This is the mail archive of the
fortran@gcc.gnu.org
mailing list for the GNU Fortran project.
Re: new proposal to vector intrinsics in Fortran
- From: Sa Liu <SALIU at de dot ibm dot com>
- To: FX <fxcoudert at gmail dot com>
- Cc: fortran at gcc dot gnu dot org, "Toon Moene" <toon at moene dot indiv dot nluug dot nl>, Ulrich Weigand <Ulrich dot Weigand at de dot ibm dot com>
- Date: Thu, 29 Nov 2007 18:51:20 +0100
- Subject: Re: new proposal to vector intrinsics in Fortran
Hi,
> Please don't start pinging to ask for answers less than 24 hours after
> your initial mail! We (maintainers) read and will reply, but you can't
> possibly expect an answer in a matter of hours.
>
> Taking a bit of time to think it through, in fact, is radically
> different than replying at the very minute you receive the email. I
> can't speak for other but I have read your mail thoroughly and have
> started a draft reply, but will send it after giving it some more
> thoughts (and after looking up a few things in the code). Probably
> other maintainers are doing the same. In any case, we're all
> volunteers with day jobs that do not involve GCC/gfortran, and will
> have different time constraints than yours: e.g., I will probably
> reply this week-end, when I get time to do it.
Very sorry for the impatient ping... I was just afraid that this thread
would die again like it's predecessor. If this is not the case, please
just take your time.
>
> We certainly will make a decision on that. However, the time window
> for that decision will be more than a few days, you have to realize
> that. (Also, talking about timescales, I hope that you're aware of GCC
> development rules and of the fact that this new feature, if included,
> will only be incorporated in the forthcoming 4.4 development branch).
I'm aware of that. I don't expect to get this upstreamed overnight. For me
it's just important to know whether this is reasonable for gfortran and
could be included in a certain furture release, so that it's worth to put
more efforts into it.
>
> In the meantime, here are a few points that I wanted to raise:
>
> -- I think that vector types are a interesting feature because of
> the recent evolution of processors, but I'd like to understand better
> what currently exist in that domain: you already told us (with links
> to appropriate documentation) what the IBM compiler does, but what
> about other compilers?
>
> -- What is the community demand for it? Who would use it? (Who
> already uses the IBM extensions?) It needs to be worth incorporating
> and maintaining.
>
> -- [**] If there is community demand for this, it will end up in a
> later Fortran standard: have there already been such proposals or
> plans? We want to make sure that, if we create a functionality, it is
> as close as possible to a future standard proposal.
> For the last two points, I think a nice thing would be to ask on
> comp.lang.fortran to gather opinions. Some very knowledgeable and
> influential people hang there, so I think that's a necessary step. How
> does that sound to you?
Thanks for pointing these out. I rarely thought of including it to Fortran
standard. I got the request from customer of IBM CELL/B.E. blade and the
ESSL library, who want to benefit from the SIMD instructions on CELL
architecture. And I think it could also be interesting for the other
platforms supporting vector computing. I have seen similar requests
appeared on comp.lang.fortran, and I'll try to ask for more opinions
there.
Thank you again for your reply!
Sa