T58: Extreme-Scale Computational Science Discovery in Fluid Dynamics and Related Disciplines I

Thu. March 7, 12:30 p.m. – 12:42 p.m. CST

205D

Portability is evermore salient as the leadership-class computing landscape becomes increasingly heterogeneous. OpenACC has stood out as a performant tool for offloading computations to accelerators, especially for Fortran codebases. Now, new hardware stresses the realized portability of tools like OpenMP and OpenACC for current compilers and infrastructure. We study the improvements and limitations of current compilers for bug-free and performant code generation from OpenACC 2.7 compliant statements in a Fortran codebase. The application is the Multi-component Flow Code (MFC, https://mflowcode.github.io); a proven-performant solver for compressible multiphase flows on NVIDIA-based leadership-class systems, including OLCF Summit, at scale. We discuss past, present, and likely future roadblocks and workarounds for current Cray and GNU compilers offloading to AMD 200-series GPU hardware. A Python metaprogramming tool, fypp, is used for Fortran code generation before compile-time to ease otherwise cumbersome workarounds. Performance studies are conducted where possible.