VIPP: Validation-Included Precision-Parametric N-Body Benchmark Suite

Abstract

Many efforts have recently been made to analyze and validate floating-point errors, particularly in mixed-precision arithmetic. However, real-world applications in approximate computing typically incorporate both model-level approximation and arithmetic-level precision. It is crucial to analyze the combined effects of both precision parameters to the extent valid in terms of approximate algorithms. In this work, we develop a benchmark suite of the practical approximate solvers of various N-body problems that parameterize both N-body approximation and arithmetic precision. It involves precision criteria to prevent us from unrestricted reduced precision and serves as a testbed to analyze the combined effects of model-level approximation and arithmetic-level reduced precision. It would help the design of precision control in approximate computing.