[MEDUSA] Code using character strings as switches cannot be put on GPU

[arporter]

The string options are a feature of the configuration interface of the external source code mocsy (Orr, J. C. and Epitalon, J.-M., 2015, "Improved routines to model the ocean carbonate system: mocsy 2.0", Geosci. Model Dev., 8, 485-499, doi:10.5194/gmd-8-485-2015) used by MEDUSA, and not part of the native MEDUSA code. It is therefore not practical to change the original source and we propose instead to implement a transformation to replace such switches with integers on a routine-by-routine basis.

[arporter]

Updating to say that @deardenchris and I have discovered that this is an issue for the halo buffer packing routines in the north-fold routines in lib_mpp in NEMO.

[sergisiso]

I am getting a segfault here and it surprised me because I thought we already prevented CHARACTERS on GPU, specially in this case that the type information is available:

    !$omp target
    !$omp teams loop collapse(1) ...
    do jn = 1, kjpt, 1
      if (cdtype(:) == 'TRA' .AND. (jn == jp_tem .OR. jn == jp_sal .AND. ln_zdfddm) .OR. cdtype(:) == 'TRC' .AND. jn == 1) then
        if (cdtype(:) == 'TRA' .AND. jn == jp_tem) then

but looking at previous issues #2854, #2510 maybe we just did it for ACC kernels and the range2loop transformation?

In any case I think we need to exclude regions where any Reference or Literal is a character in ALL gpu offloading transformations.

The more difficult question is if we should also exclude symbols that we don't know the type as they could be a CHARACTER, but that is not the case above.

[arporter]

Stumbled across this @sergisiso - it relates to the discussion we had about character sub-strings in #3041.