py-cvxopt +DSDP

[Jonathan Stickel]

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