[fix] bug in allocation of NUMA queue from the affinity mask

runtime/src/kmp_tasking.cpp

@@ -3229,6 +3229,7 @@ static inline int __kmp_execute_tasks_template(
       if (!use_own_tasks
          && (!__kmp_queue_empty(&task_team->tt.tt_task_queues[KMP_LEVEL_THREAD][tid])
 #if LIBOMP_USE_AFFINITY
+         || !__kmp_queue_empty( &task_team->tt.tt_task_private_queues[KMP_LEVEL_THREAD][tid])
          || !__kmp_queue_empty( &task_team->tt.tt_task_queues[KMP_LEVEL_NUMA][node])
          || !__kmp_queue_empty( &task_team->tt.tt_task_private_queues[KMP_LEVEL_NUMA][node])
          || !__kmp_queue_empty( &task_team->tt.tt_task_private_queues[KMP_LEVEL_CORE][cpu])
@@ -3627,6 +3628,7 @@ static int __kmp_realloc_task_threads_data(kmp_info_t *thread,
         (kmp_queue_data_t *) __kmp_allocate( __kmp_xproc * sizeof(kmp_queue_data_t));
 #endif
       for (i = 0; i < nthreads; i++)
+      {
         __kmp_init_task_deque(&task_team->tt.tt_task_queues[KMP_LEVEL_THREAD][i],
                               KMP_LEVEL_THREAD,
                               i,
@@ -3636,6 +3638,15 @@ static int __kmp_realloc_task_threads_data(kmp_info_t *thread,
                               0
 #endif
           );
+#if LIBOMP_USE_AFFINITY
+        __kmp_init_task_deque(&task_team->tt.tt_private_task_queues[KMP_LEVEL_THREAD][i],
+                              KMP_LEVEL_THREAD,
+                              i,
+                              __kmp_cpu2node(i)
+                              0
+          );
+#endif
+      }
 #if LIBOMP_USE_AFFINITY
       for (i = 0; i < __kmp_xproc; i++)
       {
@@ -3677,9 +3688,17 @@ static int __kmp_realloc_task_threads_data(kmp_info_t *thread,
     {
       KMP_CPU_FREE(task_team->tt.tt_nodes_mask);
       task_team->tt.tt_nodes_mask = new_tt_nodes_mask;
-      int setsize = KMP_CPU_SIZE(task_team->tt.tt_nodes_mask);
+      // Here we need the last bit set in order to reserve array of queue
+      //int setsize = KMP_CPU_SIZE(task_team->tt.tt_nodes_mask);
+      int setsize = 0;
+      int i;
+      KMP_CPU_SET_ITERATE(i, task_team->tt.tt_nodes_mask) {
+        if (KMP_CPU_ISSET(i, task_team->tt.tt_nodes_mask)) setsize =i;
+      }
+      setsize++;
+
       if (task_team->tt.tt_task_queues[KMP_LEVEL_NUMA])
-         __kmp_free( task_team->tt.tt_task_queues[KMP_LEVEL_NUMA] );
+        __kmp_free( task_team->tt.tt_task_queues[KMP_LEVEL_NUMA] );
       task_team->tt.tt_task_queues[KMP_LEVEL_NUMA] = (kmp_queue_data_t *) __kmp_allocate( setsize * sizeof(kmp_queue_data_t));
 
       if (task_team->tt.tt_task_private_queues[KMP_LEVEL_NUMA])
@@ -3695,7 +3714,7 @@ static int __kmp_realloc_task_threads_data(kmp_info_t *thread,
         __kmp_init_task_deque(&task_team->tt.tt_task_queues[KMP_LEVEL_NUMA][i], KMP_LEVEL_NUMA, i, phy_index);
         __kmp_init_task_deque(&task_team->tt.tt_task_private_queues[KMP_LEVEL_NUMA][i], KMP_LEVEL_NUMA, i, phy_index);
         KA_TRACE( 5, ( "__kmp_enable_tasking: phy index: %i, array index: %i\n", phy_index, i ) );
-        //printf( "[TASK_TEAM]: NUMA %i, array index: %i\n", phy_index, i );
+        printf( "[TASK_TEAM]: NUMA %i, array index: %i\n", phy_index, i );
         i++;
       }
     }