Abstract

The shift towards parallel hardware is well known, but it is currently much less clear what the corresponding changes in programming languages and compilers should be. In particular, as GPUs and other heterogeneous architectures become more popular, how can we continue to offer developers performance portability? That is, enable a single program to achieve good (if not optimal) performance on any system.

In this talk, I discuss how this was achieved for sequential systems, why parallelism offers new challenges, and how these might be overcome. Rather than proposing yet another programming language, our solution proposes a common architecture-neutral abstract machine, that can be used as a compiler intermediate representation. Our ongoing research concerns compiler techniques that map this representation to hardware effectively.