You need !$OMP SINGLE right after the start of parallel code, otherwise the tasks will get duplicated by all threads.

You need some ordinary version of omp syntax which will permit the run-time to partition the loop into threads, but you imply this "is not even possible."  For example, you could make an ordinary counted loop:

iter_set_by = nmax

logical_var = .true.

!$omp parallel do schedule(runtime) shared(logical_var, iter_set_by)

do i=1, nmax

   if(logical_var .or. i < iter_set_by)then

......

     if(exit_condition_satisfied)then

          logical_var = .false.

          iter_set_by = i

     endif

   endif

enddo

Once logical_var goes false, the iteration count for each chunk can be exhausted doing nothing, but you have to give each thread some work which doesn't depend on the others before anybody has set the exit condition, otherwise you have no useful parallelism.

You would also want to make logical_var volatile (to assure it is read on each test). Also consider using exit to the (of the) matching scope of the current do (OpenMP ought to permit slice termination by exit to matching END DO). Caution, should your do loop have a barrier, then you must not exit the loop. Use if(logical_var) as TimP shows, but place barrier outside this  then block (i.e. assure all threads reach barrier). N.B. TimP had "if(logical_expr)", I think he meant  "if(logical_var)". Also, do not use exit from within a critical section (as this would bypass the end critical section).

The TASK format may be a good alternative.

Jim Dempsey

I'm sorry for the incomplete editing of my suggestion.

As Jim says, making the shared control variables volatile (or using a critical) would assure more immediate recognition of a change, but still you would likely perform some extra loop iterations compared with the pure serial case, so you must consider whether that is valid.