blas error and crash when doing ldivide/rdivide of a scalar by a vector

[Carlo de Falco]

On 13/gen/08, at 00:18, John W. Eaton wrote:

> On 13-Jan-2008, Carlo de Falco wrote:
>
> | > I do not get a bus error or segfault but, on the mac, I see the
> | > error message and octave terminates abruptely I'm not sure whether
> | > you call this a crash or not, but I would be inclined to say  
> so; on
> | > linux, octave does not terminate after the error, so I think in
> | > that case it is not really a crash.
>
> When lapack encounters an unrecoverable error, it calls XERBLA, which
> executes a Fortran STOP statement.  Octave provides its own version of
> this function to avoid calling STOP in cases like this, so apparently
> that is not being used on your system.  There was a suggestion for how
> to fix this problem here:
>
>   http://www.nabble.com/Additions-for-OS-X-to9235901.html#a9235901
>
> but no actual patch.  I don't use OS X so I am not in a position to
> fix this problem myself.

thanks for the pointer.

> | >> For me, the breaking point seems to be N = 116.
> | >
> | > For me that is N = 740, I'm not really sure whether it is of any
> | > importance
> | >
> | >> Also, DGEBRD is not called directly from Octave, so maybe this  
> is a
> | >> bug in LAPACK?
> | >
> | > Any idea how to check?
>
> Write a small example program that does the same calculation with
> DEGLSD that Octave is doing.  I'd expect that you will see the same
> error, independent of Octave.  If not, then maybe there is an error
> int the way we are calling DGELSD in this case.

I'm not an expert in either F77 nor LAPACK (as can be clearly seen by  
the orrible f77 code I produced)
but I think the attached program does the same as the failing octave  
code.

on my mac this seems to work without problems,
I built it with f2c and linked it with VecLib as
I think this is how octave.app is built


> | doesn't matlab rely on LAPACK too?
>
> I don't think it is doing precisely the same calculation for right and
> left divide when the matrix is not square.

the result of the fortran version coincides with that given by  
matlab, though

> jwe

c.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: test.f
Type: application/octet-stream
Size: 660 bytes
Desc: not available
Url : https://www.cae.wisc.edu/pipermail/bug-octave/attachments/20080113/1d424315/attachment.obj