OpenPARF: 基于深度学习工具包的大规模异构FPGA开源布局布线框架

麦景; 王嘉睿; 邸志雄; 林亦波

doi:10.11999/JEIT230387

OpenPARF: 基于深度学习工具包的大规模异构FPGA开源布局布线框架

doi: 10.11999/JEIT230387

麦景^{1, 2},
王嘉睿^{1, 2},
邸志雄³,
林亦波^2, ,

1.
北京大学计算机学院北京 100871
2.
北京大学集成电路学院北京 100871
3.
西南交通大学信息科学与技术学院成都 611756

基金项目: 科技部重点研发计划(2021ZD0114702)

详细信息

作者简介:
麦景：男，博士生，研究方向为面向超大规模集成电路设计自动化的建模和优化、机器学习等

王嘉睿：男，博士生，研究方向为面向超大规模集成电路设计自动化的建模和优化、机器学习等

邸志雄：男，博士，副教授，硕士生导师，研究方向为数字芯片物理实现EDA算法、高性能图像编解码芯片设计和FPGA硬件加速设计

林亦波：男，博士，研究员，博士生导师，研究方向为面向超大规模集成电路设计自动化的建模和优化、深度学习及其应用和异构计算

通讯作者:
林亦波　yibolin@pku.edu.cn

1¹⁾ DREAMPlaceFPGA包括文献[18,19]两个工作，其中文献[19]在文献[18]的基础上使用GPU加速合法化算法。本文实验部分中DREAMPlaceFPGA的布局时间摘取自文献[19]。
2²⁾ 本文实验中线程数为8
中图分类号: TN47
计量
- 文章访问数: 1164
- HTML全文浏览量: 475
- PDF下载量: 175
- 被引次数: 0
出版历程
- 收稿日期: 2023-05-08
- 修回日期: 2023-08-21
- 网络出版日期: 2023-08-23
- 刊出日期: 2023-09-27

OpenPARF: An Open-source Placement and Routing Framework for Large-scale Heterogeneous FPGAs with Deep Learning Toolkit

MAI Jing^{1, 2},
WANG Jiarui^{1, 2},
DI Zhixiong³,
LIN Yibo^{2
, ,}

1.
School of Computer Science, Peking University, Beijing 100871, China
2.
School of Integrated Circuits, Peking University, Beijing 100871, China
3.
School of Information and Science and Technology, Southwest Jiaotong University, Chengdu 611756, China

Funds: Key Research and Development Program Projects of the Ministry of Science and Technology (2021ZD0114702)

摘要

摘要: 该文提出一个面向大规模可编辑逻辑门阵列(FPGA)的开源布局布线框架OpenPARF。该框架基于深度学习工具包PyTorch实现，支持GPU大规模并行计算求解。在布局算法方面，该文设计了一种新型非对称多静电场系统，对FPGA布局问题进行建模。在布线算法方面，该文支持对FPGA可编程逻辑块(CLB)内部布线资源进行准确建模，并在大规模不规则布线资源图上进行布线，提高了异构FPGA芯片布线器的性能和效率。该文在ISPD 2016和2017 FPGA竞赛数据集和工业标准级FPGA数据集上进行了实验，结果表明该框架可减少0.4%～12.7%的布线线长，并实现两倍以上布局效率提升。
- 集成电路设计与设计自动化 /
- 物理实现 /
- FPGA /
- 布局布线 /
- 机器学习
Abstract: An Open-source Placement And Routing Framework (OpenPARF) for large-scale FPGA physical design is proposed in this paper. OpenPARF is implemented with of deep learning toolkit PyTorch and supports GPU massive parallel acceleration. For placement, the framework incorporates a novel asymmetric multi-electrostatic filed system to model the FPGA placement problem. For routing, OpenPARF integrates finer-grained internal routing of FPGA Configurable Logic Blocks (CLBs) in the routing model and supports routing on large-scale irregular routing resource graph. This study can significantly improve the FPGA routing algorithm's efficiency and effectiveness. Experimental results on ISPD 2016 and ISPD 2017 FPGA conest benchmarks and industrial-level FPGA benchmarks demonstrate that OpenPARF can achieve 0.4%～12.7% improvement in routed wirelength and more than two times speedup in placement.
- Integrated circuit design and design automation /
- Physical design /
- FPGA /
- Placement and douting /
- Machine learning

