Wire length prediction in constraint driven placement

Abstract

Experiments show that lengths of individual wires are different for different placement algorithms. To achieve accurate wire length prediction, some knowledge of a placer's details is necessary. We postulate that wire length prediction should be coupled with placement flow to obtain accurate results. In this paper, we embed wire length prediction into our constraint-driven placer, developed in Fast Placer Implementation (FPI) framework [15]. We predict individual wire lengths during the clustering step. The predicted wire lengths act as constraints for the simulated annealing refinement stage, which guides the placement towards a solution fulfilling the predictions. Experimental results show that our wire length prediction process yields accurate results without quality loss at a small cost of placement effort. This is the first time that constraints have been used to guide placement and thus increase the accuracy of wire length prediction.