Generating Coupled Cluster Code for Modern Distributed-Memory Tensor Software.

Area of Science:

• Computational Chemistry
• High-Performance Computing (HPC)
• Quantum Chemistry

Background:

• Efficient execution of coupled cluster (CC) computations on GPU-based High-Performance Computing (HPC) platforms is hindered by heterogeneous hardware structures.
• Adapting software to these structures demands significant man-hours, necessitating systematized high-performance code development, especially for higher-order CC methods.

Purpose of the Study:

• To address the challenges of efficient tensor symmetry capture and hardware abstraction in developing general-order coupled cluster (CC) code generators.
• To present the design of a novel, open-source modular tensor framework, "tenpi", for CC code development.

Main Methods:

• Development of a high-level problem representation translated into low-level hardware instructions via a compiler/translator.
• Designing software to capture key tensor symmetries while maintaining hardware abstraction.
• Integration of diagrammatic derivation, visualization, symbolic algebra, and intermediate optimization within the "tenpi" framework.

Main Results:

• Systematically generated code demonstrates excellent weak scaling on up to 1200 GPUs using the ExaTENSOR distributed memory tensor library.
• The "tenpi" framework supports multiple tensor backends and facilitates higher-order CC functionality.
• Integration of "tenpi" into the DIRAC code's ExaCorr module enhances relativistic molecular calculations.

Conclusions:

• The developed general-order CC code generator and the "tenpi" framework significantly improve the efficiency and accessibility of CC computations on modern HPC platforms.
• "tenpi" provides a robust, modular solution for advanced quantum chemistry calculations, enabling higher-order CC methods on massively parallel systems.