Evolving component library for approximate high level synthesis

An approximate computing approach has recently been introduced for high level circuit synthesis (HLS) in order to make good use of approximate circuits at system and block level. It is assumed in HLS algorithms that a component library containing various implementations of elementary circuit components is available. An open problem is how to construct such a component library in the context of approximate computing, where the component's error is a new design variable and hence many compromise implementations exist for a given component. In this paper, we first introduce a multi-objective Cartesian genetic programming method to create a comprehensive component library containing hundreds of Pareto optimal implementations of approximate 8-bit adders and multipliers, where the error, area and delay are simultaneously optimized. Another multi-objective evolutionary algorithm is employed to solve the so called binding problem of HLS, in which suitable approximate components are assigned to nodes of the data flow graph describing a complex digital circuit. Two approaches are then proposed and compared in order to reduce the size of the library of approximate components. It is shown that a random sub-sampling of the component library provides satisfactory results in the context of our study. The proposed methods are evaluated using two benchmark circuits - the reduce (sum) and DCT circuits.