一种基于NoC多核系统的矩阵乘法映射技术

doi:10.16180/j.cnki.issn1007-7820.2021.05.010

摘要/Abstract

摘要：

矩阵乘法是现代信号处理的基本运算,提高数据的并行处理能力对提升矩阵乘法的运算性能具有重要现实意义。文中在基于NoC多核系统中针对不同维度的矩阵乘法的密集型计算进行任务调度以及资源分配,实现了多种适应于不同矩阵乘法的映射方案,其峰值性能可达5 078 MFLOPS。同时,文中设计的运算单元相对独立且可重构,对任意维度的矩阵乘法具有良好的扩展性和通用性,解决了通用矩阵乘法器在固定结构中受到I/O带宽和计算资源的限制而产生的运算效率较低和扩展性较差的缺陷。不同维度矩阵乘法的实验结果分析证实了文中设计的运算性能和正确性。

关键词: 矩阵乘法, 并行计算, NoC多核, 密集型, 任务调度, 资源分配, 通用性, I/O带宽

Abstract:

Matrix multiplication is the basic operation of modern signal processing. Improving the parallel processing capacity of data has important practical significance for improving the operation performance of matrix multiplication. In this study, task scheduling and resource allocation are carried out for the intensive computing of matrix multiplication in different dimensions based on NOC multi-core system, and a variety of mapping algorithms suitable for different matrix multiplication are implemented, and the peak performance can reach 5078 MFLOPS. The designed operation unit is relatively independent and reconfigurable, which has good expansibility and generality for matrix multiplication of any dimension. It overcomes the limitation of I/O bandwidth and computing resources in fixed structure, which leads to low efficiency and poor expansibility. Through the analysis of the experimental results of matrix multiplication of different dimensions, the correctness and high performance of the design are verified.

Key words: matrix multiplication, parallel computing, NoC multi-core, intensive, task scheduling, resource allocation, generality, I/O bandwidth

中图分类号:

TN492

汪杨,王晓蕾,袁子昂,袁儒明. 一种基于NoC多核系统的矩阵乘法映射技术[J]. 电子科技, 2021, 34(5): 54-60.

WANG Yang,WANG Xiaolei,YUAN Ziang,YUAN Ruming. A Matrix Multiplication Mapping Technology Based on NOC Multi-Core System[J]. Electronic Science and Technology, 2021, 34(5): 54-60.

