[SYCL][L0] Fix memory leak in pi Queue

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -1991,8 +1991,7 @@ pi_result piQueueRetain(pi_queue Queue) {
 pi_result piQueueRelease(pi_queue Queue) {
   PI_ASSERT(Queue, PI_INVALID_QUEUE);
 
-  // Lock automatically releases when this goes out of scope.
-  std::lock_guard<std::mutex> lock(Queue->PiQueueMutex);
+  Queue->PiQueueMutex.lock();
 
   if (--(Queue->RefCount) == 0) {
     // It is possible to get to here and still have an open command list
@@ -2011,6 +2010,10 @@ pi_result piQueueRelease(pi_queue Queue) {
 
     zePrint("piQueueRelease NumTimesClosedFull %d, NumTimesClosedEarly %d\n",
             Queue->NumTimesClosedFull, Queue->NumTimesClosedEarly);
+    Queue->PiQueueMutex.unlock();
+    delete Queue;
+  } else {
+    Queue->PiQueueMutex.unlock();
   }
   return PI_SUCCESS;
 }
@@ -3404,7 +3407,7 @@ piEnqueueKernelLaunch(pi_queue Queue, pi_kernel Kernel, pi_uint32 WorkDim,
   // Save the kernel in the event, so that when the event is signalled
   // the code can do a piKernelRelease on this kernel.
   (*Event)->CommandData = (void *)Kernel;
-
+  Queue->RefCount++;
   // Use piKernelRetain to increment the reference count and indicate
   // that the Kernel is in use. Once the event has been signalled, the
   // code in cleanupAfterEvent will do a piReleaseKernel to update
@@ -3677,7 +3680,7 @@ pi_result piEventRelease(pi_event Event) {
 
     auto Context = Event->Context;
     ZE_CALL(Context->decrementAliveEventsInPool(Event->ZeEventPool));
-
+    piQueueRelease(Event->Queue);
     delete Event;
   }
   return PI_SUCCESS;
@@ -3876,6 +3879,7 @@ pi_result piEnqueueEventsWaitWithBarrier(pi_queue Queue,
     (*Event)->ZeCommandList = ZeCommandList;
 
     ZeEvent = (*Event)->ZeEvent;
+    Queue->RefCount++;
   }
 
   // TODO: use unique_ptr with custom deleter in the whole Level Zero plugin for
@@ -3941,7 +3945,7 @@ enqueueMemCopyHelper(pi_command_type CommandType, pi_queue Queue, void *Dst,
                      pi_bool BlockingWrite, size_t Size, const void *Src,
                      pi_uint32 NumEventsInWaitList,
                      const pi_event *EventWaitList, pi_event *Event) {
-
+  assert(Queue);
   // Get a new command list to be used on this call
   ze_command_list_handle_t ZeCommandList = nullptr;
   ze_fence_handle_t ZeFence = nullptr;
@@ -3960,6 +3964,7 @@ enqueueMemCopyHelper(pi_command_type CommandType, pi_queue Queue, void *Dst,
     (*Event)->ZeCommandList = ZeCommandList;
 
     ZeEvent = (*Event)->ZeEvent;
+    Queue->RefCount++;
   }
 
   ze_event_handle_t *ZeEventWaitList =
@@ -4001,7 +4006,7 @@ static pi_result enqueueMemCopyRectHelper(
     size_t DstSlicePitch, pi_bool Blocking, pi_uint32 NumEventsInWaitList,
     const pi_event *EventWaitList, pi_event *Event) {
 
-  PI_ASSERT(Region && SrcOrigin && DstOrigin, PI_INVALID_VALUE);
+  PI_ASSERT(Region && SrcOrigin && DstOrigin && Queue, PI_INVALID_VALUE);
 
   // Get a new command list to be used on this call
   ze_command_list_handle_t ZeCommandList = nullptr;
@@ -4021,6 +4026,7 @@ static pi_result enqueueMemCopyRectHelper(
     (*Event)->ZeCommandList = ZeCommandList;
 
     ZeEvent = (*Event)->ZeEvent;
+    Queue->RefCount++;
   }
 
   ze_event_handle_t *ZeEventWaitList =
@@ -4184,7 +4190,7 @@ enqueueMemFillHelper(pi_command_type CommandType, pi_queue Queue, void *Ptr,
                      const void *Pattern, size_t PatternSize, size_t Size,
                      pi_uint32 NumEventsInWaitList,
                      const pi_event *EventWaitList, pi_event *Event) {
-
+  assert(Queue);
   // Get a new command list to be used on this call
   ze_command_list_handle_t ZeCommandList = nullptr;
   ze_fence_handle_t ZeFence = nullptr;
@@ -4203,6 +4209,7 @@ enqueueMemFillHelper(pi_command_type CommandType, pi_queue Queue, void *Ptr,
     (*Event)->ZeCommandList = ZeCommandList;
 
     ZeEvent = (*Event)->ZeEvent;
+    Queue->RefCount++;
   }
 
   ze_event_handle_t *ZeEventWaitList =
@@ -4288,6 +4295,7 @@ pi_result piEnqueueMemBufferMap(pi_queue Queue, pi_mem Buffer,
     (*Event)->ZeCommandList = ZeCommandList;
 
     ZeEvent = (*Event)->ZeEvent;
+    piQueueRetain(Queue);
   }
 
   // TODO: Level Zero is missing the memory "mapping" capabilities, so we are
@@ -4386,6 +4394,7 @@ pi_result piEnqueueMemUnmap(pi_queue Queue, pi_mem MemObj, void *MappedPtr,
     (*Event)->ZeCommandList = ZeCommandList;
 
     ZeEvent = (*Event)->ZeEvent;
+    piQueueRetain(Queue);
   }
 
   _pi_mem::Mapping MapInfo = {};
@@ -4517,7 +4526,7 @@ enqueueMemImageCommandHelper(pi_command_type CommandType, pi_queue Queue,
                              size_t RowPitch, size_t SlicePitch,
                              pi_uint32 NumEventsInWaitList,
                              const pi_event *EventWaitList, pi_event *Event) {
-
+  assert(Queue);
   // Get a new command list to be used on this call
   ze_command_list_handle_t ZeCommandList = nullptr;
   ze_fence_handle_t ZeFence = nullptr;
@@ -4536,6 +4545,7 @@ enqueueMemImageCommandHelper(pi_command_type CommandType, pi_queue Queue,
     (*Event)->ZeCommandList = ZeCommandList;
 
     ZeEvent = (*Event)->ZeEvent;
+    piQueueRetain(Queue);
   }
 
   ze_event_handle_t *ZeEventWaitList =
@@ -5143,6 +5153,7 @@ pi_result piextUSMEnqueuePrefetch(pi_queue Queue, const void *Ptr, size_t Size,
     (*Event)->ZeCommandList = ZeCommandList;
 
     ZeEvent = (*Event)->ZeEvent;
+    Queue->RefCount++;
   }
 
   ze_event_handle_t *ZeEventWaitList =
@@ -5203,6 +5214,7 @@ pi_result piextUSMEnqueueMemAdvise(pi_queue Queue, const void *Ptr,
     (*Event)->ZeCommandList = ZeCommandList;
 
     ZeEvent = (*Event)->ZeEvent;
+    Queue->RefCount++;
   }
 
   ZE_CALL(zeCommandListAppendMemAdvise(ZeCommandList, Queue->Device->ZeDevice,