Re: Eigensystem[] bug in Mathematica 5.0 (fixed in 5.0.1)
- To: mathgroup at smc.vnet.net
- Subject: [mg48500] Re: Eigensystem[] bug in Mathematica 5.0 (fixed in 5.0.1)
- From: Roland Franzius <roland.franzius at uos.de>
- Date: Wed, 2 Jun 2004 04:22:15 -0400 (EDT)
- Organization: Universitaet Hannover
- References: <c9ccon$6ha$1@smc.vnet.net> <c9ec93$ga3$1@smc.vnet.net>
- Sender: owner-wri-mathgroup at wolfram.com
Peltio wrote:
> "Marcus Stollsteimer" wrote
>
>
>>for those of you that work with complex matrices and
>>have not upgraded to 5.0.1 so far, this might be interesting:
>>
>>Eigensystem[] does not give the correct eigenvalues /
>>eigenvectors for complex matrices bigger than 256x256.
>>
>>(this bug has been reported to WRI and is fixed in 5.0.1)
>
>
> I wonder if someone in WRI could shed some light on the reason that
> allegedly made this a bug only on Pentiums and not on AMD's processors. Was
> the code detecting the CPU type and using the SSE instructions (faster but
> with lower internal precision w.r.t. the 80 bit x87 FPU operations)?
>
> I am insolently curious. : )
>
> cheers,
> Peltio
> invalid address in reply-to. Crafty demunging required to mail me.
>
The Athlon beats Intels Pentium 4 using a radical new floting point
achitecture with 88 named registers wich can be renamed instead of
shuffling their contents. This greatly enhances the two standard
routines used in all floating point calculations, written in terms of
mathematica
for series
Plus@@ ( data . (Power[x,#]&/@ Range[0,Length[data]-1]))
for linear algebra
Map[ Dot[#,data1]& , data2, {depth}]
Look eg
http://www.azillionmonkeys.com/qed/cpujihad.shtml
The two architectures demand different codes and probably even different
basic algorithms in order to be efficient.
--
Roland Franzius