[Coin-ipopt] Problem using lapackpp and IPOpt

[Steven Dirkse]

Sonia Singhal wrote:
> Hi,
>  
> Thanks for this info. I set the cw to the one in which my program was 
> running properly.
> Apparently, IPOpt has just the INEXACT flag on, whereas LAPACKPP has all 
> the flags you mentioned on.
>  
> If I turn on all the flags, it doesnt crash in the lapackpp function calls.
> But I was wondering if this is ok ? If there was an underflow (or 
> zerodivide, etc) in the calculations, doesnt it mean something went 
> wrong in the calculations? Even though I am getting the right answer.
>  
> Thanks,
> ~ sonia 

Interesting that your IPOPT runs with only the INEXACT exception masked. 
  I'd like to know if and why they expect DENORMAL and UNDERFLOW 
exceptions to be unmasked.

In general, you can hope that code developers set the exception mask for 
a reason.  For example, if they want to get +/- infinity when overflow 
occurs (i.e. their code is written to work with these values) they can 
just mask OVERFLOW exceptions.  If they don't expect to get reasonable 
results when working with infinity, they might not mask OVERFLOW.  So 
I'd say only the code developers (or those reading the code carefully) 
can say what exceptions ought to be masked.

-Steve

> On 12/16/06, *Steven Dirkse* <sdirkse at gams.com 
> <mailto:sdirkse at gams.com>> wrote:
> 
>     Sonia Singhal wrote:
>      > Hi,
>      >
>      > I am doing development in Microsoft Visual Studio 2003. I have been
>      > using LAPACK through a C++ interface provided in lapackpp-2.5.0.
>     I was
>      > using it for solving linear system of equations and doing SVD. This
>      > package provides the libs for lapack and blas which I used for
>     compiling
>      > their c++ interface package.
>      > Recently I used IpOpt ( 3.2.3 ) version in the code to solve an
>      > optimization problem. I havnt written the complete code yet, just
>     made a
>      > dummy optimization class just to see if I was able to link the
>     entire
>      > ipopt infrastructure with my existing code. I was able to compile and
>      > run the Cpp Example given in the IPOpt distribution.
>      > The problem is that my code crashes in the lapackpp's call to their
>      > interface to SVD (dgesdd). It crashes giving a "floating point
>      > underflow" error. This doesnt happen if I remove ipopt's libs and
>     code.
>      > I have posted this question on the lapackpp forum too. I am
>     posting here
>      > too just in case anyone has any idea of what the problem could be.
>      >
>      > Thanks,
>      > ~ sonia
>      >
> 
>     Sonia,
> 
>     Floating point underflow is a hardware exception that can be raised on
>     most CPUs.  Typically this is masked, so when underflow occurs the
>     result is stored as a denormal (if it isn't one already) if possible or
>     silently flushed to zero.  I'd guess that LAPACK is expecting this
>     setting.  Somehow running with the IPOPT code in there, this exception
>     is unmasked and you LAPACK behaves differently.
> 
>     You can experiment with your code to see what the exception mask is
>     when
>     it works and when it aborts.  Here is a snippet of code to get you
>     started:
> 
>     #if defined(_WIN32)
>     {
>       unsigned int cw;
> 
>       cw = _control87(0,0);
>       if (cw & _EM_INVALID   ) ADD2MASK(EX_INVALIDOP );
>       if (cw & _EM_DENORMAL  ) ADD2MASK(EX_DENORMAL  );
>       if (cw & _EM_ZERODIVIDE) ADD2MASK(EX_ZERODIVIDE);
>       if (cw & _EM_OVERFLOW  ) ADD2MASK(EX_OVERFLOW  );
>       if (cw & _EM_UNDERFLOW ) ADD2MASK(EX_UNDERFLOW );
>       if (cw & _EM_INEXACT   ) ADD2MASK(EX_PRECISION );
>     }
> 
>     I would google _control87 for more info.  FYI, ADD2MASK is a macro of
>     mine, but _EM_INVALID are defined by some system headers.  If you
>     replace ADD2MASK by some print statements you should see what the
>     mask is.
> 
>     -Steve
> 
> 


-- 
Steven Dirkse, Ph.D.    GAMS Development Corporation
1217 Potomac St. NW             Washington DC  20007
Voice: (202)342-0180              Fax: (202)342-0181
mailto:sdirkse at gams.com          http://www.gams.com