NVIDIA GCC Backend Gets Ready For OpenMP Offloading

[phoronix]

Phoronix: NVIDIA GCC Backend Gets Ready For OpenMP Offloading

While GCC 7 feature development is officially over, one of the late patches to land for GCC 7.1 in trunk are improvements to the NVIDIA NVPTX back-end...

http://www.phoronix.com/scan.php?page=news_item&px=GCC-NVPTX-OpenMP-Prereq

[pipe13]

Well this is interesting. With Cuda 8.0, NVCC itself doesn't work with any GCC beyond 5.4. But there is at least some OpenACC NVPTX in GCC 6, and now OpenMP 4 PTX code generation for GCC 7 and 8 as well.

There's no real reason for NVCC to support newer GCCs than are the system compilers for e.g. RedHat EL, Enterprise SuSE, CentOS, Ubuntu LTS, which afaik are all still on the GCC 4.x branch. It's not like the GCC versions are mutually link-time exclusive, and one supposes that as all new feature development must be done on the GCC dev branch Nvidia has little choice but to do OpenACC and OpenMP on the very latest versions. But one does wonder what is their long-term direction.

[pipe13]

Edit for clarity: GCC 5 objects link fine with GCC 6. GCC 4 doesn't work so well with the later versions. std::string something something.