Cholesky factorization

> I've been reading through the ublas archive, trying to
figure out the
> best way to tackle Cholesky factorization. I see that a
number of people
> have come up with custom solutions-have any of those
solutions been
> integrated into ublas? Obviously, it's preferable to
use the
> functionality if it's already in the library, rather
than cut-and-paste
> someone else's code as a custom solution.
> 
>  
> 
> Thanks,
> 
> --Steve (sgrosssjm.com)
> 

Hi!

>From a performance point of view it is adviseable to use
the boost::bindings library to link against some optimized
version of LAPACK, i.e. ATLAS.

HTH,

Benjamin

http://math-atlas.
sourceforge.net/

http://
news.tiker.net/software/boost-bindings

_______________________________________________
ublas mailing list
ublaslists.boost.org
htt
p://lists.boost.org/mailman/listinfo.cgi/ublas

On 3/19/07, Benjamin Sobotta <maydaygmx.de> wrote:

> From a performance point of view it is adviseable to
use the
> boost::bindings library to link against some optimized
version of LAPACK,
> i.e. ATLAS.

I wonder, should one expect significant performance
differences
between these three versions, while implementing something
like
Cholesky factorization:

1. F77/LAPACK using ATLAS for BLAS procedures (essentially
no uBlas)
2. uBlas using ATLAS (only for BLAS procedures)
3. pure uBlas

for

a. small (N < 10)
b. medium (N = 10-100)
c. large (N > 1000)

matrix sizes?

I remember seeing some message in the archives, saying uBlas
is very
slow (incidentally, the performance benchmark may have been
Cholesky
factorization), but I wasn't sure if the poster remembered
to turn the
debugging off and the optimizations on.
_______________________________________________
ublas mailing list
ublaslists.boost.org
htt
p://lists.boost.org/mailman/listinfo.cgi/ublas

Gunther's example implementation is probably a pretty good
start for generic 
sparse cholesky decomposition:

http://www.bauv.unibw-muenchen.de/~winkler/ublas/examp
les/

 From there, depending on how much you know about the
structure of your 
matrices, you can try to optimize sub-parts with
boost::bindings and 
BLAS/LAPACK/ATLAS.

HTH,

--nico

Gross, Steve wrote:
> I’ve been reading through the ublas archive, trying to
figure out the 
> best way to tackle Cholesky factorization. I see that a
number of people 
> have come up with custom solutions—have any of those
solutions been 
> integrated into ublas? Obviously, it’s preferable to
use the 
> functionality if it’s already in the library, rather
than cut-and-paste 
> someone else’s code as a custom solution.
> 
>  
> 
> Thanks,
> 
> --Steve (sgrosssjm.com <mailto:sgrosssjm.com>)


-- 
nico galoppo von borries     |  address: 119 fidelity st.,
carrboro
comp. graphics phd. student  |           north carolina,
27510 USA
UNC, chapel hill             |  phone: +1 (919) 962-1898
(office)
                              |         +1 (919) 942-4388
(home)
                              |  email: nico at cs dot unc
dot edu
                              |  homepage: http://www.ngaloppo.org

_______________________________________________
ublas mailing list
ublaslists.boost.org
htt
p://lists.boost.org/mailman/listinfo.cgi/ublas

Am Montag, 19. März 2007 17:18 schrieb Gross, Steve:
> I've been reading through the ublas archive, trying to
figure out the
> best way to tackle Cholesky factorization. I see that a
number of
> people have come up with custom solutions-have any of
those solutions
> been integrated into ublas? Obviously, it's preferable
to use the
> functionality if it's already in the library, rather
than
> cut-and-paste someone else's code as a custom solution.

There was a lot of discussion whether uBLAS should contain 
factorizations or not. There are different points to
consider:
1) uBLAS was originally designed as (simple but reliable)
BLAS 
replacement (this means: basic linear algebra + triangular
solvers)
2) matrix factorizations are part of more advanced libraries
as LAPACK, 
and in order to have a competitive performance one has to
consider a 
lot of things - this would exceed the scope of a BLAS lib,
would 
require a lot of work and there are already a lot of good 
implementations available. (for dense matrices: LAPACK, for
sparse 
matrices: UMFPACK, CXSPARSE)

So we decided to only support the necessary data types and
rely on 
bindings.

