[NFCI][SYCL] Prefer raw ptr/ref for queue_impl

Continuation of the refactoring in
#18715
#18748

Author: aelovikov-intel

sycl/source/detail/cg.hpp

@@ -725,14 +725,11 @@ class CGHostTask : public CG {
   std::shared_ptr<detail::context_impl> MContext;
   std::vector<ArgDesc> MArgs;
 
-  CGHostTask(std::shared_ptr<HostTask> HostTask,
-             std::shared_ptr<detail::queue_impl> Queue,
+  // TODO: ref?
+  CGHostTask(std::shared_ptr<HostTask> HostTask, detail::queue_impl *Queue,
              std::shared_ptr<detail::context_impl> Context,
              std::vector<ArgDesc> Args, CG::StorageInitHelper CGData,
-             CGType Type, detail::code_location loc = {})
-      : CG(Type, std::move(CGData), std::move(loc)),
-        MHostTask(std::move(HostTask)), MQueue(Queue), MContext(Context),
-        MArgs(std::move(Args)) {}
+             CGType Type, detail::code_location loc = {});
 };
 
 } // namespace detail

sycl/source/detail/event_impl.cpp

@@ -159,23 +159,25 @@ event_impl::event_impl(ur_event_handle_t Event, const context &SyclContext)
   }
 }
 
-event_impl::event_impl(const QueueImplPtr &Queue)
-    : MQueue{Queue}, MIsProfilingEnabled{!Queue || Queue->MIsProfilingEnabled} {
-  if (Queue)
-    this->setContextImpl(Queue->getContextImplPtr());
-  else {
-    MState.store(HES_NotComplete);
-    MHostProfilingInfo.reset(new HostProfilingInfo());
-    if (!MHostProfilingInfo)
-      throw sycl::exception(
-          sycl::make_error_code(sycl::errc::runtime),
-          "Out of host memory " +
-              codeToString(UR_RESULT_ERROR_OUT_OF_HOST_MEMORY));
-    return;
-  }
+event_impl::event_impl(queue_impl &Queue)
+    : MQueue{Queue.weak_from_this()},
+      MIsProfilingEnabled{Queue.MIsProfilingEnabled} {
+  this->setContextImpl(Queue.getContextImplPtr());
   MState.store(HES_Complete);
 }
 
+// TODO: comment about https://github.com/intel/llvm/pull/14370
+event_impl::event_impl(std::nullptr_t)
+    : MQueue{}, MIsProfilingEnabled{true} {
+  MState.store(HES_NotComplete);
+  MHostProfilingInfo.reset(new HostProfilingInfo());
+  if (!MHostProfilingInfo)
+    throw sycl::exception(sycl::make_error_code(sycl::errc::runtime),
+                          "Out of host memory " +
+                              codeToString(UR_RESULT_ERROR_OUT_OF_HOST_MEMORY));
+  return;
+}
+
 void event_impl::setQueue(queue_impl &Queue) {
   MQueue = Queue.shared_from_this();
   MIsProfilingEnabled = Queue.MIsProfilingEnabled;
@@ -212,7 +214,7 @@ void *event_impl::instrumentationProlog(std::string &Name, int32_t StreamID,
     // queue is available with the wait events. We check to see if the
     // TraceEvent is available in the Queue object.
     void *TraceEvent = nullptr;
-    if (QueueImplPtr Queue = MQueue.lock()) {
+    if (auto Queue = MQueue.lock()) {
       TraceEvent = Queue->getTraceEvent();
       WaitEvent =
           (TraceEvent ? static_cast<xpti_td *>(TraceEvent) : GSYCLGraphEvent);
@@ -281,7 +283,7 @@ void event_impl::wait_and_throw(
     std::shared_ptr<sycl::detail::event_impl> Self) {
   wait(Self);
 
-  if (QueueImplPtr SubmittedQueue = MSubmittedQueue.lock())
+  if (auto SubmittedQueue = MSubmittedQueue.lock())
     SubmittedQueue->throw_asynchronous();
 }
 
@@ -426,7 +428,7 @@ event_impl::get_backend_info<info::platform::version>() const {
                           "the info::platform::version info descriptor can "
                           "only be queried with an OpenCL backend");
   }
-  if (QueueImplPtr Queue = MQueue.lock()) {
+  if (auto Queue = MQueue.lock()) {
     return Queue->getDeviceImpl()
         .get_platform()
         .get_info<info::platform::version>();
@@ -449,7 +451,7 @@ event_impl::get_backend_info<info::device::version>() const {
                           "the info::device::version info descriptor can only "
                           "be queried with an OpenCL backend");
   }
-  if (QueueImplPtr Queue = MQueue.lock()) {
+  if (auto Queue = MQueue.lock()) {
     return Queue->getDeviceImpl().get_info<info::device::version>();
   }
   return ""; // If the queue has been released, no device will be associated so
@@ -516,21 +518,21 @@ std::vector<EventImplPtr> event_impl::getWaitList() {
   return Result;
 }
 
