Method And Its Application Of GPU Accelerated LAPW Basis Set First-principles Calculation

First-principles calculation plays an significant role in condensed matter and advanced mate-rials research.The linear augmented plane-wave(LAPW)which can be employed to expand the wave function is suit for first-principles calculation in simulating magnetic or magenetic-optical materials.However,the LAPW computation takes long a time for the complexity of its algorithm and the self-consistent process to solve Kohn-Sham equation.It needs to run on high performance computers or clusters that are expensive,and isn' t able to deal with large scale systems,which has become a bottleneck restricting related research and applications.The general purpose computing on Graphics Processing Unit(GPU)provides a platform for accelerating the LAPW basis first-principles calculation.Compared with the contemporary CPUs,the GPU devices are much better at floating-point performance,parallel processing capability.and memory bandwidth.CUDA,the general purpose computing framework by NVIDIA,help researches develop the GPU high performance computation programs conveniently.In this thesis,we propose the acceleration method of LAPW basis first-principles calculation by using GPU device and CUDA,developing a GPU accelerated computing package as application.Our works are mainly as follows:Firstly,we investigate the related technical backgrounds of LAPW basis function computing and the first-principles calculation using LAPW basis.We also studied the architecture and tech-nological frames of general purpose GPU.Secondly,we make a profound study on the three processes in self-consistent iterations involv-ing LAPW,which have a large amount of calculation.Their mathematical models and numerical algorithms have been derivated in the our thesis.Thirdly,we propose a GPU parallel accelerated method for the three key processes.Most of the calculations in the processes are moved to GPU devices through kernel functions or math libraries,and we use parallel algorithms to improve the performance.Threads and memory usage are optimized to take full advantage of the GPU,majoring in computation and storage resources.The reliability verification and performance profiling are performed on CPU-GPU heteroge-neous architecture,and the results show that the calculation is accurate and reliable.Based on exper-imental conditions,the acceleration ratios up to 4.9 and 3.1 are achieved under different calculating precisions and scales,respectively.It means that the GPU accelerated LAPW basis first-principles calculating package developed in our thesis has practical application significance and application value.