Automated Code Engine for Graphical Processing Units: Application to the Effective Core Potential Integrals and...

Area of Science:

• Computational Chemistry
• High-Performance Computing
• Electronic Structure Theory

Background:

• Electronic structure calculations require efficient computation of integrals.
• Graphical Processing Units (GPUs) offer significant computational power but require specialized optimization.
• Automated code generation can address the complexity of optimizing for diverse GPU architectures.

Purpose of the Study:

• To develop an automated code engine (ACE) for generating optimized computation kernels for electronic structure theory on GPUs.
• To enable the automatic selection of optimal code variants and GPU configurations for integral computations.
• To create a flexible code generation framework adaptable to various integral types.

Main Methods:

• Developed an automated code engine (ACE) utilizing a graph representation for code generation.
• Implemented a code generator creating multiple variants with varying memory and floating-point operation trade-offs.
• Integrated a code optimizer that scans and selects the best-performing code candidate for specific GPU platforms.

Main Results:

• ACE successfully generates optimized kernels for effective core potential integrals and gradients.
• The optimal code variant is dependent on factors like angular momentum, floating-point precision, and GPU type.
• Demonstrated that the best code candidate varies significantly across different GPUs and computational parameters.

Conclusions:

• ACE is a powerful tool for adapting electronic structure calculations to rapidly evolving GPU architectures.
• Automated optimization is crucial for maximizing computational efficiency in quantum chemistry.
• The engine's flexibility allows for extension to diverse integral types and future hardware advancements.