[NFC][SYCL][Graph] Update some maps to use raw node_impl *

Continuation of the refactoring in

#19295
#19332

Author: aelovikov-intel

sycl/source/detail/async_alloc.cpp

@@ -47,9 +47,9 @@ std::vector<std::shared_ptr<detail::node_impl>> getDepGraphNodes(
   // If this is being recorded from an in-order queue we need to get the last
   // in-order node if any, since this will later become a dependency of the
   // node being processed here.
-  if (const auto &LastInOrderNode = Graph->getLastInorderNode(Queue);
+  if (detail::node_impl *LastInOrderNode = Graph->getLastInorderNode(Queue);
       LastInOrderNode) {
-    DepNodes.push_back(LastInOrderNode);
+    DepNodes.push_back(LastInOrderNode->shared_from_this());
   }
   return DepNodes;
 }

sycl/source/detail/graph/graph_impl.cpp

@@ -255,7 +255,7 @@ void exec_graph_impl::makePartitions() {
     const std::shared_ptr<partition> &Partition = std::make_shared<partition>();
     for (auto &Node : MNodeStorage) {
       if (Node->MPartitionNum == i) {
-        MPartitionNodes[Node] = PartitionFinalNum;
+        MPartitionNodes[Node.get()] = PartitionFinalNum;
         if (isPartitionRoot(Node)) {
           Partition->MRoots.insert(Node);
           if (Node->MCGType == CGType::CodeplayHostTask) {
@@ -290,8 +290,7 @@ void exec_graph_impl::makePartitions() {
     for (auto const &Root : Partition->MRoots) {
       auto RootNode = Root.lock();
       for (node_impl &NodeDep : RootNode->predecessors()) {
-        auto &Predecessor =
-            MPartitions[MPartitionNodes[NodeDep.shared_from_this()]];
+        auto &Predecessor = MPartitions[MPartitionNodes[&NodeDep]];
         Partition->MPredecessors.push_back(Predecessor.get());
         Predecessor->MSuccessors.push_back(Partition.get());
       }
@@ -610,8 +609,7 @@ bool graph_impl::checkForCycles() {
   return CycleFound;
 }
 
-std::shared_ptr<node_impl>
-graph_impl::getLastInorderNode(sycl::detail::queue_impl *Queue) {
+node_impl *graph_impl::getLastInorderNode(sycl::detail::queue_impl *Queue) {
   if (!Queue) {
     assert(0 ==
            MInorderQueueMap.count(std::weak_ptr<sycl::detail::queue_impl>{}));
@@ -625,7 +623,7 @@ graph_impl::getLastInorderNode(sycl::detail::queue_impl *Queue) {
 
 void graph_impl::setLastInorderNode(sycl::detail::queue_impl &Queue,
                                     std::shared_ptr<node_impl> Node) {
-  MInorderQueueMap[Queue.weak_from_this()] = std::move(Node);
+  MInorderQueueMap[Queue.weak_from_this()] = &*Node;
 }
 
 void graph_impl::makeEdge(std::shared_ptr<node_impl> Src,
@@ -726,11 +724,10 @@ void exec_graph_impl::findRealDeps(
       findRealDeps(Deps, NodeImpl, ReferencePartitionNum);
     }
   } else {
-    auto CurrentNodePtr = CurrentNode.shared_from_this();
     // Verify if CurrentNode belong the the same partition
-    if (MPartitionNodes[CurrentNodePtr] == ReferencePartitionNum) {
+    if (MPartitionNodes[&CurrentNode] == ReferencePartitionNum) {
       // Verify that the sync point has actually been set for this node.
-      auto SyncPoint = MSyncPoints.find(CurrentNodePtr);
+      auto SyncPoint = MSyncPoints.find(&CurrentNode);
       assert(SyncPoint != MSyncPoints.end() &&
              "No sync point has been set for node dependency.");
       // Check if the dependency has already been added.
@@ -749,7 +746,7 @@ exec_graph_impl::enqueueNodeDirect(const sycl::context &Ctx,
                                    std::shared_ptr<node_impl> Node) {
   std::vector<ur_exp_command_buffer_sync_point_t> Deps;
   for (node_impl &N : Node->predecessors()) {
-    findRealDeps(Deps, N, MPartitionNodes[Node]);
+    findRealDeps(Deps, N, MPartitionNodes[Node.get()]);
   }
   ur_exp_command_buffer_sync_point_t NewSyncPoint;
   ur_exp_command_buffer_command_handle_t NewCommand = 0;
@@ -782,7 +779,7 @@ exec_graph_impl::enqueueNodeDirect(const sycl::context &Ctx,
       Deps, &NewSyncPoint, MIsUpdatable ? &NewCommand : nullptr, nullptr);
 
