Today, the US Department of Energy announced that it had established a partnership with Nvidia that would be enhancing the LLVM compiler collection. The goal will be to port an existing Fortran compiler that targets massively parallel GPUs. The results are expected to be released as open source in late 2016.
While Fortran isn't a mainstream language, it's still heavily used in scientific computing, and there's lots of legacy code that relies on it. A lot of that code is maintained by people at the US National Labs, and the new project is being organized by staff at Lawrence Livermore, Sandia, and Los Alamos.
They'll be working closely with a division of Nvidia that makes the PGI Accelerator toolkit. That software is designed to allow clusters to use the computation abilities of GPUs on the cluster to perform distributed computations. PGI Accelerator's syntax is intended to be similar to that of OpenMP, a parallel computing API that's also heavily used in scientific computing.
According to National Labs' announcement, LLVM was already attracting interest from the high performance computing community, but they were mostly using its CLANG C/C++ compiler. But many of the large software packages used in scientific computing have a mix of C++ and Fortran modules, providing the incentive to bring it all under one toolchain. Los Alamos' Patrick McCormick said “The goal of this effort is to fill a key gap in the LLVM infrastructure that will help it become a more complete solution for our overall capabilities in support of the advancement of science and engineering.”
In addition to porting the software, the effort will include adopting existing code so that it conforms to LLVM guidelines and integrates with existing compiler utilities. If all goes according to plan, the software will be made open source roughly a year from now.
reader comments
30