dash::copy: Pass local_chunks by the caller

This way one can call copy_impl multiple times, before triggering local
copies.

Author: bertwesarg

dash/include/dash/algorithm/Copy.h

@@ -135,10 +135,11 @@ template <
   typename ValueType,
   class GlobInputIt >
 ValueType * copy_impl(
-  GlobInputIt                  begin,
-  GlobInputIt                  end,
-  ValueType                  * out_first,
-  std::vector<dart_handle_t> * handles)
+  GlobInputIt                    begin,
+  GlobInputIt                    end,
+  ValueType                    * out_first,
+  std::vector<dart_handle_t>   * handles,
+  local_copy_chunks<ValueType> & local_chunks)
 {
   DASH_LOG_TRACE("dash::internal::copy_impl() global -> local",
                  "in_first:",  begin.pos(),
@@ -159,8 +160,6 @@ ValueType * copy_impl(
 
   ContiguousRangeSet<GlobInputIt> range_set{begin, end};
 
-  local_copy_chunks<ValueType> local_chunks;
-
   //
   // Copy elements from every unit:
   //
@@ -204,8 +203,6 @@ ValueType * copy_impl(
     num_elem_copied += num_copy_elem;
   }
 
-  do_local_copies(local_chunks);
-
   DASH_ASSERT_EQ(num_elem_copied, num_elem_total,
                  "Failed to find all contiguous subranges in range");
 
@@ -226,10 +223,11 @@ template <
   typename ValueType,
   class GlobOutputIt >
 GlobOutputIt copy_impl(
-  ValueType                  * begin,
-  ValueType                  * end,
-  GlobOutputIt                 out_first,
-  std::vector<dart_handle_t> * handles)
+  ValueType                    * begin,
+  ValueType                    * end,
+  GlobOutputIt                   out_first,
+  std::vector<dart_handle_t>   * handles,
+  local_copy_chunks<ValueType> & local_chunks)
 {
 
   DASH_LOG_TRACE("dash::copy_impl() local -> global",
@@ -254,8 +252,6 @@ GlobOutputIt copy_impl(
 
   ContiguousRangeSet<GlobOutputIt> range_set{out_first, out_last};
 
-  local_copy_chunks<value_type> local_chunks;
-
   auto in_first = begin;
 
   //
@@ -300,8 +296,6 @@ GlobOutputIt copy_impl(
     num_elem_copied += num_copy_elem;
   }
 
-  do_local_copies(local_chunks);
-
   DASH_ASSERT_EQ(num_elem_copied, num_elem_total,
                  "Failed to find all contiguous subranges in range");
 
@@ -338,9 +332,10 @@ dash::Future<ValueType *> copy_async(
   }
 
   auto handles = std::make_shared<std::vector<dart_handle_t>>();
-
-  auto out_last = dash::internal::copy_impl(in_first, in_last,
-                                            out_first, handles.get());
+  dash::internal::local_copy_chunks<ValueType> local_chunks;
+  auto out_last = dash::internal::copy_impl(in_first, in_last, out_first,
+                                            handles.get(), local_chunks);
+  dash::internal::do_local_copies(local_chunks);
 
   if (handles->empty()) {
     DASH_LOG_TRACE("dash::copy_async", "all transfers completed");
@@ -422,12 +417,15 @@ ValueType * copy(
     return out_first;
   }
 
+  dash::internal::local_copy_chunks<ValueType> local_chunks;
 #if DASH_COPY_USE_HANDLES
   std::vector<dart_handle_t> handles;
   auto out_last = dash::internal::copy_impl(in_first,
                                             in_last,
                                             out_first,
-                                            &handles);
+                                            &handles,
+                                            local_chunks);
+  dash::internal::do_local_copies(local_chunks);
   if (!handles.empty()) {
     DASH_LOG_TRACE("dash::copy", "Waiting for remote transfers to complete,",
                   "num_handles: ", handles.size());
@@ -438,7 +436,9 @@ ValueType * copy(
   auto out_last = dash::internal::copy_impl(in_first,
                                             in_last,
                                             out_first,
-                                            nullptr);
+                                            nullptr,
+                                            local_chunks);
+  dash::internal::do_local_copies(local_chunks);
   dart_flush_local(in_first.dart_gptr());
 #endif // DASH_COPY_USE_HANDLES
 
@@ -471,10 +471,13 @@ dash::Future<GlobOutputIt> copy_async(
   }
 
   auto handles  = std::make_shared<std::vector<dart_handle_t>>();
+  dash::internal::local_copy_chunks<ValueType> local_chunks;
   auto out_last = dash::internal::copy_impl(in_first,
                                             in_last,
                                             out_first,
-                                            handles.get());
+                                            handles.get(),
+                                            local_chunks);
+  dash::internal::do_local_copies(local_chunks);
 
   if (handles->empty()) {
     return dash::Future<GlobOutputIt>(out_last);
@@ -544,12 +547,15 @@ GlobOutputIt copy(
   DASH_LOG_TRACE("dash::copy()", "blocking, local to global");
   // handles to wait on at the end
 
+  dash::internal::local_copy_chunks<ValueType> local_chunks;
 #if DASH_COPY_USE_HANDLES
   std::vector<dart_handle_t> handles;
   auto out_last = dash::internal::copy_impl(in_first,
                                             in_last,
                                             out_first,
-                                            &handles);
+                                            &handles,
+                                            local_chunks);
+  dash::internal::do_local_copies(local_chunks);
 
   if (!handles.empty()) {
     DASH_LOG_TRACE("dash::copy", "Waiting for remote transfers to complete,",
@@ -562,6 +568,7 @@ GlobOutputIt copy(
                                             in_last,
                                             out_first,
                                             nullptr);
+  dash::internal::do_local_copies(local_chunks);
   dart_flush(out_first.dart_gptr());
 #endif
   return out_last;