py-cvxopt +DSDP

[Mark Brethen]

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?


Mark