Why does a DO CONCURRENT statement gives a strange result?

Question

Let us compile the following Fortran program:

program main
    implicit none
    integer::j
    complex::y(2)
    real::x(2)

    x(1)=1.0
    x(2)=2.0

    do j=1,2
        y(j)=FF(x(j))
    enddo

    write(6,*)
    write(6,*) y,"by DO"

    forall (integer::i=1:2)
        y(i)=FF(x(i))
    end forall

    write(6,*)
    write(6,*) y,"by FORALL"

    do concurrent (integer::i=1:2)
        y(i)=FF(x(i))
    enddo

    write(6,*)
    write(6,*) y,"by DO CONCURRENT"
    write(6,*)

    contains
    complex pure function FF(xx)
        real,intent(in)::xx
        FF=cmplx(xx)
    end function
end program

by ifort (ver. 19.0.4.243) with no options. Then, we get the result:

$ ./a.out

 (1.000000,0.0000000E+00) (2.000000,0.0000000E+00) by DO

 (1.000000,0.0000000E+00) (2.000000,0.0000000E+00) by FORALL

 (2.000000,0.0000000E+00) (2.000000,0.0000000E+00) by DO CONCURRENT

The first and second results are as expected. Why does not DO CONCURRENT gives the same result?

You can report bugs in the Intel compiler to Intel (using your support contact or the forums). — francescalus, Aug 21 '20 at 10:53
I tested with gfortran 7.5 and your code works properly. I would try disabling optimization during compilation with option -O0 to see if this is an optimization issue. — Emilio, Aug 21 '20 at 12:09
It seems to work when compiled with `/Qopenmp` or `/Qparallel` — , Aug 21 '20 at 14:20
I have just tried it out and I can confirm that this behaviour exists in intel compiler `ifort (IFORT) 19.1.2.254 20200623` and `ifort (IFORT) 19.0.4.243 20190416` (I have not tested any other version). It does not appear, as @Jean-ClaudeArbaut states, the `-parallel` option is given. However, just the `-qopenmp` seems to still have the bug. — chw21, Aug 22 '20 at 02:45
@chw21 That's odd. With `ifort 19.1.2.254 build 20200623` for `Windows`, it works with any of the options. It also works if y is not complex but real. There is definitely a bug, and it depends on subtle factors. As an aside, I looked up the Intel forum, and there are several questions related to problems with `do concurrent` in recent versions. — , Aug 22 '20 at 05:02
Well, I ran it on Linux. But it seems to me that that's a bug, and it's up to Intel to figure out what's going on. — chw21, Aug 22 '20 at 05:29

Why does a DO CONCURRENT statement gives a strange result?

0 Answers0