[gmx-users] Gromacs 4.6 crashes w/AVX acceleration on cygwin

[Szilárd Páll]

PS: If you have fftw in the standard location it should be picked up by
cmake; otherwise, -DCMAKE_PREFIX_PATH=/usr/local should work as well.




--
Szilárd


On Wed, Feb 27, 2013 at 9:21 PM, Szilárd Páll <szilard.pall at cbr.su.se>wrote:

> Hi,
>
> Thanks for the detailed information!
>
> The only reason why cygwin is cumbersome is that we are not testing it
> actively, have we had a developer or contributor working with cygwin there
> would be no issues.
>
> In any case, especially because there is performance involved, we should
> try to get the AVX issue solved. First and foremost, would you mind filing
> a redmine issue on the issue (or if you consider the dll problem annoying,
> feel free to file a second one).
>
> However, before filing the AVX bug report, please try the following:
> - A newer gcc, preferably 4.7; although in *nix 4.5 seems to work with
> AVX, it's best if we have things tested with the latest gcc.
> - SSE4.1 acceleration; beside that this should be faster than SSE2, but if
> that doesn't work either something is seriously broken.
>
> If you manage to get performance numbers with the above, please provide
> them, here for the sake of completeness. I would be surprised if VS2010
> would outperform gcc 4.7, though.
>
> Cheers,
>
> --
> Szilárd
>
>
> On Wed, Feb 27, 2013 at 8:31 PM, Mirco Wahab <
> mirco.wahab at chemie.tu-freiberg.de> wrote:
>
>> Out of curiosity, I installed cygwin  1.7.17-1 on my AMD/FX-8350
>> workplace box (W8Pro/x64). This is what I found out:
>>
>> - Gromacs 4.6 compiles smooth but crashes w/native(AVX_128) on
>>    simulation start with loads of fp errors (non-reproducible,
>>    sometimes stops w/no error at all, just stops)
>>
>> - Gromacs 4.6 runs fine if compiled w/GMX_CPU_ACCELERATION=SSE2,
>>    even if PME involved (see below)
>>
>> - Gromacs shared executables (runtime-dlls) didn't start due to
>>    some erratic dlopen problems with I didn't understand. By
>>    using GMX_PREFER_STATIC_LIBS=ON, this was solved
>>
>> I testet a "Martini"-system of ~1 Mio particles (a 50^3nm water
>> box with a vesicle). Performance comparison:
>>
>>   VS2010(x64)      - md/cut-off/Reaction-Field : 37.460 ns/day
>>   cygwin/gcc 4.5.3 - md/cut-off/Reaction-Field : 30.304 ns/day
>>
>>   VS2010(x64)      - md/cut-off/PME : 10.874 ns/day
>>   cygwin/gcc 4.5.3 - md/cut-off/PME :  9.283 ns/day
>>
>> The fftw3f used was installed before from source. I'll attach
>> a recipe for both builds and some more hints below.
>>
>> Resumé: cygwin installation of Gromacs 4.6 is (still) cumbersome
>> and may throw up suprising errors. For windows users, looks much
>> more promising (imho!) to build Gromacs with the Visual Studio
>> compiler (which is, through combination of VS2010 express +
>> SDK 7.1, free of cost).
>>
>> "works on my machine"-instructions: =>
>>
>>
>> --
>> gmx-users mailing list    gmx-users at gromacs.org
>> http://lists.gromacs.org/mailman/listinfo/gmx-users
>> * Please search the archive at
>> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
>> * Please don't post (un)subscribe requests to the list. Use the
>> www interface or send it to gmx-users-request at gromacs.org.
>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>>
>
>