Re: [ublas] VS2012 inconsistent Debug/Release results with lu_substitute() 1.55 & 1.52

Subject: Re: [ublas] VS2012 inconsistent Debug/Release results with lu_substitute() 1.55 & 1.52
From: oswin krause (oswin.krause_at_[hidden])
Date: 2014-10-26 07:05:23

Hi,

this error occurs most likely, because the test of the numerics is much 
too strict. This just measures an error bound between the elements of 
both matrices. most likely lu_substitute had to divide a big number by a 
small one. (you can print out the diagonal of m to see whether it 
contains very small values)

We have had problems with the UBLAS_TYPE_CHECK as well and in the end 
came to the conlusion that it should be deactivated as it is much too 
strict. It might be good for internal testing with nicely conditioned 
matrices but users can not guarantee that.

On 26.10.2014 02:54, tidswell wrote:
> Thank you Nasos.
>
> The bug Nasos identified crept in when I tried to extract a suitable 
> sample for posting.
>
> But I think I still have a problem:
>  * lu_factorize() is returning 0 (which if I understand API correctly 
> means not singular)
>  * if I reenable the type checks then the upper triangular type check 
> in lu_substitute() fails.
>     (see my inserted #error in code fragment below)
>
> Is it reasonable for the solution not to be upper triangular when the 
> matrix is not singular ?
>
> thanks,
> - JT
>
> template<class M, class E>
>     void lu_substitute (const M &m, matrix_expression<E> &e) {
>         typedef const M const_matrix_type;
>         typedef matrix<typename E::value_type> matrix_type;
>
> #if BOOST_UBLAS_TYPE_CHECK
>         matrix_type cm1 (e);
> #endif
>
>         inplace_solve (m, e, unit_lower_tag ());
>
> #if BOOST_UBLAS_TYPE_CHECK
>         BOOST_UBLAS_CHECK (detail::expression_type_check (prod 
> (triangular_adaptor<const_matrix_type, unit_lower> (m), e), cm1), 
> internal_logic ());
>         matrix_type cm2 (e);
> #endif
>
>         inplace_solve (m, e, upper_tag ());
>
> #if BOOST_UBLAS_TYPE_CHECK
> #error "This is the check that fails in debug!"
>         BOOST_UBLAS_CHECK (detail::expression_type_check (prod 
> (triangular_adaptor<const_matrix_type, upper> (m), e), cm2), 
> internal_logic ());
> #endif
>     }
>
>
> On 21/10/14 02:01, Nasos Iliopoulos wrote:
>> jT,
>> try this:
>> change line:
>> std::copy( &(input_data[0][0]), &(input_data[matSize-1][matSize-1]), 
>> data.begin() );
>>
>> to:
>> std::copy( &(input_data[0][0]), &(input_data[matSize-1][matSize-1]) + 
>> ptrdiff_t(1), data.begin() );
>>
>> (The last iterator should be one past the last element)
>>
>> You might get another response that was sent prior to this and has 
>> some further valid points you may want to keep in mind.
>>
>> -Nasos
>>
>> On 10/20/2014 07:53 AM, tidswell wrote:
>>>
>>> Folx,
>>> ublas is not my specialty, so apologies if Ive done something silly, 
>>> but as I get different results with Debug & Release builds on 
>>> identical code, I think there is a problem in ublas's interaction 
>>> with Visual Studio 2012.
>>> [There may also be problems in my code, but I think its simple 
>>> enough that I expect consistent answers (I view consistently wrong 
>>> as better than inconsistent). ]
>>>
>>> MS Windows 7
>>> Visual Studio 2012
>>> Boost 1.55 or 1.52 (I think also earlier, but I didnt go back that 
>>> far to test)
>>>
>>> Simple usage of Fredrik Orderud's LU decompisition matrix inversion
>>> http://www.crystalclearsoftware.com/cgi-bin/boost_wiki/wiki.pl?LU_Matrix_Inversion 
>>>
>>>
>>> I've disabled _SCL_SECURE_NO_WARNINGS because its not standard C++.
>>>
>>> Ive disabled BOOST_UBLAS_TYPE_CHECK because I wanted to check with 
>>> error control flow caused differences.
>>>
>>> Any guidance on why the results are inconsistent ?
>>>
>>> thanks,
>>> - JT
>>>
>>>
>>>
>>> _______________________________________________
>>> ublas mailing list
>>> ublas_at_[hidden]
>>> http://lists.boost.org/mailman/listinfo.cgi/ublas
>>> Sent to:athanasios.iliopoulos.ctr.gr_at_[hidden]
>>
>>
>>
>> _______________________________________________
>> ublas mailing list
>> ublas_at_[hidden]
>> http://lists.boost.org/mailman/listinfo.cgi/ublas
>> Sent to:tidswell_at_[hidden]
>
> _______________________________________________
> ublas mailing list
> ublas_at_[hidden]
> http://lists.boost.org/mailman/listinfo.cgi/ublas
> Sent to: Oswin.Krause_at_[hidden]