GPU acceleration of overbridging boundary matching method without Green's functions based on real-space finite-difference method

We present the graphics processing unit (GPU) acceleration of the overbridging boundary matching method for electron-transport property calculations, which is based on the density functional theory using the real-space finite-difference method. The execution of the implemented code using OpenACC and CUDA libraries on GPU is computationally more efficient and faster than a central processing unit. Furthermore, we achieve the ideal scalability in parallel execution from one to thirty-two nodes by adopting multiprocess parallelization schemes for two types of supercomputer with different configurations. To demonstrate the applicability of the accelerated code, the complex band structures of graphene and armchair carbon nanotubes with chiral indices (6, 6), (9, 9), and (12, 12) are calculated and compared with those obtained by the tight-binding method. We discuss the effects of the dispersion of evanescent waves on roll-up and axial strain in the carbon nanotubes.