Developing performance-portable molecular dynamics kernels in Open CL

This paper investigates the development of a molecular dynamics code that is highly portable between architectures. Using OpenCL, we develop an implementation of Sandia’s miniMD benchmark that achieves good levels of performance across a wide range of hardware: CPUs, discrete GPUs and integrated GPUs. We demonstrate that the performance bottlenecks of miniMD’s short-range force calculation kernel are the same across these architectures, and detail a number of platform- agnostic optimisations that improve its performance by at least 2x on all hardware considered. Our complete code is shown to be 1.7x faster than the original miniMD, and at most 2x slower than implementations individually hand-tuned for a specific architecture