Enabling Fortran Standard Parallelism in GAMESS for Accelerated Quantum Chemistry Calculations

Area of Science:

• High-performance computing
• Computational chemistry
• Parallel programming

Background:

• The Fock build is a critical computational bottleneck in quantum chemistry.
• Efficiently offloading computations to Graphics Processing Units (GPUs) is crucial for performance.
• Existing offloading models like OpenACC and OpenMP target offloading (OTO) have limitations.

Purpose of the Study:

• To evaluate the performance of Fortran 2008 DO CONCURRENT (DC) against OTO for GPU offloading.
• To assess the impact of different compilers on DC and OTO performance.
• To identify optimal programming models for accelerating quantum chemistry applications on GPUs.

Main Methods:

• Implemented DC and OTO for the Fock build in the GAMESS quantum chemistry application.
• Tested performance on NVIDIA A100 and V100 GPUs.
• Compared DC performance with OTO versions compiled using NVIDIA HPC, IBM XL, and Cray Fortran compilers.

Main Results:

• Fortran 2008 DO CONCURRENT achieved a 3.0x speedup for the Fock build compared to the OTO model.
• DC demonstrated competitive or superior performance across different compilers and GPU architectures.
• The study confirmed DC's efficacy in accelerating computationally intensive tasks.

Conclusions:

• Fortran 2008 DO CONCURRENT is a compelling and efficient programming model for offloading Fortran applications to GPUs.
• DC offers significant performance advantages over OTO for critical computational kernels like the Fock build.
• Further adoption of DC can enhance the performance of quantum chemistry software.