[gmx-developers] Branches w/working OpenCL support

[Mark Abraham]

Hi Mirco,

I've solved three correctness issues with the OpenCL implementation now,
and NPT on AMD looks quite nice. Can you please try the code at
https://gerrit.gromacs.org/#/c/4314/26 and let us know if the situation is
improved?

Thanks!

Mark

On Tue, Jun 9, 2015 at 9:34 PM Szilárd Páll <pall.szilard at gmail.com> wrote:

> Hi Mirco,
>
> Thanks for running the tests and sharing the logs. Performance looks
> reasonable and based on your log file it looks like the large OpenCL
> runtime overhead I observed on Linux is not present on Windows.
>
> On Sun, Jun 7, 2015 at 12:23 PM, Mirco Wahab
> <mirco.wahab at chemie.tu-freiberg.de> wrote:
> > Hi Szilard,
> >
> > On 31.05.2015 22:15, Szilárd Páll wrote:
> >>
> >> [...]
> >> https://gerrit.gromacs.org/#/c/4314/
> >> Use the version in review on gerrit I linked above, that's the most up
> >> to date code.
> >> [...]
> >
> >
> > I completed some tests using an AMD device (Pitcairn-GPU,
> > Radeon R9-270X), and used adh_cubic_vsites, rnase_cubic, and
> > villin_vsites from the "gromacs acceleration page".
> >
> > So far, the simulations appear to run stable and deliver correct
> > results if *no pressure coupling* is used. As I noted already with
> > the NV opencl test, this affects the AMD opencl implementation too:
> > pressure coupling just don't work so far.
> >
> > Compilation: there needs to be an additional include of
> > #include <algorithm> in src/gromacs/gmxana/gmx_wham.cpp
> > in order to provide std::min/max for VC. Another problem
> > is the use of binary numerals in src\gromacs\listed-forces\bonded.cpp,
> > which are C++14. Maybe these could be written in hexadecimal
> > numerals in order to avoid problems w/non C++14-compilers?
>
> I can't comment much on these, but I don't think we should be using C++14
> code.
>
> > Pressure coupling: Depending on the simulation setup w/pressure
> > coupling, gmx will  either crash immediately (when using pme) or
> > continuously enlarge the volume (eg. when using rf + berendsen).
>
> Yeah, that should not happen. I did very brief testing, and my guess
> is that something is wrong with the shift forces during non-energy
> calculating steps.
>
> > I put the log files an a host, these include log files and
> > input for a very simple spc water system w/pressure coupling
> > (which doesn't work w/opencl but works perfectly using -nb cpu).
> >
> http://spule.compch.tu-freiberg.de/~wahab/gromacs/gmx_opencl_2015-06_szilard.zip
> >
> > short overview:
> >
> >  adh_cubic_vsites
> >   cpu/pme:      5.738,  gpu/pme:        17.421
> >   cpu/rf:       7.913,  gpu/rf:         26.344
> >
> >  rnase_cubic
> >   cpu/pme:      15.001, gpu/pme:        39.363
> >   cpu/rf:       24.729, gpu/rf:         59.179
> >
> >  villin_vsites
> >   cpu/pme:      124.86, gpu/pme:        284.213
> >   cpu/rf:       151.05, gpu/rf:         310.794
> >
> >  (water_small_npt/nvt)
> >   cpu/pme/npt:  20.00   gpu/pme/npt:    n/a
> >   cpu/pme/nvt:  21.65   gpu/pme/nvt:    52.05
> >
> >
> > How is the current status for the planned inclusion of
> > opencl into a gromacs release?
>
> I can't comment on that, there are a few issues (including this
> pressure bug) that need to be solved before merge can be considered.
> Hopefully somebody will have time to look at these.
>
> --
> Szilárd
>
> > Regards,
> >
> >
> > M.
> >
> > --
> > Gromacs Developers mailing list
> >
> > * Please search the archive at
> > http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List before
> > posting!
> >
> > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> >
> > * For (un)subscribe requests visit
> > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers
> or
> > send a mail to gmx-developers-request at gromacs.org.
> --
> Gromacs Developers mailing list
>
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List before
> posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers
> or send a mail to gmx-developers-request at gromacs.org.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20150617/8cbd8b40/attachment-0001.html>