HTML全文

图 1 FPGA 布局布线架构示意图

下载: 全尺寸图片幻灯片

图 2 OpenPARF 总体算法框架图

下载: 全尺寸图片幻灯片

图 3 4 种情景下的 LUT 和 SHIFT 的位置关系下 LUTL 和 LUTM-AL 静电场系统的电荷密度和电势能

下载: 全尺寸图片幻灯片

表 1 ISPD2016 和 ISPD2017 benchmark 的实例数量和网表数量

ISPD 2016						ISPD 2017
FPGA设计	逻辑单元数量				网表数量 (k)	FPGA设计	逻辑单元数量				网表数量 (k)
FPGA设计	LUT (k)	FF (k)	RAM	DSP	网表数量 (k)	FPGA设计	LUT (k)	FF (k)	RAM	DSP	网表数量 (k)
FPGA01	50	55	0	0	105	CLK-FPGA01	211	324	164	75	536
FPGA02	100	66	100	100	167	CLK-FPGA02	230	280	236	112	511
FPGA03	250	170	600	500	428	CLK-FPGA03	410	481	850	395	898
FPGA04	250	172	600	500	430	CLK-FPGA04	309	372	467	224	685
FPGA05	250	174	600	500	433	CLK-FPGA05	393	469	798	150	865
FPGA06	350	352	1000	600	713	CLK-FPGA06	425	511	872	420	943
FPGA07	350	355	1000	600	716	CLK-FPGA07	254	309	313	149	565
FPGA08	500	216	600	500	725	CLK-FPGA08	212	257	161	75	470
FPGA09	500	366	1000	600	876	CLK-FPGA09	231	358	236	112	591
FPGA10	350	600	1000	600	961	CLK-FPGA10	327	506	542	255	837
FPGA11	480	363	1000	400	851	CLK-FPGA11	300	468	454	224	772
FPGA12	500	602	600	500	1111	CLK-FPGA12	277	430	389	187	710
–	–	–	–	–	–	CLK-FPGA13	339	405	570	262	749

下载: 导出CSV

表 2 在工业标准级 FPGA 数据集上的布局时间 (s)、布线时间(min)以及布线线长 (×10³)

FPGA设计	逻辑单元数量		网表数量	布局时间	布线时间	布线线长
FPGA设计	#LUT/#FF/#BRAM/#DSP	#DiRAM+#SHIFT	网表数量	布局时间	布线时间	布线线长
IND01	17k/11k/0/13	9	52492	72.36	10	90
IND02	11k/10k/0/24	6	26678	77.82	15	100
IND03	109k/12k/0/0	0	121554	109.54	108	1021
IND04	29k/17k/0/16	218	60968	69.39	19	283
IND05	64k/191k/64/928	29K	371808	126.38	109	2360
IND06	112k/65k/21/0	0	221182	88.28	176	1593
IND07	40k/156k/89/768	26K	294075	140.33	68	1450

下载: 导出CSV

表 3 在 ISPD 2016 benchmark 上的布局时间 (s)、布线时间 (min)以及布线线长 (×10⁴) 的比较

FPGA设计	RippleFPGA (CPU)			DREAMPlaceFPGA (GPU)			OpenPARF (CPU)			OpenPARF (GPU)
FPGA设计	布局时间	布线时间	布线线长	布局时间	布线时间	布线线长	布局时间	布线时间	布线线长	布局时间	布线时间	布线线长
FPGA01	41.12	3	36.44	32.08	3	31.78	422	2	31.75	38.58	3	31.72
FPGA02	64.22	5	75.29	56.82	5	68.17	719	4	67.86	59.08	5	67.73
FPGA03	245.40	17	346.91	107.96	15	299.56	760	15	294.75	119.75	15	294.75
FPGA04	337.42	22	632.96	97.22	22	569.85	727	22	577.30	111.85	22	577.30
FPGA05	391.42	57	1222.46	90.60	56	1167.60	899	58	1148.40	122.78	54	1148.72
FPGA06	593.06	25	652.41	182.62	29	571.11	975	29	573.95	218.17	27	573.54
FPGA07	782.33	46	1106.96	159.07	51	964.44	924	46	966.37	208.84	45	965.24
FPGA08	489.97	40	958.26	146.33	36	911.67	921	38	895.47	184.13	38	896.91
FPGA09	737.86	58	1327.34	190.63	54	1203.49	1036	52	1198.43	259.76	50	1198.22
FPGA10	1179.94	28	711.48	179.50	28	544.52	1082	26	542.11	258.04	35	542.02
FPGA11	721.43	58	1281.65	147.88	56	1250.49	1020	59	1254.15	220.95	59	1253.78
FPGA12	883.19	40	761.37	183.95	37	674.21	1177	40	670.31	290.43	38	670.94
平均值	2.771	1.015	1.127	0.786	1.000	1.004	6.170	0.957	1.000	1.000	1.000	1.000

