Efficient High-Level Synthesis of Approximate Computing Circuits via Multi-fidelity Modeling

High-Level Synthesis (HLS) of approximate computing circuits generates circuits based on functional units such as adders and multipliers. An approximate library containing approximate functional units is firstly built, and then an algorithm is employed to search in the library, solving the binding problem of HLS and finally generating an approximate circuit that meets design specifications. In this paper, we propose a discrete-input and multi-fidelity Bayesian optimization (BO) approach to solve this design space exploration problem. Since the input of the algorithm is discrete functional units in the approximate library, we modify the kernel function of Gaussian Process (GP) model in BO. In addition, we adopt a multi-fidelity strategy to speed up the optimization process, which is proved to be effective and efficient according to our experimental results.