Fix data race in nestkernel/per_thread_bool_indicator.cpp

Author: jprotze

nestkernel/connection_manager.cpp

@@ -876,13 +876,13 @@ nest::ConnectionManager::connect_( Node& source,
   {
 #pragma omp atomic write
     has_primary_connections_ = true;
-    check_primary_connections_[ tid ].set_true();
+    check_primary_connections_.set_true( tid );
   }
   else if ( check_secondary_connections_[ tid ].is_false() and not is_primary )
   {
 #pragma omp atomic write
     secondary_connections_exist_ = true;
-    check_secondary_connections_[ tid ].set_true();
+    check_secondary_connections_.set_true( tid );
   }
 }
 

nestkernel/event_delivery_manager.cpp

@@ -790,7 +790,7 @@ EventDeliveryManager::gather_target_data( const size_t tid )
   assert( not kernel().connection_manager.is_source_table_cleared() );
 
   // assume all threads have some work to do
-  gather_completed_checker_[ tid ].set_false();
+  gather_completed_checker_.set_false( tid );
   assert( gather_completed_checker_.all_false() );
 
   const AssignedRanks assigned_ranks = kernel().vp_manager.get_assigned_ranks( tid );
@@ -802,7 +802,7 @@ EventDeliveryManager::gather_target_data( const size_t tid )
   {
     // assume this is the last gather round and change to false
     // otherwise
-    gather_completed_checker_[ tid ].set_true();
+    gather_completed_checker_.set_true( tid );
 
 #pragma omp master
     {
@@ -819,7 +819,7 @@ EventDeliveryManager::gather_target_data( const size_t tid )
       assigned_ranks, kernel().mpi_manager.get_send_recv_count_target_data_per_rank() );
 
     const bool gather_completed = collocate_target_data_buffers_( tid, assigned_ranks, send_buffer_position );
-    gather_completed_checker_[ tid ].logical_and( gather_completed );
+    gather_completed_checker_.logical_and( tid, gather_completed );
 
     if ( gather_completed_checker_.all_true() )
     {
@@ -842,7 +842,7 @@ EventDeliveryManager::gather_target_data( const size_t tid )
 #pragma omp barrier
 
     const bool distribute_completed = distribute_target_data_buffers_( tid );
-    gather_completed_checker_[ tid ].logical_and( distribute_completed );
+    gather_completed_checker_.logical_and( tid, distribute_completed );
 
     // resize mpi buffers, if necessary and allowed
     if ( gather_completed_checker_.any_false() and kernel().mpi_manager.adaptive_target_buffers() )
@@ -864,7 +864,7 @@ EventDeliveryManager::gather_target_data_compressed( const size_t tid )
   assert( not kernel().connection_manager.is_source_table_cleared() );
 
   // assume all threads have some work to do
-  gather_completed_checker_[ tid ].set_false();
+  gather_completed_checker_.set_false( tid );
   assert( gather_completed_checker_.all_false() );
 
   const AssignedRanks assigned_ranks = kernel().vp_manager.get_assigned_ranks( tid );
@@ -874,7 +874,7 @@ EventDeliveryManager::gather_target_data_compressed( const size_t tid )
   while ( gather_completed_checker_.any_false() )
   {
     // assume this is the last gather round and change to false otherwise
-    gather_completed_checker_[ tid ].set_true();
+    gather_completed_checker_.set_true( tid );
 
 #pragma omp master
     {
@@ -891,7 +891,7 @@ EventDeliveryManager::gather_target_data_compressed( const size_t tid )
     const bool gather_completed =
       collocate_target_data_buffers_compressed_( tid, assigned_ranks, send_buffer_position );
 
-    gather_completed_checker_[ tid ].logical_and( gather_completed );
+    gather_completed_checker_.logical_and( tid, gather_completed );
 
     if ( gather_completed_checker_.all_true() )
     {
@@ -916,7 +916,7 @@ EventDeliveryManager::gather_target_data_compressed( const size_t tid )
     // all data it is responsible for to buffers. Now combine with information on whether other ranks
     // have sent all their data. Note: All threads will return the same value for distribute_completed.
     const bool distribute_completed = distribute_target_data_buffers_( tid );
-    gather_completed_checker_[ tid ].logical_and( distribute_completed );
+    gather_completed_checker_.logical_and( tid, distribute_completed );
 
