omp doesn't run properly with global variables?

[ytlee]

Hi Everyone,

I am adding omp command to run my program in a i-j nested loop. However, the result was very different from the original/serial job which I have no idea why that happened. I suspect if this issue has somethg to do with my global variable (it shouldn't be...). It will be great if I can get some help here. The following is part of my code.
(n_tot, xn, yn, zn, chk_radius, pow, G, m, de, rn and PEj are declared as global variable in the global module)

Code: Select all

subroutine selfg()
  use global
  implicit none

  integer :: i, j
  real :: delx, dely, delz, rr, r, netaccterm
  real, dimension(n_tot,n_tot) :: gxj, gyj, gzj, PEj

!$omp parallel do private(j,delx,dely,delz,rr,r,netaccterm) &
!$omp             num_threads(4)
  do i=1, n_tot
     do j=i+1, n_tot
        delx = xn(step,i)-xn(step,j)
        dely = yn(step,i)-yn(step,j)
        delz = zn(step,i)-zn(step,j)
        rr = delx**2+dely**2+delz**2
        r = sqrt(rr)
        rn(i,j) = r
        if (r < chk_radius) then
           !models of self-gravitation
           netaccterm = G*m*(-1/rr + de**(pow-2)/r**pow)/r  !inverse law
           gxj(i,j) = delx*netaccterm
           gyj(i,j) = dely*netaccterm           
           gzj(i,j) = delz*netaccterm           
           PEj(i,j) = G*m*(-1/r + de**(pow-2)*r**(1-pow)/(pow-1))
        endif
     enddo
  enddo
!$omp end parallel do
.
.
.
end subroutine selfg

I appreciate for your help...

-yt

[ftinetti]

Hi yt,

I've been looking at your code, and I didn't see anything wrong... maybe I lost something. Would you send a complete example that reproduces the wrong behaviour to play around with?

Fernando.

[ytlee]

Hi Fernando,

After further checking, I think the code is now working... Seems like it works after initializing all the arrays' entries to zero. Apparently, without initializing give me some sort of error with(/without?) openMP. I can't believe I have to do this for Fortran. I assumed the entries are zero after doing: real, dimension(n_tot,n_tot) :: gxj, gyj, gzj, PEj but seems like it is not (sometimes it gives wield values)...

Really appreciate for looking at my code!

-yt

[ftinetti]

Hi again,

I see, I tend to always overlook this kind of errors. Actually, I would not wonder too much about

I can't believe I have to do this for Fortran. I assumed the entries are zero after doing: real, dimension(n_tot,n_tot) :: gxj, gyj, gzj, PEj

because I think that zero-initialization is not defined in any of the Fortran standards, even when many (most? all?) compilers actually implement such initialization.

It's good to know your code works now.

Fernando.