LAPACK Archives

[Lapack] issue with DGGSVD in LAPACK


There is definetely a problem.
With 8 GB, you could get in the 10,000 easily.
Two things:

(*) how do you allocate your matrices?
Do you declare static arrays as:
DOUBLE PRECISION A(500,500)
or do you allocate them dynamically:
DOUBLE PRECISION, DIMENSION(:,:), ALLOCATABLE :: A
then ALLOCATE, then DEALLOCATE ?
The latter is preferable for your size of arrays.

(*) This should not be an issue but how much workspace do you use for 
DGESVD or DGESDD? You need to give the workspace returned by a workspace 
query using LWORK=-1.

--julien.



On Fri, 10 Jul 2009, Eric Loomis wrote:

I was wondering if you could give me an estimate of what size matrix I should 
be able to perform SVD using either dgesvd or dgesdd on a machine with
a few Gb of RAM?? On both a 2.5 Gb and an 8 Gb machine the largest matrix I 
can use is 500 x 500, any higher and I get a segmentation fault.? I use
the work space size as suggested in the dgesvd and dgesdd routines.


At 8:51 AM -0600 7/9/09, julie langou wrote:
      Eric,

      I will wait for your output with the new flags.

      For info, we just released CLAPACK 3.1.2 a couple of weeks ago, it 
includes all the changes made to LAPACK 3.2.1.

      Regards

      Julie

      On Jul 9, 2009, at 8:40 AM, Eric Loomis wrote:


            Thank you very much for your reply.


            Regarding your previous email, the reason I chose CLAPACK is 
because I know it best.? I have very limited experience with
            fortran in general and I have not looked into MKL.


            I am using the native BLAS that came with the CLAPACK 3.1.1.1 
distribution, which works on my local machine using gcc, but I
            want to get it running on a larger platform and uses icc.? The 
larger platform has ATLAS BLAS which I hope to use in the
            future.? I would also like to get this running in a parallel 
cluster because I am dealing with very large matrices and
            require a few hundred singular values.


            As for the error, I run the exact same code on my local machine 
(with gcc) and it works perfectly.


            I will try your suggested flags and thanks again for looking into 
it.


            Eric

            At 7:22 AM -0600 7/9/09, julie langou wrote:
                  Eric,

                  Icc is a picky compiler.?Usually we use the following 
flags,-fltconsistency -fp_port

                  I am not sure this will help much anyway...

                  Which BLAS library are you using? MKL?

                  and which LAPACK (c version)? CLAPACK 3.2.1?

                  I will try to reproduce your problem .


                  Regarding the error code, if INFO =-4, he is complaining 
about the parameter N.

                  ????? SUBROUTINE DGESVD( JOBU, JOBVT, M, N, A, LDA, S, U, 
LDU, VT, LDVT,
                  ???? $?????????????????? WORK, LWORK, INFO )

                  In the code, we are just making sure N is greater than 0.

                  Maybe you put L in place of N, no?.

                  ????? ELSE IF( N.LT.0 ) THEN
                  ???????? INFO = -4


                  julie



                  On Jul 8, 2009, at 11:45 AM, Eric Loomis wrote:


                        Hello, I ran into a problem building the LAPACK (c 
version) on a high performance computing platform
                        using an icc compiler. ?The libraries seemed to build 
okay, but when I ran the testing executables I
                        ran into some errors in the SVD routine. ?I have 
attached the output file with the errors. ?Do you
                        know what may have caused these errors?

                        I have probably a related problem when I try to run 
some of my code using the DGESVD routine in that
                        I get an error
                        ** On entry to DGESVD, parameter number ?4 had an 
illegal value

                        INFO = -4
                        LWORK = -1

                        Can you help me with any of these issues?

                        thanks.

                        At 10:21 AM -0700 6/25/09, Zhaojun Bai wrote:
                              Hi Eric,


                              I would be happy to try to help out here. Can 
you let me

                              the exact formulation of the constrained LS

                              problem you want to solver? I can take a look 
and

                              see whether we're on the right-track to use the 
DGGSVD

                              from LAPACK.


                              Apologize the delay for response. We just 
finished the spring

                              quarter term.


                              Regards,

                              Zhaojun Bai

                              UC Davis





                              Eric Loomis wrote:
                                    Hello,


                                    I am using the routine DGGSVD (CLAPACK) 
to solve constrained least squares
                                    problems. ?To do this I need to form a 
matrix X from Q and R as discussed in
                                    GSVD literature. ?R appears to be nearly 
diagonal. The result of using the
                                    GSVD to solve a system of equations is 
that the first half of the solution is
                                    nearly zero, but the second half gives 
the correct answer. ?Can you put me in
                                    contact with someone familiar with this 
routine to work out this issue? ?I
                                    can supply the source code if required.




                        --
                        Eric Loomis
                        P-24 plasma physics
                        Los Alamos National Laboratory
                        mailstop E526
                        505-665-3196
                        
loomis@Domain.Removed<dsvd.out>_______________________________________________

                        Lapack mailing list
                        Lapack@Domain.Removed
                        http://lists.cs.utk.edu/listinfo/lapack


                  **********************************************

                  Julie Langou; Research Associate in Computer Science

                  Innovative Computing Laboratory;

                  University of Tennessee from Denver, Colorado ;-)

                  julie@Domain.Removed;?http://www.cs.utk.edu/~julie/



            --

            Eric Loomis
            P-24 plasma physics
            Los Alamos National Laboratory
            mailstop E526
            505-665-3196
            loomis@Domain.Removed


      **********************************************

      Julie Langou; Research Associate in Computer Science

      Innovative Computing Laboratory;

      University of Tennessee from Denver, Colorado ;-)

      julie@Domain.Removed;?http://www.cs.utk.edu/~julie/



 -- 
Eric Loomis
P-24 plasma physics
Los Alamos National Laboratory
mailstop E526
505-665-3196
loomis@Domain.Removed



<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