[Mpi3-rma] Call for errata

Hubert Ritzdorf Hubert.Ritzdorf at EMEA.NEC.COM
Wed Jan 30 08:19:46 CST 2013


Hi Jim

I think that the ticket 355 is not correct:

(*) The original text was correct:
      The target buffer must fit into the target window.

(*) The second part of ticket 355 is not correct.
       A call

    MPI_Get_accumulate (origin_addr, 1, MPI_INT,
                                            result_addr, 1, MPI_INT,
                                            target_rank, target_disp, 1000, MPI_INT, 
                                            MPI_SUM, win)

      would violate

and the *target* buffer must fit in the result buffer without truncation.

     This is part of 2nd part of Ticket 355.  
     The size of the target buffer is 1000 int's, the origin and result buffer 1 int.

My proposal for the text would be:

The target buffer must fit in the target window or in attached memory in a dynamic window.
target_datatype and result_datatype must not specify overlapping entries.
The origin buffer must fit in target buffer and result buffer without truncation, if op is not equal to MPI_NO_OP..
Otherwise, the target buffer must fit in the result buffer without truncation.

Hubert.
________________________________________
From: mpi3-rma-bounces at lists.mpi-forum.org [mpi3-rma-bounces at lists.mpi-forum.org] on behalf of Hubert Ritzdorf [Hubert.Ritzdorf at emea.nec.com]
Sent: Tuesday, January 29, 2013 9:24 PM
To: MPI 3.0 Remote Memory Access working group
Subject: Re: [Mpi3-rma] Call for errata

Hi Jim,

you are right. Datatypes created by MPI_TYPE_CREATE_F90_{INTEGER,REAL,COMPLEX}
are defined as predefined nevertheless they are handles. MPI_WCHAR was only a joke to
a previous discussion.

I think ticket 355 doesn't clarify this. It doesn't specify that target_count items of target_datatype
are transferred if MPI_NO_OP is specified. It says only something about (minimal/maximal) size of buffers
nothing about the items transferred and is independent of MPI_NO_OP.
You could interpret it that always target_count items of target_datatype are returned to
the result buffer independent on origin_count and origin datatype (even if op is not MPI_NO_OP).

I think that it is much more clear for the reader if ticket 250 specifies:

When MPI_NO_OP is specified as the operation, the origin buffer arguments are ignored
and target_count elements of type target_datatype are returned from the buffer at offset target_disp
to result buffer result_addr.

Best regards

Hubert

________________________________________
From: mpi3-rma-bounces at lists.mpi-forum.org [mpi3-rma-bounces at lists.mpi-forum.org] on behalf of Jim Dinan [dinan at mcs.anl.gov]
Sent: Monday, January 28, 2013 11:51 PM
To: mpi3-rma at lists.mpi-forum.org
Subject: Re: [Mpi3-rma] Call for errata

Hi Hubert,

Please correct me if I'm wrong, but as far as I can tell, every type
that is internal to the MPI implementation (including Fortran and pair
types) are "predefined".

The F90 types are described as predefined in the standard (pg. 617-618).

MPI_WCHAR is also predefined (pg. 665).

Tables 3.x say "predefined" in the table caption.

This wording was copied from the original MPI_Accumulate text in MPI
2.2, so if predefined is not all-encompassing, we may not be able to fix
this as an erratum.

Re: Ticket 350 -- I think we fixed this with ticket 355.  Could you take
a look at that ticket and see if this clears up your concerns?

https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/355

Thanks for your detailed feedback!
  ~Jim.

On 1/28/13 1:04 PM, Hubert Ritzdorf wrote:
> Hi,
>
> Page 429, Line 41-42 specifies for MPI_Fetch_and_op:
>
> The datatype argument must be a predefined datatype.
>
> I think it should be:
>
> The datatype argument must be a predefined or MPI basic datatype (cf. Section 5.9.2 on page 176).
>
> Reasons:
>
> (*) This would be consistent to the reduce functions and MPI_Compare_and_swap
>        and would allow handles returned by MPI_TYPE_CREATE_F90_INTEGER,
>        MPI_TYPE_CREATE_F90_REAL, or MPI_TYPE_F90_COMPLEX.
>        (which are currently excluded since they are not predefined).
>
> (*) MPI_Compare_and_swap accepts handles created by
>        MPI_TYPE_CREATE_F90_INTEGER (they are contained in Fortran integer category).
>
> (*) This proposed change support also atomic gets of MPI_WCHAR.
>
>
> Additionally, on   Page 430, Line 39:
>                 predefined datatypes
> should be replaced by
>                 basic datatypes
> since Section 5.9.2 defines the  MPI basic datatypes and not the predefined
> datatypes (which are defined in Tables 3.x).
>
> Ticket 350 (MPI_NO_OP in new accumulate functions, origin buffer arguments are ignored) should also
> state that the number of entries transferred is defined by target_count
>
> Hubert
>
> ________________________________________
> From: mpi3-rma-bounces at lists.mpi-forum.org [mpi3-rma-bounces at lists.mpi-forum.org] on behalf of Jim Dinan [dinan at mcs.anl.gov]
> Sent: Monday, January 28, 2013 4:54 PM
> To: MPI 3.0 Remote Memory Access working group
> Subject: [Mpi3-rma] Call for errata
>
> Hi All,
>
> We are assembling the list of RMA errata to be voted on at the March
> meeting.  Currently, we have tickets #347, #348, #350, and #355 (I'm
> working on official wording for these changes and will circulate it soon).
>
> Are there any other errata items that should be included on the agenda
> for the upcoming meeting?
>
> Thanks,
>    ~Jim.
> _______________________________________________
> mpi3-rma mailing list
> mpi3-rma at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
>
>
>   Click https://www.mailcontrol.com/sr/cqZpjZJWCcHGX2PQPOmvUsQN7C0BBstkddyjT4P!KFPMxAFFaa8ieyScSFKfD4GCvgRNeNI84u45W2BPFL9JFA==  to report this email as spam.
>
> _______________________________________________
> mpi3-rma mailing list
> mpi3-rma at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma
>
_______________________________________________
mpi3-rma mailing list
mpi3-rma at lists.mpi-forum.org
http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma

_______________________________________________
mpi3-rma mailing list
mpi3-rma at lists.mpi-forum.org
http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-rma




More information about the mpiwg-rma mailing list