LI75389: INCORRECT OUTPUT WITH OMP_SCHEDULE

May 2010 Update for XL Fortran for Linux, V12.1
January 2012 Update for XL Fortran for Linux, V12.1
January 2011 Update for XL Fortran for Linux, V12.1
June 2011 Update for XL Fortran for Linux, V12.1
XL Fortran for Linux Fix Pack 8 (January 2014 Update) for 12.1

APAR status

• Closed as program error.

Error description

• The following test case produces incorrect output for
  OMP_SCHEDULE=DYNAMIC and OMP_SCHEDULE=GUIDED.
  
  
  
  $cat workshare.f90
  program workshare
  !$ use OMP_LIB
    implicit none
    integer, parameter           :: n=100
    integer                      :: i, j, num_th
    real(kind=8), dimension(n,n) :: a
  !$ integer                     :: nb_taches
    !$OMP PARALLEL
    !$ nb_taches = OMP_GET_NUM_THREADS()
    !$OMP END PARALLEL
    !$ print '(3X,"Parallel execution with ",i2,"
  threads")',nb_taches
    a(:,:)=1
    !$OMP PARALLEL PRIVATE(num_th)
    !$  num_th= OMP_GET_THREAD_NUM()
    !$OMP WORKSHARE
    forall (i=2:n, j=1:n-1, i>j) a(j,i) = a(i,j)
    forall (i=1:n) a(i,i) = a(i,i) + 800.0_8
    !$OMP END WORKSHARE
    if (num_th == 0) print*,"Sum = ",sum(a)
  !$OMP END PARALLEL
  end program workshare
  
  $cat submit.sh
  cd ./
  make all
  echo Monoprocessor
  echo
  ./workshare_mono.exe
  echo
  echo OMP_SCHEDULE=STATIC
  export OMP_SCHEDULE="STATIC"
  ./workshare_omp.exe
  echo
  echo OMP_SCHEDULE=DYNAMIC
  export OMP_SCHEDULE="DYNAMIC"
  ./workshare_omp.exe
  echo
  echo OMP_SCHEDULE=GUIDED
  export OMP_SCHEDULE="GUIDED"
  ./workshare_omp.exe
  echo
  echo The END.
  
  $cat Makefile
  FC =xlf90_r
  
  
  LDFLAGS =
  OMP = -qsmp=omp
  CMD = workshare
  SRC     = workshare.f90
  OBJ     = workshare.o
  FFLAGS = -qfree=f90 -O3
  all: mono para
  mono:
          $(FC) $(FFLAGS) -c $(SRC)
          $(FC) $(LDFLAGS) $(OBJ) -o $(CMD)_mono.exe
  para:
          $(FC) $(FFLAGS) $(OMP) -c $(SRC)
          $(FC) $(LDFLAGS) $(OMP) $(OBJ) -o $(CMD)_omp.exe
  clean:
          rm *.o *.exe
  
  $./submit.sh
  
  Monoprocessor
  
   Sum =  90000.0000000000000
  
  OMP_SCHEDULE=STATIC
     Parallel execution with  4 threads
   Sum =  90000.0000000000000
  
  OMP_SCHEDULE=DYNAMIC
     Parallel execution with  4 threads
   Sum =  -0.392987997483186611E+233
  
  OMP_SCHEDULE=GUIDED
     Parallel execution with  4 threads
   Sum =  72378.0000000000000
  

Local fix

• N/A
  

Problem summary

• PROBLEM DESCRIPTION:
  When a forall statement that performs array self-assignment
  appears in a worksharing construct, the self-assigned array may
  get corrupted under the dynamic or guided OMP scheduling.
  
  USERS AFFECTED:
  Users of forall statements that perform array self-assignment
  inside a worksharing construct and run their application under
  the dynamic or guided OMP scheduling.
  
  The test case can be simplied further:
  
  program workshare
  use OMP_LIB
  implicit none
  integer, parameter           :: n=10
  integer                      :: i, j, num_th
  real(kind=4), dimension(n,n) :: a
  
  a(:,:)=1
  
  !$OMP PARALLEL num_threads(4) private(i,j)
  !$OMP WORKSHARE
  forall (i=2:n, j=1:n-1, i>j) a(j,i) = a(i,j)
  !$OMP END WORKSHARE
  !$OMP END PARALLEL
  print *, sum(a)
  
  end program workshare
  

Problem conclusion

• Made compiler temporary variable created for self-assignment
  inside forall non-private; the problem is fixed.
  

Temporary fix

Comments

APAR Information

• APAR is sysrouted FROM one or more of the following:

  IZ69674

• APAR is sysrouted TO one or more of the following:

Fix information

• Fixed component name

  XL FORTRAN RHEL

• Fixed component ID

  5724U8410

Applicable component levels

• RC10 PSY

     UP