Re: Octave for Mac with Reference Lapack

[Lukas Reichlin]

On 31.05.2012, at 18:48, Alexander Hansen wrote:

> On 5/30/12 2:41 PM, Alexander Hansen wrote:
>> On 5/30/12 4:03 AM, c. wrote:
>>> 
>>> On 29 May 2012, at 23:19, Lukas Reichlin wrote:
>>> 
>>>> 
>>>> I'm surprised to see it working, also on my machine. I fear it doesn't 
>>>> help much, as Accelerate 10.6 used to be the best out of three:
>>>> 
>>>> 1. Accelerate OSX 10.6
>>>> 2. ATLAS 3.9 from MacPorts
>>>> 3. Accelerate OSX 10.7
>>>> 
>>>> I don't have access to 10.7 this evening, and when I build the variant 
>>>> 
>>>>    octave-devel +atlas +fltk +gcc45
>>>> 
>>>> from the attached MacPorts portfile, mkoctfile returns
>>>> 
>>>> nothing for  $(shell $(MKOCTFILE) -p LAPACK_LIBS)
>>>> framework veclib for  $(shell $(MKOCTFILE) -p BLAS_LIBS)
>>> 
>>> this seems to indicate that although you specified atlas as an option your 
>>> Octave was linked 
>>> against the Accelerate framework rather than ATLAS, I'm not an expert in 
>>> portfile syntax so 
>>> I don't see where the problem is there.
>>> 
>>> To further check whether Octave is actually linking to vecLib you should do 
>>> something like:
>>> 
>>> otool -L /opt/octave/3.7/lib/octave/3.7.0+/liboctave.dylib 
>>> 
>>> what I get is the following:
>>> 
>>> octave/3.7.0+/liboctave.dylib:
>>>     /opt/octave/3.7/lib/octave/3.7.0+/liboctave.1.dylib (compatibility 
>>> version 2.0.0, current version 2.1.0)
>>>     /opt/octave/3.7/lib/octave/3.7.0+/libcruft.1.dylib (compatibility 
>>> version 2.0.0, current version 2.0.0)
>>>     /opt/arpackng/3.0.2/lib/libarpack.2.dylib (compatibility version 3.0.0, 
>>> current version 3.0.0)
>>>     /opt/qrupdate/1.1.1/lib/libqrupdate.1.dylib (compatibility version 
>>> 0.0.0, current version 0.0.0)
>>>     /sw/lib/libreadline.5.dylib (compatibility version 5.0.0, current 
>>> version 5.2.0)
>>>     /sw/lib/ncurses/libncurses.5.dylib (compatibility version 5.0.0, 
>>> current version 5.0.0)
>>>     /opt/pcre/8.20/lib/libpcre.0.dylib (compatibility version 1.0.0, 
>>> current version 1.1.0)
>>>     /usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current 
>>> version 125.2.11)
>>>     /System/Library/Frameworks/Accelerate.framework/Versions/A/Accelerate 
>>> (compatibility version 1.0.0, current version 4.0.0)
>>>     /usr/lib/libstdc++.6.dylib (compatibility version 7.0.0, current 
>>> version 7.9.0)
>>>     /System/Library/Frameworks/vecLib.framework/Versions/A/vecLib 
>>> (compatibility version 1.0.0, current version 268.0.1)
>>> 
>>> I think your problem might rather be with mixing up different LAPACK 
>>> implementations rather than choosing the wrong one.
>>> 
>>> 
>>>> Lukas
>>> c.
>> 
>> I'm in the process of collecting data from my own systems
>> (control-2.3.50 and Octave 3.6.1)  to post.  Real life keeps getting in
>> the way, though. :-)
>> 
>> I've mandated in Fink that Octave Forge packages which use BLAS or
>> LAPACK use the same provider as does Octave, so there shouldn't be an
>> opportunity for any mixing.
> 
> I've uploaded the test_control log files here:
> 
> http://akh.users.finkproject.org/test_control/
> 
> Notes:
> 
> 1)  They cover Fink's currently supported platforms:  10.5/powerpc,
> 10.5/i386, 10.6/i386, 10.6/x86_64, 10.7/x86_64 (10.5/x86_64 is
> quasi-supported and I'm in the process getting data for that, too).
> 
> 2)  Octave 3.4.3+ won't build in Fink on 10.5/x86_64 or 10.6/x86_64
> using Accelerate because of configure-time failures involving the
> interaction of gfortran and Accelerate.  powerpc, i386, and 10.7 don't
> seem to have this issue.
> 
> 3)  Most of the atlas-based results are for atlas-3.9.11 because that
> was our current version until a couple of days ago.  The exception for
> 10.7 was because I have a machine that was too new for atlas-3.9.11.
> I'll upload results from 10.7/x86_64/atlas-3.9.11 for completeness,
> however, since I have an older machine where atlas-3.9.11 works.
> 
> -- 
> Alexander Hansen, Ph.D.
> Fink User Liaison
> http://finkakh.wordpress.com/2012/02/21/got-job/

Wow, thanks for your work! In general, your results look familiar to me. I 
experienced failing algorithms with Accelerate 10.7 as in

http://akh.users.finkproject.org/test_control/control-10.7-x86_64-Accelerate.txt

I hope we can circumvent these problems with a reference LAPACK variant once 
and for all.

Best regards,
Lukas