template<class Xpr, class T, class Reducer>
T fold(Xpr expr, T init, Reducer reducer)
 {
   #pragma omp parallel
   {
     T pinit = init;
     #pragma omp for nowait
     for(int n=1; n<=size; ++n)
       pinit = reducer(pinit, run(expr, n) );
     #pragma omp atomic
     init = reducer(pinit,init);
   }
   return init;
 }