Abstract:

Lammps is a widely used open source software for molecular dynamics simulations and related issues, through which we can understand solid and liquid nature. It enables the CUDA and the OpenCL to accelerate calculation. Since the OpenCL has crossplatform features, it becomes a research focus. We summarize the OpenCL kernel programming principles and propose an improved Amdahl's law which can measure the acceleration performance on heterogeneous platforms. The shortrange force algorithm in Lammps on Y485P platform is tested. We achieve better acceleration effect by optimizing the kernel source codes of the key parts, such as neighborlist building and shortrange force calculation.

Key words: GPU;Lammps;OpenCL;molecular dynamics;short-range force;kernel optimization