Re: [OMPI devel] Bug report: single processor MPI_Allgatherv

[George Bosilca]

Daniel,




  Thanks,
    george.

On Jan 18, 2008, at 4:57 PM, Daniel G. Hyams wrote:

Thanks for the fix. It is indeed the right solution. I'll make sure it gets into the trunk asap.

Sorry to reply to my own mail, but the bug only affects MPI_Allgatherv.

In this changeset:

https://svn.open-mpi.org/trac/ompi/changeset/16360



fix is just to multiply
"extent" by disps[0], on line 50.  I've verified that this fixes the
problem.



Daniel G. Hyams wrote:

In coll_self_allgatherv.c, the "extent" variable is never used. So the

I don't think that the displacements (disps) are being handled correctly

in MPI_Allgatherv, for a single process case.
The disps are being handled as byte offsets, instead of 'item'
offsets...they need to be multiplied by the size, in bytes,
of the MPI_Datatype being sent.

This bug seems to be closely related this:
https://svn.open-mpi.org/trac/ompi/changeset/16360
and probably is present in all of the routines listed there.


process only.  The input and output should
be the same, but that is not the case.

==================================================

#include <stdio.h>
#include <stdlib.h>
#include <mpi.h>

void tcalloc(int **p,int n)
{
 *p = (int*)calloc(n,sizeof(int));
}


int main(int argc, char * argv[])
{
 int i;
 int nto_global = 0;
 int np = 1;
 int *propto_global = NULL;
 int *propto = NULL;
 int  nto = 10;
 int counts[1];
 int displs[1];

 MPI_Init(&argc,&argv);

 tcalloc(&propto,nto+1);

 for (i = 1; i <= nto; i++) propto[i] = i;

 counts[0] = nto;
 displs[0] = 1;


the displacements
 // are supposed to be the number of items, not bytes.

 for (i = 0; i < np; i++)
   {
     nto_global += counts[i];
   }


 tcalloc(&propto_global, nto_global+1);

 for (i = 1; i <= nto; i++) printf("INPUT %d --> %d\n",i,propto[i]);

 MPI_Allgatherv(propto+1,
                nto,
                MPI_INT,
                propto_global,
                counts,
                displs,
                MPI_INT,
                MPI_COMM_WORLD);

 for (i = 1; i <= nto_global; i++) printf("OUTPUT %d -->
%d\n",i,propto_global[i]);

 free(propto_global);
 free(propto);

 return(0);
}

The following code demonstrates the problem; make sure you run it as one // BUG: if I set displs[0] above to 4, things work as expected. But,

--
=========================================================
Daniel G. Hyams
Associate Research Professor
UT SimCenter at Chattanooga
Email: daniel-hy...@utc.edu
Phone: 423-425-5491
Fax:   423-425-5517
=========================================================

_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel

smime.p7s
Description: S/MIME cryptographic signature