   if (MIsUpdatable) {
-    MCommandMap[Node] = NewCommand;
+    MCommandMap[Node.get()] = NewCommand;
   }
 
   if (Res != UR_RESULT_SUCCESS) {
@@ -805,7 +802,7 @@ exec_graph_impl::enqueueNode(ur_exp_command_buffer_handle_t CommandBuffer,
 
   std::vector<ur_exp_command_buffer_sync_point_t> Deps;
   for (node_impl &N : Node->predecessors()) {
-    findRealDeps(Deps, N, MPartitionNodes[Node]);
+    findRealDeps(Deps, N, MPartitionNodes[Node.get()]);
   }
 
   sycl::detail::EventImplPtr Event =
@@ -814,7 +811,7 @@ exec_graph_impl::enqueueNode(ur_exp_command_buffer_handle_t CommandBuffer,
           /*EventNeeded=*/true, CommandBuffer, Deps);
 
   if (MIsUpdatable) {
-    MCommandMap[Node] = Event->getCommandBufferCommand();
+    MCommandMap[Node.get()] = Event->getCommandBufferCommand();
   }
 
   return Event->getSyncPoint();
@@ -830,7 +827,8 @@ void exec_graph_impl::buildRequirements() {
                          Node->MCommandGroup->getRequirements().begin(),
                          Node->MCommandGroup->getRequirements().end());
 
-    std::shared_ptr<partition> &Partition = MPartitions[MPartitionNodes[Node]];
+    std::shared_ptr<partition> &Partition =
+        MPartitions[MPartitionNodes[Node.get()]];
 
     Partition->MRequirements.insert(
         Partition->MRequirements.end(),
@@ -877,10 +875,10 @@ void exec_graph_impl::createCommandBuffers(
                     Node->MCommandGroup.get())
                     ->MStreams.size() ==
             0) {
-      MSyncPoints[Node] =
+      MSyncPoints[Node.get()] =
           enqueueNodeDirect(MContext, DeviceImpl, OutCommandBuffer, Node);
     } else {
-      MSyncPoints[Node] = enqueueNode(OutCommandBuffer, Node);
+      MSyncPoints[Node.get()] = enqueueNode(OutCommandBuffer, Node);
     }
   }
 
@@ -1726,7 +1724,7 @@ void exec_graph_impl::populateURKernelUpdateStructs(
   auto ExecNode = MIDCache.find(Node->MID);
   assert(ExecNode != MIDCache.end() && "Node ID was not found in ID cache");
 
-  auto Command = MCommandMap.find(ExecNode->second);
+  auto Command = MCommandMap.find(ExecNode->second.get());
   assert(Command != MCommandMap.end());
   UpdateDesc.hCommand = Command->second;
 
@@ -1756,7 +1754,7 @@ exec_graph_impl::getURUpdatableNodes(
 
     auto ExecNode = MIDCache.find(Node->MID);
     assert(ExecNode != MIDCache.end() && "Node ID was not found in ID cache");
-    auto PartitionIndex = MPartitionNodes.find(ExecNode->second);
+    auto PartitionIndex = MPartitionNodes.find(ExecNode->second.get());
     assert(PartitionIndex != MPartitionNodes.end());
     PartitionedNodes[PartitionIndex->second].push_back(Node);
   }

sycl/source/detail/graph/graph_impl.hpp

@@ -294,8 +294,7 @@ class graph_impl : public std::enable_shared_from_this<graph_impl> {
   /// @param Queue In-order queue to find the last node added to the graph from.
   /// @return Last node in this graph added from \p Queue recording, or empty
   /// shared pointer if none.
-  std::shared_ptr<node_impl>
-  getLastInorderNode(sycl::detail::queue_impl *Queue);
+  node_impl *getLastInorderNode(sycl::detail::queue_impl *Queue);
 
   /// Track the last node added to this graph from an in-order queue.
   /// @param Queue In-order queue to register \p Node for.
@@ -466,14 +465,13 @@ class graph_impl : public std::enable_shared_from_this<graph_impl> {
   /// @param[in] BarrierNodeImpl The created barrier node.
   void setBarrierDep(std::weak_ptr<sycl::detail::queue_impl> Queue,
                      std::shared_ptr<node_impl> BarrierNodeImpl) {
-    MBarrierDependencyMap[Queue] = BarrierNodeImpl;
+    MBarrierDependencyMap[Queue] = &*BarrierNodeImpl;
   }
 
