为什么实际上生成并行进程的 OpenMP do-loop 不能被 OMP_GET_THREAD_NUM() 检测到？

Question

我不明白为什么!$OMP DO实际上是将任务分配给不同的线程，但是使用openMP内部函数无法检测到OMP_GET_THREAD_NUM()。

program test
implicit none

integer :: i,su
double precision a(10), b(10),c
INTEGER OMP_GET_THREAD_NUM

su=0


!$OMP DO
do i=1,10
   b(i) = 10*i;
   c = b(i);   
   write(*,*)'in the loop, rank =',c,OMP_GET_THREAD_NUM()
enddo
!$OMP END DO

!$OMP PARALLEL
write(*,*) 'Rank = ',OMP_GET_THREAD_NUM()
!$OMP END PARALLEL

end

结果是：

 in the loop, rank =   10.000000000000000                0
 in the loop, rank =   20.000000000000000                0
 in the loop, rank =   30.000000000000000                0
 in the loop, rank =   40.000000000000000                0
 in the loop, rank =   50.000000000000000                0
 in the loop, rank =   60.000000000000000                0
 in the loop, rank =   70.000000000000000                0
 in the loop, rank =   80.000000000000000                0
 in the loop, rank =   90.000000000000000                0
 in the loop, rank =   100.00000000000000                0
 Rank =            0
 Rank =            6
 Rank =            1
 Rank =            7
 Rank =            2
 Rank =            5
 Rank =            4
 Rank =            3

看到了吗？ DO-LOOP中public好像只能看到Master线程。这是不公平的，因为这不是他唯一的贡献。

Answer 1

您的 do 循环不在并行区域中，因此未并行化 -- 所有循环索引都是由线程 0 处理的。

如果我更改您的程序以包含并行区域

...
!$OMP PARALLEL
!$OMP DO
do i=1,10
   b(i) = 10*i;
   c = b(i);
   write(*,*)'in the loop, rank =',c,OMP_GET_THREAD_NUM()
enddo
!$OMP END DO
!$OMP END PARALLEL
...

然后我得到 OMP 线程数的正确输出：

 in the loop, rank =   50.000000000000000                8
 in the loop, rank =   20.000000000000000                3
 in the loop, rank =   20.000000000000000                7
 in the loop, rank =   20.000000000000000                4
 in the loop, rank =   20.000000000000000                5
 in the loop, rank =   20.000000000000000                9
 in the loop, rank =   20.000000000000000                6
 in the loop, rank =   20.000000000000000                0
 in the loop, rank =   20.000000000000000                1
 in the loop, rank =   30.000000000000000                2

这个特定的输出也暴露了您的代码中的一个缺陷，即 c 是共享的，所以它的值被每个线程破坏了。此外，如果 do 循环是并行区域中唯一的东西，您可以组合 OMP 指令。最后，如果我们将您的代码更改为：

!$OMP PARALLEL DO private(c)
do i=1,10
   b(i) = 10*i;
   c = b(i);
   write(*,*)'in the loop, rank =',c,OMP_GET_THREAD_NUM()
enddo
!$OMP END PARALLEL DO

那么输出就正确了

 in the loop, rank =   100.00000000000000                9
 in the loop, rank =   30.000000000000000                2
 in the loop, rank =   20.000000000000000                1
 in the loop, rank =   70.000000000000000                6
 in the loop, rank =   60.000000000000000                5
 in the loop, rank =   50.000000000000000                4
 in the loop, rank =   80.000000000000000                7
 in the loop, rank =   90.000000000000000                8
 in the loop, rank =   10.000000000000000                0
 in the loop, rank =   40.000000000000000                3

为什么实际上生成并行进程的 OpenMP do-loop 不能被 OMP_GET_THREAD_NUM() 检测到？

Why OpenMP do-loop that actually generates parallel process, can NOT be detected by OMP_GET_THREAD_NUM()?

fortran

openmp