-void event_impl::flushIfNeeded(const QueueImplPtr &UserQueue) {
+void event_impl::flushIfNeeded(queue_impl *UserQueue) {
   // Some events might not have a native handle underneath even at this point,
   // e.g. those produced by memset with 0 size (no UR call is made).
   auto Handle = this->getHandle();
   if (MIsFlushed || !Handle)
     return;
 
-  QueueImplPtr Queue = MQueue.lock();
+  auto Queue = MQueue.lock();
   // If the queue has been released, all of the commands have already been
   // implicitly flushed by urQueueRelease.
   if (!Queue) {
     MIsFlushed = true;
     return;
   }
-  if (Queue == UserQueue)
+  if (Queue.get() == UserQueue)
     return;
 
   // Check if the task for this event has already been submitted.
@@ -568,7 +570,7 @@ void event_impl::setSubmissionTime() {
   if (!MIsProfilingEnabled && !MProfilingTagEvent)
     return;
 
-  if (QueueImplPtr Queue = MQueue.lock()) {
+  if (auto Queue = MQueue.lock()) {
     try {
       MSubmitTime = Queue->getDeviceImpl().getCurrentDeviceTime();
     } catch (sycl::exception &e) {

sycl/source/detail/event_impl.hpp

@@ -31,7 +31,6 @@ class Adapter;
 class context_impl;
 using ContextImplPtr = std::shared_ptr<sycl::detail::context_impl>;
 class queue_impl;
-using QueueImplPtr = std::shared_ptr<sycl::detail::queue_impl>;
 class event_impl;
 using EventImplPtr = std::shared_ptr<sycl::detail::event_impl>;
 
@@ -48,7 +47,7 @@ class event_impl {
   /// If the constructed SYCL event is waited on it will complete immediately.
   /// Normally constructs a host event, use std::nullopt to instead instantiate
   /// a device event.
-  event_impl(std::optional<HostEventState> State = HES_Complete)
+  event_impl(std::optional<HostEventState> State)
       : MIsFlushed(true), MState(State.value_or(HES_Complete)),
         MIsDefaultConstructed(!State), MIsHostEvent(State) {
     // Need to fail in event() constructor  if there are problems with the
@@ -66,7 +65,9 @@ class event_impl {
   /// \param Event is a valid instance of UR event.
   /// \param SyclContext is an instance of SYCL context.
   event_impl(ur_event_handle_t Event, const context &SyclContext);
-  event_impl(const QueueImplPtr &Queue);
+  event_impl(queue_impl &Queue);
+  // TODO: what about the very first ctor overload?
+  event_impl(std::nullptr_t);
 
   /// Sets a queue associated with the event
   ///
@@ -209,7 +210,7 @@ class event_impl {
   /// Performs a flush on the queue associated with this event if the user queue
   /// is different and the task associated with this event hasn't been submitted
   /// to the device yet.
-  void flushIfNeeded(const QueueImplPtr &UserQueue);
+  void flushIfNeeded(queue_impl *UserQueue);
 
   /// Cleans dependencies of this event_impl.
   void cleanupDependencyEvents();
@@ -229,7 +230,9 @@ class event_impl {
   ///
   /// @return shared_ptr to MWorkerQueue, please be aware it can be empty
   /// pointer
-  QueueImplPtr getWorkerQueue() { return MWorkerQueue.lock(); };
+  std::shared_ptr<sycl::detail::queue_impl> getWorkerQueue() {
+    return MWorkerQueue.lock();
+  };
 
   /// Sets worker queue for command.
   ///
@@ -258,7 +261,9 @@ class event_impl {
   /// @return Submission time for command associated with this event
   uint64_t getSubmissionTime();
 
-  QueueImplPtr getSubmittedQueue() const { return MSubmittedQueue.lock(); };
+  std::shared_ptr<sycl::detail::queue_impl> getSubmittedQueue() const {
+    return MSubmittedQueue.lock();
+  };
 
   /// Checks if this event is complete.
   ///

sycl/source/detail/graph_impl.cpp

@@ -404,7 +404,9 @@ std::shared_ptr<node_impl> graph_impl::addNodesToExits(
   // Add all the new nodes to the node storage
   for (auto &Node : NodeList) {
     MNodeStorage.push_back(Node);
-    addEventForNode(std::make_shared<sycl::detail::event_impl>(), Node);
+    addEventForNode(std::make_shared<sycl::detail::event_impl>(
+                        sycl::detail::event_impl::HES_Complete),
+                    Node);
   }
 
   return this->add(Outputs);
@@ -494,7 +496,9 @@ graph_impl::add(std::vector<std::shared_ptr<node_impl>> &Deps) {
 
   addDepsToNode(NodeImpl, Deps);
   // Add an event associated with this explicit node for mixed usage
-  addEventForNode(std::make_shared<sycl::detail::event_impl>(), NodeImpl);
+  addEventForNode(std::make_shared<sycl::detail::event_impl>(
+                      sycl::detail::event_impl::HES_Complete),
+                  NodeImpl);
   return NodeImpl;
 }
 
