On Tue, May 28, 2024 at 09:23:41PM +0200, Tobias Burnus wrote:
> -fopenmp-force-usm can be useful for some badly written code. Explicity
> using 'omp requires' makes more sense but still. It might also make sense
> for testing purpose.
>
> Unfortunately, I did not see a simple way of testing it. When trying it
> manually, I looked at the 'a.xamdgcn-amdhsa.c' -save-temps file, where
> gcn_data has the omp_requires_mask as second argument and testing showed
> that an explicit pragma and the -f... argument have the same result.
>
> Alternative would be to move this code later, e.g. to lto-cgraph.cc's
> omp_requires_mask, which might be safer (as it avoids changing as many
> locations). On the other hand, it might require more special cases
> elsewhere.*
>
> Comment, suggestions?
>
> Tobias
>
> *I am especially thinking about a global variable and "#pragma omp declare
> target". At least with 'omp requires self_maps' of OpenMP 6, it seems as if
> 'declare target enter(global_var)' should become 'link(global_var)' where
> the global_var pointer is updated to point to the host version.
How is that option different from
echo '#pragma omp requires unified_shared_memory' > omp-usm.h
gcc -include omp-usm.h
?
I mean with -include you can add anything you want, not just one particular
directive, and adding a separate option for each is just weird.
Jakub
