[MPIWG Fortran] [mpiwg-p2p] FP16 16-bit-floating-point

[Jeff Hammond]

No change to the MPI standard is required here.  MPI_Type_create_f90_real
is already sufficient.

"This function returns a *predefined* MPI datatype that matches a REAL
variable of KIND selected_real_kind(p, r)." - I do not see why we need to
enumerate a predefined datatype name associated with float16 in Fortran for
this statement to be true.  An implementation can return a datatype with
the appropriate properties without being enumerated like MPI_REAL.

"It is erroneous to supply values for p and r not supported by the
compiler." - Conversely, the user should be allowed to supply values that
are supported by the compiler (although the standard doesn't guarentee it
today).  MPI implementations might not support every type the compiler
does, as is already the case for Fortran float128 (because MPI libraries
are written in C, and do not call back into Fortran to do math on these
types, which may not be identical to C extensions like __float128), but I
consider that an implementation defect.

Thus, I propose to merely add the text that allows users to rely on
MPI_Type_create_f90_real for float16 in Fortran, provided that the compiler
supports it.  It will then be up to implementations to figure out how to do
math on these types internally.

While we are on this subject, we should be considering the more general
inadequacy of the current state of the MPI type system from the perspective
of reductions.  See https://github.com/mpi-forum/mpi-issues/issues/18 for
details.

As for ISO C/C++, I can talk to Clark Nelson if a better candidate is not
identified, but Lawrence Crowl is leading the C++ effort and I suspect the
C folks will follow, although I will confirm this with Clark.

Jeff

On Fri, Jun 16, 2017 at 6:16 AM, Kawashima, Takahiro <
t-kawashima at jp.fujitsu.com> wrote:

> Hi,
>
> I don't know the status of C standardization nor contact person
> but the suggestion of 'short float' was seem to made in a WG.
>
>   http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0192r0.pdf
>   http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2016.pdf
>
> Regards,
> Takahiro Kawashima,
> Fujitsu
>
> > Tony,
> >
> > we try to find the type names used in C and Fortran before we define
> > the new predefined MPI datatype handles for representing those
> > language types.
> >
> > Who knows, whom we can contact on the C standardization side?
> >
> > Best regards
> > Rolf
> >
> > ----- Original Message -----
> > > From: "Anthony Skjellum" <tony at runtimecomputing.com>
> > > To: "MPI WG Point To Point Communications working group" <
> mpiwg-p2p at lists.mpi-forum.org>
> > > Cc: "Reinhold Bader" <Reinhold.Bader at lrz-muenchen.de>, "Atsushi Hori"
> <ahori at riken.jp>, "fortran standards email list
> > > for J3" <j3 at j3-fortran.org>, "MPI-WG Fortran working group" <
> mpiwg-fortran at lists.mpi-forum.org>
> > > Sent: Friday, June 16, 2017 2:50:00 PM
> > > Subject: Re: [MPIWG Fortran] [mpiwg-p2p] FP16 16-bit-floating-point
> >
> > > Rolf, apart from the Fortan angle, may I strongly encourage that we
> adopt this
> > > datatype for MPI-4. It is part of real MPI codes already, but now we
> have to
> > > fake transfer datatypes as 16-bit integers or bytes... which means not
> really
> > > right if heterogeneous interoperability were to be maintained. And
> this is a
> > > canonical IEEE type now, as you note. There is also the "2 half-float"
> datatype
> > > (32-bits) at least in CUDA to be looked at; maybe ignored or maybe
> supported,
> > > but it is there too. Key to this MPI users and applications use half
> floats and
> > > 2 half-floats :-) in practice.
> > >
> > > May I also suggest a forward look at the types in AVX-512, in case any
> of these
> > > become particularly relevant in future.
> > > At least 8-bit integers are conceptually new, apart from BYTE. [
> > > https://en.wikipedia.org/wiki/AVX-512 | https://en.wikipedia.org/wiki/
> AVX-512 ]
> > >
> > > Regards,
> > > Tony
> > >
> > >
> > > On Fri, Jun 16, 2017 at 6:52 AM, Rolf Rabenseifner < [
> > > mailto:rabenseifner at hlrs.de | rabenseifner at hlrs.de ] > wrote:
> > >
> > >
> > > Hi all together,
> > >
> > > I'm sitting in the MPI Forum Meeting and we discuss the new IEEE FP16
> > > datatype.
> > >
> > > Questions to the Fortran atandardization body:
> > >
> > > - Does the Fortran Standardization Body has already a proposal for
> > > the new name in Fortran for this IEEE FP16?
> > >
> > > - Will be REAL*2 identical to IEEE FP16?
> > >
> > > Best regards
> > > Rolf
> _______________________________________________
> mpiwg-p2p mailing list
> mpiwg-p2p at lists.mpi-forum.org
> https://lists.mpi-forum.org/mailman/listinfo/mpiwg-p2p
>



-- 
Jeff Hammond
jeff.science at gmail.com
http://jeffhammond.github.io/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpiwg-fortran/attachments/20170616/88d8d9de/attachment.html>