New Sparse Packages

classic Classic list List threaded Threaded
4 messages Options
Reply | Threaded
Open this post in threaded view
|

New Sparse Packages

David Bateman-3
Dear All,

Just for information, I just received the attached e-mail from Tim
Davis. It describes several developments that are useful for the octave
sparse toolbox, including better reordering code, and long versions of
the existing code (ie. support 64 bit), block triangular factorization
code and new cholesky factorization code. This means that the missing
symbfact, chol, dmperm functions in the octave sparse code can be
implemented with this code. It also appears that this is one of the
fastest sparse cholesky solvers out there.. Tim Davis also supplied mex
interfaces to his code that is going to make writing the octave code
much easier...

The downside is that all of his code is now in a single package
UFSparse, and so this will mean the upstream package of in particular
UMFPACK will need to be looked into. Additionally UFSparse introduces a
dependency on the package METIS.

Cheers
David

--
David Bateman                                [hidden email]
Motorola Labs - Paris                        +33 1 69 35 48 04 (Ph)
Parc Les Algorithmes, Commune de St Aubin    +33 1 69 35 77 01 (Fax)
91193 Gif-Sur-Yvette FRANCE

The information contained in this communication has been classified as:

[x] General Business Information
[ ] Motorola Internal Use Only
[ ] Motorola Confidential Proprietary


I thought you might like a short heads-up about a new
code I just released: CHOLMOD, a sparse supernodal Cholesky
factorization and update/downdate package.  Details will be in an
upcoming NA-Digest article ( http://www.netlib.org/na-digest-html/ ).
It will be the new x=A\b in MATLAB when A is sparse and symmetric
positive definite.  The package is under a mix of GNU LGPL and GPL;
other licenses are also available.

Thanks,
Tim Davis
http://www.cise.ufl.edu/~davis
Reply | Threaded
Open this post in threaded view
|

Re: New Sparse Packages

Andy Adler
David,

This sounds like good news. That new code, especially with 64 bit
versions will be a nice advantage to octave. While I agree that
moving to a new package is a little bit of pain. It will be nice
to have all the sparse routines in one package.

Have you yet had time to figure out wither this will change any
of the Sparse class in octave. ie. will those of us who write
sparse code need to plan for any changes?

Thanks
--
Andy Adler <[hidden email]> 1(613)562-5800x6218

On Thu, 1 Sep 2005, David Bateman wrote:

> Dear All,
>
> Just for information, I just received the attached e-mail from Tim
> Davis. It describes several developments that are useful for the octave
> sparse toolbox, including better reordering code, and long versions of
> the existing code (ie. support 64 bit), block triangular factorization
> code and new cholesky factorization code. This means that the missing
> symbfact, chol, dmperm functions in the octave sparse code can be
> implemented with this code. It also appears that this is one of the
> fastest sparse cholesky solvers out there.. Tim Davis also supplied mex
> interfaces to his code that is going to make writing the octave code
> much easier...
>
> The downside is that all of his code is now in a single package
> UFSparse, and so this will mean the upstream package of in particular
> UMFPACK will need to be looked into. Additionally UFSparse introduces a
> dependency on the package METIS.
>
> Cheers
> David
>
> --
> David Bateman                                [hidden email]
> Motorola Labs - Paris                        +33 1 69 35 48 04 (Ph)
> Parc Les Algorithmes, Commune de St Aubin    +33 1 69 35 77 01 (Fax)
> 91193 Gif-Sur-Yvette FRANCE
>
> The information contained in this communication has been classified as:
>
> [x] General Business Information
> [ ] Motorola Internal Use Only
> [ ] Motorola Confidential Proprietary
>
>

Reply | Threaded
Open this post in threaded view
|

Re: New Sparse Packages

David Bateman-3
Andy Adler a écrit :

>David,
>
>This sounds like good news. That new code, especially with 64 bit
>versions will be a nice advantage to octave. While I agree that
>moving to a new package is a little bit of pain. It will be nice
>to have all the sparse routines in one package.
>
>  
>
Apparently UMFPACK will continue as a seperate package as well, but I
especially want to use the new CHOLMOD package that is not clear whether
it will be seperate or not. Inclusion of a dependency on METIS is also a
bit of a pain. But hey I don't do packaging, so I tell myself that is
someone elses problem :-)

There was already a 64 bit version of UMFPACK, but what is missing is a
64-bit COLAMD function.

>Have you yet had time to figure out wither this will change any
>of the Sparse class in octave. ie. will those of us who write
>sparse code need to plan for any changes?
>  
>
Not at all.. The only visible change to existing functionality will be
that it will add the Cholesky factorization to the "\" operator, and so
the warning about cholesky factorization not being present will go away,
and it won't fall through to UMFPACK in that case. Apart from that, the
other things this code allows is
 
* 64-bit COLAMD
* implementation of the chol, symbfact and dmperm functions
* Inclusion of the new CCOLAMD functions

So if what is there is what you use, then there should be no need for
changes...

>Thanks
>  
>
I haven't touched the code yet, just downloaded it, and probably won't
touch it for a little while, so thanks are premature :-)

Cheers
David

Reply | Threaded
Open this post in threaded view
|

Re: New Sparse Packages

David Bateman-3
In reply to this post by Andy Adler
Andy Adler a écrit :

>David,
>
>This sounds like good news. That new code, especially with 64 bit
>versions will be a nice advantage to octave. While I agree that
>moving to a new package is a little bit of pain. It will be nice
>to have all the sparse routines in one package.
>
>Have you yet had time to figure out wither this will change any
>of the Sparse class in octave. ie. will those of us who write
>sparse code need to plan for any changes?
>  
>
Ok, so I lied it will introduce another user visible change. The
interface to colamd and symamd changed, in that the definition of the
second argument to colamd and symamd changed. This is used when
determining which rows/cols are to be considered as dense and moved to
the end of the reordering.

I checked in an updated version of Fcolamd, Fsymand and Fetree that can
be built for 64-bit systems. Check the help of colamd and symamd for the
new definition of "knobs"...

Cheers
David