Remove usage of Fortran copy of field data, part 1

[travissluka]

Description

Part 1 of 3 to address

• Remove Fortran copy of field values. #1126

This refactors Model2GeoVaLs and Balance so that:

• the Atlas fieldset is used on the Fortran side for accessing field data, no longer using the Fortran-only internal field%val structure
• The sync_to_atlas() and sync_from_atlas() calls in the interface for these classes were removed, which should make these classes go faster
• Also, I fixed the loop order in Balance, it was wrong, and there were switch statements inside the loops, eeek. Should be faster now

2 more PRs after this before we can remove the Fortran-only internal field data and free up some memory!

Testing

There are no change in any answers.

I also tested with the 1/4 deg, 1/2 deg minimization on my computer, 24 cores, there are no changes in answers and runtime for these components is faster (not much, and would be even less with more PEs, but every little bit counts)

develop branch

OOPS_STATS Name                                                :   total (ms)   count    time/call (ms)
OOPS_STATS soca::Balance::Balance                              :      2835.24       1         2835.2442
OOPS_STATS soca::Balance::multiply                             :      5452.57      12          454.3808
OOPS_STATS soca::Balance::multiplyAD                           :      4556.61      12          379.7176
OOPS_STATS soca::Model2GeoVaLs::changeVar                      :      1482.66       3          494.2216

this branch

OOPS_STATS Name                                                :   total (ms)   count    time/call (ms)
OOPS_STATS soca::Balance::Balance                              :      2231.90       1         2231.8992
OOPS_STATS soca::Balance::multiply                             :       448.99      12           37.4157
OOPS_STATS soca::Balance::multiplyAD                           :       467.96      12           38.9964
OOPS_STATS soca::Model2GeoVaLs::changeVar                      :         3.82       3            1.2744

[guillaumevernieres]

nice 👍

[shlyaeva]

🎉 looks good to me!
a couple of unrelated naive questions for my understanding of this code in the comments below.

[shlyaeva]

Double-checking since I don't know this code (and don't feel like thinking, I guess): both in develop and here is it intended to have data_m = data_m + ...? (I see all other variables updated as data_m = data_a + ....

[travissluka]

yes, it's different here because 3D variables are combined into a 2D variable. There is an initial data_m = data_a on line 299 (2D) then the loop is a summation over 3D data putting into the 2D data_m

[travissluka]

... in fact, when I refactored this i did have data_m = data_a + at first and couldn't figure out why it was giving completely wrong answers !

[shlyaeva]

similar question here: intended to have data_a = data_a + ... as opposed to data_a = data_m + ...