Is there anything like "THREADPRIVATE" for OpenMP?

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hello,

I like to switch my code from OpenMP to OpenACC. In my OpenMP code, I pass “thread-dependent” arrays to designated threads. I am wondering whether I can do the same thing using OpenACC.

I outline the function I need in the following code snippet: I want to pass two more variables, “istart” and “iend”, into the subroutine “work” which only takes one input argument. Thanks.

Lee

module model
  integer :: istart,iend
  !$omp threadprivate(istart,iend)
  implicit none
contains    
subroutine work(iarray)
    ...
    integer iarray(10000)
    do i = istart, iend
       iarray(i) = i * i
    enddo
end subroutine work
end module model

program main
    use parameters !not completely necessary here
    use model
    implicit none
    ...
    integer iarray(10000)
    N = 10000
!$omp parallel private(iam, nthreads, chunk)
    nthreads = omp_get_num_threads()
    iam      = omp_get_thread_num()
    chunk    = (N + nthreads – 1)/nthreads
    istart   = iam * chunk + 1
    iend     = min((iam + 1) * chunk, N)
    call work(iarray)
!$omp end parallel
end program main
2

Lee,

I think OpenACC can intelligently chunk up your array for you without all the explicit OpenMP chunking you are doing here. Try something like this:

subroutine work(iarray) 
    ... 
    integer iarray(10000) 
!$acc kernels loop copyin(iarray(1:10000))
    do i = istart, iend 
       iarray(i) = i * i 
    enddo 
end subroutine work

Hope this helps,

+chris

3

@Chris,

Thanks for the reply. I need to rephrase my question so that you can know what I came across in my project. I have the following two files. I use a do loop over a threaded subroutine “tool” and am restricted to pass a single- variate function to the subroutine. However, in my mathematical model, the function have two arguments, so I need to feed thread-dependent variables to respective threads.

In OpenMP, I can simply add the directive “!$OMP THREADPRIVATE(i).” I am wondering how to port this code to OpenACC. Thanks.


1. MAIN.F90
program main
    use toolbox
    real :: a(5),c(5)
    integer :: j
    
    a = [((j),j=1,9,2)]
    b = [((j),j=2,10,2)]
    
    !$OMP PARALLEL DO SHARED(b)
    do j=1,5
        i=j
        call tool(fun1,a(j),c(j))
    enddo
    !$OMP END PARALLEL DO
    write(6,'5(f3,x)') c
end program main

2. MODEL.F90
module toolbox
    integer :: i
    !$OMP THREADPRIVATE(i)
    real :: b(5)
  contains
    subroutine tool(func,e,f)
    interface
    real function func(x)
        real :: x
    end function func
    end interface
    real :: e,f
    f=func(e)
    end subroutine tool
    
    function fun1(z)
    real :: fun1,z
    fun1=z+b(i)
    end function fun1
end module toolbox
4

Lee,

I’m checking on something related to this, and I will follow up soon.

Best regards,

+chris