面向深度神经网络加速芯片的高效硬件优化策略

张萌; 张经纬; 李国庆; 吴瑞霞; 曾晓洋

doi:10.11999/JEIT210002

面向深度神经网络加速芯片的高效硬件优化策略

doi: 10.11999/JEIT210002

1.
东南大学电子学院国家专用集成电路系统工程技术研究中心南京 210096
2.
复旦大学专用集成电路与系统国家重点实验室上海 200433

基金项目: 国家重点研发计划(2018YFB2202703)，江苏省自然科学基金(BK20201145)

详细信息

作者简介:
张萌：男，1964年生，研究员，研究方向为数字信号处理、深度学习算法及硬件加速

张经纬：男，1997年生，硕士生，研究方向为深度学习硬件加速器设计

李国庆：男，1991年生，博士生，研究方向为计算机视觉和深度学习硬件加速器设计

吴瑞霞：女，1996年生，硕士生，研究方向为深度学习算法

曾晓洋：男，1972年生，教授，研究方向为高能效系统芯片(SoC)

通讯作者:
张经纬　zhangjingwei@seu.edu.cn

中图分类号: TN79.1
计量
- 文章访问数: 1451
- HTML全文浏览量: 503
- PDF下载量: 192
- 被引次数: 0
出版历程
- 收稿日期: 2021-01-04
- 修回日期: 2021-04-21
- 网络出版日期: 2021-04-29
- 刊出日期: 2021-06-18

Efficient Hardware Optimization Strategies for Deep Neural Networks Acceleration Chip

1.
National ASIC Engineering Center, School of Electronic Sci. and Eng., Southeast University, Nanjing 210096, China
2.
National ASIC Key Laboratory, Fudan University, Shanghai 200433, China

Funds: The National Key R&D Program of China(2018YFB2202703), Jiangsu Province of Natural Science and Technology(BK20201145)

摘要

摘要: 轻量级神经网络部署在低功耗平台上的解决方案可有效用于无人机(UAV)检测、自动驾驶等人工智能(AI)、物联网(IOT)领域，但在资源有限情况下，同时兼顾高精度和低延时来构建深度神经网络(DNN)加速器是非常有挑战性的。该文针对此问题提出一系列高效的硬件优化策略，包括构建可堆叠共享计算引擎(PE)以平衡不同卷积中数据重用和内存访问模式的不一致；提出了可调的循环次数和通道增强方法，有效扩展加速器与外部存储器之间的访问带宽，提高DNN浅层网络计算效率；优化了预加载工作流，从整体上提高了异构系统的并行度。经Xilinx Ultra96 V2板卡验证，该文的硬件优化策略有效地改进了iSmart3-SkyNet和SkrSkr-SkyNet类的DNN加速芯片设计。结果显示，优化后的加速器每秒处理78.576帧图像，每幅图像的功耗为0.068 J。
- 深度神经网络 /
- 目标检测 /
- 神经网络加速器 /
- 低功耗 /
- 硬件优化
Abstract: Lightweight neural networks deployed on low-power platforms have proven to be effective solutions for Artificial Intelligence (AI) and Internet Of Things (IOT) domains such as Unmanned Aerial Vehicle (UAV) detection and unmanned driving. However, in the case of limited resources, it is very challenging to build Deep Neural Networks (DNN) accelerator with both high precision and low delay. In this paper, a series of efficient hardware optimization strategies are proposed, including stackable shared Processing Engine (PE) to balance the inconsistency of data reuse and memory access patterns in different convolutions; Regulable loop parallelism and channel augmentation are proposed to increase effectively the access bandwidth between accelerator and external memory. It also improve the efficiency of DNN shallow layers computing; Pre-Workflow is applied to improve the overall parallelism of heterogeneous systems. Verified by Xilinx Ultra96 V2 board, the hardware optimization strategies in this paper improve effectively the design of DNN acceleration chips like iSmart3-SkyNet and SkrSkr-SkyNet. The results show that the optimized accelerator processes 78.576 frames per second, and the power consumption of each picture is 0.068 Joules.
- Deep Neural Networks (DNN) /
- Object detection /
- Neural network accelerator /
- Low power consumption /
- Hardware optimization

HTML全文

图 1 iSmart3-SkyNet加速器上的SkyNet Roofline模型分析

下载: 全尺寸图片幻灯片

图 2 系统-计算模块-线性缓冲区结构示意图

下载: 全尺寸图片幻灯片

图 3 通道增强流程说明图

下载: 全尺寸图片幻灯片

图 4 3种工作流比较图

下载: 全尺寸图片幻灯片

图 5 优化后加速器上的SkyNet Roofline模型分析

下载: 全尺寸图片幻灯片

图 6 iSmart3和Skrskr加速优化前后性能对比

下载: 全尺寸图片幻灯片

表 1 SkyNet的体系结构和每个捆绑包的推理速度表格

捆绑包	层数	输入尺寸	操作类型	计算量、计算量占比(%)	延迟占比(%)
#1	1	3×160×320	DW-Conv3	119.61M, 20.6	33.90
	2	3×160×320	PW-Conv1
	3	48×160×320	POOLING
#2	4	48×80×160	DW-Conv3	86.02M, 14.42	16.54
	5	48×80×160	PW-Conv1
	6	96×80×160	POOLING
#3	7	96×40×80	DW-Conv3	61.75M, 10.36	6.23
	8	96×40×80	PW-Conv1
	9	192×40×80	POOLING
#4	10	192×20×40	DW-Conv3	60.36M, 10.13	4.92
#4	11	192×20×40	PW-Conv1	60.36M, 10.13	4.92
#5	12	384×20×40	DW-Conv3	160.05M, 26.85	12.43
#5	13	384×20×40	PW-Conv1	160.05M, 26.85	12.43
#6	–	合并第9层输出		107.52M, 18.04	20.08
	14	1280×20×40	[旁路] DW-Conv3
	15	1280×20×40	PW-Conv1
#7	16	96×20×40	PW-Conv1	0.77M, 0.14	0.10
–	17	10×20×40	计算回归框		0.16
CPU	–	–	–	–	5.64

