<br><br><div class="gmail_quote">On Wed, Jun 30, 2010 at 7:13 PM, <span dir="ltr"><<a href="mailto:hess@sbc.su.se">hess@sbc.su.se</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div class="im">> On Wed, Jun 30, 2010 at 5:58 PM, Szilárd Páll<br>
> <<a href="mailto:szilard.pall@cbr.su.se">szilard.pall@cbr.su.se</a>>wrote:<br>
><br>
>> Hi,<br>
>><br>
>> > -Kieee. Does someone know whether this is required for GROMACS? One<br>
>> can<br>
>> also<br>
>> > gain with GCC by adding -ffast-math. Is this safe?<br>
>><br>
>> When testing/debugging the strange performance hit we experienced a<br>
>> while ago with gcc 4.3.3/4 - which turned out to be a problem with<br>
>> expf - I also tried the -ffast-math flag and as far as I remember I<br>
>> got no relevant performance improvement (tested on x86_64). With some<br>
>> gcc versions it even seemed to produce a small performance hit,<br>
>> overall showing -1-2% "improvement".<br>
>><br>
> Yes. I did the same test back then and got 5% improvement for only the<br>
> solve_pme part. So this agrees with your result.<br>
> Of course this is not that much, but since it is no work to activate it, I<br>
> think it is still interested.<br>
><br>
> The question is whether it is safe. Thus is there some code which requires<br>
> strict accordance of the IEEE rules? My understanding is that fast-math<br>
> usually doesn't affect the accuracy. Only if bitwise results are important<br>
> or things are unstable it should be important. I'm not aware of any code<br>
> like this, but I'm not sure at all. If we have any code like this we might<br>
> consider to activate fast-math only for the solve_pme function.<br>
><br>
> Roland<br>
><br>
<br>
</div>I don't think we need ieee for anything.<br>
<br>
But do you still get 5% on solve pme after I added the SSE exp?<br>
I also had an SSE 1/x which gave 1 or 2%.<br>
But solve pme is now so little of the total time that such things<br>
become irrelevant.<br></blockquote><div><br></div><div>Just tested it and I get 30% speed improvement in PME solve with the latest HEAD version when using -ffast-math. This is with GCC 4.4.2.</div><div>And I checked that the new SSE exp function is used.</div>
<div>If you are interested I can run it with profiling to see which code lines cause the difference.</div><div><br></div><div>Roland</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<br>
Berk<br>
<div><div></div><div class="h5"><br>
>><br>
>> Take these metrics with a grain of salt though as my benchmarking was<br>
>> focused on a 10-20% performance hit on PME. Moreover it was done ~1.5<br>
>> months ago so It might be worth to test again.<br>
>><br>
>> --<br>
>> Szilárd<br>
>> --<br>
>> gmx-developers mailing list<br>
>> <a href="mailto:gmx-developers@gromacs.org">gmx-developers@gromacs.org</a><br>
>> <a href="http://lists.gromacs.org/mailman/listinfo/gmx-developers" target="_blank">http://lists.gromacs.org/mailman/listinfo/gmx-developers</a><br>
>> Please don't post (un)subscribe requests to the list. Use the<br>
>> www interface or send it to <a href="mailto:gmx-developers-request@gromacs.org">gmx-developers-request@gromacs.org</a>.<br>
>><br>
><br>
><br>
><br>
> --<br>
> ORNL/UT Center for Molecular Biophysics <a href="http://cmb.ornl.gov" target="_blank">cmb.ornl.gov</a><br>
> 865-241-1537, ORNL PO BOX 2008 MS6309<br>
> --<br>
> gmx-developers mailing list<br>
> <a href="mailto:gmx-developers@gromacs.org">gmx-developers@gromacs.org</a><br>
> <a href="http://lists.gromacs.org/mailman/listinfo/gmx-developers" target="_blank">http://lists.gromacs.org/mailman/listinfo/gmx-developers</a><br>
> Please don't post (un)subscribe requests to the list. Use the<br>
> www interface or send it to <a href="mailto:gmx-developers-request@gromacs.org">gmx-developers-request@gromacs.org</a>.<br>
<br>
--<br>
gmx-developers mailing list<br>
<a href="mailto:gmx-developers@gromacs.org">gmx-developers@gromacs.org</a><br>
<a href="http://lists.gromacs.org/mailman/listinfo/gmx-developers" target="_blank">http://lists.gromacs.org/mailman/listinfo/gmx-developers</a><br>
Please don't post (un)subscribe requests to the list. Use the<br>
www interface or send it to <a href="mailto:gmx-developers-request@gromacs.org">gmx-developers-request@gromacs.org</a>.<br>
</div></div></blockquote></div><br><br clear="all"><br>-- <br>ORNL/UT Center for Molecular Biophysics <a href="http://cmb.ornl.gov">cmb.ornl.gov</a><br>865-241-1537, ORNL PO BOX 2008 MS6309<br>