Logic optimization using rule-based randomized search

In this paper we describe a new logic synthesis approach based on rule-based randomized search using simulated annealing. Our work is motivated by two observations: (1) traditional logic synthesis applies literal count as the primary quality metric during the technology independent optimization phase. This simplistic metric often leads to poor circuit structures as it cannot foresee the impact of early choices on the final area, delay, power consumption, etc. (2) Although powerful, global Boolean optimization is not robust and corresponding algorithms cannot be used in practice without artificially restricting the application window. Other techniques, such as algebraic methods scale well but provide weaker optimization power. To address both problems, we use randomized search that is based on a simple circuit graph representation and a complete set of local transformations that include algebraic and Boolean optimization steps. The objective of the search process can be tuned to complex cost functions, combining area, timing, mutability, and power. Our experimental results on benchmark functions demonstrate the significant potential of the presented approach.