I tried to compile the latest version of SuperLU-DIST library with OpenMP enabled. I got the following error
“/home/xxxxx/petsc/traverse-nvidia-real/externalpackages/git.superlu_dist/SRC/pdgstrs_lsum.c”, line 1072: error: this OpenMP construct is not supported in NVIDIA subset: The ‘taskloop’ construct is not supported
However, it compiles well using the legacy PGI compiler.
I don’t know why we only get a subset of features in NVCC by upgrading. Is there a way to release this limit?