on the use of mikt/u/v...

[smasson]

I don't know if you noticed the introduction of mikt/u/v in the code...

I must say that I liked this change from the beginning :-)

I first discovered at the merge party that to get the SST it was no more tsn(:,:,1,jp_tem) but

         DO jj = 1, jpj
            DO ji = 1, jpi
               z2d(ji,jj) = tsn(ji,jj,mikt(ji,jj),jp_tem)
            END DO
         END DO

Did I see "code simplification" somewhere???

Next I saw this is affecting 21 routines... so good for a development sold as orthogonal to other parts of the code.

And today, spending some time in zdftke, i discovered that because of the use of mikt, all 3D loops have been re-ordered with the k loop inside the j and i loops!!

     DO jj = 2, jpjm1
         DO ji = fs_2, fs_jpim1   ! vector opt.
            DO jk = mikt(ji,jj)+1, jpkm1

I hope you noticed the use of fs_2 and fs_jpim1 for a potential improvement of the code performance!

Did I see "HPC" somewhere??? Any reviewer of this? What is doing the leader of the HPC working group (me)?

First conclusions

=> We must add performance tests in addition to sette tests. At least when committing a new release.

=> a small "do and don'ts for HPC" seams to be required... maybe ​this

[mathiot]

In order to close this ticket, I created a development branch and the associated ticket #1473.

[mathiot]

Development branch ISFCLEAN (see #1473) was merged to the trunk in the revision 5120 (2015-03-03).

Log of the merge: #1473. Re-organisation and optimisation of ice shelf cavity option. This commit merges changes from the dev_r5094_UKMO_ISFCLEAN branch onto the trunk. Results will change, even with ln_isfcav=F, due to a return to original definitions of the vertical metrics. All changes have been reviewed and SETTE tested.

By the way, if ln_isfcav = false the coast line is preserved (issue highlighted by G. Reffray by email the 18th Feb. 2015).