图/表 15

图1

图2

图3

图4

图5

图6

图7

图8

图9

图10

图11

图12

表1

表2

表3

参考文献 17

[1]	张天作. 基于FPGA的矩阵乘法实现方案在全连接深度神经网络前向传播中的性能评估[D]. 北京:北京邮电大学, 2018.
	Zhang Tianzuo. Performance evaluation of matrix multiplication based on FPGA in forward propagation of full connection deep neural network[D]. Beijing:Beijing University of Posts and Telecommunications, 2018.
[2]	王阳, 陶华敏, 肖山竹, 等. 基于脉动阵列的矩阵乘法器硬件加速技术研究[J]. 微电子学与计算机, 2015(11):120-124.
	Wang Yang, Tao Huamin, Xiao Shanzhu, et al. Research on hardware acceleration technology of matrix multiplier based on pulse array[J]. Microelectronics and Computer, 2015(11):120-124.
[3]	Kung H T. Systolic communication[C]. San Diego:Proceedings of the International Conference on Systolic Array, 1988.
[4]	田翔, 周凡, 陈耀武, 等. 基于FPGA的实时双精度浮点矩阵乘法器设计[J]. 浙江大学学报(工学版), 2008,42(9):1611-1615.
	Tian Xiang, Zhou Fan, Chen Yaowu, et al. Design of real-time double precision floating-point matrix multiplier based on FPGA[J]. Journal of Zhejiang University (Engineering Edition), 2008,42(9):1611-1615.
[5]	周磊涛, 陶耀东, 刘生, 等. 基于FPGA的Systolic乘法技术研究[J]. 计算机工程与科学, 2015,37(9):1632-1636.
	Zhou Leitao, Tao Yaodong, Liu Sheng, et al. Research onSystolic multiplication technology based on FPGA[J]. Computer Engineering and Science, 2015,37(9):1632-1636.
[6]	费敏锐, 熊南, 李韬. 网络化系统时钟同步算法[J]. 中国科学信息科学, 2016,46(11):1527-1541.
	Fei Minrui, Xiong Nan, Li Tao. Clock synchronization algorithm for networked systems[J]. China Science and Information Science, 2016,46(11):1527-1541.
[7]	Tan C Y, Ewetz R, Koh C K. Clustering of flip-flops for useful-skew clock tree synconfproc[C]. Jeju lsland:Design Automation Conference, 2018.
[8]	Shirwaikar M G, Kadayinti N, Sharma D K. Clock skew measurement using an all-digital Sigma-Delta time to digital converter[C]. Hyderabad: International Conference on Embedded Systems, 2017.
[9]	谢盈, 吴尽昭. 一种多核系统任务调度算法动态度量方法[J]. 计算机应用研究, 2019(1):132-135.
	Xie Ying, Wu Jingzhao. A dynamic measurement method of multi-core system task scheduling algorithm[J]. Computer Application Research, 2019(1):132-135.
[10]	汪健, 张磊, 赵忠惠, 等. 多核系统中NoC通讯架构的关键技术[J]. 电子科技, 2012,25(6):47-52.
	Wang Jian, Zhang Lei, Zhao Zhonghui, et al. Key technologies of NOC communication architecture in multi-core system[J]. ElectronicScience and Technology, 2012,25(6):47-52.
[11]	张保岗. Mesh拓扑片上网络映射方法研究[D]. 郑州:战略支援部队信息工程大学, 2019.
	Zhang Baogang. Research on network mapping method on Mesh topology chip[D]. Zhengzhou:University of Information Engineering, Strategic Support Force, 2019.
[12]	钱庆松. 异构多核片上网络布局优化研究与实现[D]. 合肥:合肥工业大学, 2017.
	Qian Qingsong. Research and implementation of network layout optimization on heterogeneous multi-core chip[D]. Hefei:Hefei University of Technology, 2017.
[13]	Song Y, Jiao R, Zhang D, et al. Performance analysis for matrix-multiplication based on an heterogeneous multi-core SoC[C]. Chengdu:International Conference on ASIC, 2015.
[14]	全钊锋. NoC边界扫描测试系统硬件设计[J]. 电子科技, 2015,28(9):63-66.
	Quan Zhaofeng. Hardware design of NoC boundary scan test system[J]. Electronic Science and Technology, 2015,28(9):63-66.
[15]	陈国良. 并行计算: 结构·算法·编程[M]. 北京: 高等教育出版社, 2011.
	Chen Guoliang. Parallel computing: structure, algorithm and programming[M]. Beijing: Higher Education Press, 2011.
[16]	贾迅, 邬贵明, 谢向辉, 等. 双精度浮点矩阵乘协处理器研究[J]. 计算机研究与发展, 2019,56(2):186-196.
	Jia Xun, Wu Guiming, Xie Xianghui, et al. Research on double precision floating-point matrix multiplication coprocessor[J]. Computer Research and Development, 2019,56(2):186-196.
[17]	Cong J, Wang J. Automatic interior I/O elimination in systolic array architecture[C]. Boulder:Annual International Symposium on Field-Programmable Custom Computing Machines, 2018.

内部资源	资源使用量/个	总资源占用比/%
LUT	99 044	8.10
BRAM	96	14.86
DSP48E	32	1.48

阶数	运算周期
阶数	方案1	方案2	方案3	脉动阵列
32	2 243	2 277	3 074	2 055
64	16 835	16 983	21 016	16 391
128	132 127	133 032	136 971	131 079
256	1 148 982	1 151 318	1 071 563	1 048 572
512	9 375 040	9 383 437	8 404 794	8 388 615
1 024	74 696 304	74 918 735	67 137 058	67 108 871

阶数	运算周期
阶数	方案1	方案2	方案3	脉动阵列
32	1.82	1.79	1.33	1.99
64	1.94	1.92	1.55	1.99
128	1.98	1.97	1.91	1.99
256	1.82	1.82	1.95	1.99
512	1.79	1.79	1.99	1.99
1024	1.78	1.78	1.99	1.99