OpenMP question

[Richard Guenther]

Hi!

Any OpenMP experts on the table? How would one do manual reduction
parallelized with OpenMP like from Evaluator/ReductionEvaluator.h:

  template<class T, class Op, class Expr, class Domain>
  inline static void evaluate(T &ret, const Op &op, const Expr &e,
    const Domain &domain, WrappedInt<1>)
  {
    Expr localExpr(e);
    int e0 = domain[0].length();

    T answer = ReductionTraits<Op, T>::identity();
    for (int i0 = 0; i0 < e0; ++i0)
      op(answer, localExpr.read(i0));

    ret = answer;
  }

So the first part may be easy, just(?)

    T answer = ReductionTraits<Op, T>::identity();
#pragma omp parallel for private (answer)
    for (int i0 = 0; i0 < e0; ++i0)
      op(answer, localExpr.read(i0));

but how do the final reduction on the multiple private answer's? One cant
use the reduction functionality from OpenMP with these C++ constructs
here, sadly. The stranges version I could come up with is (I'm sure it
still won't work):

#pragma omp parallel shared (answer)
    {
      T answer[omp_get_num_threads()]; // is probably private now... :/
      int n = omp_get_thread_num();
      answer[n] = ReductionTraits<Op, T>::identity();
#pragma omp for
      for (int i0 = 0; i0 < e0; ++i0)
        op(answer[n], localExpr.read(i0));
#pragma omp master
      {
        for (int i = 1; i<omp_get_num_threads(); ++i)
          op(answer[0], answer[i]);

        ret = answer[0];
      }
    }


Any hints on doing OpenMP reductions with C++ operators here from anyone?

Thanks in advance!

Richard.