下载: 导出CSV

表 2 优化策略效果对比

加速器	iSmart3 ^[9]	SEUer A	Skrskr ^[10]	SEUer B
网络模型	SkyNet	SkyNet	SkyNet	SkyNet
量化精度	A9/W11	A9/W11	A8/W6	A8/W6
硬件平台	Ultra96V2	Ultra96V2	Ultra96V2	Ultra96V2
准确率(DJI)	0.716	0.724	0.731	0.731
时钟频率(MHz)	215	215	300	300
DSP数量	329	287	360	360
LUT数量(k)	54	54	56	46
FF数量(k)	60	70	68	51
帧率(fps)	25.05	37.393	52.429	78.576
GOPS/W	3.21	5.95	7.22	11.19
Energy/Pic.(J)	0.289	0.135	0.129	0.068

下载: 导出CSV

参考文献(17)

[1]	王巍, 周凯利, 王伊昌, 等. 基于快速滤波算法的卷积神经网络加速器设计[J]. 电子与信息学报, 2019, 41(11): 2578–2584. doi: 10.11999/JEIT190037 WANG Wei, ZHOU Kaili, WANG Yichang, et al. Design of convolutional neural networks accelerator based on fast filter algorithm[J]. Journal of Electronics &Information Technology, 2019, 41(11): 2578–2584. doi: 10.11999/JEIT190037
[2]	ZHANG Xiaofan, WANG Junsong, ZHU Chao, et al. DNNBuilder: An automated tool for building high-performance DNN hardware accelerators for FPGAs[C]. 2018 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), San Diego, USA, 2018: 1–8.
[3]	LI Huimin, FAN Xitian, JIAO Li, et al. A high performance FPGA-based accelerator for large-scale convolutional neural networks[C]. The 26th International Conference on Field Programmable Logic and Applications (FPL), Lausanne, Switzerland, 2016: 1–9.
[4]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, real-time object detection[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 779–788.
[5]	REN Shaoqing, HE Kaiming, GIRSHICK R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137–1149. doi: 10.1109/TPAMI.2016.2577031
[6]	TAN Mingxing, PANG Ruoming, and LE Q V. EfficientDet: Scalable and efficient object detection[C]. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, USA, 2020: 10781–10790.
[7]	YU Yunxuan, WU Chen, ZHAO Tiandong, et al. OPU: An FPGA-based overlay processor for convolutional neural networks[J]. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 2020, 28(1): 35–47. doi: 10.1109/TVLSI.2019.2939726
[8]	YU Yunxuan, ZHAO Tiandong, WANG Kun, et al. Light-OPU: An FPGA-based overlay processor for lightweight convolutional neural networks[C]. 2020 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays, Seaside, USA, 2020: 122–132.
[9]	ZHANG Xiaofan, LU Haoming, HAO Cong, et al. SkyNet: A hardware-efficient method for object detection and tracking on embedded systems[J]. arXiv: 1909.09709, 2019.
[10]	JIANG W, LIU X, SUN H, et al. Skrskr: Dacsdc. 2020 2nd place winner in fpga track[EB/OL]. https://github.com/jiangwx/SkrSkr/, 2020.
[11]	ZHANG Chen, LI Peng, SUN Guangyu, et al. Optimizing FPGA-based accelerator design for deep convolutional neural networks[C]. 2015 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays, Monterey, USA, 2015: 161–170.
[12]	HAO Cong, ZHANG Xiaofan, LI Yuhong, et al. FPGA/DNN Co-Design: An efficient design methodology for 1ot intelligence on the edge[C]. The 56th ACM/IEEE Design Automation Conference (DAC), Las Vegas, USA, 2019: 1–6.
[13]	MOTAMEDI M, GYSEL P, AKELLA V, et al. Design space exploration of FPGA-based deep convolutional neural networks[C]. The 21st Asia and South Pacific Design Automation Conference (ASP-DAC), Macao, China, 2016: 575–580.
[14]	FAN Hongxiang, LIU Shuanglong, FERIANC M, et al. A real-time object detection accelerator with compressed SSDLite on FPGA[C]. 2018 International Conference on Field-Programmable Technology (FPT), Naha, Japan, 2018: 14–21.
[15]	LI Fanrong, MO Zitao, WANG Peisong, et al. A system-level solution for low-power object detection[C]. 2019 IEEE/CVF International Conference on Computer Vision Workshops, Seoul, Korea (South), 2019: 2461–2468.
[16]	DONG Zhen, WANG Dequan, HUANG Qijing, et al. CoDeNet: Efficient deployment of input-adaptive object detection on embedded FPGAs[J]. arXiv: 2006.08357, 2020.
[17]	WU Di, ZHANG Yu, JIA Xijie, et al. A high-performance CNN processor based on FPGA for MobileNets[C]. The 29th International Conference on Field Programmable Logic and Applications (FPL), Barcelona, Spain, 2019: 136–143.