Automated Search-Based Functional Approximation for Digital Circuits

The problem of developing an approximate implementation of a given combinational circuit can be formulated as a multi-objective design problem and solved by means of a search algorithm. This approach usually provides many solutions showing high-quality tradeoffs between key design objectives; however, it is very computationally expensive. This chapter presents a general-purpose method based on genetic programming for an automated functional approximation of combinational circuits at the gate and register-transfer levels. It surveys relevant error metrics and circuit parameters that are typically optimized by genetic programming. A special attention is given to the techniques capable of providing formal guarantees in terms of error bounds and accelerating the search process. Case studies dealing with approximate implementations of arithmetic circuits and image operators are presented to highlight the quality of results obtained by the search-based functional approximation in completely different application domains.