GPU acceleration of a canonical molecular orbital calculation program by the third-generation density-functional-theory-based method
Toshiyuki Hirano, Fumitoshi Sato
The University of Tokyo, Institute of Industrial Science,
4-6-1, Komaba, Meguro-ku, Tokyo, 153-8505, Japan
We have developed a canonical molecular orbitals (CMOs) calculation program for large-scale molecules such as proteins, ProteinDF. In order to perform massively parallel computing based on the MPI/OpenMP hybrid parallelization in the current distributed-memory machines, we had studied the third-generation density functional calculation method (3G method) [1] and implemented it to the ProteinDF. In the 3G method, the time-consuming molecular integrals are evaluated before the SCF loop, and the Kohn-Sham (or Fock) matrix is estimated only by matrix operations. Matrix operations are easier to equalize computing tasks than molecular integrals in the distributed-memory type parallel computer. Furthermore, it is expected that the whole CMOs calculation will be time-shortened by accelerating the matrix operation.
In this study, implementation of the linear algebra routines with the use of the ViennaCL [2], which supports OpenCL, was carried out. By using OpenCL, it was possible to use not only nVidia’s GPUs but also AMD’s. The 3G method provided an easy way to implement the GPU computing on the large-scale CMO calculation. And, the fast matrix multiplication with the GPU had a beneficial effect on the 3G method.
The ProteinDF has been distributed under the GPL v3 via the Internet [3]. Here we describe development status of the ProteinDF in detail, and we show some CMO calculation results by using the application.
[1] T. Hirano, F. Sato, Phys. Chem. Chem. Phys., 16, 14496 (2014).
[2] K. Rupp, Ph. Tillet, F. Rudolf, J. Weinbub, A. Morhammer, T. Grasser, A. Jüngel, S. Selberherr, SIAM J. Sci. Comput., 38, S412 (2016).
[3] https://proteindf.github.io/