Electron repulsion integral evaluation over f-type functions on GPUs via OpenMP offloading

Abstract

Evaluation of electron repulsion integrals over f-type basis functions on graphics processing units (GPUs) in the LibERI library is presented. The GPU architecture is targeted via the OpenMP programming model, re-utilizing pre-existing central processing unit-oriented code as the backbone for automatic GPU kernel generation. The electron repulsion integral calculator and Rys quadrature methods are utilized to accommodate different degrees of contraction and angular momenta. Upon comparison against GPU-native compute unified device architecture programs, the implementation presented here is observed to be competitive with state-of-the-art software for small to medium-sized systems. Multi-GPU performance is also presented. Parallel efficiencies higher than 80% are observed for large systems on 4 GPUs (i.e., a full node), and the multi-node performance of LibERI is demonstrated for up to 128 GPUs (32 nodes). The limitations of the implementation presented here are also reported.