[MITgcm-devel] [MITgcm-cvs] MITgcm/model/src CVS Commit

[Martin Losch]

Hi Jean-Michel,
I am surprised that you got this email at all! After all the sx8-mailing problems seem to be sorted out.

The reason for this error is as follows:
for better vectorization I extended the loop ranges to the full range, but if the caller does not provide non-zero values for b3d, then the code returns the errorcode=1 on the halos.
The safe solution is to remove the extension of the loop ranges, but having b3d and bet = 0. on the halos does not hurt. (In impldiff.F, the coefficients are computed also for the halos if TARGET_NEC_SX and there is no problem.)

I guess you prefer to keep the safety catch. In that case I'd have to bit the bullet and take the performance hit (it's actually surprisingly big, makes the model run longer by up to 10%, so if we do not need this catch, I'd be happy).

Same is true for solve_pentadiagnogal.F

Martin


On Oct 24, 2011, at 2:45 PM, Jean-Michel Campin wrote:

> Hi Martin,
> 
> Looks like there is a problem on sx8 for test exp. front_relax.bvp :
> it stops in gmredi_calc_psi_bvp.F because of an error code return
> value from SOLVE_TRIDIAGONAL
> I guess this test is using TARGET_NEC_SX ? (I did not see any
> problem for the test run on other platform).
> 
> Cheers,
> Jean-Michel
> 
> On Thu, Oct 13, 2011 at 10:48:44AM -0400, Martin Losch wrote:
>> Update of /u/gcmpack/MITgcm/model/src
>> In directory forge:/tmp/cvs-serv29041/model/src
>> 
>> Modified Files:
>> 	solve_tridiagonal.F solve_pentadiagonal.F 
>> Log Message:
>>  - improve vectorization also for adjoint code of 
>>    solve_tridiagonal and solve_pentadiagonal. As this requires up to
>>    5 new 3D-fields, this code is only turned on for the combination of
>>    ALLOW_AUTODIFF_TAMC and TARGET_NEC_SX.
>>  - remove hard-wiring of CPP-flag ALLOW_SOLVERS_KLOOPINSIDE to
>>    ALLOW_AUTODIFF.  !!!!! THIS FLAG NOW NEEDS TO BE SET OUTSIDE OF
>>    SOLVE_TRI/PENTADIAGONAL.F !!!!
>> 
>> 
>> _______________________________________________
>> MITgcm-cvs mailing list
>> MITgcm-cvs at mitgcm.org
>> http://mitgcm.org/mailman/listinfo/mitgcm-cvs
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel