超软赝势密度泛函分子动力学计算中的若干优化算法

Several Optimized Algorithms for Ultra-soft Psedupotential DFT Molecular Dynamics Simulation

基于密度泛函理论的第一性原理计算是科学计算中重要的领域,被广泛用于材料计算和量子化学等研究。我们在PEtot软件的基础上研发了GPU加速软件PWMat,本文主要介绍:1)针对超软赝势密度泛函算法部分的GPU算法优化与实现,取得了2-3倍的加速;2)针对超软赝势密度泛函分子动力学计算部分的若干算法优化,包括原子受力求解算法的修正、电荷密度和波函数的插值算法的改进等。其中原子受力的收敛精度获得了1-2个数量级的提高,电荷密度的收敛精度获得了2-4个数量级的提高。

DFT First principle computation is one important field in scientific computation. It has been widely used in material computation, quantum chemistry and so on. We developed the GPU software PWMat based on PEtot. This paper is focused on the following： 1） Using GPU to speed up and optimize the codes of ultra-soft DFT in PWMat, and a 2-3x speedup is achieved. 2） Optimizing the algorithm of MD part in PWMat. Through correcting the force calculation and changing the way of charge density and wave function interpolation, a speedup of 1-2 orders is achieved for the atom force calculation and a speedup of 2-4 orders is achieved for the charge density calculation.1. Computer Network Information Center of Chinese Academy of Sciences, Beijing 100190, China 2. Lawrence Berkeley National Laboratory, USA