Experience in porting Fortran code to GPUs with OpenACC
Speakers: Mike Zingale and Adam Jacobs
This talk was given at the 9-23-15 OLCF User Group Conference Call.
In the slides and video below, we describe our efforts to put large physics modules written in Fortran onto the GPUs. Our main target is integrating stiff systems of ODEs describing nuclear reactions. Our strategy is to put the entire integration: righthand side evaluations, Jacobian construction, and the actual timestepping onto the GPUs. This minimizes the communication costs — only the initial state vector is passed in, and the new state vector comes out at the end. We will discuss the challenges we had in getting the compilers to accept modern Fortran constructions with OpenACC and discuss preliminary results.