MLBalance returns NaN's when not using the model's resolution for B

[guillaumevernieres]

Description

There is something wrong in the MLBalance saber block. It returns NaN's in the first inner iteration. I suspect something related to the trajectory.

To Reproduce:

Run a variational application using the MLBalance block at a different resolution than the deterministic.

[travissluka]

you are using this parameter, right? (

soca/test/testinput/3dvar_lowres.yml

Line 56 in 3dd464a

change background resolution: true

)

[guillaumevernieres]

you are using this parameter, right? (

soca/test/testinput/3dvar_lowres.yml

Line 56 in 3dd464a

change background resolution: true
)

yes

[guillaumevernieres]

Skipping the mask/ghost cells like it is done in the other saber blocks does the trick. I'll try to squeeze in a code cleanup and PR in between v17 work duties.

[travissluka]

Skipping the mask/ghost cells like it is done in the other saber blocks does the trick. I'll try to squeeze in a code cleanup and PR in between v17 work duties.

great. You probably had junk in your halos. For future reference, if you are using information in the ghost cells, make sure you always call halo exchange before accessing them. atlas will only do an actual halo exchange if an individual field has been previously marked dirty. Related: if you change this to skip ghost cells, be sure to call set_dirty() on any field you write to