下载: 导出CSV

表 4 在ISPD 2017 benchmark上的布局时间 (s) 、布线时间 (min)以及布线线长 (×10⁴ )的比较

FPGA设计	RippleFPGA (CPU)			OpenPARF (CPU)			OpenPARF (GPU)
FPGA设计	布局时间	布线时间	布线线长	布局时间	布线时间	布线线长	布局时间	布线时间	布线线长
CLK-FPGA01	277.78	10	238.54	864	9	205.79	130.97	10	205.44
CLK-FPGA02	249.99	15	261.85	782	13	247.73	126.83	14	246.65
CLK-FPGA03	537.36	24	648.69	963	26	593.15	205.98	24	594.00
CLK-FPGA04	346.45	18	440.09	860	19	419.74	156.97	19	420.30
CLK-FPGA05	501.15	25	560.18	962	23	510.30	201.01	23	510.62
CLK-FPGA06	545.09	28	678.43	988	26	617.94	217.75	28	617.28
CLK-FPGA07	288.26	13	276.29	795	13	256.62	136.39	13	256.62
CLK-FPGA08	234.69	10	213.06	672	9	196.63	119.32	10	196.67
CLK-FPGA09	311.68	13	297.02	807	14	251.27	148.38	14	250.98
CLK-FPGA10	464.83	23	544.07	930	25	449.52	194.63	14	451.28
CLK-FPGA11	421.12	24	516.67	897	23	422.01	181.58	30	421.52
CLK-FPGA12	377.65	18	403.59	862	19	335.13	167.02	20	336.03
CLK-FPGA13	393.25	21	464.78	880	20	427.86	177.86	19	428.41
平均值	2.251	1.037	1.125	5.305	1.036	1.000	1.000	1.000	1.000

下载: 导出CSV

参考文献(41)

