Re: [atlas-devel] ATLAS for Freescale T4240: patches.

[José L. G. P. <jgp...@gm...>]

2013/4/17 <wh...@cs...>

> Brooks,
>
> >Here are the patches that I used to get ATLAS to compile on our
> >Freescale T4240 -- hopefully they'll be useful to someone, and I expect
> >at least the first one or two of these are useful to include in the
> >ATLAS trunk.
> >
> >The first patch addresses the fact that, on the T4240, the processor
> >speed lines in /proc/cpuinfo look like this:
> >
> >> clock           : 1666.666650MHz
> >
> >ATLAS currently expects to find an integer at the end of the relevant
> >line, and so it detects "666650" as the speed, and rightly considers
> >that an error.
> >
> >This patch adjusts the detection code to look for a float value at the
> >end of the relevant line, and thus correctly detects the right value. To
> >do that, I needed to add a GetLastDouble function (since there wasn't
> >already one); it is based closely on GetFirstDouble.
> >
> >
> >The second patch simply adds the e6500 architecture type to the list of
> >architectures, adds code to detect it, and code in a couple of places to
> >treat it correctly as a Power architecture.
>
> I have applied all these (not exactly, but things that should be
> equivalent).
> However, I have no access to any such machine to verify these patches.
> Does someone have access to this type of machine that can confirm/deny
> my patches work when I release 3.11.9 (not yet, because I can't presently
> pass make check on my own desktop, due to some unrelated work I'm in the
> middle of)?
>

Hello:

Sorry for the offtopic, but as you mention the next unstable release
3.11.9, will you consider to add the feature requested here?:
http://sourceforge.net/p/math-atlas/feature-requests/79/

Thanks


>
> >The third patch I mentioned a couple of weeks ago, as a way to get ATLAS
> >to correctly detect the number of available registers since the
> >performance drop when it oversubscribes them is smaller than the
> >detected threshold.  This is undoubtably a hack, but at least a
> >functional one.
>
> You for got the dis in front of functional :)  The threshhold you put is
> so low it will trigger by accident on most machines, unfortunately.
>
> Usually, when you don't see a significant drop in performance once the
> number
> of register is exceeded, this is due to the compiler producing such
> terribly
> optimized code for the architecture, that you can't see the cost of
> additional
> spills inside the loop.  In gcc's case for a normal POWER, I think it
> should
> produce decent enough code that you can see it, so I'm guessing that maybe
> you need to find some better flags for the search.  This might help:
>    http://math-atlas.sourceforge.net/atlas_install/node20.html
>
> In the meantime, this is certainly not the only architecture for which
> ATLAS's
> nreg search is not reliable, so what I've done instead of the above is add
> a
>   -r #
> flag to configure.  If this is set, ATLAS will just use that number of
> registers rather than probing for it.  I'm currently testing that this
> fully works, and it looks good so far.  So, this functionally should fix
> allow you to install once I release the next dev.
>
> Cheers,
> Clint
>
>
> Thanks,
> - Brooks
>
> **************************************************************************
> ** R. Clint Whaley, PhD ** Assoc Prof, UTSA ** www.cs.utsa.edu/~whaley<http://www.cs.utsa.edu/%7Ewhaley> **
> **************************************************************************
>
>
> ------------------------------------------------------------------------------
> Precog is a next-generation analytics platform capable of advanced
> analytics on semi-structured data. The platform includes APIs for building
> apps and a phenomenal toolset for data science. Developers can use
> our toolset for easy data analysis & visualization. Get a free account!
> http://www2.precog.com/precogplatform/slashdotnewsletter
> _______________________________________________
> Math-atlas-devel mailing list
> Mat...@li...
> https://lists.sourceforge.net/lists/listinfo/math-atlas-devel
>



-- 
*****************************************
José Luis García Pallero
jgp...@gm...
(o<
/ / \
V_/_
Use Debian GNU/Linux and enjoy!
*****************************************