There’s a conflict between the clang math forwards for cuda and libstdc++. At least, cuda 10.0 plus libstdc++ 7.2 or 7.4 error on duplicate, incompatible versions.
For those that have the proprietary nvcc toolchain working with clang openmp, please share your gnu toolchain + cuda version, along with any unexpected config tricks you needed to use to get it it work. Plus what the host arch / GPU are please.
My config doesn’t really work so I can’t start us off.
gcc 6 toolchain used to work for me and I never got gcc >=7 working due the massive mess in math header.
I configured clang with -DGCC_INSTALL_PREFIX=/opt/gcc/6.5 to force clang always picking up my own build of gcc 6.5.
My fresh build of clang today can compile my programs although I did workaround std::fpclassify in the application source code.
gcc 6.5 works + CUDA 10.1 build clang openmp offload properly.
The OS is ubuntu 18.04 and I used CUDA 10.1.
The machine has Ryzen 1700x and GTX1050Ti (pascal sm_61).
Jon Chesterfield via Openmp-dev <email@example.com> 于2019年11月5日周二 上午8:02写道：