[Mitgcm-support] Re: testscript FAIL

[mitgcm-support at dev.mitgcm.org]

Alistair thanks for suggestion.  IEEE compliance for SGI f77 is specified
using "-OPT:IEEE_arithmetic=1" and is on by default.  Adding this option
and/or turning off all optimizations, "-O0", does not change the results on
the origin 2000.  The only option that seems to make the test results closer
to those of g77 is the removal of "-r8", which forces REAL*8 as the default
for real variables.  Here is a partial list of test results:

default release1_p8 on sgi origin 2000 (castor):
FFLAGS = ( '-extend_source -mp -mpio -bytereclen -r10000 -mips4 -r8 -static' )
FOPTIM = ( '-O3' )
Y Y Y Y  5  5  7  7  7  8 10  9  6  6  6  6  7  5  7  5  7 FAIL  lab_sea

FFLAGS = ( '-extend_source -mp -mpio -bytereclen -r10000 -mips4 -r8 -static' )
FOPTIM = ( '-O0 -OPT:IEEE_arithmetic=1' )
Y Y Y Y  5  5  7  7  7  8 10  9  6  6  6  6  7  5  7  5  7 FAIL  lab_sea

FFLAGS = ( '-extend_source -mp -mpio -bytereclen -r10000 -mips4 -static' )
FOPTIM = ( '-O3' )
Y Y Y Y  7  9  7  9  8  8 10 11  8  8  7  7  8  7  7  7  8 FAIL  lab_sea

D.


>> It's probably optimization that's breaking the results.
>> 
>> It would be helpful to add an IEEE compliant option to genmake
>> for the SGIs. Don't know what the option is but look under
>> case Linux*: for lines of the form
>>    if ($IEEE) set FFLAGS = ($FFLAGS option)
>> to see how to add it to genmake. The option should limit optimization
>> where it will change the numerical results. It shouldn't be the
>> default (i.e. with $IEEE unset) since it normally slows the code
>> down a lot.