Hello Michael, Thank you very much for this information. I am forwarding this to the macports-dev mailing list for others who are involved in MacPorts more closely and for much longer than I have been.
On Friday, October 16, 2020 at 6:32 PM, Mike Pique wrote: > I have a DTK and have built OpenMP both for arm64 an x86_64. You need to > compile with “-Xclang -fopenmp” and to include the OMP header files. You need > to link with -l omp and include the path to the run-time library directory. > Example: > cc -I /usr/local/include -Xclang -fopenmp omp_hello.c -L /usr/local/lib -l > omp -o omp_hello > For arm64, when you build the openmp run-time library you need to change 5 > lines in the assembly language file Z_Linux_asm.S (disregard the historic > “Linux” - it has #ifdefs for many platforms). On 24 September, I submitted my > 5-line patch to the OpenMP group (differential ID D88252) where it is under > review. The 5-line patch affects only arm64 builds and corrects an assumption > in the code that MacOS (Mach-O) object files are never built for arm64; the > patched version is completely x86_64-compatible. > I’ve not been able to figure out how to interest HomeBrew and MacPorts about > this, presumable when eventually OpenMP adopts it (or something similar) that > will flow downstream. But I had several pieces of our lab’s scientific > software needing OpenMP so I went ahead for my work without waiting. > Let me know any suggestions on how, or whether, these changes should be put > into MacPorts and/or HomeBrew. Thank you. MacPorts is likely interested in this. It currently has some support for OpenMP: the recommended approach would be for portfiles to specify `compiler.openmp_version`, so that an OpenMP-capable compiler and a MacPorts-provided libomp are used. But plenty of ports which could use OpenMP aren't doing so yet, or are not using it by default. One concert was the fact an alternative MacPorts-provided compiler was needed; this new info may help with using Xcode clang instead, which I wonder could be as simple as tweaking the paths to use MacPorts' libomp. There are still some other concerns, such as how using OpenMP in library ports affects building their dependents. I'm also not aware whether MacPorts has considered OpenMP on ARM64 yet either. Searching online for "-Xclang" "-fopenmp" reveals there are a few other interesting posts out there on the topic. Christopher A. Chavez
