Development of novel, highly scalable, performant and portable multi-GPU quantum chemistry algorithms is crucial for solving chemistry problems on modern and future hardware architectures. Many computational software packages allow for studying reaction mechanisms, protein-ligand binding interactions, chemistry of excited state molecules and many more. However, due to limitations in computational power, very accurate quantum chemistry calculations are often limited to a small number of atoms/molecules. In addition, most modern quantum chemistry software packages are still optimized to run efficiently on CPUs. With the emergence of new hardware architectures, specifically heterogeneous CPU-GPU architectures, it is important to develop novel and efficient algorithms to fully utilize the computational resources.<br/>    This work focuses on development of new multi-GPU Hartree-Fock (HF) codes in GAMESS via two different programming models. Our approach consists of a new implementation of the HF code in a C++/CUDA library, which is the high-performance engine in GAMESS, as well as offloading of the older legacy Fortran Hartree-Fock code via OpenMP. We tackle the bottlenecks of quantum chemistry methods, specifically the evaluation of the two-electron integrals (2EIs) and their digestion into the Fock matrix on GPUs for both codes. The new Fock builds exhibit excellent performance compared to the parallel hybrid MPI/OpenMP CPU code in GAMESS as well as near ideal scaling on up to 6 V100 GPUs on Summit. For the CUDA implementation, we additionally offloaded the one-electron integrals and diagonalization step to GPUs, therefore running all the stages of the Hartree-Fock method on GPUs. GAMESS users will be able to run the highly performant CUDA code on NVIDIA GPUs and use the newly developed OpenMP code as a portable solution to run accelerated calculations on other GPU architectures.<br/>    The new GPU accelerated codes were used as the basis to build the correlated RI-MP2 calculations to elucidate the mechanism of the Pd-catalyzed carboacylation reaction of esters, which goes through a catalytic cycle with multiple elementary steps. The potential energy surface of this reaction was found to be somewhat flat, therefore, necessitating the use of very accurate as well as fast quantum chemical methods such as RI-MP2 built upon the new Hartree-Fock codes.

Development of highly scalable multi-GPU quantum chemistry codes in GAMESS and applications to homogeneous catalysis

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.