[mpiwg-p2p] FP16 Support

William Gropp wgropp at illinois.edu
Thu Jun 22 00:03:31 CDT 2017

On timing, note that MPI already defined optional types.  One of these is MPI_REAL2, which is a 2-byte floating point type - that is, FP16. (See p25, line 36, 177, line 2, 540 line 10, and 674 line 38).  Was MPI_REAL2 discussed?


William Gropp
Interim Director and Chief Scientist, NCSA
Thomas M. Siebel Chair in Computer Science
University of Illinois Urbana-Champaign

> On Jun 21, 2017, at 9:17 PM, Atsushi HORI <ahori at riken.jp> wrote:
> Hello,
> Since I am a newcomer to this working group (and any of other WGs), I would like to start this e-mail with self introduction. I am Atsushi Hori (just call me 'Atsushi') at Riken and in charge of Flagship-2020 project to develop the post-K computer. Note that the project name do NOT suggest the year when the post-K computer will be installed.
> For some reason, I am responsible to push FP16 into the MPI standard and the straw vote in the MPI Forum meeting in Aachen ended up to have a new ticket for FP16.
> I was very amazed that hot discussions were carried out immediately after the Forum meeting was over. I appreciate Rolf, Kawashima-san, Jeff and others who gave me very important information.
> Let me clarify several, non-technical, points before start making FP16 draft;
> * WG
>  It is the Martin's decision to discuss this ticket in this PT2PT WG.  This is just because there is no appropriate WG to discuss this issue.  Rolf discussed this with the Fortran WG too.  Shall we discuss this issue with the Fortran WG too ?
> * Timing
>  Rolf suggested that this issue must be "voted in only after the C standardization gets confirmed."  In my opinion, all standardization effort of C (and Fortran), IEEE 754 (binary representation), and MPI are "independent."  The compiler implementations are also independent. So, in theory, MPI standard may have FP16 independent from the others.  In practice and my opinion, (final?) voting must take place when GCC supports FP16.
> * FP128
>  Jeff mentioned about FP128.  Unlike FP16, FP128 is already supported by GCC on many CPU architectures.  Were there any discussions on FP128 ever ?  I think it might be a good idea to have both FP16 and FP128 in the future release of MPI standard. Or should it be discussed in a separate ticket ?
> Any comments are welcome.
> -----
> Atsushi HORI
> ahori at riken.jp
> http://aics-sys.riken.jp
> _______________________________________________
> mpiwg-p2p mailing list
> mpiwg-p2p at lists.mpi-forum.org
> https://lists.mpi-forum.org/mailman/listinfo/mpiwg-p2p

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-p2p/attachments/20170622/c7919202/attachment.html>

More information about the mpiwg-p2p mailing list