Mixed-precision evaluation of two-electron integrals by Rys quadrature

Abstract

A mixed precision implementation of two-electron integrals is demonstrated to have two benefits: (a) computations can be performed reliably in 32-bit precision on architectures for which 32-bit precision is significantly faster than 64-bit precision (e.g. graphical processing units), and (b) numerical results that match those using higher than 64-bit precision can be recovered without a significant penalty associated with performing the entire computation in higher precision. A justification is presented for using mixed precision in the Rys two-electron integral quadrature algorithm, together with timings and numerical results using a variety of floating-point types. The code discussed here presents a systematic way to control the accuracy of the Rys algorithm, regardless of the types and numbers of integrals.