re PR libstdc++/33893 ([parallel mode] Algorithms rely on omp_set_dynamic(false))

2007-11-22  Johannes Singler  <singler@ira.uka.de>

        PR libstdc++/33893
        * include/parallel/multiway_merge.h: made omp_dynamic-safe
        * include/parallel/workstealing.h: made omp_dynamic-safe
        * include/parallel/base.h: infrastructure, cleanup
        * include/parallel/par_loop.h: made omp_dynamic-safe
        * include/parallel/features.h: activate loser tree variant
        * include/parallel/quicksort.h: made omp_dynamic-safe
        * include/parallel/compiletime_settings.h: settings overridable
        * include/parallel/equally_split.h: made omp_dynamic-safe
        * include/parallel/omp_loop_static.h: made omp_dynamic-safe
        * include/parallel/random_shuffle.h: made omp_dynamic-safe
        * include/parallel/balanced_quicksort.h: made omp_dynamic-safe
        * include/parallel/set_operations.h: made omp_dynamic-safe
        * include/parallel/unique_copy.h: made omp_dynamic-safe
        * include/parallel/multiway_mergesort.h: made omp_dynamic-safe
        * include/parallel/search.h: made omp_dynamic-safe
        * include/parallel/partition.h: made omp_dynamic-safe
        * include/parallel/partial_sum.h: made omp_dynamic-safe
        * include/parallel/find.h: made omp_dynamic-safe
        * include/parallel/omp_loop.h: made omp_dynamic-safe
        * include/parallel/losertree.h: avoid default constructor

From-SVN: r130347

1 files changed, 56 insertions, 45 deletions

diff --git a/libstdc++-v3/include/parallel/par_loop.h b/libstdc++-v3/include/parallel/par_loop.h
index 98604cf1da4..6954e745076 100644
--- a/libstdc++-v3/include/parallel/par_loop.h
+++ b/libstdc++-v3/include/parallel/par_loop.h
@@ -41,69 +41,80 @@
 
 #include <omp.h>
 #include <parallel/settings.h>
+#include <parallel/base.h>
 
 namespace __gnu_parallel
 {
 
-  /** @brief Embarrassingly parallel algorithm for random access
-   * iterators, using hand-crafted parallelization by equal splitting
-   * the work.
-   *
-   *  @param begin Begin iterator of element sequence.
-   *  @param end End iterator of element sequence.
-   *  @param o User-supplied functor (comparator, predicate, adding
-   *  functor, ...)
-   *  @param f Functor to "process" an element with op (depends on
-   *  desired functionality, e. g. for std::for_each(), ...).
-   *  @param r Functor to "add" a single result to the already
-   *  processed elements (depends on functionality).
-   *  @param base Base value for reduction.
-   *  @param output Pointer to position where final result is written to
-   *  @param bound Maximum number of elements processed (e. g. for
-   *  std::count_n()).
-   *  @return User-supplied functor (that may contain a part of the result).
-   */
-  template<typename RandomAccessIterator, typename Op, typename Fu, typename Red, typename Result>
+/** @brief Embarrassingly parallel algorithm for random access
+  * iterators, using hand-crafted parallelization by equal splitting
+  * the work.
+  *
+  *  @param begin Begin iterator of element sequence.
+  *  @param end End iterator of element sequence.
+  *  @param o User-supplied functor (comparator, predicate, adding
+  *  functor, ...)
+  *  @param f Functor to "process" an element with op (depends on
+  *  desired functionality, e. g. for std::for_each(), ...).
+  *  @param r Functor to "add" a single result to the already
+  *  processed elements (depends on functionality).
+  *  @param base Base value for reduction.
+  *  @param output Pointer to position where final result is written to
+  *  @param bound Maximum number of elements processed (e. g. for
+  *  std::count_n()).
+  *  @return User-supplied functor (that may contain a part of the result).
+  */
+template<
+    typename RandomAccessIterator,
+    typename Op,
+    typename Fu,
+    typename Red,
+    typename Result>
   Op
-  for_each_template_random_access_ed(RandomAccessIterator begin,
-				     RandomAccessIterator end, Op o, Fu& f,
-				     Red r, Result base, Result& output,
-				     typename std::iterator_traits<RandomAccessIterator>::difference_type bound)
+  for_each_template_random_access_ed(
+              RandomAccessIterator begin,
+              RandomAccessIterator end,
+              Op o, Fu& f, Red r, Result base, Result& output,
+              typename std::iterator_traits<RandomAccessIterator>::
+                  difference_type bound)
   {
     typedef std::iterator_traits<RandomAccessIterator> traits_type;
     typedef typename traits_type::difference_type difference_type;
 
     const difference_type length = end - begin;
-    const difference_type settings_threads = static_cast<difference_type>(get_max_threads());
-    const difference_type dmin = settings_threads < length ? settings_threads : length;
-    const difference_type dmax = dmin > 1 ? dmin : 1;
+    Result *thread_results;
 
-    thread_index_t num_threads = static_cast<thread_index_t>(dmax);
+    thread_index_t num_threads =
+        __gnu_parallel::min<difference_type>(get_max_threads(), length);
 
+#   pragma omp parallel num_threads(num_threads)
+      {
+#       pragma omp single
+          {
+            num_threads = omp_get_num_threads();
+            thread_results = new Result[num_threads];
+          }
 
-    Result *thread_results = new Result[num_threads];
+        thread_index_t iam = omp_get_thread_num();
 
-#pragma omp parallel num_threads(num_threads)
-    {
-      // Neutral element.
-      Result reduct = Result();
+        // Neutral element.
+        Result reduct = Result();
 
-      thread_index_t p = num_threads;
-      thread_index_t iam = omp_get_thread_num();
-      difference_type start = iam * length / p;
-      difference_type limit = (iam == p - 1) ? length : (iam + 1) * length / p;
+        difference_type
+            start = equally_split_point(length, num_threads, iam),
+            stop = equally_split_point(length, num_threads, iam + 1);
 
-      if (start < limit)
-	{
-	  reduct = f(o, begin + start);
-	  start++;
-	}
+        if (start < stop)
+          {
+            reduct = f(o, begin + start);
+            ++start;
+          }
 
-      for (; start < limit; start++)
-	reduct = r(reduct, f(o, begin + start));
+        for (; start < stop; ++start)
+          reduct = r(reduct, f(o, begin + start));
 
-      thread_results[iam] = reduct;
-    }
+        thread_results[iam] = reduct;
+      } //parallel
 
     for (thread_index_t i = 0; i < num_threads; i++)
       output = r(output, thread_results[i]);