     // resize mpi buffers, if necessary and allowed
     if ( gather_completed_checker_.any_false() and kernel().mpi_manager.adaptive_target_buffers() )

nestkernel/per_thread_bool_indicator.cpp

@@ -50,62 +50,47 @@ PerThreadBoolIndicator::initialize( const size_t num_threads, const bool status
   kernel().vp_manager.assert_single_threaded();
   per_thread_status_.clear();
   per_thread_status_.resize( num_threads, BoolIndicatorUInt64( status ) );
+  size_ = num_threads;
+  if ( status )
+    are_true_ = num_threads;
+  else
+    are_true_ = 0;
 }
 
 bool
 PerThreadBoolIndicator::all_false() const
 {
 #pragma omp barrier
-  for ( auto it = per_thread_status_.begin(); it < per_thread_status_.end(); ++it )
-  {
-    if ( it->is_true() )
-    {
-      return false;
-    }
-  }
-  return true;
+  bool ret = ( are_true_ == 0 );
+#pragma omp barrier
+  return ret;
 }
 
 bool
 PerThreadBoolIndicator::all_true() const
 {
 #pragma omp barrier
-  for ( auto it = per_thread_status_.begin(); it < per_thread_status_.end(); ++it )
-  {
-    if ( it->is_false() )
-    {
-      return false;
-    }
-  }
-  return true;
+  bool ret = ( are_true_ == size_ );
+#pragma omp barrier
+  return ret;
 }
 
 bool
 PerThreadBoolIndicator::any_false() const
 {
 #pragma omp barrier
-  for ( auto it = per_thread_status_.begin(); it < per_thread_status_.end(); ++it )
-  {
-    if ( it->is_false() )
-    {
-      return true;
-    }
-  }
-  return false;
+  bool ret = ( are_true_ < size_ );
+#pragma omp barrier
+  return ret;
 }
 
 bool
 PerThreadBoolIndicator::any_true() const
 {
 #pragma omp barrier
-  for ( auto it = per_thread_status_.begin(); it < per_thread_status_.end(); ++it )
-  {
-    if ( it->is_true() )
-    {
-      return true;
-    }
-  }
-  return false;
+  bool ret = ( are_true_ > 0 );
+#pragma omp barrier
+  return ret;
 }
 
 } // namespace nest

nestkernel/per_thread_bool_indicator.h

@@ -52,15 +52,17 @@ class BoolIndicatorUInt64
   bool is_true() const;
   bool is_false() const;
 
+
+protected:
   void set_true();
   void set_false();
-
   void logical_and( const bool status );
 
 private:
   static constexpr std::uint_fast64_t true_uint64 = true;
   static constexpr std::uint_fast64_t false_uint64 = false;
   std::uint_fast64_t status_;
+  friend class PerThreadBoolIndicator;
 };
 
 inline bool
@@ -106,6 +108,34 @@ class PerThreadBoolIndicator
 
   BoolIndicatorUInt64& operator[]( const size_t tid );
 
+  void
+  set_true( const size_t tid )
+  {
+    if ( per_thread_status_[ tid ].is_false() )
+    {
+      are_true_++;
+      per_thread_status_[ tid ].set_true();
+    }
+  }
+  void
+  set_false( const size_t tid )
+  {
+    if ( per_thread_status_[ tid ].is_true() )
+    {
+      are_true_--;
+      per_thread_status_[ tid ].set_false();
+    }
+  }
+  void
+  logical_and( const size_t tid, const bool status )
+  {
+    if ( per_thread_status_[ tid ].is_true() && !status )
+    {
+      are_true_--;
+      per_thread_status_[ tid ].set_false();
+    }
+  }
+
   /**
    * Resize to the given number of threads and set all elements to false.
    */
@@ -133,6 +163,7 @@ class PerThreadBoolIndicator
 
 private:
   std::vector< BoolIndicatorUInt64 > per_thread_status_;
+  std::atomic< int > size_ { 0 }, are_true_ { 0 };
 };
 
 } // namespace nest

nestkernel/source_table.h

@@ -372,7 +372,7 @@ SourceTable::clear( const size_t tid )
     it->clear();
   }
   sources_[ tid ].clear();
-  is_cleared_[ tid ].set_true();
+  is_cleared_.set_true( tid );
 }
 
 inline void
@@ -412,15 +412,15 @@ SourceTable::save_entry_point( const size_t tid )
       assert( current_positions_[ tid ].lcid == -1 );
       saved_positions_[ tid ].lcid = -1;
     }
-    saved_entry_point_[ tid ].set_true();
+    saved_entry_point_.set_true( tid );
   }
 }
 
 inline void
 SourceTable::restore_entry_point( const size_t tid )
 {
   current_positions_[ tid ] = saved_positions_[ tid ];
-  saved_entry_point_[ tid ].set_false();
+  saved_entry_point_.set_false( tid );
 }
 
 inline void