Re: [OMPI devel] fortran types alignment

[George Bosilca]

I think I like d the most but it is not a perfect solution. With d all
real8 types in a common will be badly aligned and the Open MPI
internal datatype will be incorrect.  So I will vote for a combo: b +
d.

  George.

On Fri, May 30, 2014 at 4:57 AM, Gilles Gouaillardet
 wrote:
> Folks,
>
> i recently had to solve a tricky issue that involves alignment of fortran
> types.
>
> the attached program can be used and ran on two tasks in order to evidence
> the issue.
>
> if gfortran is used (to build both openmpi and the test case), then the test
> is successful
> if ifort (Intel compiler) is used (to build both openmpi and the test case),
> then the test fails.
>
> this was mentionned in the openmpi users list quite a while ago at
> http://www.open-mpi.org/community/lists/users/2010/07/13857.php
>
> the root cause is gfortran considers mpi_real8 must be aligned on 8 bytes
> whereas ifort considers mpi_real8 does not need to be aligned.
> consequently, the derived data type ddt is built with an extent of 16
> (gfortran) or 12 (ifort)
>
>
> in order to determine the type aligment, configure builds a simple program
> with c and fortran that involves common.
> the default behaviour of ifort is to :
> - *not* align common
> - align records (aka the real8_int fortran type)
> hence the mismatch and the failure.
>
> the default behaviour of gfortran is to align both common and records, hence
> the success.
>
> /* i "extracted" from configure conftest.c and conftestf.f that can be used
> to build the conftest binary. conftest will store the alignment in the
> conftestval file */
>
> i am wondering how this should be dealt by OpenMPI.
>
>
> here is a non exhaustive list of option :
>
> a) do nothing, this is not related to openmpi, and even if we do something,
> application built with -noalign will break.
> b) advise ifort users to configure with FCFLAGS="-align zcommons" since it
> is likely this is what they want
> c) advise ifort users to build their application with "-noalign" to be on
> the safe side (modulo a performance penalty)
> d) update OpenMPI so fortran type alignment is determined via a record
> instead of a common if fortran >= 90 is used
> (so far, i could not find any drawback in doing that)
> e) advise ifort users to create ddt with MPI_DOUBLE instead of mpi_real8
> (because this works (!), i did not dig to find out why)
> f) other ...
>
> any thoughts ?
>
> Cheers,
>
> Gilles
>
> ___
> devel mailing list
> de...@open-mpi.org
> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel
> Link to this post:
> http://www.open-mpi.org/community/lists/devel/2014/05/14911.php

[OMPI devel] fortran types alignment

[Gilles Gouaillardet]

Folks,

i recently had to solve a tricky issue that involves alignment of fortran
types.

the attached program can be used and ran on two tasks in order to evidence
the issue.

if gfortran is used (to build both openmpi and the test case), then the
test is successful
if ifort (Intel compiler) is used (to build both openmpi and the test
case), then the test fails.

this was mentionned in the openmpi users list quite a while ago at
http://www.open-mpi.org/community/lists/users/2010/07/13857.php

the root cause is gfortran considers mpi_real8 must be aligned on 8 bytes
whereas ifort considers mpi_real8 does not need to be aligned.
consequently, the derived data type ddt is built with an extent of 16
(gfortran) or 12 (ifort)


in order to determine the type aligment, configure builds a simple program
with c and fortran that involves common.
the default behaviour of ifort is to :
- *not* align common
- align records (aka the real8_int fortran type)
hence the mismatch and the failure.

the default behaviour of gfortran is to align both common and records,
hence the success.

/* i "extracted" from configure conftest.c and conftestf.f that can be used
to build the conftest binary. conftest will store the alignment in the
conftestval file */

i am wondering how this should be dealt by OpenMPI.


here is a non exhaustive list of option :

a) do nothing, this is not related to openmpi, and even if we do something,
application built with -noalign will break.
b) advise ifort users to configure with FCFLAGS="-align zcommons" since it
is likely this is what they want
c) advise ifort users to build their application with "-noalign" to be on
the safe side (modulo a performance penalty)
d) update OpenMPI so fortran type alignment is determined via a record
instead of a common if fortran >= 90 is used
(so far, i could not find any drawback in doing that)
e) advise ifort users to create ddt with MPI_DOUBLE instead of mpi_real8
(because this works (!), i did not dig to find out why)
f) other ...

any thoughts ?

Cheers,

Gilles


bcast_types.f90
Description: Binary data


conftestf.f
Description: Binary data
#include 
#include 

#ifdef __cplusplus
extern "C" {
#endif
void align_(char *w, char *x, char *y, char *z)
{   unsigned long aw, ax, ay, az;
FILE *f=fopen("conftestval", "w");
if (!f) exit(1);
aw = (unsigned long) w;
ax = (unsigned long) x;
ay = (unsigned long) y;
az = (unsigned long) z;
if (! ((aw%16)||(ax%16)||(ay%16)||(az%16))) fprintf(f, "%d\n", 16);
else if (! ((aw%12)||(ax%12)||(ay%12)||(az%12))) fprintf(f, "%d\n", 12);
else if (! ((aw%8)||(ax%8)||(ay%8)||(az%8))) fprintf(f, "%d\n", 8);
else if (! ((aw%4)||(ax%4)||(ay%4)||(az%4))) fprintf(f, "%d\n", 4);
else if (! ((aw%2)||(ax%2)||(ay%2)||(az%2))) fprintf(f, "%d\n", 2);
else fprintf(f, "%d\n", 1);
fclose(f);
}
#ifdef __cplusplus
}
#endif