Halo Update Fill Implemention Bug with Padded Blocks

[apcraig]

In ice_boundary.F90 (both mpi and serial), the fill implementation in the halo update is carried out assuming blocks are NOT padded. This is a bug. The current implementation looks like

    do j = 1,nghost
       array(1:nx_block,           j,:) = fill
       array(1:nx_block,ny_block-j+1,:) = fill
    enddo
    do i = 1,nghost
       array(i,           1:ny_block,:) = fill
       array(nx_block-i+1,1:ny_block,:) = fill
    enddo

but it should look more like

       do j = 1,nghost
          array(1:nx_block, jlo-j,iblk) = fill
          array(1:nx_block, jhi+j,iblk) = fill
       enddo
       do i = 1,nghost
          array(ilo-i, 1:ny_block,iblk) = fill
          array(ihi+i, 1:ny_block,iblk) = fill
       enddo

One particular challenge is that ilo/ihi/jlo/jhi do not seem to be accessible in ice_boundary due to circular logic, so some data may need to be moved around in the code.

The good thing is that this doesn't seem to affect any results. This was detected by initializing all arrays with signalling nans and then trapping. Without that, you seem to fortuitously get a 0 multiplied by the uninitialized values in all cases. The fact that padded and non-padded decomps produce bit-for-bit results suggests this has not affected answers in actual model runs.

In the one place where the signalling nans was trapped (to_ugrid in ice_grid.F90), we have worked around it by initializing tarea to zero everywhere. When the halo fill is fixed, that fix should be removed, and we should confirm that the original issue is resolved.

See also #560.

[phil-blain]

For easy reference, one such loop is at:

CICE/cicecore/cicedynB/infrastructure/comm/mpi/ice_boundary.F90

Lines 1288 to 1295 in d529df8

	do j = 1,nghost
	array(1:nx_block, j,:) = fill
	array(1:nx_block,ny_block-j+1,:) = fill
	enddo
	do i = 1,nghost
	array(i, 1:ny_block,:) = fill
	array(nx_block-i+1,1:ny_block,:) = fill
	enddo