[MITgcm-support] problems on alpha machine
Hezi Gildor
hezi.gildor at weizmann.ac.il
Tue Dec 6 14:04:11 EST 2005
Dear Ed and Martin,
Thanks for trying to help.
I have tried all the -fpeX options as well as reducing the optimization
level but it didn't help.
It seems that phlocal=-Infinity at certain location after 2 iteration
and then the code try to execute 10**(-Infinity). Still didn't find when
and why phlocal gets this value.
as for the other problem, in diagnostics_main_init.F :
i tried to define a separate character variable and assign
rUnit2c//'^2/s^2' to it but it didn't help. same when trying ' m'
instead of 'm '
Thanks again..., hezi.
Ed Hill wrote:
>On Sun, 2005-12-04 at 12:03 +0100, Martin Losch wrote:
>
>
>>Hezi,
>>
>>has someone answered you, yet?
>>
>>The first problem looks like perfect F77 syntax to me, so I have no
>>clue why your compiler doesn't take it. Can it compile rUnit2c//'^2'?
>>Maybe one has to define a separate character variable, assign
>>rUnit2c//'^2/s^2' to it and then pass it to diags_mk_units? This is
>>really stupid but have you tried to set rUnit2c = ' m' instead of 'm '?
>>
>>
>
>Hi Hezi,
>
>I think Martin is right and that it is valid syntax. I'm sorry your
>compiler doesn't like it. Perhaps there is a compiler option that will
>help?
>
>
>
>
>>Maybe you can try reducing your optimization level. Again I don't know
>>your compiler, but -O5 sounds like aggressive optimization.
>>
>>
>
>Some compilers for the alpha architecture have options that will allow
>you to turn off some or all floating-point exceptions (or make them only
>issue warnings instead of stopping your program). The options are
>usually called "-fpeX" where X is a number in the range 0--4. So please
>take a look at your compiler documentation and see if there are any
>options that can help.
>
>Ed
>
>ps - As a last resort, one can extend the recently-added signal
> handler code (MITgcm/eesupp/src/sigreg.c) to trap and ignore
> any signals caused by floating point exceptions. [This is
> what the "-fpe" options do.]
>
>
>
More information about the MITgcm-support
mailing list