[MITgcm-devel] adjoint and vectorizing options

[Jean-Michel Campin]

Hi Martin,

I don't see problems in make the changes you propose. 

Regarding timing, I would like to check with Gael why his AD glacier
latest test are all failing before further changes in pkg/seaice.

Cheers,
Jean-Michel

On Fri, Jun 12, 2015 at 12:52:00PM +0200, Martin Losch wrote:
> Hi Patrick, Jean-Michel,
> 
> I have returned to the adjoint of seaice_lsr. I cannot claim success, but I have tried to clean up  the tridiagonal solvers, so that their logic corresponds very much to ???model/src/solver_tridiagonal.F??? 
> 
> The adjoint still explodes (for cost_ice_flag=7, which is bascially uice**2), but I can get rid of the extra stores in seaice_lsr_tridiagu/v, so I would like to check in this code probably even today, as a new starting point.
> 
> There are issues: 
> - the adjoint simulations of lab_sea changes
> 
> - I modified the action of the SEAICE_VECTORIZE_LSR_ZEBRA flag, so that this ???zebra??? algorithm can also be used without SEAICE_VECTORIZE_LSR. This means that the name is misleading and I would like to rename it to SEAICE_LSR_ZEBRA (shorter and more accurate). How do we deal with this? Should I just do it (becaues basically nobody except me and maybe some AWI people uses this flag anyway, or should I make sure that the model stops if the obsolete flat SEAICE_VECTORIZE_LSR_ZEBRA is defined?
> 
> - I would like my ???zebra??? code to be tested. Since lab_sea will change anyway, this may be good option. What do you think?
> 
> And then: should I wait until after the weekend for any of these changes, or should I go ahead now?
> 
> Martin
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel