v1.3.3: pool optimization, broadcast

tugrul512bit · Jun 15, 2017 · 2e8347c · 2e8347c
1 parent 0cac6fc
commit 2e8347c
Showing 1 changed file with 29 additions and 9 deletions.
diff --git a/Cekirdekler/Cekirdekler/ClPipeline.cs b/Cekirdekler/Cekirdekler/ClPipeline.cs
@@ -3252,7 +3252,7 @@ public enum ClTaskType:int
                 TASK_MESSAGE_DEFAULT = 0,
 
                 /// <summary>
-                /// <para>not implemeneted</para>
+                /// <para></para>
                 /// <para>enables single device mode and selects a device for the following tasks until SELECT_END is reached.</para>
                 /// <para>if SELECT_END is quickly enqueued on device, remaining tasks can be enqueued to other devices concurrently</para>
                 /// <para>multiple select groups can run on different devices</para>
@@ -3262,7 +3262,7 @@ public enum ClTaskType:int
                 TASK_MESSAGE_DEVICE_SELECT_BEGIN = 1,
 
                 /// <summary>
-                /// <para>not implemeneted</para>
+                /// <para></para>
                 /// <para>deselects device of single device mode if it is enabled</para>
                 /// <para>this does not ensure any synchronization between devices. Other devices can compute other groups of tasks</para>
                 /// <para>SELECT_END must exist between two SELECT_BEGIN. Multiple begin-end ranges must not overlap</para>
@@ -3282,8 +3282,10 @@ public enum ClTaskType:int
                 TASK_MESSAGE_GLOBAL_SYNCHRONIZATION_LAST = 8,
 
                 /// <summary>
-                /// <para>not implemeneted</para>
-                /// <para>runs this task on all devices when device pool intercepts this(writing results to host should be done by only 1 device)</para>
+                /// <para>runs this task on all devices when device pool intercepts this</para>
+                /// <para>must not write to host(array.write or array.writeAll must not be enabled)</para>
+                /// <para>meant for only initializing all devices' buffers, read-only, even on kernel side</para>
+                /// <para></para>
                 /// </summary>
                 TASK_MESSAGE_BROADCAST = 16,
 
@@ -4158,9 +4160,27 @@ void produceTasksComputeAtWill()
                                     // if single device mode has ended
                                     handleSingleDeviceModeEnd(data.task.type);
 
+                                    // if task is to be run on all devices
+                                    if ((data.task.type & ClTaskType.TASK_MESSAGE_BROADCAST) > 0)
+                                    {
+                                        for (int i = 0; i < numDevices; i++)
+                                        {
+                                            ClPoolTaskIdPair broadcastedTask = new ClPoolTaskIdPair();
+                                            broadcastedTask.id = data.id;
+                                            broadcastedTask.serialMode = data.serialMode;
+                                            broadcastedTask.deviceIndex = i;
+                                            broadcastedTask.task = data.task.duplicate();
+                                            pipe.push(broadcastedTask);
+                                        }
+                                    }
+                                    else
+                                    {
 
-                                    // sends to common queue that all devices consume
-                                    pipe.push(data);
+                                        // sends to common queue, an idle device will consume
+                                        pipe.push(data);
+                                    }
+
+                                    // reactivate device threads to process new data 
                                     wakeDevices();
 
                                     // if global synchronization_last is taken
@@ -4685,7 +4705,7 @@ void consumeTasksComputeAtWill()
                         }
 
 
-
+                        // getting all tasks quickly to cache if all are prepared for this device only
                         bool tasksForThisDeviceExist = true;
                         int tmpCtr = 0;
                         while(tasksForThisDeviceExist && tmpCtr<20)
@@ -4695,7 +4715,7 @@ void consumeTasksComputeAtWill()
                             if (newCacheData != null)
                             {
                                 cachePipe.push(newCacheData);
-                                if(newCacheData.deviceIndex==deviceIndex)
+                                if((newCacheData.deviceIndex==deviceIndex) || (newCacheData.serialMode))
                                 {
                                     // this is single device mode enabled series of tasks
                                     tasksForThisDeviceExist = true;
@@ -4758,7 +4778,7 @@ void consumeTasksComputeAtWill()
                                 {
                                     float probability = (((float)(data.task.totalTasks- data.task.remainingTasks)) / ((float)(data.task.totalTasks + 1)));
                                     float testing = (float)rand.NextDouble();
-                                    if ((testing < (probability*probability)) /*|| ((data.task.totalTasks - data.task.remainingTasks) > (data.task.totalTasks - 3))*/)
+                                    if ((testing < (probability*probability) || (data.task.remainingTasks <=2 )) /*|| ((data.task.totalTasks - data.task.remainingTasks) > (data.task.totalTasks - 3))*/)
                                     {
                                         numberCruncher.fineGrainedQueueControl = true;
                                     }