[Scilab-Dev] Scilab5.3 SSE3

[Eduardo Tarasiuk]

*This message was transferred with a trial version of CommuniGate(r) Pro*
Thanks for your answer Sylvestre

It was a very good idea to allow the download of this SSE version,
I will continue my tests, you are right, it is not so simple, but I have
experience in programs I compiled with the SSE4.2 flag (in addition to
the optimization) using the Intel compiler and received an improvement
of at least 5 times in comparison without this single flag, Scilab seems
to behave in the same direction.
Please, keep this SSE version, I think that it is very important for the
vast majority of the users.

       Regards

       Eduardo 

-----Original Message-----
From: Sylvestre Ledru [mailto:sylvestre.ledru at scilab.org] 
Sent: ו 07 ינואר 2011 18:23
To: dev at lists.scilab.org
Subject: Re: [Scilab-Dev] Scilab5.3 SSE3

*This message was transferred with a trial version of CommuniGate(r)
Pro*
Le vendredi 07 janvier 2011 à 07:52 +0200, Eduardo Tarasiuk a écrit :
> *This message was transferred with a trial version of CommuniGate(r)
Pro*
> Hello,
> 
> This week I downloaded the Last Scilab 5.3 version compiled with the
> SSE3  commands.
> According to my very preliminary performance tests for the commands I
> checked, I can see that the improvement is amazing, even better than
in
> Matlab.
> My question is, why this binary is not the default to be downloaded ?
> 
> If the source is unmodified, different builds, in order to receive the
> best possible performance (IMHO) is very recommended, and can give
> Scilab a push over other Mathematical tools.
> 
> SSE3 for every modern existing CPU,  and SSE4.2 for the i3,i5,i7 and
> Xeon55XX, Xeon56XX  on both Windows and Linux will give a considerable
> performance  improvement to Scilab, and is only a different build
(i.e.:
> compilation flags) keeping the same central source untouched.
> I think that to the most of the users there is no important from where
> the performance comes : 1] Algorithm Improvement 2] Compilation Flags,
> isn’t ?
Performance analyzes are very complicated. It depends which operating
system you run, which BLAS/LAPACK libraries you are using (MKL is very
good but proprietary), which compilation flags have been used, etc

For Scilab binaries, we have to find a good compromise between
performances and portability.
We wished we could set SSE3 extensions as mandatory but we receive, from
to time, complaint from people who think forcing Pentium III (SSE) is
already too much...

In the future, we might release an HPC version of Scilab...

Sylvestre



-- 
-------------------------
Sylvestre Ledru
-------------------------
The Scilab Consortium
Digiteo
Domaine de Voluceau
Rocquencourt - B.P. 105
78153 Le Chesnay Cedex
France