non-coarray vector lhs produces crashes or wrong results with gets

[reinh-bader]

The attached source code was built with

• gfortran 6.3 + opencoarrays 1.6.2
• gfortran 7.0 (preview) + opencoarrays 1.8.2
  and produces wrong results when executed with any number of images. The version build with gfortran -fcoarray=single executes fine, hence I suspect the problem to be in the run time library.

The program checks its output.
For opencoarrays 1.8.2, I get the run time error

Fatal error in MPI_Win_unlock_all: Invalid MPI_Win, error stack:
MPI_Win_unlock_all(127): MPI_Win_unlock_all(win=0x55444f4d) failed
MPI_Win_unlock_all(82).: Invalid MPI_Win
Fatal error in MPI_Win_unlock_all: Invalid MPI_Win, error stack:
MPI_Win_unlock_all(127): MPI_Win_unlock_all(win=0x55444f4d) failed
MPI_Win_unlock_all(82).: Invalid MPI_Win

PMR.zip

[zbeekman]

Can you please include a little bit more information: MPI implementation, system (OS, arch, virtualized?) and how you built and installed OpenCoarrays (with install log if possible). Also how did you compile and invoke your code? How many procs?

Thanks!

[reinh-bader]

Base MPI is Intel MPI 5.1.3
The system is a X86-based shared memory node with SLES11 SP4 as OS.
Here is the command line used to build:

CC=gcc FC="gfortran -I/lrz/sys/intel/impi/5.1.3.181/intel64/include/gfortran/7.0.0" cmake /lrz/noarch/src/gcc/opencoarrays/opencoarrays-1.8.2 -DBUILD_SHARED_LIBS:BOOL=ON -DCMAKE_INSTALL_PREFIX=${PWD}

The only log file I found was CMakeFiles/CMakeOutput.log which I have attached (adding a .txt extension, to enable the upload).

The example code was compiled with

caf vector_subscript_pos_01.f90
and executed with
cafrun -n 2 ./a.out

CMakeOutput.log.txt

Hope this helps.

[zbeekman]

Thanks so much @reinh-bader... I think I know of a machine or two that has Intel MPI 5.1.3 if the problem does not show up with MPICH/OpenMPI

[zbeekman]

Sorry I have been swamped, and have not had a chance to fully investigate yet, but this is still on my radar

[reinh-bader]

No trouble, take your time.

[zbeekman]

This is still broken as of 0c2dce7 on 12/26/2017

Here is the test case inlined:

program mod_vecsub_01
  implicit none
  integer, parameter :: ndim = 5, vdim = 2
  real :: vec(ndim), res(ndim)[*]
  integer :: idx(vdim)
  integer :: i, me
  logical :: ok[*]

  res = [ (real(i), i=1, ndim) ]
  vec = 0.0
  idx = [ ndim, 1 ]
  ok = .true.
  sync all
  me = this_image()
  vec(idx) = res(1:2)[1]
  if (vec(1) /= 2.0 .or. vec(5) /= 1.0) then
    critical
      ok[1] = .false.
      write(*, *) 'FAIL on image ',me,vec(idx)
    end critical
  end if
  if (me == 1) then
     if (ok) then
       write(*, *) 'OK'
     end if
  end if
end program

[vehre]

That's got nothing to do with opencoarrays. It's a compiler bug. Gfortran is not respecting the indexed addressing for the lhs after the get(). It's trying to do it before.

[zbeekman]

@vehre awesome, thanks for the response! @rouson we should look and see if this bug has been filed, and if it has not, make sure we file one and post the tracking number here. Perhaps @afanfa or Sorren might be interested in working on a fix for this?

[rouson]

This has been submitted as Bug 83606 on gfortran with the following slightly simplified code:

  integer, parameter :: ndim=5
  integer :: i,vec(ndim)=0, res(ndim)[*]=[ (i, i=1, ndim) ]
  vec([ndim,1]) = res(1:2)[1]
  if (vec(1) /= res(2) .or. vec(ndim) /= res(1)) &
    print *," wrong result ",vec([ndim,1])," on image ",this_image()
end

[zbeekman]

See also #427.

[vehre]

Fixed on all gcc's from 6 on.

[zbeekman]

@vehre is the fix in 6.4 and 7.3 or will it be in 6.5 and 7.4?