LAPACK Archives

[Lapack] diagonalization of a big matrix


Dear Carolina,

Glad to hear that LAPACK served you well for six years.

I suspect N=20,000 (about) is the limit for our Divide and Conquer 
interface unfortunately. This needs some more investigation. There is 
a related post at: 
https://icl.cs.utk.edu/lapack-forum/viewtopic.php?f=2&t=1418
And this is bug report #0020 at 
http://www.netlib.org/lapack/Errata/

If you do a worspace query (LWORK=-1) what is the workspace needed? ( 
WORK(1) ). Is this value not above the integer overflow (2^31)? So that 
when you feed it to the 32-bit integer LWORK for the real run actually it 
does not work. I do not quite get it. Because that should not be the case 
for N=20,000.

Another possible problem is that it is not clear that WORK(1) (double 
precision) convert correctly to LWORK (integer), you might be missing a 
"one". If you do a workspace query can you try LWORK = LWORK + 1 (and 
allocate with this plus one!).

I am positive that you can not use dpsevd for matrices bigger than 46,000
(sqrt(2^31)). LAPACK should be using 64-bit integer for this.

(There is another limitation from packed format itself but it will hit us 
later at N=2^16=65,536.)

Please keep in touch, to summarize, you can:
* check that the value LWORK before calling LAPACK "makes some sense" by
   printing it
* add +1 to the LWORK before allocating the workspace

Yes, yes, 64-bit integers in LAPACK ...

Best wishes,
Julien


On Sat, 13 Feb 2010, Carolina Brito wrote:

Hi!

I have been using CLAPACK in the past 6 years to diagonalize matrices up to 
size 2000 x 2000 and I did not have any problem.
I am currently? using the routine? dspevd_()

I now want to diagonalize a matrix of size 20000 x 20000 and the routine 
returns "segmentation fault".
I would like to know if there is a limit of size that this routine can 
diagonalize of if this is a limitation of my computer (memory problems). This 
is
important because I do not know if I should keep tying to use CLAPACK or if 
it has some limitation that would not allow me to diagonalize such a big
matrix.

Thanks a lot.
carolina

________________________________________________________________________________________________________________________________________________________
Veja quais s?o os assuntos do momento no Yahoo! + Buscados: Top 10 - 
Celebridades - M?sica - Esportes


<Prev in Thread] Current Thread [Next in Thread>


For additional information you may use the LAPACK/ScaLAPACK Forum.
Or one of the mailing lists, or