Unable to build OpenMP examples with Clang 13

[rseidler2019]

While using clang 13.0.0 (here AOCC, but it was confirmed that clang 13 in general produces this bug) as Host-Compiler for the samples, the build process terminates with the following error in any OpenMP related example: /opt/AMD/aocc-compiler-3.2.0/lib/clang/13.0.0/include/omp.h(488): error: function "omp_is_initial_device" has already been defined

Affected Samples:

• Samples/0_Introduction/cudaOpenMP
• Samples/0_Introduction/UnifiedMemoryStreams

Solution: The Bug is present, when using OpenMP 5.0 in clang (which is the default in 13.0.0). When the OpenMP version is lowered to 4.5 (e.g. with -Xcompliler -fopenmp-version=45) the examples are compiled as expected.

I am currently working on incorporating the changes to the corresponding Makefiles and will create a PR soon.

[robogast]

I'm running into this issue when compiling magma 2.6.2 with cuda 11.7 and icx/icpx compilers

[mgravey]

after a quick investigation, it's the way this function is declared for version of openmp >=5

a simple fix is to add in the omp.h file a condition to not declare this function again, something like that for me:

1. first to a copy of your omp.h file before you destroy it :)
2. use sed -i '492s/$/ \&\& !defined(__NVCC__)/' /opt/intel/oneapi/compiler/2023.0.0/linux/bin-llvm/../compiler/include/omp.h

for me I add the condition at the line 492, the error was at line 497, so it looks to be always 5 before, if you had the error at 488 use sed -i '483s/$/ \&\& !defined(__NVCC__)/' /path/to/your/omp.h and so on