pgf90 -ta=vector option to manage openacc scheduling ?

Hello .

Do you have a ( beta/hiden ) pgf90 option to manage openacc scheduling ?

For example something like :

pgf90 -ta=nvidia,vector(32)

Or a ‘pgi’ directive in the fortran code for the entire file ?

The goal is to avail to put this information in all the openacc directive

In my test for example the grid size which are allocated a run time is not visible
by the compiler , which generate vector(128) or vector(256) directive
but the grid tested are mush smaller due to scalability test with MPI and domain decomposition .

1 MPI rank = grid 128x128x128
2 MPI rank = local size 64x128x128
etc …

Thanks .

Juan

Hi Juan,

No, sorry there’s no compiler option to set the default vector width. This is part of the OpenACC API and set via the loop directive’s “vector ( width )” clause. Though, I can see where it would add benefits so I added a feature request (TPR#19615) and sent it on to engineering.

  • Mat