Synthesizing Efficient Low-Precision Kernels

In this paper, we present a fully automated approach for synthesizing fast numerical kernels with guaranteed error bounds. The kernels we target contain elementary functions such as sine and logarithm, which are widely used in scientific computing, embedded as well as machine-learning programs. However, standard library implementations of these functions are often overly accurate and therefore unnecessarily expensive. Our approach trades superfluous accuracy against performance by approximating elementary function calls by polynomials and by implementing arithmetic operations in low-precision fixed-point arithmetic. Our algorithm soundly distributes and guarantees an overall error budget specified by the user. The evaluation on benchmarks from different domains shows significant performance improvements of 2.23 $$\times $$ on average compared to state-of-the-art implementations of such kernel functions.