Fixed memory leak enqueuing marker

Author: gwoltman

src/Queue.cpp

@@ -92,7 +92,7 @@ void Queue::queueMarkerEvent() {
     }
     // Enqueue a marker for nVidia GPUs
     else {
-      clEnqueueMarkerWithWaitList(get(), 0, NULL, &markerEvent);
+      markerEvent = enqueueMarker(get());
       markerQueued = true;
       queueCount = 0;
     }
@@ -103,7 +103,7 @@ void Queue::waitForMarkerEvent() {
   if (!markerQueued) return;
   // By default, nVidia finish causes a CPU busy wait.  Instead, sleep for a while.  Since we know how many items are enqueued after the marker we can make an
   // educated guess of how long to sleep to keep CPU overhead low.
-  while (getEventInfo(markerEvent) != CL_COMPLETE) {
+  while (getEventInfo(markerEvent.get()) != CL_COMPLETE) {
     // There are 4, 7, or 10 kernels per squaring.  Don't overestimate sleep time.  Divide by much more than the number of kernels.
     std::this_thread::sleep_for(std::chrono::microseconds(1 + queueCount * squareTime / squareKernels / 2));
   }

src/Queue.h

@@ -55,7 +55,7 @@ class Queue : public QueueHolder {
 
 private:                            // This replaces the "call queue->finish every 400 squarings" code in Gpu.cpp.  Solves the busy wait on nVidia GPUs.
   int MAX_QUEUE_COUNT;              // Queue size before a marker will be enqueued.  Typically, 100 to 1000 squarings.
-  cl_event markerEvent;             // Event associated with an enqueued marker placed in the queue every MAX_QUEUE_COUNT entries and before r/w operations.
+  EventHolder markerEvent;          // Event associated with an enqueued marker placed in the queue every MAX_QUEUE_COUNT entries and before r/w operations.
   bool markerQueued;                // TRUE if a marker and event have been queued
   int queueCount;                   // Count of items added to the queue since last marker
   int squareTime;                   // Time to do one squaring (in microseconds)

src/clwrap.cpp

@@ -363,6 +363,12 @@ EventHolder fillBuf(cl_queue q, vector<cl_event>&& waits,
   return genEvent ? EventHolder{event} : EventHolder{};
 }
 
+EventHolder enqueueMarker(cl_queue q) {
+  cl_event event{};
+  CHECK1(clEnqueueMarkerWithWaitList(q, 0, 0, &event));
+  return EventHolder{event};
+}
+
 void waitForEvents(vector<cl_event>&& waits) {
   if (!waits.empty()) {
     CHECK1(clWaitForEvents(waits.size(), waits.data()));

src/clwrap.h

@@ -106,6 +106,8 @@ EventHolder copyBuf(cl_queue queue, vector<cl_event>&& waits, const cl_mem src,
 
 EventHolder fillBuf(cl_queue q, vector<cl_event>&& waits, cl_mem buf, const void *pat, size_t patSize, size_t size, bool genEvent);
 
+EventHolder enqueueMarker(cl_queue q);
+
 void waitForEvents(vector<cl_event>&& waits);