How odd!
Yes, I guess I could see this happening. You might also want to look
at the resulting assembly to see if this is the case, too.
But I agree with Ralf -- please report this to PGI, and let's get their
take on it as well.
On Jul 4, 2005, at 5:17 AM, Ole Holm Nielsen wrote:
> The LAM-MPI integer*2 test dtyp/fint2_f.f fails with PGI 6.0-4
> as described below. I added a couple of print-statements in the
> code as requested, and now the code works as expected !
>
> So I went back to the original code and added a -O0 (no
> optimization) to FFLAGS in the dtyp/Makefile.
> Now the error is absent even with the original code.
> Then I changed the optimization to -O1 (the PGI default level),
> and the error returns.
>
> I have repeated this test on another system with the
> Portland Group PGI compiler 5.2-4, and the exact same
> behavior is observed: The error goes away when I insert
> a print statement.
>
> If I copy the receiver's test
> if (int2array(2) .ne. int2) then
> call lamtest_error_f('MPI_INTEGER2 receive error')
> endif
> into the sender's code (rank .eq. 0) after the MPI_SEND,
> the error exists also on the sender side.
>
> I've experimented with this a bit, and it seems to me that
> the PGI optimizer has removed the variable named "int2".
> If I make a "call dummy(int2)" subroutine call, or if I
> turn int2 into an array(3), the error goes away.
>
> It seems to me that an "optimizer fooler" code such as
> "call dummy(int2)" or better should be added to dtyp/fint2_f.f.
> On the other hand, I'm not enough of an expert to judge
> whether the present error is really a compiler bug or not.
> I'd be happy to report the error to the Portland Group, in case
> someone is dead sure that it's a compiler bug.
>
> Best regards,
> Ole
>
> Jeff Squyres wrote:
>> This is quite odd; the test is relatively simple. I've thought about
>> this all day and I can't see how it would go wrong without the others
>> going wrong as well. :-(
>>
>> If I could impose, could you stick a few print* statements in there
>> and
>> examine a few values to see what's happening at the relevant steps?
>>
>> Then again, if your applications do not use MPI_INTEGER2, don't worry
>> about it! ;-) (however, I would like to figure this out and
>> potentially
>> fix whatever the problem is, if you've got a few free cycles)
>>
>> On Jun 29, 2005, at 10:44 AM, Ole Holm Nielsen wrote:
>>
>>> I managed to build the lamtests-7.1.1 with the Portland Group
>>> PGI compilers version 6.0-4, employing a fix as discussed in
>>> http://www.lam-mpi.org/MailArchives/lam/2005/06/10866.php
>>>
>>> When running the lamtests on 4 nodes over Ethernet, "make check"
>>> gives a single unexpected error in a Fortran module:
>>>
>>> mpirun -x TEST C -ssi rpi crtcp
>>> /home/camp/ohnielse/lamtests-7.1.1/dtyp/./fint2_f
>>> [**ERROR**]: LAM/MPI MPI_COMM_WORLD rank 1, file fint2_f:
>>> MPI_INTEGER2 receive error
>>> FORTRAN STOP
>>> FORTRAN STOP
>>> FORTRAN STOP
>>> ---------------------------------------------------------------------
>>> --
>>> ------
>>> One of the processes started by mpirun has exited with a nonzero exit
>>> code. This typically indicates that the process finished in error.
>>> If your process did not finish in error, be sure to include a "return
>>> 0" or "exit(0)" in your C code before exiting the application.
>>>
>>> PID 21610 failed on node n1 (10.2.131.139) with exit status 1.
>>> ---------------------------------------------------------------------
>>> --
>>> ------
>>>
>>> This error occurs for all 5 modes: crtcp, lamd, sysv, tcp, usysv.
>>> All other tests pass (except for the ones where errors are expected).
>>> As requested in the README file, I attach a gzipped tar-file with
>>> the files check.out config.log laminfo.log configure.log.
>>>
>>> What can be done to remedy this lamtests error ?
>>>
>>> Ole Holm Nielsen
>>> Department of Physics, Technical University of Denmark
>
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
>
--
{+} Jeff Squyres
{+} jsquyres_at_[hidden]
{+} http://www.lam-mpi.org/
|
