Complex left-divide and SVD give ireproducible results

[Miroslaw Kwasniak]

On Mon, Jul 28, 2008 at 11:38:26AM -0400, John W. Eaton wrote:
> Of the systems I have easy access to, I'm only able to reproduce this
> on one with Octave 2.1.73.  The failure happens on a 32-bit AMD system
> running Debian.  I'm unable to reproduce the problem with Octave 3.0.0
> on the same system.  The difference appears to be that Octave 3.0 is
> linked with
> 
>   liblapack.so.3gf => /usr/lib/liblapack.so.3gf (0xa633d000)
>   libblas.so.3gf => /usr/lib/libblas.so.3gf (0xa627c000)
> 
> which are provided by
> 
>   ii  libblas3gf                 1.2-1.5
>   ii  liblapack3gf               3.1.1-0.4

Reproducible with 3.0.1 on debian lenny amd32 

-  octave3.0 1:3.0.1-4
        liblapack.so.3gf => /usr/lib/atlas/liblapack.so.3gf (0xb638f000)
        libblas.so.3gf => /usr/lib/atlas/libblas.so.3gf (0xb6013000)
- libatlas3gf-base 3.6.0-21.5

This error disappears when I run with libraries provided by:
  libblas3gf     1.2-1.6  
  liblapack3gf   3.1.1-0.5  

LD_PRELOAD=/usr/lib/libblas.so.3gf.0:/usr/lib/liblapack.so.3gf.0 \
/usr/bin/octave-3.0.1

I found an anothor issue - inefficiency of debian's atlas on my Athlon XP
1600+:

A=rand(200,40)+1i*rand(200,40); b=rand(200,1); 
t0=cputime;
x0=A\b;for iter=1:1000, x=A\b; err(iter)=norm(x0-x,Inf); end;
sum(err~=0),cputime-t0

atlas ~10.5 seconds
blas+lapack  ~9.2 seconds

> I'd guess that the problem is in the linear algebra package, not
> Octave.

Yes but I can't remove atlas on my machine because harminv forced it.

Mirek