This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: valarray
>>>>> «David», David Banks <dbanks@cislunar.com> wrote:
David> Hi,
David> Do you have a reference or further explaination for this statement?
David> As near as I can tell, Blitz could be extremely useful as the compilers
David> are catching up to it. Maybe you see a problem in using it with unstructured/
David> adaptivly refined meshes while it could come in very handy for structured
David> and overset/chimera) paradigms? Or is there something I'm missing and the
David> results quoted really aren't as good as they appear due to some paralyzing
David> constraint?
Blitz behaves very well for *medium arrays* and is very *clean* from a pure
C++ programmer view point. But (like valarray) blitz was not primarly
designed to serve the scientific computing community *but* to provide
an answer to a problem in C++. In that it succeeded (illustrating use
of expression template). The purpose of valarray was to provide
building blocks but it failed (for ex. almost every binary fonction
taking a valarray object an a builtin type does take a *const
reference* on the builtin type which is the most inefficient way to do
things. One might have expected to see a use of traits to decide
value/const ref according to sizeof of the type). Furthermore some
*fondamental* functions are missing (a mathematically sound inner
product on valarrays; why do we have valarray<>::sum() and not
valarray<>::product() ?)
Blitz choises (for reasons I can undestand) not to give the same
meaning to the copy-ctor and the copy-assignment operator. This does
not only encourages aliasing (which prevents parallelism) but also
is error-prone.
Below, I included an e-mail form Dag giving his impression on the
start of the art.
David> -Dave Banks
David> Cislunar Aerospace, inc
David> dbanks@cislunar.com
--
Gabriel Dos Reis | Ecole Normale Superieure de Cachan
INRIA, Unite de Recherche de | Centre de Mathematiques et de Leurs
Sophia Antipolis | Applications
Projet SAGA | Equipe de Geometrie
------- start of forwarded message (RFC 934 encapsulation) -------
Return-Path: <dag@dynasim.se>
Received: from net.dynasim.se (root@net.dynasim.se [192.16.137.60])
by piano.dptmaths.ens-cachan.fr (8.8.8/jtpda-5.2) with ESMTP id PAA24587
for <Gabriel.Dos-Reis@dptmaths.ens-cachan.fr>; Wed, 18 Mar 1998 15:51:49 +0100 (MET)
Received: from dag-nt (dag-nt.dynasim.se [194.103.53.10])
by net.dynasim.se (8.8.8/8.8.8) with SMTP id PAA02644;
Wed, 18 Mar 1998 15:54:55 +0100
Message-Id: <3.0.5.32.19980318155725.00960d30@194.103.53.1>
X-Sender: dag-pop@194.103.53.1
X-Mailer: QUALCOMM Windows Eudora Light Version 3.0.5 (32)
In-Reply-To: <199803171653.RAA12584@piano.dptmaths.ens-cachan.fr>
Mime-Version: 1.0
Content-Type: text/plain; charset="iso-8859-1"
Content-Transfer-Encoding: quoted-printable
X-Filter: mailagent [version 3.0 PL44] for dosreis@piano.ens-cachan.fr
From: Dag Bruck <dag@dynasim.se>
To: Gabriel Dos Reis <Gabriel.Dos-Reis@dptmaths.ens-cachan.fr>
Cc: bs@research.att.com
Subject: Re: Numerics with C++
Date: Wed, 18 Mar 1998 15:57:25 +0100
Dear Gabriel,
At 17:53 1998-03-17 +0100, you wrote:
>
>Hello,
>=09
> I am working on the class valarray from the C++ Standard
>Library. I discussed with Bjarne Stroustrup in Nice (France) last week
>about the implementation of valarray. He gave me your email suggesting=20
>to get in touch with you because you're specialist of numerics in both
>Fortran and C++. So I'm there.
I think that Bjarne Stroustrup was perhaps a little too enthusiastic about
my capabilities, but the company I work for is heavily involved in
efficient numerical simulation and we are right now looking at C/C++ matrix
libraries.
> I would like to know if you're aware of an implementation of
>numerics library (other than Blitz++, TNT) in C++ that exhibits good
>performance. Would it be possible you give an explanation of technics
>you use ?
What is the right library of course depends on your particular needs, and
ours are quite general, including LU and QR decomposition, handling of
sparse matrices etc. Another important aspect is a FORTRAN-compatible
storage format, because our customers want to interface to existing
libraries or code generated by f2c. We have looked at several libraries,
and currently we are probably most interested in a package called Meschach
(ftp://ftpmaths.anu.edu.au/pub/meschach/meschach.html) because of its
overall functionality.
The Blitz++ package is probably the most exciting from a C++ perspective,
and should give very good performance for medium-sized matrices. Maybe
there are problems to get it to run on current compilers. The TNT library
is also quite good, and I know people who have used it with great success.
Here are some additional comments by a colleague of mine, Martin Otter:
- ----------------------------------------------------------------------------=
- -
I downloaded the TNT library when searching for sparse matrix=20
packages, but presently it seems to be that
no useful operations are provided (no sparse LU decomposition,=20
even no dense LU decomposition).
I looked at the packages on which TNT is based. These are packages
for dense matrices (LAPACK) and iterative sparse matrix packages
but no direct sparse matrix package for general matrices.=20
I therefore have the feeling,
that direct sparse matrix support will not be available or if
they plan to do it, it will come quite late, because they will first
implement the stuff they are familiar with.
The problem with iterative sparse matrix schemes is that no methods
seems to be known which work for all types of systems (which is of course
very important for a tool such as Dymola). Direct sparse matrix schemes
seem to be much more robust with respect to the system to be solveed.
On the other hand, very big systems (dim > 100 000) seem to be only
solvable by iterative methods.
- ----------------------------------------------------------------------------=
- -
In our application, symbolic transformations that take advantage of zero
elements results in runtime performance that a pure numerical approach
cannot match. Our product is an object-oriented modeling and simulation
environment for large continuous systems (with discrete events).
If we are primarily interested in implementing valarray, TNT is probably a
good choice.
Best regards,
Dag Br=FCck
- --
Dynasim AB Phone: +46 46 182500
Research Park Ideon FAX: +46 46 182501
S-223 70 Lund E-mail: dag@Dynasim.se
Sweden URL: http://www.Dynasim.se
------- end -------