|
|
Ublas : |
From: Nico Galoppo (nico_at_[hidden])
Date: 2006-02-15 22:38:40
Just as a heads up, I for one would be very much in favour of having
block-matrix computations to be more natural with ublas. I've subclassed some
specialized matrix classes (such as 6x6), but always find myself having to use
complicated subrange() and project() calls very often.
--nico
Vardan Akopian wrote:
> A quick note and a question.
> For block matrices to be possible with ublas, we also need to make
> sure that there are no implicit assumptions of the element (scalar)
> product being commutative, i.e. a*b = b*a.
> Can anyone confirm or deny the presence for these kind of assumptions?
>
> -Vardan
>
> On 2/15/06, Riccardo Rossi <rrossi_at_[hidden]> wrote:
>
>>Dear Gunther,
>> first of all thanks for your quick answer, i have however a doubt
>>concerning your proposal:
>>
>>first of all let's consider the code
>>Matrix A(3,3);
>>
>>Vector b(3);
>>
>>Vector c(3);
>>
>>c = A * b;
>>
>>using boost 1.32.0 ( i still didn't upgrade to 1.33.1) i get an error
>>message due to the existance of two candidates for the overloaded operator *
>>... the error message reads
>>
>>..\..\..\..\external_libraries\boost\boost_1_32_0\boost\numeric\ublas\vector_expression.hpp(1468):
>>could be
>>'boost::numeric::ublas::vector_binary_scalar1_traits<E1,E2,F,C1>::result_type
>>boost::numeric::ublas::operator
>>*<Kratos::Matrix,boost::numeric::ublas::vector<T>>(const T1 &,const
>>boost::numeric::ublas::vector_expression<E> &)'
>>
>>with
>>
>>[
>>
>>E1=const Kratos::Matrix,
>>
>>E2=boost::numeric::ublas::vector<double>,
>>
>>F=boost::numeric::ublas::scalar_multiplies<Kratos::Matrix,double>,
>>
>>C1=boost::numeric::ublas::scalar_reference<const Kratos::Matrix>,
>>
>>T=double,
>>
>>T1=Kratos::Matrix,
>>
>>E=boost::numeric::ublas::vector<double>
>>
>>]
>>
>>..\..\..\..\external_libraries\boost\boost_1_32_0\boost\numeric\ublas\matrix_expression.hpp(3631):
>>or
>>'boost::numeric::ublas::matrix_binary_scalar2_traits<E1,E2,F,C2>::result_type
>>boost::numeric::ublas::operator
>>*<boost::numeric::ublas::matrix<T>,Kratos::Vector>(const
>>boost::numeric::ublas::matrix_expression<E> &,const T2 &)'
>>
>>with
>>
>>[
>>
>>E1=boost::numeric::ublas::matrix<double>,
>>
>>E2=const Kratos::Vector,
>>
>>F=boost::numeric::ublas::scalar_multiplies<double,Kratos::Vector>,
>>
>>C2=boost::numeric::ublas::scalar_reference<const Kratos::Vector>,
>>
>>T=double,
>>
>>E=boost::numeric::ublas::matrix<double>,
>>
>>T2=Kratos::Vector
>>
>>]
>>
>>while trying to match the argument list '(Kratos::Matrix, Kratos::Vector)'
>>
>>\kra
>>
>>
>>in any case even if i define univocally my operator *, as i understand the
>>operation
>>
>>c = A * b
>>
>>is not efficient as a temporary is created for each multiplication, while
>>the correct form would be
>>
>>noalias(c) = prod(A,b);
>>
>>for me efficiency is crucial ...
>>
>>concerning the Symmetric Product wouldn't it work if i define i binary prod
>>Vector Vector as a scalar prod?
>>i mean
>>trans(a)*a = inner_prod(a,a)?
