[NotiAMCA] Fwd: [Abmec-l] Solvers

[Victorio Sonzogni]

----------  Mensaje reenviado  ----------

Asunto: [Abmec-l] Solvers
Fecha: Lun 16 Feb 2009
De: rssr en lncc.br
Para: abmec-l en dees.ufmg.br, cpg-l en lists.lncc.br, cmc-l en lists.lncc.br



Caros 

Os links abaixo podem ser uteis

Sairam na 
NA Digest  Sunday, February 15, 2009  Volume 08 : Issue 07

Abracos 
Renato Silva

-------------------------------------------------------

From:
Yvan Notay <ynotay en ulb.ac.be>
Date: Fri, 13 Feb 2009 07:53:53
-0500
Subject: Software for solving eigenvalue problems

We
are pleased to announce release 2.0 of JADAMILU software package.

JADAMILU is a Fortran 77 program to compute selected eigenvalues and
associated eigenvectors of large scale real symmetric or complex
Hermitian matrices.  Generalized eigenvalue problems with positive
definite mass matrix can also be solved. The eigenvalues sought can
either be the smallest ones or those closest to some specified target
inside the spectrum.  The software is based on the Jacobi-Davidson
method. Key features are modest memory requirements and robust
convergence to accurate solutions.  What makes JADAMILU different
from
many other excellent eigenvalue solvers is that it provides
automatic
setup of an in-built preconditioner.  Moreover, if needed,
the
preconditioner is automatically adapted as the computation
proceeds.

The package is easy to install (we provide
pre-compiled libraries for
most architectures, including Windows
based systems), and the driver
routines have be designed to be
easy-to-use by non-experts.
If the matrix is sparse and explicitly
available, the user can simply 
pass it to the program and let
JADAMILU compute the desired eigenvalues. 
If not, then a crude
sparse approximation which may even be diagonal, 
can be passed. In
the latter case the user only needs to define a routine 
that
performs matrix-vector products.  

New features in
JADAMILU_2.0:
- Routines for generalized eigenvalue problems
-
Support for complex arithmetic to solve Hermitian (generalized or
standard)
  eigenvalue problems 
- All routines available in
double precision, double complex, single precision
  and single
complex 

For further details and download, see
http://homepages.ulb.ac.be/~jadamilu/

Matthias Bollhoefer &
Yvan Notay

-------------------------------------------------------

From:
Sergey Golobov <sergey.golobov en gmail.com>
Date: Tue, 10 Feb
2009 22:09:37 -0500
Subject: New efficient cluster solvers for
separable PDE are available

New Intel(R) Cluster Poisson Solver
Library (Intel(R) CPSL) is a powerful 
tool set for solving separable
Partial Differential Equations (PDE). It 
offers a set of 2D solvers
for Poisson-like problems, as well as an example 
of a 3D Poisson
solver. Intel(R) CPSL provides an advanced implementation of 
the
modern algorithms. For more experienced users, Intel(R) CPSL offers 
insight into the solvers sufficient to finer tune them for better 
performance. The Intel(R) CPSL is currently available only for the
Linux* 
OS. 

Intel(R) CPSL offers:
2D Helmholtz
solver 
2D Poisson solver 
2D Laplace solver 
Example of a
3D Poisson solver 

The solvers employ fast Fourier transforms
for solving PDE problems with 
separable variables. The main goal for
releasing the solvers to the public  
is to get feedback on what you
do or do not like about them. 

Free version of the Intel(R)
CPSL can be downloaded from 
http://whatif.intel.com , or from
http://software.intel.com/en-us/articles/intel-cluster-poisson-solver-library

-------------------------------------------------------

From: "Field G. Van Zee" <field en cs.utexas.edu>
Date: Tue, 10 Feb 2009 19:14:34 -0600
Subject: libflame: The
Complete Reference

We are happy to announce the birth of a
user's guide to our libflame
software package. libflame is a dense
linear algebra library that has
resulted from the FLAME project. It
is a significant step towards a
modern replacement of the classic
LAPACK library and can be used as a
substitution for the major
routines in that library and/or as an
infrastructure for developing
new functionality.  While it boasts
performance that rivals or
surpasses that of LAPACK, it emphasizes
programmability. libflame
also includes prototype routines that allow
matrices to be stored by
blocks (even with arbitrary-depth
hierarchies), as well as a runtime
system that extracts high-level,
dependency-aware parallelism from
major operations on multithreaded
architectures. We believe libflame
is also a useful pedagogical tool.

This document, titled
"libflame: The Complete Reference", can be
downloaded for
free from

 
http://www.cs.utexas.edu/users/flame/publications/

Complete
instructions on how to download, configure, build, link with
the
library can be found in the document, as well as detailed
descriptions of all supported routines.

Comments on the
document as well as the library itself are welcomed
and
appreciated.

Also, we are pleased to announce that we now
support both GNU/Linux and
Microsoft Windows.

Field G. Van
Zee
for the FLAME team