[1]	MARKOV I L, HU Jin, and KIM M C. Progress and challenges in VLSI placement research[J]. Proceedings of the IEEE, 2015, 103(11): 1985–2003. doi: 10.1109/JPROC.2015.2478963
[2]	CHEN Deming, CONG J, and PAN Peichan. FPGA design automation: A survey[M]. BOX P O. Foundations and Trends in Electronic Design Automation. Hanover: Now Publishers Inc. , 2006.
[3]	MURRAY K E, WHITTY S, LIU Suya, et al. Timing-driven Titan: Enabling large benchmarks and exploring the gap between academic and commercial CAD[J]. ACM Transactions on Reconfigurable Technology and Systems, 2015, 8(2): 1–18. doi: 10.1145/2629579
[4]	MURRAY K E, PETELIN O, ZHONG Suya, et al. VTR 8: High-performance CAD and customizable FPGA architecture modelling[J]. ACM Transactions on Reconfigurable Technology and Systems, 2020, 13(2): 1–55. doi: 10.1145/3388617
[5]	UltraScale architecture configurable logic block user guide (UG574)[EB/OL].https://china.xilinx.com/content/dam/xilinx/support/documents/user_guides/ug574-ultrascale-clb.pdf, 2017.
[6]	YANG S, GAYASEN A, MULPURI C, et al. Routability-driven FPGA placement contest[C]. The 2016 on International Symposium on Physical Design, Santa Rosa, USA, 2016: 139–143.
[7]	ZHANG Niansong, CHEN Xiang, and KAPRE N. RapidLayout: Fast hard block placement of FPGA-optimized systolic arrays using evolutionary algorithm[J]. ACM Transactions on Reconfigurable Technology and Systems, 2022, 15(4): 38. doi: 10.1145/3501803
[8]	ZHOU Yun, MAIDEE P, LAVIN C, et al. RWRoute: An open-source timing-driven router for commercial FPGAs[J]. ACM Transactions on Reconfigurable Technology and Systems, 2022, 15(1): 8. doi: 10.1145/3491236
[9]	YANG S, MULPURI C, REDDY S, et al. Clock-aware FPGA placement contest[C]. The 2017 ACM on International Symposium on Physical Design, Portland, USA, 2017: 159–164.
[10]	MARTIN T, BARNES C, AREIBI S, et al. An adaptive sequential decision making flow for FPGAs using machine learning[C]. 2022 International Conference on Microelectronics (ICM), Casablanca, Morocco, 2022: 34–37.
[11]	LIANG Tingyuan, CHEN Gengjie, ZHAO Jieru, et al. AMF-placer: High-performance analytical mixed-size placer for FPGA[C]. 2021 IEEE/ACM International Conference On Computer Aided Design (ICCAD), Munich, Germany, 2021: 1–9.
[12]	CHEN T C, JIANG Zhewei, HSU T C, et al. NTUplace3: An analytical placer for large-scale mixed-size designs with preplaced blocks and density constraints[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2008, 27(7): 1228–1240. doi: 10.1109/TCAD.2008.923063
[13]	LU Jingwei, CHEN Pengwen, CHANG C C, et al. ePlace: Electrostatics based placement using Nesterov’s method[C]. The the 51st Annual Design Automation Conference, San Francisco, USA, 2014: 1–6.
[14]	CHENG C K, KAHNG A B, KANG I, et al. RePlAce: Advancing solution quality and routability validation in global placement[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2019, 38(9): 1717–1730. doi: 10.1109/TCAD.2018.2859220
[15]	LIU Lixin, FU Bangqi, WONG M D F, et al. Xplace: An extremely fast and extensible global placement framework[C]. The 59th ACM/IEEE Design Automation Conference, San Francisco, USA, 2022: 1309–1314.
[16]	AGNESINA A, RAJVANSHI P, YANG Tian, et al. AutoDMP: Automated DREAMPlace-based macro placement[C]. The 2023 International Symposium on Physical Design, Virtual Event, USA, 2023: 149–157.
[17]	MAI Jing, MENG Yibai, DI Zhixiong, et al. Multi-electrostatic FPGA placement considering SLICEL-SLICEM heterogeneity and clock feasibility[C]. The 59th ACM/IEEE Design Automation Conference, San Francisco, USA, 2022: 649–654.
[18]	RAJARATHNAM R S, ALAWIEH M B, JIANG Zixuan, et al. DREAMPlaceFPGA: An open-source analytical placer for large scale heterogeneous FPGAs using deep-learning toolkit[C]. 2022 27th Asia and South Pacific Design Automation Conference (ASP-DAC), Taipei, China, 2022: 300–306.
[19]	RAJARATHNAM R S, JIANG Zixuan, IYER M A, et al. DREAMPlaceFPGA-PL: An open-source GPU-accelerated packer-legalizer for heterogeneous FPGAs[C]. The 2023 International Symposium on Physical Design, Virtual Event, USA, 2023: 175–184.
[20]	MENG Yibai, LI Wuxi, LIN Yibo, et al. elfPlace: Electrostatics-based placement for large-scale heterogeneous FPGAs[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2022, 41(1): 155–168. doi: 10.1109/TCAD.2021.3053191
