|
|
Ublas : |
From: Brian Budge (brian.budge_at_[hidden])
Date: 2006-08-04 12:23:33
Just to chime in... I also would like to see something like the
classes from tvmet in ublas. tvmet is pretty efficient and a pretty
nice library to work with. I think if the classes could be integrated
in such a way that they would work in the existing ublas framework
would be great. For example, tiny matrices/vectors could work with
other tiny matrices/vectors to be super fast, and could work with
larger matrices/vectors and run at the slower speed.
Brian
On 8/2/06, Riccardo Rossi <rrossi_at_[hidden]> wrote:
> Hi All,
> i admit i am not familiar with how the ublas works at the "low level"
>
> let's consider for example one problem of the type
>
> TinyMatrix<4,3> A;
> TinyMatrix<5,3> B;
> TinyMatrix<4,5> C;
>
> suppose that the operation we would like to do is
> noalias(C) += prod(A,trans(B));
>
> as i understand to obtain maximal performance i should write a
> specialization of the product of the type
>
> prod(TinyMatrix<dim1,dim2>::RowOriented,
> TinyMatrix<dim1,dim2>::ColOriented)
>
> now my first problem is that i don't know WHERE to add this specialized
> product (in matrix_expressions.hpp??????) and how to make it to work in
> practice (i mean the real code for the definition of the function, not
> what is done inside). What is done inside should be less complex once i
> understand this.
>
> there is in any case another thing i am not sure of:
> once in some way i optimized the product operation, am i still paying
> any cost for the
>
> noalias += ...
>
> or i can assume that this is optimized away once i give to the compiler
> fixed size matrices? In a ideal word this should completely disappear as
> all the operations should be done directly on C...but i guess this is
> not the case.
>
> i personally think that the subject is very interesting as it would
> bring a sensible performance gain without any code rewriting to the end
> user...which is clearly the dream of any user (yessssss ... also my
> dream :-D).
>
> ok,
> greetings to everyone
>
> ciao
>
> Riccardo
>
>
>
>
> On Tue, 2006-08-01 at 20:27 +0200, Gunter Winkler wrote:
> > emijah_s_at_[hidden] schrieb:
> > > Could you please elaborate on the last part? We are using uBlas for a robotics application and the
> > > typical size of a matrix tends to be pretty small(around 6x6 typical, 35x35 maximum). So it doesn't
> > > really make sense to use atlas which tends to be effective for large matrices.
> > >
> > > I'd like to try out what you mentioned to see if we can use it to boost performance a bit.
> > >
> > >
> >
> > There were a lot of "mul_bench" examples on this list some time ago. The
> > typical problem of ublas is, that it is hard to give the correct
> > (partial) specilization because the expression templates produce very
> > ugly and long types. The first try would be to write a better
> >
> > rowmatrix prod(const rowmatrix& a, const colmatrix& b)
> >
> > function, but ,as I told, it is hard to chose the correct types for
> > rowmatrix and colmatrix.
> >
> > this would be much easier if rowmatrix and colmatrix have a fixed size
> > which is part of the type.
> >
> > 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
>