LibERI-A portable and performant multi-GPU accelerated library for electron repulsion integrals via OpenMP offloading...

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• High-Performance Computing

Background:

• Electron repulsion integrals (ERIs) are computationally intensive in quantum chemistry.
• Existing methods for ERI evaluation face challenges with large molecular systems and GPU acceleration.

Purpose of the Study:

• To develop a portable and performant GPU-accelerated library for ERI evaluation named LibERI.
• To enhance computational efficacy for large molecular systems through optimized GPU offloading and data handling.

Main Methods:

• Implementation of LibERI using directive-based (OpenMP, OpenACC) and standard language parallelism (Fortran DO CONCURRENT).
• Utilized rotated-axis and Rys quadrature methods for integrals over s, p, and d functions.
• Factorized GPU codes with integral screening and quartet presorting, moving density screening to the GPU.
• Separated L-shells in Pople basis sets for improved ERI homogeneity and reduced memory footprint.

Main Results:

• LibERI demonstrates comparable performance to state-of-the-art GPU-accelerated codes like TeraChem and GMSHPC.
• Outperforms conventional ERI CUDA kernels (e.g., QUICK) in certain cases while maintaining portability.
• Successful integration with GAMESS software package for benchmark calculations across various GPU architectures and molecular systems.

Conclusions:

• LibERI provides a portable and efficient GPU-accelerated solution for ERI evaluation.
• The developed methods enhance computational efficacy for large-scale quantum chemistry calculations.
• LibERI is compatible with quantum chemistry drivers supporting the MolSSI Driver Interface.