[MITgcm-devel] MITgcm/pkg/streamice changes

[Jean-Michel Campin]

Hi Dan,

There is this question you raised that might be of interest for others.

> I saw that it did change testreport. I am surprised, it might mean the
> matrix wasn't symmetric before.

I think it might just be different machine truncation error: 
When I take a look at how halfpipe_streamice results (and especially the solver 
output that testreport checks for), it seems sensitive to the compiler used:
e.g. fwd test with pathscale, pgi or intel compiler agree on the first 7 digits
(compared to the reference, gfortran). And regarding AD results, seems sensitive
too: with pgi (or on the old geo cluster, intel or open64) we get only 10 digits
for the AD gradient. 
After you made these changes, the results were not more different than when
switching to an other compiler.

Cheers,
Jean-Michel