An Iterative Concave-Convex Wirelength Model for Analytical Placement

Abstract

Analytical placer very often uses half-perimeter wirelength (HPWL) model as an objective function for solving placement problem. State-of-the-art iterative wirelength models for HPWL are log-sum-exp (LSE) [1], weighted average (WA) [2], (γ, q) [3] and CHMAX [4] wirelength models. In this paper, we propose an iterative wirelength model for HPWL providing smooth approximation to max function. The new max function is the difference of two convex functions. The error upper bound of proposed model is tighter than upper bound errors of existing iterative wirelength models. Integration of new model in analytical placement engine NTUplacer, reduces the final placement wirelength by 13%, 10%, 3% and 1.5% on an average on ISPD 2005 benchmark circuits as oppose to LSE, WA, (γ, q) and CHMAX models respectively.