   /// Get the last barrier node that was submitted to the queue.
   /// @param[in] Queue The queue to find the last barrier node of. An empty
   /// shared_ptr is returned if no barrier node has been recorded to the queue.
-  std::shared_ptr<node_impl>
-  getBarrierDep(std::weak_ptr<sycl::detail::queue_impl> Queue) {
+  node_impl *getBarrierDep(std::weak_ptr<sycl::detail::queue_impl> Queue) {
     return MBarrierDependencyMap[Queue];
   }
 
@@ -553,7 +551,7 @@ class graph_impl : public std::enable_shared_from_this<graph_impl> {
   /// Map for every in-order queue thats recorded a node to the graph, what
   /// the last node added was. We can use this to create new edges on the last
   /// node if any more nodes are added to the graph from the queue.
-  std::map<std::weak_ptr<sycl::detail::queue_impl>, std::shared_ptr<node_impl>,
+  std::map<std::weak_ptr<sycl::detail::queue_impl>, node_impl *,
            std::owner_less<std::weak_ptr<sycl::detail::queue_impl>>>
       MInorderQueueMap;
   /// Controls whether we skip the cycle checks in makeEdge, set by the presence
@@ -568,7 +566,7 @@ class graph_impl : public std::enable_shared_from_this<graph_impl> {
 
   /// Mapping from queues to barrier nodes. For each queue the last barrier
   /// node recorded to the graph from the queue is stored.
-  std::map<std::weak_ptr<sycl::detail::queue_impl>, std::shared_ptr<node_impl>,
+  std::map<std::weak_ptr<sycl::detail::queue_impl>, node_impl *,
            std::owner_less<std::weak_ptr<sycl::detail::queue_impl>>>
       MBarrierDependencyMap;
   /// Graph memory pool for handling graph-owned memory allocations for this
@@ -886,14 +884,13 @@ class exec_graph_impl {
   std::shared_ptr<graph_impl> MGraphImpl;
   /// Map of nodes in the exec graph to the sync point representing their
   /// execution in the command graph.
-  std::unordered_map<std::shared_ptr<node_impl>,
-                     ur_exp_command_buffer_sync_point_t>
+  std::unordered_map<node_impl *, ur_exp_command_buffer_sync_point_t>
       MSyncPoints;
   /// Sycl queue impl ptr associated with this graph.
   std::shared_ptr<sycl::detail::queue_impl> MQueueImpl;
   /// Map of nodes in the exec graph to the partition number to which they
   /// belong.
-  std::unordered_map<std::shared_ptr<node_impl>, int> MPartitionNodes;
+  std::unordered_map<node_impl *, int> MPartitionNodes;
   /// Device associated with this executable graph.
   sycl::device MDevice;
   /// Context associated with this executable graph.
@@ -909,8 +906,7 @@ class exec_graph_impl {
   /// Storage for copies of nodes from the original modifiable graph.
   std::vector<std::shared_ptr<node_impl>> MNodeStorage;
   /// Map of nodes to their associated UR command handles.
-  std::unordered_map<std::shared_ptr<node_impl>,
-                     ur_exp_command_buffer_command_handle_t>
+  std::unordered_map<node_impl *, ur_exp_command_buffer_command_handle_t>
       MCommandMap;
   /// List of partition without any predecessors in this exec graph.
   std::vector<std::weak_ptr<partition>> MRootPartitions;

sycl/source/handler.cpp

@@ -888,8 +888,9 @@ event handler::finalize() {
       // node can set it as a predecessor.
       std::vector<std::shared_ptr<ext::oneapi::experimental::detail::node_impl>>
           Deps;
-      if (auto DependentNode = GraphImpl->getLastInorderNode(Queue)) {
-        Deps.push_back(std::move(DependentNode));
+      if (ext::oneapi::experimental::detail::node_impl *DependentNode =
+              GraphImpl->getLastInorderNode(Queue)) {
+        Deps.push_back(DependentNode->shared_from_this());
       }
       NodeImpl = GraphImpl->add(NodeType, std::move(CommandGroup), Deps);
 
@@ -898,13 +899,14 @@ event handler::finalize() {
       // queue.
       GraphImpl->setLastInorderNode(*Queue, NodeImpl);
     } else {
-      auto LastBarrierRecordedFromQueue =
-          GraphImpl->getBarrierDep(Queue->weak_from_this());
+      ext::oneapi::experimental::detail::node_impl
+          *LastBarrierRecordedFromQueue =
+              GraphImpl->getBarrierDep(Queue->weak_from_this());
       std::vector<std::shared_ptr<ext::oneapi::experimental::detail::node_impl>>
           Deps;
 
       if (LastBarrierRecordedFromQueue) {
-        Deps.push_back(LastBarrierRecordedFromQueue);
+        Deps.push_back(LastBarrierRecordedFromQueue->shared_from_this());
       }
       NodeImpl = GraphImpl->add(NodeType, std::move(CommandGroup), Deps);