Thanks Jeff for the quick response.
my knowledge of mpi datatypes is actually limited to mpi_type_vector and
mpi_type_hvector. I'm sure there is more to datatypes than just that. I'll
have to look into it, I appreciate the advice.
Here is a snippet of a test code that illustrates where I'm stuck right now
integer recvbuf(10,5), sendbuf(5,5), rcounts(2), displ(2)
include "mpif.h"
displ=(/ 0,1 /)
rcounts=1
...
...
sendbuf=rank
call mpi_type_vector(5,5,5,mpi_integer,stride,err)
call mpi_type_commit(stride,err)
call mpi_gatherv(sendbuf,25,mpi_integer,&
recvbuf, &
rcounts,displ,stride, &
0,mpi_comm_world,err)
and here is the output
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1
any idea?
----- Original Message -----
From: "Jeff Squyres" <jsquyres_at_[hidden]>
To: "General LAM/MPI mailing list" <lam_at_[hidden]>
Sent: Friday, April 15, 2005 8:47 PM
Subject: Re: LAM: MPI_gather for 3D topology
> MPI_GATHER can be used to put data in this order, but you need to think
> about your data layout in memory, and structure your use datatypes and
> MPI_GATHER accordingly.
>
> Keep in mind that the datatype that you gather *from* does not need to be
> the same datatype that you gather *to*. Specifically, the datatypes
> provided on the root and non-root processes do not need to be the same.
> They must be equivalent -- essentially meaning that the resulting number
> of bytes and basic data elements are the same, but the layout in memory
> could be different.
>
> I realize that I'm not giving you much of an answer :-(, but datatypes are
> a rather complex issue; I would strongly suggest spending a little time
> with an MPI book or a tutorial to get at least the basics of complex
> datatypes. They can give much more insight and detail than I can in an
> e-mail.
>
> Hope that helps!
>
>
> On Apr 15, 2005, at 7:05 PM, Khaled Al Salem wrote:
>
>> Hello,
>> I wrote an f90 mpi code to solve a 3D problem. The topology is
>> obtained using mpi_cart_creat. At the end of the calculations I wish
>> to get the results gathered from all processors into one big 3
>> dimensional array that can be directly visualized without going
>> through post-processing. The problem is that mpi_gather doesn't
>> gather the results from the different proccessors as blocks.
>> for example, in 2D, if proc. 1 has the following
>>
>> 1 1 1
>> 1 1 1
>> 1 1 1
>>
>> and proc. 2
>>
>> 2 2 2
>> 2 2 2
>> 2 2 2
>>
>> then mpi_gather results in the following
>>
>> 1 1 1 1 1 1
>> 1 1 1 2 2 2
>> 2 2 2 2 2 2
>>
>> the results that I'm after would be
>>
>> 1 1 1 2 2 2
>> 1 1 1 2 2 2
>> 1 1 1 2 2 2
>>
>> is there an easy way to do this?
>> _______________________________________________
>> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
>>
>
> --
> {+} Jeff Squyres
> {+} jsquyres_at_[hidden]
> {+} http://www.lam-mpi.org/
>
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
>
>
|
