摘要
本文提出一种新型的基于深度学习的FPGA快速布局算法,将FPGA布局转化为动态的进行逻辑单元块的选择和逻辑单元块位置确定的过程,从而实现电路网表在FPGA上的逐步布局.其中每一个逻辑单元块的位置确定由训练好的深度学习网络预测实现,所有逻辑单元块位置确定之后采用基于交换的快速详细布局算法进行优化.实验中使用MCNC基准电路进行测试,将测试结果与VPR中基于模拟退火的布局算法进行对比,结果表明:在关键路径延时平均9.8%布线后的损失代价下,整个布局过程的运行速度平均提升了24.54倍,其中处理十万量级大规模电路实现64.9倍的速度提升.
This paper proposes a new deep learning based FPGA placement algorithm,which implements the FPGA placement flow through a dynamic sequential process of selecting the logical element and determining the position of the logic element,thus achieving FPGA placement result of the netlist step-by-step.The choosen logic element block every time is placed on where the well-trained deep learning network predicts,and after the above process is completed,a fast detailed placement is used to optimize the previous placement result.Experiment is carried out to test the performance of the algorithm in MCNC benchmarks.For comparison,we use the simulated annealing based VPR to run these circuits.The result shows that under 9.8%increase of the critical path delay after routing,the entire runtime speed up is 24.54 times on average and for circuit netlist with about a hundred thousand exchangeable logical element blocks it rises to 64.9 times.
作者
刘伟
王伶俐
周灏
LIU Wei;WANG Lingli;ZHOU Hao(State Key Laborlatory of ASIC&Systems,Fudan University,Shanghai 201203,China)
出处
《复旦学报（自然科学版）》
CAS
CSCD
北大核心
2019年第6期687-695,共9页
Journal of Fudan University：Natural Science
