topic I just tried a practical in Intel® Moderncode for Parallel Architectures

Effeciently parallelizing the code in fortran

Ajay_N_ — Wed, 12 Feb 2014 17:57:26 GMT

Hi,

I am trying to parallelize a certain section of my code which is written in fortran. The code snippet looks as below:

do i=1,array(K)

j = K
... if conditions on K...
....write and reads on j...
... do lot of things ...
K = K+1

So I tried to parallelize using the below code.. which was obviously not as it should have been

!$OMP PARALLEL DO PRIVATE(j)
do i=1,50
j = K
... if conditions on K...
....write and reads on j...
... do lot of things ...
K = K+1
!$OMP END PARALLEL DO

The obvious mistake being, all the threads race to the same K. What would be the best way to ensure every thread gets assigned an incremental K and the threads run in parallel.

Thanks
Ajay

If work on K=1 is not

Andrey_Vladimirov — Wed, 12 Feb 2014 19:04:07 GMT

If work on K=1 is not dependent on the results of K=0, and work on K=2 is not dependent on the results of K=1, etc., then you should just express a mapping between the value of i and the value of j, rather than using the shared counter K.

However, if you must start processing K=0 before you can start processing K=1, and must start K=1 before you can start K=2, etc., then you should look into the ORDERED clause of OMP DO: http://software.intel.com/en-us/node/466858

Thanks Andrev,

Ajay_N_ — Wed, 12 Feb 2014 21:16:52 GMT

Thanks Andrev,

I was able to map j to values of i. Now I face a slightly different problem pertaining to openMP i think.

!$OMP PARALLEL DO PRIVATE(j)
do i=1,50
j = somefunction(i)
print thread_id, i
print thread_id, j

When you write a openMP like above, i noticed that say j gets computer for a value of i that is spawned by the next thread. Say the function is i+7
Example
Thread 0 i: 5
Thread 0 j: 9
Thread 1 i: 2
Thread 1 j: 11

Any reason why this should happen ?

Thanks
Ajay

It is hard to say without a

Andrey_Vladimirov — Wed, 12 Feb 2014 22:32:29 GMT

It is hard to say without a reproducer code. Perhaps some variables that should have been private were declared or assumed shared. Here is a code that worked for me:

[fortran]

program omploop

include 'omp_lib.h'

integer :: i, j, thread_id

write(*,'(3A10)') "Thread", "i", "j"

!$OMP PARALLEL DO PRIVATE(i, thread_id, j)
do i = 1, 10
     thread_id = omp_get_thread_num()
     j = i_to_j(i)
     write(*,'(3I10)') thread_id, i, j
end do

contains

pure integer function i_to_j(i)
integer, intent(in) :: i
i_to_j = i + 7
end function i_to_j

end program omploop

[/fortran]

Compilation and execution:

[bash]

ifort -openmp omploop.f90 -o omploop

./omploop

    Thread         i         j
         0         1         8
         2         7        14
         1         4        11
...

[/bash]

I just tried a practical

Ajay_N_ — Wed, 12 Feb 2014 22:35:00 GMT

I just tried a practical example for four threads:


!$OMP PARALLEL PRIVATE(k,id)

    !$OMP DO

    DO i = 1, 10

       count = i+5 

       print *,"id: ",OMP_GET_THREAD_NUM()," i    : ",i

       print *,"id: ",OMP_GET_THREAD_NUM()," count: ",count

    end do

    !$OMP END DO

    !$OMP END PARALLEL

Notice for the first thread 2, is 7 and the count for thread 2 should have been i+5 so 12 but the output is 14.

Output:
 id:            2  i    :            7
 id:            1  i    :            4
 id:            0  i    :            1
 id:            3  i    :            9
 id:            2  count:           14
 id:            1  count:           13
 id:            0  count:           10
 id:            3  count:            7
 id:            2  i    :            8
 id:            1  i    :            5
 id:            0  i    :            2
 id:            3  i    :           10
 id:            2  count:           15
 id:            1  count:           15
 id:            0  count:           11
 id:            3  count:            8
 id:            1  i    :            6
 id:            0  i    :            3
 id:            1  count:            8
 id:            0  count:            8

Thanks
Ajay

Sorry for the formatting, for

Ajay_N_ — Wed, 12 Feb 2014 22:36:37 GMT

Sorry for the formatting, for some reason the code quotes dont seem to work for me !!