@@ -552,7 +556,9 @@ graph_impl::add(std::function<void(handler &)> CGF,
       this->add(NodeType, std::move(Handler.impl->MGraphNodeCG), Deps);
 
   // Add an event associated with this explicit node for mixed usage
-  addEventForNode(std::make_shared<sycl::detail::event_impl>(), NodeImpl);
+  addEventForNode(std::make_shared<sycl::detail::event_impl>(
+                      sycl::detail::event_impl::HES_Complete),
+                  NodeImpl);
 
   // Retrieve any dynamic parameters which have been registered in the CGF and
   // register the actual nodes with them.
@@ -651,7 +657,9 @@ graph_impl::add(std::shared_ptr<dynamic_command_group_impl> &DynCGImpl,
       add(NodeType, ActiveKernel, Deps);
 
   // Add an event associated with this explicit node for mixed usage
-  addEventForNode(std::make_shared<sycl::detail::event_impl>(), NodeImpl);
+  addEventForNode(std::make_shared<sycl::detail::event_impl>(
+                      sycl::detail::event_impl::HES_Complete),
+                  NodeImpl);
 
   // Track the dynamic command-group used inside the node object
   DynCGImpl->MNodes.push_back(NodeImpl);
@@ -897,7 +905,7 @@ exec_graph_impl::enqueueNode(ur_exp_command_buffer_handle_t CommandBuffer,
 
   sycl::detail::EventImplPtr Event =
       sycl::detail::Scheduler::getInstance().addCG(
-          Node->getCGCopy(), MQueueImpl,
+          Node->getCGCopy(), *MQueueImpl,
           /*EventNeeded=*/true, CommandBuffer, Deps);
 
   if (MIsUpdatable) {
@@ -1037,8 +1045,7 @@ exec_graph_impl::enqueue(sycl::detail::queue_impl &Queue,
       PartitionsExecutionEvents;
 
   auto CreateNewEvent([&]() {
-    auto NewEvent =
-        std::make_shared<sycl::detail::event_impl>(Queue.shared_from_this());
+    auto NewEvent = std::make_shared<sycl::detail::event_impl>(Queue);
     NewEvent->setContextImpl(Queue.getContextImplPtr());
     NewEvent->setStateIncomplete();
     return NewEvent;
@@ -1122,7 +1129,7 @@ exec_graph_impl::enqueue(sycl::detail::queue_impl &Queue,
                 CommandBuffer, nullptr, std::move(CGData));
 
         NewEvent = sycl::detail::Scheduler::getInstance().addCG(
-            std::move(CommandGroup), Queue.shared_from_this(),
+            std::move(CommandGroup), Queue,
             /*EventNeeded=*/true);
       }
       NewEvent->setEventFromSubmittedExecCommandBuffer(true);
@@ -1142,7 +1149,7 @@ exec_graph_impl::enqueue(sycl::detail::queue_impl &Queue,
           .MQueue = Queue.shared_from_this();
 
       NewEvent = sycl::detail::Scheduler::getInstance().addCG(
-          NodeImpl->getCGCopy(), Queue.shared_from_this(),
+          NodeImpl->getCGCopy(), Queue,
           /*EventNeeded=*/true);
     }
     PartitionsExecutionEvents[CurrentPartition] = NewEvent;
@@ -1424,7 +1431,7 @@ void exec_graph_impl::update(
     // other scheduler commands
     auto UpdateEvent =
         sycl::detail::Scheduler::getInstance().addCommandGraphUpdate(
-            this, Nodes, MQueueImpl, std::move(UpdateRequirements),
+            this, Nodes, MQueueImpl.get(), std::move(UpdateRequirements),
             MExecutionEvents);
 
     MExecutionEvents.push_back(UpdateEvent);

sycl/source/detail/graph_impl.hpp

@@ -288,7 +288,7 @@ class node_impl : public std::enable_shared_from_this<node_impl> {
 
       return std::make_unique<sycl::detail::CGHostTask>(
           sycl::detail::CGHostTask(
-              std::move(HostTaskSPtr), CommandGroupPtr->MQueue,
+              std::move(HostTaskSPtr), CommandGroupPtr->MQueue.get(),
               CommandGroupPtr->MContext, std::move(NewArgs), std::move(Data),
               CommandGroupPtr->getType(), Loc));
     }

sycl/source/detail/helpers.hpp

@@ -21,7 +21,6 @@ class event;
 namespace detail {
 class CGExecKernel;
 class queue_impl;
-using QueueImplPtr = std::shared_ptr<sycl::detail::queue_impl>;
 class RTDeviceBinaryImage;
 
