The LLVM Fortran Performance Is Beating Out GCC But Losing To PGI

At last month's EuroLLVM conference, NVIDIA provided an update on the "Flang" project for offering first-rate Fortran support within LLVM, including some initial benchmark figures.

LLVM has accepted the "f18" project to become an official LLVM sub-project albeit much code work is ahead before it's expected to appear in an LLVM release. The actual name is still to be decided of the sub-project as well. But whatever it ends up being called, the performance is looking quite good for the code in its initial state.

This code, which started off from the PGI Fortran front-end and then adapted to LLVM, is performing quite well and continuing to support new Fortran language features. With NVIDIA doing a lot of the work on this Fortran compiler support, GPU/OpenMP offloading support is also a big goal in mind for eventually being able to punt the work off to their GPUs with this Fortran LLVM compiler.

As for the limited benchmarks they shared at EuroLLVM 2019, the performance is coming in short of the PGI 19.1 release but is performing better than the Fortran support within GCC8.

The slide deck in full covering this Fortran/Flang update here.