public void Foreach(T[] array, Action action)
{
   int cpus = 2 * Environment.ProcessorCount;
   int nitems = array.Length;
   int chunk = nitems / cpus;
   int counter = cpus;

   using (AutoResetEvent signal = new AutoResetEvent(false))
   {
      for (int i = 1; i <= cpus; i++)
      {
         ThreadPool.QueueUserWorkItem(delegate(object o)
         {
            int unit = (int)o;

            for (int j = (unit - 1) * chunk;
                  j < (unit == cpus ? nitems : unit * chunk);
                  j++)
            {
               action(array[j]);
            }

            if (Interlocked.Decrement(ref counter) == 0)
            signal.Set();
         }, i);
      }

      signal.WaitOne();
   }
}

public void ForeachWithResults(T[] array, T1[] results, Action action)
{
   int cpus = 2 * Environment.ProcessorCount;
   int nitems = array.Length;
   int chunk = nitems / cpus;
   int counter = cpus;

   using (AutoResetEvent signal = new AutoResetEvent(false))
   {
      for (int i = 1; i <= cpus; i++)
      {
         ThreadPool.QueueUserWorkItem(delegate(object o)
         {
            int unit = (int)o;

            for (int j = (unit - 1) * chunk;
                  j < (unit == cpus ? nitems : unit * chunk);
                  j++)
            {
               action(j, array[j], results);
            }

            if (Interlocked.Decrement(ref counter) == 0)
               signal.Set();
         }, i);
      }

      signal.WaitOne();
   }
}