 #ifdef __INTEL_PREVIEW_BREAKING_CHANGES

sycl/source/detail/memory_manager.hpp

@@ -28,7 +28,6 @@ class queue_impl;
 class event_impl;
 class context_impl;
 
-using QueueImplPtr = std::shared_ptr<detail::queue_impl>;
 using EventImplPtr = std::shared_ptr<detail::event_impl>;
 using ContextImplPtr = std::shared_ptr<detail::context_impl>;
 

sycl/source/detail/queue_impl.cpp

@@ -114,10 +114,10 @@ queue_impl::get_backend_info<info::device::backend_version>() const {
 }
 #endif
 
-static event prepareSYCLEventAssociatedWithQueue(
-    const std::shared_ptr<detail::queue_impl> &QueueImpl) {
+static event
+prepareSYCLEventAssociatedWithQueue(detail::queue_impl &QueueImpl) {
   auto EventImpl = std::make_shared<detail::event_impl>(QueueImpl);
-  EventImpl->setContextImpl(detail::getSyclObjImpl(QueueImpl->get_context()));
+  EventImpl->setContextImpl(detail::getSyclObjImpl(QueueImpl.get_context()));
   EventImpl->setStateIncomplete();
   return detail::createSyclObjFromImpl<event>(EventImpl);
 }
@@ -462,7 +462,7 @@ event queue_impl::submitMemOpHelper(const std::vector<event> &DepEvents,
         return createDiscardedEvent();
       }
 
-      event ResEvent = prepareSYCLEventAssociatedWithQueue(shared_from_this());
+      event ResEvent = prepareSYCLEventAssociatedWithQueue(*this);
       const auto &EventImpl = detail::getSyclObjImpl(ResEvent);
       {
         NestedCallsTracker tracker;

sycl/source/detail/queue_impl.hpp

@@ -648,9 +648,12 @@ class queue_impl : public std::enable_shared_from_this<queue_impl> {
   // for in order ones.
   void revisitUnenqueuedCommandsState(const EventImplPtr &CompletedHostTask);
 
-  static ContextImplPtr getContext(const QueueImplPtr &Queue) {
+  static ContextImplPtr getContext(queue_impl *Queue) {
     return Queue ? Queue->getContextImplPtr() : nullptr;
   }
+  static ContextImplPtr getContext(const std::shared_ptr<queue_impl> &Queue) {
+    return getContext(Queue.get());
+  }
 
   // Must be called under MMutex protection
   void doUnenqueuedCommandCleanup(
@@ -663,7 +666,7 @@ class queue_impl : public std::enable_shared_from_this<queue_impl> {
   /// will wait for the completion of all work in the queue at the time of the
   /// insertion, but will not act as a barrier unless the queue is in-order.
   EventImplPtr insertMarkerEvent() {
-    auto ResEvent = std::make_shared<detail::event_impl>(shared_from_this());
+    auto ResEvent = std::make_shared<detail::event_impl>(*this);
     ur_event_handle_t UREvent = nullptr;
     getAdapter()->call<UrApiKind::urEnqueueEventsWait>(getHandleRef(), 0,
                                                        nullptr, &UREvent);
@@ -687,9 +690,8 @@ class queue_impl : public std::enable_shared_from_this<queue_impl> {
 protected:
   template <typename HandlerType = handler>
   EventImplPtr insertHelperBarrier(const HandlerType &Handler) {
-    auto &Queue = Handler.impl->get_queue();
-    auto ResEvent =
-        std::make_shared<detail::event_impl>(Queue.shared_from_this());
+    queue_impl &Queue = Handler.impl->get_queue();
+    auto ResEvent = std::make_shared<detail::event_impl>(Queue);
     ur_event_handle_t UREvent = nullptr;
     getAdapter()->call<UrApiKind::urEnqueueEventsWaitWithBarrier>(
         Queue.getHandleRef(), 0, nullptr, &UREvent);

sycl/source/detail/reduction.cpp

@@ -207,7 +207,7 @@ __SYCL_EXPORT size_t reduGetPreferredWGSize(std::shared_ptr<queue_impl> &Queue,
 __SYCL_EXPORT void
 addCounterInit(handler &CGH, std::shared_ptr<sycl::detail::queue_impl> &Queue,
                std::shared_ptr<int> &Counter) {
-  auto EventImpl = std::make_shared<detail::event_impl>(Queue);
+  auto EventImpl = std::make_shared<detail::event_impl>(*Queue);
   EventImpl->setContextImpl(detail::getSyclObjImpl(Queue->get_context()));
   EventImpl->setStateIncomplete();
   ur_event_handle_t UREvent = nullptr;