Hari work 2

[science-enthusiast]

Bonjour
I still have doubts that I have marked as "DOUBT_HARI". We will discuss and address the issues.

Merci

Hari

[pledac]

Hello Harri, thanks for the PR. Finally, the rebase can't be done for the moment. So just:
a) Fix the GPU build as intended
b) Please can you keep the syntax for the rewritten kernel and avoid using view_tab instead of tab except if necessary ? Example:

         for (int num_face = ndeb; num_face < nfin; num_face++)
            {
              const int elem1 = face_voisins(num_face, 0), elem2 = face_voisins(num_face, 1);
              int elem = elem1 == -1 ? elem2 : elem1;
              const double surf = zp1b.surface(num_face);
              tab_flux_faces(num_face) = val_flux(num_face - ndeb, 0) * surf / zp1b.volumes(elem); // TODO multiple elements!! units val_flux(num_face-ndeb,0) *surf [kg.s-1] => gives [kg.m-3.s-1]
            }

Would be rewritten like that:

          CDoubleArrView volumes_entrelaces = zp1b.volumes_entrelaces().view_ro();
          CIntTabView face_voisins = zp1b.face_voisins().view_ro();
          CIntTabView face_sommets = zp1b.face_sommets().view_ro();
          Kokkos::parallel_for(start_gpu_timer(__KERNEL_NAME__), Kokkos::RangePolicy<>(ndeb, nfin), KOKKOS_LAMBDA (const int num_face)
          {
            const int elem1 = face_voisins(num_face, 0), elem2 = face_voisins(num_face, 1);
            int elem = elem1 == -1 ? elem2 : elem1;
            const double surf = face_surfaces(num_face);
            flux_faces(num_face) = val_flux(num_face - ndeb, 0) * surf / volumes(elem); // TODO multiple elements!! units val_flux(num_face-ndeb,0) *surf [kg.s-1] => gives [kg.m-3.s-1]
          });
          end_gpu_timer(Objet_U::computeOnDevice, __KERNEL_NAME__);

It is for readability and to keep the Kokkos port the same on the whole code. Thanks !

[science-enthusiast]

Hi Pierre
I agree with you about maintaining consistent style. Thanks for the descriptive example. I will modify my code accordingly.