>>
>>okkk
>>
>>thanks for the attention
>>
>>grettings
>> Riccardo
>>
>>
>>
>>----- Original Message -----
>>From: "Gunter Winkler" <guwi17_at_[hidden]>
>>To: "ublas mailing list" <ublas_at_[hidden]>
>>Sent: Wednesday, February 15, 2006 12:53 PM
>>Subject: Re: [ublas] blocked products and relation UBLAS/GLAS
>>
>>
>>
>>>On Tuesday 14 February 2006 19:37, Riccardo Rossi wrote:
>>>
>>>>dear list,
>>>> i am more or less a Newbie at using ublas for sparse operations
>>>>and i would love a little advice on some delicate questions.
>>>>
>>>>in particular i would like to work with a sparse matrix in which all of
>>>>the
>>>>single terms are vectors or matrices instead of double. All the
>>>>containers
>>>>are templated matrices which implies that i am free to do it
>>>>nevertheless...
>>>
>>>You can use the matrix and vector classes as a container for any type
>>>provided
>>>the type is default and zero constructable. That means you need a
>>>constructor
>>>like
>>>
>>>my_type(const int zero) : ...
>>>{ assert( 0 == zero ); }
>>>
>>>which is necessary to let the sparse types know what "zero element" means.
>>>
>>>
>>>>how can i make the library to know how to perform the multiplication
>>>>between all of the different terms? the case is for example of a
>>>>SparseMatrix of Matrices to be multiplied by a vector of vectors.
>>>
>>>the easiest way is to use a custom type as value type which behaves like a
>>>scalar. That means you have to overload the usual arithmetic operators for
>>>your type.
>>>
>>>
>>>>Similarly i would like to create a matrix doing something like A =
>>>>trans(B)*B (i think it is called Symmetric Product)... is it available
>>>>in
>>>>ublas?
>>>
>>>this is more complicated because ublas assumes that a scalar is symmetric
>>>and
>>>therefore never calls trans(x) for any scalar x. Maybe you can create a
>>>new
>>>specialization of
>>>
>>>matrix_unary2_traits (in matrix_expression.hpp, lines 1648ff)
>>>
>>>that defines a new result type.
>>>
>>> // (trans m) [i] [j] = m [j] [i]
>>> template<class E>
>>> BOOST_UBLAS_INLINE
>>> typename matrix_unary2_traits<const E,
>>> scalar_identity<typename E::value_type> >::result_type
>>> trans (const matrix_expression<E> &e) {
>>> typedef typename matrix_unary2_traits<const E,
>>> scalar_identity<typename E::value_type>
>>> >::expression_type expression_type;
>>> return expression_type (e ());
>>> }
>>>
>>>(you have to replace scalar_identity by scalar_transpose and write you own
>>>scalar_transpose, see functional.hpp - ask me if you need help doing
>>>this.)
>>>
>>>
>>>>finally i would like to know what is the relation of ublas to GLAS
>>>
>>>GLAS is (will be) a general linear algebra library supporting different
>>>back
>>>ends, e.g. ublas. Maybe Karl can better comment on that.
>>>
>>>mfg
>>>Gunter
>>>_______________________________________________
>>>ublas mailing list
>>>ublas_at_[hidden]
>>>http://lists.boost.org/mailman/listinfo.cgi/ublas
>>>
>>>
>>
>>
>>_______________________________________________
>>ublas mailing list
>>ublas_at_[hidden]
>>http://lists.boost.org/mailman/listinfo.cgi/ublas
>>
>
>
> _______________________________________________
> ublas mailing list
> ublas_at_[hidden]
> http://lists.boost.org/mailman/listinfo.cgi/ublas
--
nico galoppo von borries @ address: 105 creel st., chapel hill
comp. graphics phd. student @ north carolina, 27516 USA
UNC, chapel hill @ phone: +1 (919) 962-1898 (office)
@ +1 (919) 942-7609 (home)
@ email: nico at cs dot unc dot edu
@ homepage: http://www.cs.unc.edu/~nico
--- debian linux :: vim powered