[MITgcm-devel] vectorizing cg2d

[Martin Losch]

Hi all,
this is a fairly general question:
Jens-Olaf has found a way to make cg2d more than 30% fast on our SX8  
by inserting 3 compiler directives (!CDIR OUTERUNROLL=10) for a  
243x170 configuration. You'll find the modified cg2d in  
eddy.csail.mit.edu:~mlosch/scratch/MITgcm/model/src/cg2d.F.
I find that this is very valuable and useful information, but how  
does one put this into the general code, so that this information is  
not lost.
Do we want compiler directives ( maybe within ifdef TARGET_NEC_SX )  
in the main code? Also the number 10 may be subject to details of the  
configuration. Jens-Olaf tells me that he normally uses 16. Is there  
a good solution for this type of platform (and configuration)  
dependent optimization?

Martin