Right, you need to make

Andrey_Vladimirov — Wed, 12 Feb 2014 22:39:22 GMT

Right, you need to make "count" a private varible.

Hi Andrev,

Ajay_N_ — Fri, 14 Feb 2014 17:24:52 GMT

Hi Andrev,

A follow up question. if there are global variables that I have made private to an openMP do loop. And there is a function that is being called within the do loop which makes use of the same global variables that I have made private. Please note that I am not passing the variables in the function, global variables are accessible to all the functions.

Does the function that is being called in the do loop get the private copy of the calling loop or does it behave unexpectedly? I am not making any changes to the called function.

I think I am observing ambiguity for this case and is there a way to tackle this?

Thanks

Ajay

Pass the private copy of the

jimdempseyatthecove — Sat, 15 Feb 2014 13:44:43 GMT

Pass the private copy of the (otherwise) global variable as an argument. Not doing so results in the called routine using the actual global variable and not the thread's copy.

If the(se) global variable(s) are thread specific, as opposed to being application global, and you are using global variables as a means to reduce the number of arguments, then consider either passing a reference to a thread context (containing the collection of variables you want specific to each thread) or use threadprivate variables. Threadprivate will have least impact on code.

Jim Dempsey

Hi Jim,

Ajay_N_ — Mon, 17 Feb 2014 18:20:46 GMT

Hi Jim,

Thanks for the reply, yes the arguments are not being passed lets say because it was built that way by some one ! And there are lot of functions that I will have to go and make the change. Is there a way to make the section of the parallelized code ensure that the calling function gets a private copy without me going to all functions and make the changes?

Ajay

The only way to do this is to

jimdempseyatthecove — Mon, 17 Feb 2014 22:39:41 GMT

The only way to do this is to declare the variables as threadprivate (global) variables. These variables have global namespace but thread local address space.

module temp
real(8) :: A(3),B(3),C(3)
!$OMP THREADPRIVATE(A,B,C)
...
end module temp
...
subroutine usingTemp(X,F)
use temp
real :: X(3)
A=X ! copies X to threadprivate A
...
end subroutine usingTemp

subroutine foo
!$OMP PARALLEL DO
do i=1,nObjs
call usingTemp(Obj(i)%pos) ! pass in objects position

The caveat here is the thread private variables are thread private. i.e. they are not global singularities, they are global multiplicities.

Jim Dempsey

Hi Jim,

Ajay_N_ — Wed, 05 Mar 2014 21:20:31 GMT

Hi Jim,

Thanks for your example, but my case is probably a little different.
gloabl header file header.f90
integer :: X;

Main function:
do i=1,n:
call subroutine(i)
end do

function subroutine:
X = some function on (i)

Now the issue is I want X to have a private copy when called within an openmp thread:

!$OMP PARALLEL DO (private X)

only ensures X is private within the do loop but not on the called functions. The moment a function is called from the openMP thread, the function accesses the global variable.
I know passing the variables solves this issue, but I am trying to parallelize an already existing code which is very huge and looking for a quick workaround if possible !

Thanks
Ajay

If you make X treadprivate,

jimdempseyatthecove — Thu, 06 Mar 2014 13:53:09 GMT

If you make X treadprivate, as outlined earlier, each thread, including the main thread, will have a global variable X that is private to that thread. When the X's are truly independent then there will be no issue. The only issue you may have is that you may be required to reduce the multiple instances of X into a preferred choice for use by the thread outside the region. An example of this could be constructed out of finding the minimum X in an array (using the threadprivate global X's).

gloabl header file header.f90
integer :: X;
!$OMP THREADPRIVATE(X)
----
integer :: Xcopy ! in source with main function

Main function:
Xcopy = 999999 ! some value on the other side of the solution HUGE(X)
!$OMP PARALLEL
!$OMP DO
do i=1,n:
call subroutine(i)
end do
!$OMP CRITICAL ! or !$OMP ATOMIC for newer version of OpenMP
if(X .lt. Xcopy) Xcopy = X
!$OMP END CRITICAL
!$OMP END PARALLEL
X = Xcopy

function subroutine:
X = some function on (I) ! X is threadprivate

Note, newer versions of OpenMP may have a MIN and MAX reduction operator

Jim Dempsey

Thank you so much !! I did

Ajay_N_ — Fri, 07 Mar 2014 16:55:01 GMT

Thank you so much !! I did not think threadprivate and private are different until now. I thought both were same !! Will experiment with this a bit more.

Thank you again.
Ajay