From: Martin Kronbichler <kronbichler@lnm.mw.tum.de>
Date: Fri, 8 May 2020 09:39:15 +0000 (+0200)
Subject: Fix matrix-free pre/post loops with threads
X-Git-Tag: v9.2.0-rc1~67^2~1
X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=dbe71e890d6351a1aeadd326058a161363dbb7d5;p=dealii.git

Fix matrix-free pre/post loops with threads
---

diff --git a/include/deal.II/matrix_free/matrix_free.h b/include/deal.II/matrix_free/matrix_free.h
index 4427698ade..67c262b017 100644
--- a/include/deal.II/matrix_free/matrix_free.h
+++ b/include/deal.II/matrix_free/matrix_free.h
@@ -4568,13 +4568,26 @@ namespace internal
         {
           const internal::MatrixFreeFunctions::DoFInfo &dof_info =
             matrix_free.get_dof_info(dof_handler_index_pre_post);
-          AssertIndexRange(range_index,
-                           dof_info.cell_loop_pre_list_index.size() - 1);
-          for (unsigned int id = dof_info.cell_loop_pre_list_index[range_index];
-               id != dof_info.cell_loop_pre_list_index[range_index + 1];
-               ++id)
-            operation_before_loop(dof_info.cell_loop_pre_list[id].first,
-                                  dof_info.cell_loop_pre_list[id].second);
+          if (range_index == numbers::invalid_unsigned_int)
+            {
+              // Case with threaded loop -> currently no overlap implemented
+              parallel::apply_to_subranges(
+                0U,
+                dof_info.vector_partitioner->local_size(),
+                operation_before_loop,
+                internal::VectorImplementation::minimum_parallel_grain_size);
+            }
+          else
+            {
+              AssertIndexRange(range_index,
+                               dof_info.cell_loop_pre_list_index.size() - 1);
+              for (unsigned int id =
+                     dof_info.cell_loop_pre_list_index[range_index];
+                   id != dof_info.cell_loop_pre_list_index[range_index + 1];
+                   ++id)
+                operation_before_loop(dof_info.cell_loop_pre_list[id].first,
+                                      dof_info.cell_loop_pre_list[id].second);
+            }
         }
     }
 
@@ -4585,14 +4598,26 @@ namespace internal
         {
           const internal::MatrixFreeFunctions::DoFInfo &dof_info =
             matrix_free.get_dof_info(dof_handler_index_pre_post);
-          AssertIndexRange(range_index,
-                           dof_info.cell_loop_post_list_index.size() - 1);
-          for (unsigned int id =
-                 dof_info.cell_loop_post_list_index[range_index];
-               id != dof_info.cell_loop_post_list_index[range_index + 1];
-               ++id)
-            operation_after_loop(dof_info.cell_loop_post_list[id].first,
-                                 dof_info.cell_loop_post_list[id].second);
+          if (range_index == numbers::invalid_unsigned_int)
+            {
+              // Case with threaded loop -> currently no overlap implemented
+              parallel::apply_to_subranges(
+                0U,
+                dof_info.vector_partitioner->local_size(),
+                operation_after_loop,
+                internal::VectorImplementation::minimum_parallel_grain_size);
+            }
+          else
+            {
+              AssertIndexRange(range_index,
+                               dof_info.cell_loop_post_list_index.size() - 1);
+              for (unsigned int id =
+                     dof_info.cell_loop_post_list_index[range_index];
+                   id != dof_info.cell_loop_post_list_index[range_index + 1];
+                   ++id)
+                operation_after_loop(dof_info.cell_loop_post_list[id].first,
+                                     dof_info.cell_loop_post_list[id].second);
+            }
         }
     }
 
diff --git a/source/matrix_free/task_info.cc b/source/matrix_free/task_info.cc
index cbae4ea508..831b23fc0b 100644
--- a/source/matrix_free/task_info.cc
+++ b/source/matrix_free/task_info.cc
@@ -338,7 +338,13 @@ namespace internal
     void
     TaskInfo::loop(MFWorkerInterface &funct) const
     {
-      if (scheme == none)
+      // If we use thread parallelism, we do not currently support to schedule
+      // pieces of updates within the loop, so this index will collect all
+      // calls in that case and work like a single complete loop over all
+      // cells
+      if (scheme != none)
+        funct.cell_loop_pre_range(numbers::invalid_unsigned_int);
+      else
         funct.cell_loop_pre_range(
           partition_row_index[partition_row_index.size() - 2]);
 
@@ -618,7 +624,10 @@ namespace internal
             }
         }
       funct.vector_compress_finish();
-      if (scheme == none)
+
+      if (scheme != none)
+        funct.cell_loop_post_range(numbers::invalid_unsigned_int);
+      else
         funct.cell_loop_post_range(
           partition_row_index[partition_row_index.size() - 2]);
     }