Evaluating the Impact of Proposed OpenMP 5.0 Features on Performance, Portability and Productivity

We investigate how specialization mechanisms proposed for OpenMP 5.0 -- specifically, the metadirective and declare variant directives -- may be deployed in a real-life code, using the miniMD benchmark from the Mantevo suite. Additionally, we develop an OpenMP 4.5 implementation of miniMD that achieves a performance portability of 59.35% across contemporary CPU and GPU hardware, discuss the processes of porting and enabling this code, and show that the use of specialization would enable our code to be expressed in a significantly more compact form, with implications for productivity.