A global postsynthesis optimization method for combinational circuits

Abstract

Abstract-A genetic programming-based circuit synthesis method is proposed that enables to globally optimize the number of gates in circuits that have already been synthesized using common methods such as ABC and SIS. The main contribution is a proposal for a new fitness function that enables to significantly reduce the fitness evaluation time in comparison to the state of the art. The fitness function performs optimized equivalence checking using a SAT solver. It is shown that the equivalence checking time can significantly be reduced when knowledge of the parent circuit and its mutated offspring is taken into account. For a cost of a runtime, results of conventional synthesis conducted using SIS and ABC were improved by 20-40% for the LGSynth93 benchmarks.

Cartesian genetic programming as local optimizer of logic networks

“…This method led to a 25% area improvement with respect to the circuits heavily optimized by ABC [10]. As far as we know, no local evolutionary circuit resynthesis/optimization method has been proposed so far.…”

“…It was shown that CGP can design and optimize various digital circuits [9] and improve results of conventional circuit synthesis and optimization tools [10]. In this work, we will utilize the standard CGP for combinational circuit evolution.…”

“…Authors of [10] suggest to utilize knowledge of genes which have been modified by the mutation operator to calculate a 'difference' between the parent individual and its offspring, and to submit only the 'difference' to the SAT solver.…”

“…We used the setup which was identified as the most advantageous during the initial experiments: λ = 1, h = 2, ρ = 2.5, 1 CGP run with g max = 10 6 for a subcircuit, sg min = 10, sg max = 50, si min = 1, and si max = 10. As we prefer the quality of results over the time of optimization and we are interested in limits of the optimization methods, we allowed 12-hour LO runs, which is also a typical setup used for the global optimization in [10], [21]. Each experiment was repeated 5 times.…”

“…As CGP is intended to iteratively optimize small, randomly chosen subcircuits, the scalability problems known from the CGPbased circuit evolution will be eliminated. This work should be understood as a complementary approach to the already existing CGP-based global optimization methods developed for combinational circuits [7], [10].…”