--- a/src/parallel_execution/MultiDimIterator.cbc	Wed Sep 06 22:01:27 2017 +0900
+++ b/src/parallel_execution/MultiDimIterator.cbc	Thu Sep 14 02:35:20 2017 +0900
@@ -53,14 +53,16 @@
     return task1;
 }
 
-__code execMultiDimIterator(struct MultiDimIterator* iterator, struct TaskManager* taskManager, struct Context* task, __code next(...)) {
-    int x = iterator->counterX;
-    int y = iterator->counterY;
-    int z = iterator->counterZ;
-    struct Context* iterateTask = createMultiDimIterateTask(task, x, y, z);
+__code execMultiDimIterator(struct MultiDimIterator* iterator, struct TaskManager* taskManager, struct Context* task, int numGPU, __code next(...)) {
+    // No GPU device
+    if (numGPU == 0) {
+        goto meta(context, C_execMultiDimIterator1);
+    }
+    task->iterate = 1;
+    task->gpu = 1;
     taskManager->taskManager = (union Data*)task->taskManager;
-    taskManager->context = iterateTask;
-    taskManager->next = C_execMultiDimIterator1;
+    taskManager->context = task;
+    taskManager->next = next;
     goto meta(context, task->taskManager->spawn);
 }
 
@@ -68,11 +70,31 @@
     MultiDimIterator* iterator = (MultiDimIterator*)GearImpl(context, Iterator, iterator);
     TaskManager* taskManager = Gearef(context, TaskManager);
     Context* task = Gearef(context, Iterator)->task;
+    int numGPU = Gearef(context, Iterator)->numGPU;
     enum Code next = Gearef(context, Iterator)->next;
-    goto execMultiDimIterator(context, iterator, taskManager, task, next);
+    goto execMultiDimIterator(context, iterator, taskManager, task, numGPU, next);
 } 
 
-__code execMultiDimIterator1(struct MultiDimIterator* iterator, struct Context* task, __code next(...)) {
+__code execMultiDimIterator1(struct MultiDimIterator* iterator, struct TaskManager* taskManager, struct Context* task, __code next(...)) {
+    int x = iterator->counterX;
+    int y = iterator->counterY;
+    int z = iterator->counterZ;
+    struct Context* iterateTask = createMultiDimIterateTask(task, x, y, z);
+    taskManager->taskManager = (union Data*)task->taskManager;
+    taskManager->context = iterateTask;
+    taskManager->next = C_execMultiDimIterator2;
+    goto meta(context, task->taskManager->spawn);
+}
+
+__code execMultiDimIterator1_stub(struct Context* context) {
+    MultiDimIterator* iterator = (MultiDimIterator*)GearImpl(context, Iterator, iterator);
+    TaskManager* taskManager = Gearef(context, TaskManager);
+    Context* task = Gearef(context, Iterator)->task;
+    enum Code next = Gearef(context, Iterator)->next;
+    goto execMultiDimIterator1(context, iterator, taskManager, task, next);
+} 
+
+__code execMultiDimIterator2(struct MultiDimIterator* iterator, struct Context* task, __code next(...)) {
     if (++iterator->counterX >= iterator->x) {
         iterator->counterX = 0;
         if (++iterator->counterY >= iterator->y) {
@@ -83,7 +105,7 @@
             }
         }
     }
-    goto meta(context, C_execMultiDimIterator);
+    goto meta(context, C_execMultiDimIterator1);
 }
 
 __code barrierMultiDimIterator(struct MultiDimIterator* iterator, struct Context* task, __code next(...), __code whenWait(...)) {