K. Grace Johnson, Stanford University
As computational power increases, quantum chemistry becomes an increasingly important tool for understanding the behavior of known chemical systems as well as for chemical discovery and design. The modern trend of increased computational power, however, also is one of increased heterogeneity in compute architectures. This presents new challenges for the quantum chemistry community, as code will need to be adapted to new architectures and scaled to larger systems. Evaluating two-electron repulsion integrals (ERIs) is the most computationally intensive step for a popular family of quantum chemistry methods‚ Hartree-Fock (HF) and density functional theory (DFT). ERIs scale formally as N^4, where N is the number of atom-centered Gaussian basis functions used to represent the chemical system. The contributions of ERIs to the Fock matrix are of Coulomb (J) and exchange (K) type, and require separate algorithms to compute matrix elements efficiently. We previously implemented highly efficient GPU-accelerated J-matrix and K-matrix algorithms in the electronic structure code TeraChem. However, these implementations do not support multi-node architectures, which are key to enabling cutting-edge ab initio simulations of large systems, e.g. excited-state dynamics of photoactive proteins. We present our implementation of multi-node, multi-GPU J and K-matrix algorithms in TeraChem using the Regent programming language. Regent uses a task-based model with logical regions for high-performance computing and can generate code that can run on a variety of architectures, including NVIDIA GPUs and multiple nodes. We demonstrate multi-node scaling and benchmark against the hand-coded TeraChem integral code, as well as highlight the ability to metaprogram the Regent code to generate routines for arbitrary angular momentum integrals.
Abstract Author(s): K. Grace Johnson, Seema Mirchandaney, Ellis Hoag, Alex Aiken, Todd J. Martinez