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

Area of Science:

• Computational Chemistry
• High-Performance Computing

Background:

• Accurate calculation of electron repulsion integrals (ERIs) is crucial for quantum chemistry.
• Efficient computation of ERIs is essential for large molecular systems.
• Leveraging GPU acceleration can significantly speed up computational chemistry tasks.

Purpose of the Study:

• To evaluate the performance of the LibERI library for calculating ERIs over f-type basis functions on GPUs.
• To assess the scalability of the implementation across multiple GPUs and nodes.

Main Methods:

• Utilized OpenMP programming model for GPU targeting.
• Re-used existing CPU-oriented code for automatic GPU kernel generation.
• Employed electron repulsion integral calculator and Rys quadrature methods.
• Compared performance against GPU-native compute unified device architecture (CUDA) programs.

Main Results:

• The LibERI implementation demonstrates competitive performance with state-of-the-art software for small to medium-sized systems.
• Achieved parallel efficiencies exceeding 80% for large systems on 4 GPUs.
• Demonstrated multi-node performance up to 128 GPUs (32 nodes).

Conclusions:

• The presented GPU implementation of LibERI offers a competitive solution for ERI calculations.
• The library shows excellent scalability for large-scale quantum chemistry simulations.
• Limitations of the current implementation are also reported.