Acceleration of the Relativistic Dirac-Kohn-Sham Method with GPU: A Pre-Exascale Implementation of BERTHA and...

Area of Science:

• Computational Physics
• Quantum Chemistry
• High-Performance Computing

Background:

• The Dirac-Kohn-Sham (DKS) method is crucial for relativistic electronic structure calculations.
• Efficient computation of DKS is essential for tackling complex molecular systems.
• Porting computational codes to modern hardware like GPUs is vital for performance gains.

Purpose of the Study:

• To present advances in the computation of the DKS method within the BERTHA code.
• To demonstrate efficient porting of the FORTRAN code to Graphics Processing Units (GPUs).
• To accelerate DKS and real-time DKS (RT-TDDKS) calculations through hybrid CPU/GPU implementation.

Main Methods:

• Utilizing a hybrid OpenMP/OpenACC programming model for CPU/GPU implementation.
• Offloading computationally intensive DKS matrix evaluation (three-center two-electron integrals) to GPUs via OpenACC.
• Employing GPU-optimized linear algebra libraries (cuBLAS, cuSOLVER).
• Porting a Python-based RT-TDDKS implementation (PyBERTHART) using a FORTRAN integral kernel.

Main Results:

• Achieved significant speedups: up to 30x for single-point DKS energy calculations (Au16) and up to 10x for RT-TDDKS calculations (Au8).
• Demonstrated a highly efficient hybrid CPU/GPU implementation leveraging FORTRAN and OpenACC.
• Successfully ported a Python API to GPUs based on a FORTRAN kernel for two-electron integral evaluation.

Conclusions:

• The presented hybrid CPU/GPU approach offers substantial acceleration for DKS calculations.
• This work represents a novel porting of a Python API to GPUs using a FORTRAN kernel for integral evaluation.
• The methodology is general and paves the way for future exascale implementations.