-3640.
29	ZHU X Z, SU W J, LU L W, et al. Deformable DETR: deformable transformers for end-to-end object detection[EB/OL]. [2026-03-28]. https://arxiv.org/abs/2010.04159.
30	PICRON C, CHAKRAVARTY P, TUYTELAARS T. FQDet: fast-converging query-based detector[EB/OL]. [2026-03-28]. https://arxiv.org/abs/2210.02318.
31	HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, USA: IEEE Press, 2017: 2261-2269.
32	HOWARD A, SANDLER M, CHEN B, et al. Searching for MobileNetV3[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Seoul, Republic of Korea: IEEE Press, 2019: 1314-1324.
33	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, USA: IEEE Press, 2016: 770-778.
34	DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16×16 words: transformers for image recognition at scale[EB/OL]. [2026-03-28]. https://arxiv.org/abs/2010.11929.
35	ZAGORUYKO S, KOMODAKIS N. Wide residual networks[EB/OL]. [2026-03-28]. https://arxiv.org/abs/1605.07146.
36	DEALCALA D, MORALES A, FIERREZ J, et al. AttZoom: attention zoom for better visual features[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). Honolulu, USA: IEEE Press, 2026: 4893-4902.
37	LEANDRE S H, SHIFERAW N A, ROUT D. Enhancing compact convolutional transformers with super attention[EB/OL]. [2026-03-28]. https://arxiv.org/abs/2508.18960.

Please choose a citation manager

Content to export