Hybrid programming-model strategies for GPU offloading of electronic structure calculation kernels.

Area of Science:

• Computational Physics
• Materials Science
• Quantum Chemistry

Background:

• Electronic structure calculations are computationally intensive.
• Achieving performance portability across diverse hardware (CPUs, GPUs) is a significant challenge.
• Existing libraries may lack unified interfaces or broad architectural support.

Purpose of the Study:

• To develop and present the Basic Matrix Library (BML) and the Parallel, Rapid O(N), and Graph-based Recursive Electronic Structure Solver (PROGRESS) library.
• To enable efficient and portable implementation of electronic structure solvers.
• To address performance challenges in computational materials science and quantum chemistry.

Main Methods:

• Developed the Basic Matrix Library (BML) for unified linear algebra operations on dense and sparse matrices across CPUs and GPUs.
• Implemented the PROGRESS library, leveraging BML, for electronic structure solvers, focusing on density functional theory and tight-binding models.
• Utilized OpenMP target functionalities and third-party libraries for GPU acceleration and performance-critical numerical kernels.

Main Results:

• Demonstrated the performance portability of the BML and PROGRESS libraries across different computer architectures.
• Showcased efficient computation of the single-particle density matrix using the developed libraries.
• Validated the approach through benchmark problems, confirming its effectiveness.

Conclusions:

• The BML and PROGRESS libraries offer a unified and portable solution for electronic structure calculations.
• The implementation strategy using OpenMP target functionalities enables efficient GPU acceleration.
• These libraries facilitate the development and deployment of advanced computational methods in physics and chemistry.