OLCF User Group Call: Hackathon User Expereince OpenAcc with GCC6

User Driven Content

This talk was originally given on 12-17-15 at OLCF User Group Call.

User Philip LoCascio speaks about his experience with testing the GCC5 compiler and OpenACC at the last OpenACC Hackathon.

He states, “I tested OpenACC with the GCC5/6 compiler. There are a number of codes that have yet to be GPU optimized and GCC5/6 appeared a widely deployed platform and development route for other applications. Despite the compiler being available for a few months, there were not any good examples using it due to the build dependencies. Therefore, I wrote an auto build script, suitable for testing the compiler on non-root systems e.g. Supercomputers using an example OpenACC code that calculated pi as an objective test.

I added the OpenACC code to deploy on a GPU in Fortran, testing both SM_30 and SM_50 architectures, vector lengths and compared this with the same problem implemented in OpenMP with 32 CPU threads. OpenACC looks as if it may well be powerful enough to augment CUDA in some apps (e.g. GROMACS), and maybe replace it entirely in others (e.g. BLAST). Full scale application deployment has not yet been tested with GCC5/6.”

Video of Philip’s talk: https://vimeo.com/149295776

Slides: OLCF-GCC5-hackathon-December