[CPU] Refactor Infer update strategy

by using a template function with a functor template argument
to reduce Infer overhead

Author: EgorDuplensky

src/plugins/intel_cpu/src/graph.cpp

@@ -1137,16 +1137,11 @@ void Graph::InferStatic(SyncInferRequest* request) {
 
 namespace {
 
-class IUpdateNodes {
-public:
-    virtual void run(size_t stopIndx) = 0;
-    virtual ~IUpdateNodes() = default;
-};
-
-class UpdateNodesSeq : public IUpdateNodes {
+class UpdateNodesSeq {
 public:
     explicit UpdateNodesSeq(std::vector<NodePtr>& executableGraphNodes) : m_executableGraphNodes(executableGraphNodes) {}
-    void run(size_t stopIndx) override {
+
+    void operator()(size_t stopIndx) {
         for (; prepareCounter < stopIndx; ++prepareCounter) {
             const auto& node = m_executableGraphNodes[prepareCounter];
             if (node->isDynamicNode()) {
@@ -1177,7 +1172,7 @@ class UpdateNodesSeq : public IUpdateNodes {
 #        define ov_memory_order_acquire std::memory_order::memory_order_acquire
 #    endif
 
-class UpdateNodesBase : public IUpdateNodes {
+class UpdateNodesBase {
 public:
     explicit UpdateNodesBase(std::vector<NodePtr>& executableGraphNodes) : m_executableGraphNodes(executableGraphNodes) {}
     void updateShapes(size_t node_indx, size_t stop_indx) {
@@ -1248,7 +1243,8 @@ class AsyncTask : public tbb::detail::d1::task {
 class UpdateNodes : public UpdateNodesBase {
 public:
     using UpdateNodesBase::UpdateNodesBase;
-    void run(size_t stopIndx) override {
+
+    void operator()(size_t stopIndx) {
         m_completion.store(false);
         auto startCounter = m_prepareCounter.load();
         tbb::detail::d1::wait_context wait_ctx(2);
@@ -1289,7 +1285,7 @@ class AsyncTask : public tbb::task {
 class UpdateNodes : public UpdateNodesBase {
 public:
     using UpdateNodesBase::UpdateNodesBase;
-    void run(size_t stopIndx) override {
+    void operator()(size_t stopIndx) {
         m_completion.store(false);
         auto startCounter = m_prepareCounter.load();
         tbb::task& root = *new(tbb::task::allocate_root()) tbb::empty_task;
@@ -1317,7 +1313,7 @@ class UpdateNodes : public UpdateNodesBase {
 class UpdateNodes : public UpdateNodesBase {
 public:
     using UpdateNodesBase::UpdateNodesBase;
-    void run(size_t stopIndx) override {
+    void operator()(size_t stopIndx) {
         m_completion.store(false);
         auto startCounter = m_prepareCounter.load();
 
@@ -1340,20 +1336,14 @@ class UpdateNodes : public UpdateNodesBase {
 #endif
 } // namespace
 
-
-void Graph::InferDynamic(SyncInferRequest* request) {
+template<typename UpdateStrategy>
+void Graph::InferDynamic(SyncInferRequest* request, UpdateStrategy&& update) {
     dnnl::stream stream(getEngine());
 
-    std::unique_ptr<IUpdateNodes> updateNodes{};
-    if (parallel_get_max_threads() > 1) {
-        updateNodes.reset(new UpdateNodes(m_executableGraphNodes));
-    } else {
-        updateNodes.reset(new UpdateNodesSeq(m_executableGraphNodes));
-    }
-
     size_t inferCounter = 0;
     for (auto stopIndx : m_executableSyncNodesInds) {
-        updateNodes->run(stopIndx);
+        update(stopIndx);
+
         for (; inferCounter < stopIndx; ++inferCounter) {
             auto& node = m_executableGraphNodes[inferCounter];
             VERBOSE(node, getConfig().debugCaps.verbose);
@@ -1461,7 +1451,11 @@ void Graph::Infer(SyncInferRequest* request) {
     }
 
     if (Status::ReadyDynamic == status) {
-        InferDynamic(request);
+        if (parallel_get_max_threads() > 1) {
+            InferDynamic(request, UpdateNodes(m_executableGraphNodes));
+        } else {
+            InferDynamic(request, UpdateNodesSeq(m_executableGraphNodes));
+        }
     } else if (Status::ReadyStatic == status) {
         InferStatic(request);
     } else {

src/plugins/intel_cpu/src/graph.h

@@ -228,7 +228,9 @@ class Graph {
     void ExecuteNode(const NodePtr& node, const dnnl::stream& stream) const;
     void CreatePrimitivesAndExecConstants() const;
     void InferStatic(SyncInferRequest* request);
-    void InferDynamic(SyncInferRequest* request);
+
+    template<typename UpdateStrategy>
+    void InferDynamic(SyncInferRequest* request, UpdateStrategy&& update);
     void ParalleMtNuma(size_t num_nodes,
                        ov::threading::CPUStreamsExecutor::Ptr executor,
                        const std::function<void(size_t, size_t)>& func) const;