Problem with OpenMP

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hi,

I originally posted this in the forum on OpenMP’s website, but they said to try to post it here. The original question is here: http://openmp.org/forum/viewtopic.php?f=3&t=215.

I am trying to compile this with the PGI Fortran compiler version 6.1 on a 64 bit system. But it seems that it has some problems with OpenMP.

The code is shown below, and to get a full description of the problem, please see the above link…

	Amax = 0.0_dbl
	!$OMP PARALLEL DEFAULT(SHARED) PRIVATE(j,alpha,beta,nu,A,priv_max,priv_numax)
	priv_max = 0.0_dbl
	!$OMP DO
	do j = 1,Nspec
		nu = nu_min+dF*(j-1)
	
		! Calculate amplitude:
		call CalcAlphaBeta(t,f,w,nu,alpha,beta)
		A = sqrt(alpha**2 + beta**2)
        
		if (A > priv_max) then
			priv_max = A
			priv_numax = nu
		endif
	enddo
	!$OMP END DO NOWAIT
	!$OMP FLUSH
	if (priv_max > Amax) then
		!$OMP CRITICAL
			if (priv_max > Amax) then
				print *, priv_max
				Amax = priv_max
				!$OMP FLUSH
				numax = priv_numax
			endif
		!$OMP END CRITICAL
	endif
	!$OMP END PARALLEL
	
! If this line is included it works?! WHY?!
!	print *, 'Result: ', Amax

Please help… I will be really grateful!

Best regards,
/Rasmus

2

Hi Rasmus,

What happens if you remove the outer if statement for “priv_max > Amax” or place it under the critical section? Having Amax outside of the critical section is most likely cause of the error.

  • Mat
3

No, that doesn’t help… Same thing…

/Rasmus

4

Ok. I’ll need to be able to compile the source and run the exe to determine what’s happening. Can you send and example to PGI customer support (trs@pgroup,com) along with instructions and ask them to forward it to me?

Thanks,
Mat

5

I have now sent you the source code.
Hope you can find the error…

/Rasmus

6

Hi Rasmus,

The issue here is that the compiler’s dead store elimination optimization is removing the store to Amax (“Amax = priv_max”) since Amax is not used later in the subroutine. (This also explains why the program works as expected when you add the print statement), While fine in a serial context, this optimization is not correct in a parallel context. Hence, I’ve filed a technical problem report (TPR#15143) and sent it to our engineers.

To work around the issue, please compile this file at “-O1” instead of “-O3” since dead store elimination occurs at “-O2”. Of course, your print statement works as well.

We appreciate you bring this problem to our attention and we should be able to have a fix soon.

Mange tak,
Mat

7

Great! Thanks a million! I really appreciate that you could take the time to help me, and it seems that it was beneficial for both of us. :-)

Thumbs up to you and the rest of the PGI-boys!

Best regards,
Rasmus

PS: Selv tak, du var en stor hjælp!

8

Hi Rasmus,

TPR#15143 has been fixed in PGI 7.2-5.

Thanks,
Mat