Re: [AMBER-Developers] Request for testing new cuda code updates

From: David Cerutti <dscerutti.gmail.com>
Date: Tue, 29 Jan 2019 15:58:56 -0500

We've now removed the -volta flag, and there will be no -turing flag
either. The code underneath understands how to convey these optimizations
without having the user ask for it.

On Tue, Jan 29, 2019 at 3:41 PM Jamie Smith <
jsmith.crackofdawn.onmicrosoft.com> wrote:

>
> If you can make the updates to configure2, I'll merge them into the CMake
> system. Does this mean we're increasing the min required CUDA version or
> removing the -volta flag?
> ________________________________
> From: David Cerutti <dscerutti.gmail.com>
> Sent: Tuesday, January 29, 2019 12:18 PM
> To: AMBER Developers Mailing List
> Subject: Re: [AMBER-Developers] Request for testing new cuda code updates
>
> Thanks for spurring action on this. We should even update one more thing
> in configure2, that is commitment to consistent code/arch settings for each
> compiled executable. Thanks to Peng Wang's commits, we are now in
> compliance with the new thread model and no longer have to do this
> sm60/sm70 or sm61/sm75 workaround. There is no speed advantage to it
> anymore, either. Overall, because we were always so good about writing
> code that executed well under warp-synchronous constraints, we lose a bit
> due to the versatility CUDA 9 and 10 have introduced. __syncwarp() is
> tedium for us, but a vital component of many other codes that people would
> like to write on GPUs. The loss is minor, and in time the cards will just
> keep getting bigger and faster, so we'll take this all in stride.
>
> Dave
>
>
> On Tue, Jan 29, 2019 at 10:52 AM David A Case <david.case.rutgers.edu>
> wrote:
>
> > Hi everyone:
> >
> > Dave Cerutti has added some new CUDA code (commit
> > 4596dbe1571a33581f04f), that adds Turing and Volta-related updates to
> > the CUDA codes. It adds support for CUDA 10, and gets optimized codes
> > for Volta and Turing without the need for any user flags (like the
> '-volta'
> > flag that used to be needed).
> >
> > I'm asking developers that use CUDA codes to update their master branch,
> > recompile pmemd.cuda, and see if you find any problems on jobs you are
> > running as a part of your ordinary work. These changes will probably
> > become an update to Amber18, but we want to give them a workout
> > ourselves before unleashing them on the world. (They pass all the test
> > cases, but I'd like to get them exposed to a wider variety of problems as
> > well.)
> >
> > If you are interested in performance issues and have the time, reported
> > before/after timings, especially on Volta and Turning, would be useful.
> >
> > ...thanks...dac
> >
> >
> > _______________________________________________
> > AMBER-Developers mailing list
> > AMBER-Developers.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber-developers
> >
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Tue Jan 29 2019 - 13:00:03 PST
Custom Search