LAPACK/ScaLAPACK Development

I am trying to calculate the eigenvalues and eigenvectors of matrices of different sizes. I am using a piece of very simple Fortran90 code and I am compiling it linking to the appropriate Lapack libraries included in the Intel MKL package, available in my machine, which runs in Ubuntu. The code "matrix_diag_01.f90" is attached at the end of the message. The "random" module just includes the "ran" random number generator from Numerical Recipes. The code compiles well using

Code: Select all

ifort -I $(MKLPATH) -o matrix_diag_01 matrix_diag_01.f90
      random.f90 $(MKLPATH)libmkl_lapack95.a -Wl,--start-group 
      $(MKLPATH)libmkl_intel_lp64.a $(MKLPATH)libmkl_lapack.a
      $(MKLPATH)libmkl_intel_thread.a $(MKLPATH)libmkl_core.a
       -Wl,--end-group -lguide -lpthread

The executable works nicely when smallish matrices are given. However, for matrices of size 3000x3000 it produces some strange behaviour. First it gives this error

Code: Select all

MKL ERROR : Parameter 8 was incorrect on entry to SSYEVD


However, there are only 3 parameters in the call to SSYEVD. Second, it returns the eigenvectors but not the eigenvalues. I have checked by compiling in another machine with larger memory but the outcome was the same.

Can anyone please help?

Thanks!

Code: Select all

PROGRAM matrix_diag_01

USE random

IMPLICIT NONE

INTERFACE
   SUBROUTINE diag(mat,n)
      INTEGER n
      REAL,DIMENSION(n,n) :: mat
   END SUBROUTINE
END INTERFACE

INTEGER n,i,j,iseed
REAL, DIMENSION(:), ALLOCATABLE  :: w
REAL, DIMENSION(:,:), ALLOCATABLE :: mat

write (*,*) ' Please enter size of matrix'
read (*,*) n
write (*,*) ' Please type seed'
read (*,*) iseed

allocate (mat(n,n))

do i = 1,n
   do j = 1,n
      mat(i,j) = ran(iseed)
   end do
end do

call diag(mat,n)

stop
END
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
SUBROUTINE diag(mat,n)

  USE mkl95_lapack
  USE mkl95_precision

  IMPLICIT NONE

  CHARACTER(len=1) :: jobz = 'V'

  INTEGER n,i

  REAL,DIMENSION(n,n) :: mat
  REAL,DIMENSION(:,:),ALLOCATABLE :: matt,a
  REAL,DIMENSION(:),ALLOCATABLE :: w

  allocate (matt(n,n),a(n,n),w(n))

  matt = mat*transpose(mat)
  a = sqrt(matt)
  open (unit=7,file="matrix.dat",status="unknown")
  do i = 1,n
     write (7,100) a(i,:)
  end do
  close (unit=7)

  call syevd(a,w,jobz)

  open (unit=8,file="eig_val.dat",status="unknown")
  do i = 1,n
     write (8,100) w(i)
  end do
  close (unit=8)

  open (unit=9,file="eig_vec.dat",status="unknown")
  do i = 1,n
     write (9,100) a(i,:)
  end do
  close (unit=9)

  return
 100 format(5000f16.5)
end


Have you checked your stack size? If your matrices are taking up more memory than your default stack then you can run into some weird errors; typically segmentation faults but not always.

You are using the syevd subroutines from LAPACK95 supported by MKL.
Not an expert with this interface on my end. Maybe contact MKL folks directly.
Julien.

Thank you. It really looks like a problem in Fortran95 interfaces.
I'll try to investigate it. You should post this problem on Intel MKL forum.