[21]	KIM M C, HU Jin, LEE D J, et al. A SimPLR method for routability-driven placement[C]. 2011 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), San Jose, USA, 2011: 67–73.
[22]	LI Wuxi, DHAR S, and PAN D Z. UTPlaceF: A routability-driven FPGA placer with physical and congestion aware packing[C]. The 2016 IEEE/ACM International Conference on Computer-Aided Design, Austin, USA, 2016: 1–7.
[23]	CHEN Gengjie, PUI C W, CHOW W K, et al. RippleFPGA: Routability-driven simultaneous packing and placement for modern FPGAs[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2018, 37(10): 2022–2035. doi: 10.1109/TCAD.2017.2778058
[24]	HSU M K, CHOU S, LIN T H, et al. Routability-driven analytical placement for mixed-size circuit designs[C]. 2011 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), San Jose, USA, 2011: 80–84.
[25]	CHEN Jianli, LIN Zhifeng, KUO Y C, et al. Clock-aware placement for large-scale heterogeneous FPGAs[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2020, 39(12): 5042–5055. doi: 10.1109/TCAD.2020.2968892
[26]	PUI C W, CHEN Gengjie, MA Yuzhe, et al. Clock-aware ultrascale FPGA placement with machine learning routability prediction: (Invited paper)[C]. 2017 IEEE/ACM International Conference on Computer-Aided Design (ICCAD), Irvine, USA, 2017: 929–936.
[27]	LI Wuxi, LIN Yibo, LI Meng, et al. UTPlaceF 2.0: A high-performance clock-aware FPGA placement engine[J]. ACM Transactions on Design Automation of Electronic Systems, 2018, 23(4): 42. doi: 10.1145/3174849
[28]	LI Wuxi, DEHKORDI M E, YANG S, et al. Simultaneous placement and clock tree construction for modern FPGAs[C]. The 2019 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays, Seaside, USA, 2019: 132–141.
[29]	ZHU Ziran, MEI Yangjie, LI Zijun, et al. High-performance placement for large-scale heterogeneous FPGAs with clock constraints[C]. The 59th ACM/IEEE Design Automation Conference, San Francisco, USA, 2022: 643–648.
[30]	FENG Wenyi. K-way partitioning based packing for FPGA logic blocks without input bandwidth constraint[C]. 2012 International Conference on Field-Programmable Technology, Seoul, Korea (South), 2012: 8–15.
[31]	BETZ V and ROSE J. VPR: A new packing, placement and routing tool for FPGA research[M]. LUK W, CHEUNG P Y K and GLESNER M. Field-Programmable Logic and Applications. Berlin, Heidelberg: Springer, 1997: 213–222.
[32]	LI Wuxi and PAN D Z. A new paradigm for FPGA placement without explicit packing[J]. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2019, 38(11): 2113–2126. doi: 10.1109/TCAD.2018.2877017
[33]	DI Zhixiong, TAO Runzhe, CHEN Lin, et al. Imbalanced large graph learning framework for FPGA logic elements packing prediction[EB/OL]. Available: http: //arxiv. org/abs/2308. 03231, 2023.
[34]	MCMURCHIE L and EBELING C. PathFinder: A negotiation-based performance-driven router for FPGAs[C]. Third International ACM Symposium on Field-Programmable Gate Arrays, Napa Valley, USA, 1995: 111–117.
[35]	MURRAY K E, ZHONG Sheng, and BETZ V. AIR: A fast but lazy timing-driven FPGA router[C]. 2020 25th Asia and South Pacific Design Automation Conference (ASP-DAC), Beijing, China, 2020: 338–344.
[36]	ZHOU Yun, VERCRUYCE D, and STROOBANDT D. Accelerating FPGA routing through algorithmic enhancements and connection-aware parallelization[J]. ACM Transactions on Reconfigurable Technology and Systems, 2020, 13(4): 1–26. doi: 10.1145/3406959
[37]	ZHA Yue and LI Jing. Revisiting pathfinder routing algorithm[C]. Proceedings of the 2022 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays, Virtual Event, USA, 2022: 24–34.
[38]	SHEN Minghua and LUO Guojie. Corolla: GPU-accelerated FPGA routing based on subgraph dynamic expansion[C]. The 2017 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays, Monterey, USA, 2017: 105–114.
[39]	WANG Jiarui, MAI Jing, DI Zhixiong, et al. A Robust FPGA router with concurrent intra-CLB rerouting[C]. The 28th Asia and South Pacific Design Automation Conference, Tokyo, Japan, 2023: 529–534.
[40]	RAY B N B, TRIPATHY A R, SAMAL P, et al. Half-perimeter wirelength model for VLSI analytical placement[C]. 2014 International Conference on Information Technology, Bhubaneswar, India, 2014: 287–292.
[41]	SPINDLER P and JOHANNES F M. Fast and accurate routing demand estimation for efficient routability-driven placement[C]. 2007 Design Automation & Test in Europe Conference & Exhibition, Nice, France, 2007: 1–6.