When copying data in and out of the MARBL instance prior to calling surface_flux_compute(), the loops are not ordered in an efficient manner for memory management. The existing code is
!-----------------------------------------------------------------------
! Copy data from slab data structure to column input for marbl
!-----------------------------------------------------------------------
do index_marbl = 1, marbl_col_cnt(iblock)
i = marbl_col_to_pop_i(index_marbl,iblock)
j = marbl_col_to_pop_j(index_marbl,iblock)
do n = 1,size(surface_flux_forcings)
marbl_instances(iblock)%surface_flux_forcings(n)%field_0d(index_marbl) = &
surface_flux_forcings(n)%field_0d(i,j,iblock)
end do
do n = 1,ecosys_tracer_cnt
marbl_instances(iblock)%tracers_at_surface(index_marbl,n) = &
p5*(tracers_at_surface_old(i,j,n) + tracers_at_surface_cur(i,j,n))
end do
do n=1,size(surface_flux_saved_state)
marbl_instances(iblock)%surface_flux_saved_state%state(n)%field_2d(index_marbl) = &
surface_flux_saved_state(n)%field_2d(i,j,iblock)
end do
end do
but we would be better served having the n loops outside of the (i,j) loop. This came up because @klindsay28 and I were looking at the the tracers_at_surface loop
do n = 1,ecosys_tracer_cnt
marbl_instances(iblock)%tracers_at_surface(index_marbl,n) = &
p5*(tracers_at_surface_old(i,j,n) + tracers_at_surface_cur(i,j,n))
end do
and talking about how we don't actually need the surface values of every tracer... so if MARBL could tell POP which tracer indices it actually cares about, we could have an if (nth tracer not required for surface flux computation) cycle line (and it would be much better to have that if statement outside the (i,j) loop)
Note that this is not applicable to interior_tendency_compute() because MARBL needs to make that call column-by-column (and also there is a transpose happening when we copy data into the MARBL structure), but it's something to keep in mind when we introduce marbl_instance%reset()
Description of the issue:
When copying data in and out of the MARBL instance prior to calling
surface_flux_compute(), the loops are not ordered in an efficient manner for memory management. The existing code isbut we would be better served having the
nloops outside of the(i,j)loop. This came up because @klindsay28 and I were looking at the thetracers_at_surfaceloopand talking about how we don't actually need the surface values of every tracer... so if MARBL could tell POP which tracer indices it actually cares about, we could have an
if (nth tracer not required for surface flux computation) cycleline (and it would be much better to have that if statement outside the(i,j)loop)Note that this is not applicable to
interior_tendency_compute()because MARBL needs to make that call column-by-column (and also there is a transpose happening when we copy data into the MARBL structure), but it's something to keep in mind when we introducemarbl_instance%reset()Version:
Machine/Environment Description:
N/A
Any xml/namelist changes or SourceMods:
N/A