py-cvxopt +DSDP

[Jonathan Stickel]

On 7/31/14 17:27 , Mark Brethen wrote:
>
> On Jul 31, 2014, at 3:10 PM, Jonathan Stickel <jjstickel at gmail.com>
> wrote:
>
>>
>>
>> On 7/29/14 19:48 , Mark Brethen wrote:
>>>
>>> On Jul 28, 2014, at 11:56 AM, Jonathan Stickel
>>> <jjstickel at gmail.com> wrote:
>>>
>>>>> Date: Sun, 27 Jul 2014 13:47:20 -0500 From: Mark Brethen
>>>>> <mark.brethen at gmail.com> Subject: py-cvxopt +DSDP
>>>>>
>>>>> py-cvxopt installed fine without any errors. However, when I
>>>>> tried running example code it reports:
>>>>>
>>>>> Symbol not found: _ATL_cGetNB Referenced from:
>>>>> /opt/local/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/cvxopt/lapack.so
>>>>>
>>>>>
>>>>>
>>
>>>>>
Why does numpy and cvxopt use Apple's Accelerate framework by default 
and DSDP use macport's ATLAS?
>>>>>
>>>>> Ticket #37097 describes issues building DSDP and other ports
>>>>> that depend on it (like cvxopt) due to linker issues with
>>>>> building against macports' ATLAS. I uninstalled DSDP and
>>>>> ATLAS and then patched DSDP as suggested in the ticket, but
>>>>> when I tried to install DSDP again it failed because it could
>>>>> not find Apple's Accelerate framework:
>>>>>
>>>>> :info:build /usr/bin/clang -O3 -arch x86_64 -o maxcut
>>>>> maxcut.o
>>>>> /opt/local/var/macports/build/_Users_marbre_ports_math_DSDP/DSDP/work/DSDP5.8/lib/libdsdp.a
>>>>>
>>>>>
-latlas -lm :info:build ld: library not found for -latlas
>>>>> :info:build clang: error: linker command failed with exit
>>>>> code 1 (use -v to see invocation)
>>>>>
>>>>> I guess if you do this, you also need to provide the path to
>>>>> Apple's Accelerate framework?
>>>>
>>>>
>>>> It seems that DSDP does not work with Apple's Accelerate
>>>> framework, and hence if you use the +dsdp variant for
>>>> py-cvxopt (not required), then you must use also use the +atlas
>>>> variant for py-cvxopt. I suppose it is a bug that this conflict
>>>> is not flagged in the py-cvxopt portfile. You could open a
>>>> ticket for that part, if you like. In the meantime, just
>>>> reinstall with consistent variants (I recommend using +atlas
>>>> for all of the above, including numpy and scipy).
>>>>
>>>> Regards, Jonathan
>>>
>>> I was able to build DSDP with Apple's Accelerate.framework by
>>> patching 'make.include' at this line:
>>>
>>> -LAPACKBLAS  = -llapack -lblas -lg2c -lm +LAPACKBLAS  =
>>> -Wl,-framework -Wl,Accelerate -lm
>>>
>>> and test that it works with the following python program:
>>>
>>> from cvxopt.base import matrix,spmatrix from cvxopt import
>>> solvers
>>>
>>> clst=[0.0, 0.0, 0.0, 0.0, 0.0, 2.0, 2.0, 2.0] hlst=[0.0, 0.0,
>>> 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
>>> 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
>>> 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -1.0] Gr=[0, 7, 14, 21, 28, 1,
>>> 5, 8]+[35]*5 Gc=[0, 1, 2, 3, 4, 5, 6, 7]+range(5)
>>> Gmtr=spmatrix([1.0]*8+[-1.0]*5,Gr,Gc,(36,8)) c=-matrix(clst)
>>> Gs=[-Gmtr] h=[-matrix(hlst,(6,6))] try:
>>> sol=solvers.sdp(c,Gs=Gs,hs=h,solver='dsdp') print 'optimal
>>> value: %.2f'%(1+(matrix(clst,(1,8))*sol['x'])[0],) except: print
>>> 'could not use dsdp' sol=solvers.sdp(c,Gs=Gs,hs=h) print 'optimal
>>> value: %.2f'%(1+(matrix(clst,(1,8))*sol['x'])[0],) raise
>>>
>>> which returns: 'optimal value: 2.00'
>>>
>>>
>>> I'd suggest updating the port to use Apple's
>>> accelerated.framework by default (as does numpy and cvxopt).
>>>
>>>
>>> Mark
>>
>> Based on this info, I patched the DSDP portfile to use accelerate
>> framework by default (i.e., no variants) and then to have an atlas
>> variant. I uploaded patches to the ticket:
>>
>> https://trac.macports.org/ticket/37097
>>
>> Jonathan
>
>
> Was there a particular reason you added the reinplace in the octave
> variant? Won't this be replaced by the patch?
>

Its not in the octave variant block (the way trac shows the patch makes 
it look that way). See the ticket for more text. I suggest you cc your 
email on the ticket, and we can continue to discuss (as needed) there.

Thanks,
Jonathan