Automatic implementation of arithmetic functions in high-level synthesis

Implementation of floating-point arithmetic functions is an essential task in high-level synthesis (HLS). However, most of the existing HLS EDA tools cannot well synthesize floating-point functions. In this paper, we present an automatic method to implement general floating-point functions using piecewise polynomial approximation. Based on the proposed hardware architecture, our method achieves compact implementations, which have low delay and high throughput. Experiment shows that compared with existing methods, the proposed method reduces register usage by about 10%.