Computational Science Stack Exchange is a question and answer site for scientists using computers to solve scientific problems. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

After going to all possible "Introductions" to ScaLapack, I still can't understand how to carry out a simple PDGEMV operation using it.

Here is what I must do :

I have to generate a matrix using

do i=1,n
  x(i) = i*i*1.0D+00
  do j=1,n
    A(i,j) = (i+j)*j*1.00D+00
  end do
end do

and then simply multiply Matrix A by Vector x (b=Ax).

For Lapack, I simply use CALL DGEMV('N',n,n,1.0D+00,a,n,x,1,0.0D+00,b,1)

How do I achieve this in PBLAS? I have run the sample (example1.f from the website) program and it works.

What is partitioning of the matrix? How do I do it?

Everyone (in tutorials) seem to be concentrating on DGESV while I want DGEMV/DGEMM i.e. BLAS

share|improve this question
Google is your friend. Look at (somewhere at the middle) – stali Mar 20 '12 at 17:03
up vote 3 down vote accepted

The data distribution scalapack requires is a block-cyclic distribution. (A little interactive calculator is here). Once you get the hang of that, Scalapack is pretty straightforward.

For a given blocksize (which is a free parameter) and decomposition of the processors into a grid (eg, 4 processors -> 2 rows, 2 columns of processors) you can go from the local indices of your chunk of the matrix to the global indices of the matrix using l2g or g2l from this NERSC example; that's a pretty straightforward way to get started. Row or column vectors are treated as nx1 or 1xn matrices.

Below is an example of multiplying the diagonal matrix A (1,2,3...) by a unit vector X and getting the results into Y; the only real trick here is that not all processors have a chunk of the X or Y vectors.

program gemv1
      use mpi
      implicit none

      integer :: n, nb    ! problem size and block size
      integer :: myArows, myAcols   ! size of local subset of global matrix
      integer :: myXrows, myXcols   ! size of local subset of global vector
      integer :: i,j, myi, myj
      real, dimension(:,:), allocatable :: myA,myX,myY
      integer :: ierr

      integer, external :: numroc   ! blacs routine
      integer :: me, procs, icontxt, prow, pcol, myrow, mycol  ! blacs data
      integer :: info    ! scalapack return value
      integer, dimension(9)   :: ides_a, ides_x, ides_y ! matrix descriptors
      integer, dimension(2) :: dims
      real :: error, globerror

! Initialize blacs processor grid

      call blacs_pinfo   (me,procs)

! create as square as possible a grid of processors

      dims = 0
      call MPI_Dims_create(procs, 2, dims, ierr)
      prow = dims(1)
      pcol = dims(2)

! create the BLACS context

      call blacs_get     (0, 0, icontxt)
      call blacs_gridinit(icontxt, 'R', prow, pcol)
      call blacs_gridinfo(icontxt, prow, pcol, myrow, mycol)

! Construct local arrays
! Global structure:  matrix A of n rows and n columns

      n = int(25000.*sqrt(dble(procs)))

! blocksize - a free parameter.

      nb = 100

! how big is "my" chunk of matrix A?

      myArows = numroc(n, nb, myrow, 0, prow)
      myAcols = numroc(n, nb, mycol, 0, pcol)

! how big is "my" chunk of vector x?

      myXrows = numroc(n, nb, myrow, 0, prow)
      myXcols = 1

! Initialize local arrays    


      myA = 0.
      do myj=1,myAcols
          ! get global index from local index
          call l2g(myj,mycol,n,pcol,nb,j)
          do myi=1,myArows
              ! get global index from local index
              call l2g(myi,myrow,n,prow,nb,i)
              if (i == j) myA(myi,myj) = i

      myX = 0.
      call l2g(1,mycol,n,pcol,nb,j)
      if (j == 1) then
          do myi=1,myXrows
              call l2g(myi,myrow,n,prow,nb,i)
              myX(myi,1) = 1.

      myY = 0.

! Prepare array descriptors for ScaLAPACK 

      call descinit( ides_a, n, n, nb, nb, 0, 0, icontxt, myArows, info )
      call descinit( ides_x, n, 1, nb, nb, 0, 0, icontxt, myXrows, info )
      call descinit( ides_y, n, 1, nb, nb, 0, 0, icontxt, myXrows, info )

! Call ScaLAPACK library routine

      call psgemv('N',n,n,1.,mya,1,1,ides_a,myx,1,1,ides_x,1,0.,myy,1,1,ides_y,1)
      if (me == 0) then
        if (info /= 0) then
             print *, 'Error -- info = ', info

! Deallocate the local arrays

      deallocate(myA, myX)

! Print results - Y should be 1,2,3...

      error = 0.
      call l2g(1,mycol,n,pcol,nb,j)
      if (j == 1) then
          do myi=1,myXrows
              call l2g(myi,myrow,n,prow,nb,i)
              error = error + (myY(myi,1)-1.*i)**2.

      call MPI_Reduce(error, globerror, 1, MPI_REAL, MPI_SUM, 0, MPI_COMM_WORLD, ierr)

      if (me == 0) then
        print *,'Y l2 error = ', sqrt(globerror/n)


! End blacs for processors that are used

      call blacs_gridexit(icontxt)
      call blacs_exit(0)


! convert global index to local index in block-cyclic distribution

   subroutine g2l(i,n,np,nb,p,il)

   implicit none
   integer, intent(in) :: i    ! global array index, input
   integer, intent(in) :: n    ! global array dimension, input
   integer, intent(in) :: np   ! processor array dimension, input
   integer, intent(in) :: nb   ! block size, input
   integer, intent(out):: p    ! processor array index, output
   integer, intent(out):: il   ! local array index, output
   integer :: im1   

   im1 = i-1
   p   = mod((im1/nb),np)
   il  = (im1/(np*nb))*nb + mod(im1,nb) + 1

   end subroutine g2l

! convert local index to global index in block-cyclic distribution

   subroutine l2g(il,p,n,np,nb,i)

   implicit none
   integer :: il   ! local array index, input
   integer :: p    ! processor array index, input
   integer :: n    ! global array dimension, input
   integer :: np   ! processor array dimension, input
   integer :: nb   ! block size, input
   integer :: i    ! global array index, output
   integer :: ilm1   

   ilm1 = il-1
   i    = (((ilm1/nb) * np) + p)*nb + mod(ilm1,nb) + 1

   end subroutine l2g

end program gemv1
share|improve this answer

Use the pblas GEMV.

share|improve this answer
I know I have to use that. My question is how? What is a submatrix, how do I generate it? – Inquest Mar 20 '12 at 15:02

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.