However, you can use my examples and extend it - so we may
be able to 
include it into ublas. Currently I am about to finish a LU
and ILU 
implementation and to update the triangular solvers.

mfg
Gunter

PS: I am moving my homepage to http://freenet-hom
epage.de/guwi17/
_______________________________________________
ublas mailing list
ublaslists.boost.org
htt
p://lists.boost.org/mailman/listinfo.cgi/ublas

On Tue, Mar 20, 2007 at 09:23:40AM +0100, Gunter Winkler
wrote:
<snip> 
> So we decided to only support the necessary data types
and rely on 
> bindings.

This would be a lot more reasonable (meant in a polite way)
if the bindings
were part of the distribution rather than hidden away in the
sandbox and
often broken.

Jeff

-- 
------------------------------------------------------------
----------------
Hofstadter's Law states: It always takes longer than you
expect, even when
you take into account Hofstadter's Law.
------------------------------------------------------------
----------------

_______________________________________________
ublas mailing list
ublaslists.boost.org
htt
p://lists.boost.org/mailman/listinfo.cgi/ublas

On Tue, Mar 20, 2007 at 03:38:21PM +0100, Gunter Winkler
wrote:
> Jeffrey Brent McBeth schrieb:
> > On Tue, Mar 20, 2007 at 09:23:40AM +0100, Gunter
Winkler wrote:
> > <snip> 
> >   
> >> So we decided to only support the necessary
data types and rely on 
> >> bindings.
> >>     
> >
> > This would be a lot more reasonable (meant in a
polite way) if the bindings
> > were part of the distribution rather than hidden
away in the sandbox and
> > often broken.
> >   
> 
> Any volunteers? I gladly accept patches 

I do have patches, and will get them sent in.

Jeff

-- 
------------------------------------------------------------
----------------
Hofstadter's Law states: It always takes longer than you
expect, even when
you take into account Hofstadter's Law.
------------------------------------------------------------
----------------

_______________________________________________
ublas mailing list
ublaslists.boost.org
htt
p://lists.boost.org/mailman/listinfo.cgi/ublas

In principle, it would be ok to have the
ublas-specializations
(numeric/bindings/traits/ublas_vector.hpp and matrix.hpp) of
the
bindings in the ublas distribution. The bindings themselves
can stay in
the boost-sandbox.

One of the reasons is that people who have write access to
the
boost-sandbox do not necessarily have access to ublas
(myself for example).

Karl



Quoting Jeffrey Brent McBeth <mcbethbroggs.org>:
> On Tue, Mar 20, 2007 at 03:38:21PM +0100, Gunter
Winkler wrote:
> > Jeffrey Brent McBeth schrieb:
> > > On Tue, Mar 20, 2007 at 09:23:40AM +0100,
Gunter Winkler wrote:
> > > <snip> 
> > >   
> > >> So we decided to only support the
necessary data types and rely
> on 
> > >> bindings.
> > >>     
> > >
> > > This would be a lot more reasonable (meant in
a polite way) if
> the bindings
> > > were part of the distribution rather than
hidden away in the
> sandbox and
> > > often broken.
> > >   
> > 
> > Any volunteers? I gladly accept patches 
> 
> I do have patches, and will get them sent in.
> 
> Jeff
> 
> -- 
>
------------------------------------------------------------
----------------
> Hofstadter's Law states: It always takes longer than
you expect, even
> when
> you take into account Hofstadter's Law.
>
------------------------------------------------------------
----------------
> 



_______________________________________________
ublas mailing list
ublaslists.boost.org
htt
p://lists.boost.org/mailman/listinfo.cgi/ublas

Quoting Kresimir Fresl <freslgrad.hr>:
> Mike Hayford wrote:
> 
> > There is an additional issue that stands in the
way of usage for
> some of 
> > us, and that it that the bindings don't appear to
be covered by the
> 
> > Boost license. The header files have the right
kind of reassuring
> words 
> > in them, but what is reassuring to me and what is
reassuring to my
> 
> > company's attorneys are two different things.
Would Kresimir Fresl,
> Toon 
> > Knapen and Karl Meerbergen consider changing the
bindings to the
> Boost 
> > license?
> 
> I agree.
> 
> fres
> 

OK for me.

Karl
_______________________________________________
ublas mailing list
ublaslists.boost.org
htt
p://lists.boost.org/mailman/listinfo.cgi/ublas