NEC Is Looking To Contribute SX-Aurora VE Accelerator Support To LLVM

The newest compiler back-end proposed for merging into the LLVM compiler code-base is for the NEC SX-Aurora VE (Vector Engine) accelerator card.

NEC launched the SX-Aurora a year ago as a "Vector Engine" PCI Express accelerator card supporting up to eight of these vector processors per server. The NEC SX-Aurora features eight cores clocked up to 1.6GHz and six HBM2 memory modules on a silicon interposer. The CPU double precision performance is rated up to 2.45 TFLOPS based upon the model or 4.9 single-precision TFLOPS. The six HBM2 stacks yield up to 48GB of working memory at 1.2TB/s memory bandwidth.

Since the SX-Aurora VEs began shipping a year ago, NEC has provided a proprietary Fortran/C/C++ compilers and a proprietary MPI implementation for offloading work to these vector engines. OpenMP is also supported.

NEC HPC Europe has been working on an open-source LLVM back-end for the SX-Aurora VE and has reached a point of maturity where they are hoping to merge this code to mainline. This back-end would end up allowing LLVM offloading to these PCIe accelerator cards for supported instructions.

The code is thorough enough that it does allow users to utilize the VEs without any proprietary software. As part of this back-end are also patches improving LLVM's vectorization abilities.

More details via the mailing list post proposing this new back-end while the current code is hosted on GitHub and sx-aurora.github.io.