search

GEOS-ESM / GFDL_atmos_cubed_sphere

The GFDL atmos_cubed_sphere dynamical core code
Other
0 stars 2 forks source link

WIP: Fix for GCC 14. Always allocate NH arrays #104

Closed mathomp4 closed 1 month ago

mathomp4 commented 1 month ago

This fixes issue with GCC 14 with Debugging flags.

Testing with Intel shows this is zero-diff when running hydrostatic v11.

Also, as expected, in v12, GCC 14 Debug runs just fine since that is non-hydrostatic by default.

Testing with GCC 14 with Debugging on found a bug when running as hydrostatic:

Error termination. Backtrace:
At line 505 of file /discover/nobackup/mathomp4/SystemTests/builds/AGCM_GNU/CURRENT/GEOSgcm/src/Components/@GEOSgcm_GridComp/GEOSagcm_GridComp/GEOSsuperdyn_GridComp/@FVdycoreCubed_GridComp/@fvdycore/model/dyn_core.F90
Fortran runtime error: Index '2' of dimension 3 of array 'w' outside of expected range (1:1)

This is on the call to c_sw: https://github.com/GEOS-ESM/GFDL_atmos_cubed_sphere/blob/0dd21f5426b602a698f5dfd73b44da3a62c81fe4/model/dyn_core.F90#L502-L510

The current code for allocating w is:

    if ( Atm%flagstruct%hydrostatic ) then
       !Note length-one initialization if hydrostatic = .true.
       allocate (    Atm%w(isd:isd, jsd:jsd  ,1) )
       allocate ( Atm%delz(isd:isd, jsd:jsd  ,1) )
       allocate (  Atm%ze0(is:is, js:js  ,1) )
    else
       allocate (    Atm%w(isd:ied, jsd:jed  ,npz  ) )
       allocate ( Atm%delz(isd:ied, jsd:jed  ,npz) )
       if( Atm%flagstruct%hybrid_z ) then
          allocate (  Atm%ze0(is:ie, js:je ,npz+1) )
       else
          allocate (  Atm%ze0(is:is, js:js  ,1) )
       endif
       !         allocate ( mono(isd:ied, jsd:jed, npz))
    endif

The issue is that because I was running hydrostatic, w is w(1,1,1). But the call to c_sw is do k = 1, npz and when it gets to the second loop, boom.

So, the simple fix is to just always allocate these arrays. It is wasteful, but solves the issue.

NOTE 1: As I'm not sure how to test the hybrid_z bits of this, I've kept it as is. I suppose it should probably always be allocated as well

NOTE 2: From my reading of the current mothership fvdycore (which we have diverged from, but share a common ancestor), they could have the same issue there (see https://github.com/NOAA-GFDL/GFDL_atmos_cubed_sphere/blob/44e20a7d45ac1d2c3b84a6ffaa0c88b4cc3d5175/model/fv_arrays.F90#L1546-L1563 and https://github.com/NOAA-GFDL/GFDL_atmos_cubed_sphere/blob/44e20a7d45ac1d2c3b84a6ffaa0c88b4cc3d5175/model/dyn_core.F90#L440-L448). For that reason, I mention @bensonr as this might be seen by them if/when they test with GCC 14. (Though if they don't see it, I guess that says something interesting too!)