[SYCL][Graph] try to opt duplicateNodes() function.

sycl/source/detail/graph/graph_impl.cpp

@@ -705,6 +705,14 @@ void graph_impl::beginRecording(sycl::detail::queue_impl &Queue) {
     addQueue(Queue);
   }
 }
+void graph_impl::endRecording() {
+  // Collect attribute(s) for recorded Graph
+  if (MRecordingQueues.size() == 1) {
+    if (auto ValidQueue = MRecordingQueues.begin()->lock(); ValidQueue) {
+      MIsLinearRecorded = ValidQueue->isInOrder();
+    }
+  }
+}
 
 // Check if nodes do not require enqueueing and if so loop back through
 // predecessors until we find the real dependency.
@@ -1262,6 +1270,11 @@ exec_graph_impl::enqueue(sycl::detail::queue_impl &Queue,
 }
 
 void exec_graph_impl::duplicateNodes() {
+  if (MGraphImpl->isLinearRecorded() && !MGraphImpl->hasSubGraph()) {
+    MNodeStorage = MGraphImpl->MNodeStorage;
+    return;
+  }
+
   // Map of original modifiable nodes (keys) to new duplicated nodes (values)
   std::map<node_impl *, node_impl *> NodesMap;
 
@@ -1965,6 +1978,7 @@ void modifiable_command_graph::begin_recording(
 
 void modifiable_command_graph::end_recording() {
   graph_impl::WriteLock Lock(impl->MMutex);
+  impl->endRecording();
   impl->clearQueues();
 }
 

sycl/source/detail/graph/graph_impl.hpp

@@ -474,6 +474,11 @@ class graph_impl : public std::enable_shared_from_this<graph_impl> {
   /// of recording queues associated with this graph.
   /// @param[in] Queue The queue to be recorded from.
   void beginRecording(sycl::detail::queue_impl &Queue);
+  void endRecording();
+
+  bool isLinearRecorded() const { return MIsLinearRecorded; }
+  void setHasSubGraph() { MOriginGraphHasSubGraph = true; }
+  bool hasSubGraph() const { return MOriginGraphHasSubGraph; }
 
   /// Store the last barrier node that was submitted to the queue.
   /// @param[in] Queue The queue the barrier was recorded from.
@@ -600,6 +605,11 @@ class graph_impl : public std::enable_shared_from_this<graph_impl> {
   // The number of live executable graphs that have been created from this
   // modifiable graph
   std::atomic<size_t> MExecGraphCount = 0;
+
+  // True if the graph is recorded from a single in-order queue
+  bool MIsLinearRecorded = false;
+  // True if the graph contains subgraph when constructed by recording
+  bool MOriginGraphHasSubGraph = false;
 };
 
 /// Class representing the implementation of command_graph<executable>.

sycl/source/handler.cpp

@@ -788,6 +788,7 @@ event handler::finalize() {
         ParentLock = ext::oneapi::experimental::detail::graph_impl::WriteLock(
             ParentGraph->MMutex);
       }
+      ParentGraph->setHasSubGraph();
       impl->CGData.MRequirements = impl->MExecGraph->getRequirements();
       // Here we are using the CommandGroup without passing a CommandBuffer to
       // pass the exec_graph_impl and event dependencies. Since this subgraph CG