6:19 p.m.
Hi,
with a fairly recent snapshot (July) I'm noticing that an MPI block is
failing if I replace a "funcerr" command line (and if-block) with an
equivalent errorif() call.
The obvious workaround is to stick with the funcerr command, but I guess
at some point the more recent syntax should also work with MPI. BTW, I
am aware that issuing funcerr inside MPI is also kind of dubious,
because in my experience the error message usually doesn't make it
through to the world of the caller, but well, that's a different thing.
Example:
<hansl>
mpi
err = 0
# errorif( $mpirank == 0 && err, "just checking") # fails
if $mpirank == 0 && err # works
funcerr "just checking"
endif
end mpi --np=2
</hansl>
thanks
sven
6:56 p.m.
On Tue, 10 Aug 2021, Sven Schreiber wrote:
with a fairly recent snapshot (July) I'm noticing that an MPI
block is
failing if I replace a "funcerr" command line (and if-block) with an
equivalent errorif() call.
It works if you respect the documented stipulation that errorif()
can only be called within a function.
<hansl>
function void humho (void)
err = 0
errorif($mpirank == 0 && err, "Oops!")
end function
mpi --send-functions
humho()
printf "%d: on the way out\n", $mpirank
end mpi --np=2
</hansl>
Allin
9:42 p.m.
Am 10.08.2021 um 18:56 schrieb Allin Cottrell:
On Tue, 10 Aug 2021, Sven Schreiber wrote:
> with a fairly recent snapshot (July) I'm noticing that an MPI block is
> failing if I replace a "funcerr" command line (and if-block) with an
> equivalent errorif() call.
It works if you respect the documented stipulation that errorif() can
only be called within a function.
OK, my example was stupid in that respect. (Although interesting that
"funcerr" still worked.)
But in fact, my original problem _did_ occur within a function. Only the
nesting order was reversed: The MPI block itself was inside the
function. Like this:
function ... main...
mpi
errorif(...)
end mpi
end function
So what's the recommended solution? Wrap the errorif call in another
function, like this?:
function ... errorcall...
errorif()
end function
function ... main ...
mpi --send-functions
errorcall()
end mpi
end function
Although obviously, the extra function wrap would defeat the purpose of
the concise errorif() idiom then.
thanks
sven
4:08 a.m.
On Tue, 10 Aug 2021, Sven Schreiber wrote:
Am 10.08.2021 um 18:56 schrieb Allin Cottrell:
> On Tue, 10 Aug 2021, Sven Schreiber wrote:
>
>> with a fairly recent snapshot (July) I'm noticing that an MPI block is
>> failing if I replace a "funcerr" command line (and if-block) with an
>> equivalent errorif() call.
>
> It works if you respect the documented stipulation that errorif() can
> only be called within a function.
OK, my example was stupid in that respect. (Although interesting that
"funcerr" still worked.
In the example you posted funcerr didn't "work", it was never
actually invoked (blocked by a false conditional).
But in fact, my original problem _did_ occur within a function. Only
the
nesting order was reversed: The MPI block itself was inside the
function. Like this:
function ... main...
mpi
errorif(...)
end mpi
end function
OK, that's tricky. So far as gretlmpi (a separate program) is
concerned, we're not inside a function at the point where errorif()
is called. Maybe the "only inside a function" requirement could be
waived within an mpi block. Or maybe the whole program could be
rewritten as something to be run by gretlmpi from the start?
Allin
8:45 a.m.
Am 11.08.2021 um 04:08 schrieb Allin Cottrell:
On Tue, 10 Aug 2021, Sven Schreiber wrote:
> OK, my example was stupid in that respect. (Although interesting that
> "funcerr" still worked.
In the example you posted funcerr didn't "work", it was never actually
invoked (blocked by a false conditional).
Ah, of course.
> But in fact, my original problem _did_ occur within a function.
Only the
> nesting order was reversed: The MPI block itself was inside the
> function. Like this:
OK, that's tricky. So far as gretlmpi (a separate program) is
concerned, we're not inside a function at the point where errorif() is
called.
Yeah, that's what I had guessed.
Maybe the "only inside a function" requirement could be
waived within
an mpi block. Or maybe the whole program could be rewritten as
something to be run by gretlmpi from the start?
No, because we're talking about a function package, so everything is a
function.
There are two problems here in the broad context: First, the workings of
funcerr and errorif inside MPI. Secondly, how to get printout to appear
from MPI whenever there's an error. This latter issue is probably only
relevant on Windows. (AFAIK on Linux the printed output is captured OK.
On Windows, in my experience what helps is something like mpibarrier()
followed by flush, but this of course can only be done if the error can
be anticipated - during debugging, for example.)
Since MPI is an advanced thing and it's a separate program, maybe a new
thing like mpierror() could be warranted. (Plus documenting that funcerr
and errorif shouldn't be used there.) OTOH, if in the end it does the
same thing as errorif, then maybe you're right and it simply should be
allowed.
thanks
sven
1462
days inactive
1463
days old
4 comments
2 participants
participants (2)
-
Allin Cottrell -
Sven Schreiber