From 121760476a8f914eefa296bc6bba02d1887f4239 Mon Sep 17 00:00:00 2001
From: Vladislav Volkov <vladislav.volkov@intel.com>
Date: Fri, 5 Mar 2021 16:28:41 +0300
Subject: [PATCH] Fix for MKLDNN constant layers execution (#4632)

* Single mkldnn::engine for all MKLDNN graphs

* Fix for MKLDNN constant layers execution
---
 .../src/mkldnn_plugin/mkldnn_edge.cpp         |  4 +--
 .../src/mkldnn_plugin/mkldnn_graph.cpp        | 26 +++++++++++++------
 .../src/mkldnn_plugin/mkldnn_graph.h          |  6 ++---
 .../nodes/mkldnn_tensoriterator_node.cpp      |  3 +--
 4 files changed, 24 insertions(+), 15 deletions(-)

diff --git a/inference-engine/src/mkldnn_plugin/mkldnn_edge.cpp b/inference-engine/src/mkldnn_plugin/mkldnn_edge.cpp
index bfda87ff812..b4c9434cd9f 100644
--- a/inference-engine/src/mkldnn_plugin/mkldnn_edge.cpp
+++ b/inference-engine/src/mkldnn_plugin/mkldnn_edge.cpp
@@ -169,9 +169,9 @@ void MKLDNNEdge::allocate(const void* mem_ptr) {
 }
 
 std::string MKLDNNEdge::name() const {
-    auto childPtr = getChild();
     auto parentPtr = getParent();
-    return childPtr->getName() + "<->" + parentPtr->getName();
+    auto childPtr = getChild();
+    return parentPtr->getName() + std::to_string(parent_port) + "<->" + childPtr->getName() + std::to_string(child_port);
 }
 
 void MKLDNNEdge::externalAllocate(MKLDNNWeightsSharing::Ptr weightsCache) {
diff --git a/inference-engine/src/mkldnn_plugin/mkldnn_graph.cpp b/inference-engine/src/mkldnn_plugin/mkldnn_graph.cpp
index 8087ba7b627..df071706ce4 100644
--- a/inference-engine/src/mkldnn_plugin/mkldnn_graph.cpp
+++ b/inference-engine/src/mkldnn_plugin/mkldnn_graph.cpp
@@ -6,6 +6,7 @@
 #include <string>
 #include <map>
 #include <vector>
+#include <tuple>
 #include <unordered_set>
 #include <limits>
 #include <fstream>
@@ -67,6 +68,8 @@ using namespace InferenceEngine::details;
 typedef std::unordered_set<MKLDNNEdgePtr> edge_cluster_t;
 typedef std::vector<edge_cluster_t> edge_clusters_t;
 
+mkldnn::engine MKLDNNGraph::eng(mkldnn::engine::kind::cpu, 0);
+
 template<typename NET>
 void MKLDNNGraph::ApplyUnrollPasses(NET &net) {
     OV_ITT_SCOPED_TASK(itt::domains::MKLDNNPlugin, "MKLDNNGraph::ApplyUnrollPasses");
@@ -453,15 +456,24 @@ void MKLDNNGraph::ExecuteConstantNodesOnly() {
 
     auto acquireSharedOutputs = [this](MKLDNNNodePtr & graphNode) {
         std::vector<shared_memory_ptr> outputs;
+        bool hasLocalAllocatedEdges = false;
+        bool hasExternalInvalidEdges = false;
 
         for (size_t i = 0; i < graphNode->getChildEdges().size(); ++i) {
             auto edgePtr = graphNode->getChildEdgeAt(i);
-            if (edgePtr && edgePtr->isUseExternalMemory()) {
-                outputs.emplace_back(weightsCache->get(edgePtr->name()));
+            if (edgePtr) {
+                if (edgePtr->isUseExternalMemory()) {
+                    auto ptr = weightsCache->get(edgePtr->name());
+                    outputs.emplace_back(ptr);
+                    if (!ptr->isValid())
+                        hasExternalInvalidEdges = true;
+                } else {
+                    hasLocalAllocatedEdges = true;
+                }
             }
         }
 
-        return outputs;
+        return std::make_tuple(hasExternalInvalidEdges, hasLocalAllocatedEdges, outputs);
     };
 
     for (auto &graphNode : graphNodes) {
@@ -471,12 +483,10 @@ void MKLDNNGraph::ExecuteConstantNodesOnly() {
         if (weightsCache) {
             auto sharedOutputs = acquireSharedOutputs(graphNode);
 
-            if (std::find_if(sharedOutputs.begin(), sharedOutputs.end(),
-                [](const shared_memory_ptr & ptr) {
-                    return !ptr->isValid();
-                }) != sharedOutputs.end()) {
+            if (std::get<0>(sharedOutputs) || std::get<1>(sharedOutputs)) {
                 graphNode->execute(stream);
-                for (auto & output : sharedOutputs)
+
+                for (auto & output : std::get<2>(sharedOutputs))
                     output->valid(true);
             }
         } else {
diff --git a/inference-engine/src/mkldnn_plugin/mkldnn_graph.h b/inference-engine/src/mkldnn_plugin/mkldnn_graph.h
index 0df95cbe86b..7834d08a7d3 100644
--- a/inference-engine/src/mkldnn_plugin/mkldnn_graph.h
+++ b/inference-engine/src/mkldnn_plugin/mkldnn_graph.h
@@ -30,7 +30,7 @@ public:
         Ready = 1,
     };
 
-    MKLDNNGraph(mkldnn::engine eng = mkldnn::engine(mkldnn::engine::kind::cpu, 0)) : status(NotReady), eng(eng) {}
+    MKLDNNGraph() = default;
 
     Status GetStatus() {
         return status;
@@ -172,7 +172,7 @@ protected:
         graphEdges.clear();
         _meanImages.clear();
     }
-    Status status;
+    Status status { NotReady };
     Config config;
 
     // For dumping purposes. -1 - no counting, all other positive
@@ -191,7 +191,7 @@ protected:
     std::map<std::string, MeanImage> _meanImages;
     std::string _name;
 
-    mkldnn::engine eng;
+    static mkldnn::engine eng;
 
     void Replicate(const InferenceEngine::CNNNetwork &network, const MKLDNNExtensionManager::Ptr& extMgr);
     void Replicate(const InferenceEngine::TensorIterator::Body &subgraph, const MKLDNNExtensionManager::Ptr& extMgr);
diff --git a/inference-engine/src/mkldnn_plugin/nodes/mkldnn_tensoriterator_node.cpp b/inference-engine/src/mkldnn_plugin/nodes/mkldnn_tensoriterator_node.cpp
index 173718e122e..b7898dfb992 100644
--- a/inference-engine/src/mkldnn_plugin/nodes/mkldnn_tensoriterator_node.cpp
+++ b/inference-engine/src/mkldnn_plugin/nodes/mkldnn_tensoriterator_node.cpp
@@ -187,8 +187,7 @@ private:
 }  // namespace MKLDNNPlugin
 
 MKLDNNTensorIteratorNode::MKLDNNTensorIteratorNode(InferenceEngine::CNNLayerPtr layer, const mkldnn::engine& eng, MKLDNNWeightsSharing::Ptr &cache) :
-        MKLDNNNode(layer, eng, cache),
-        sub_graph(eng) {}
+        MKLDNNNode(layer, eng, cache) {}
 
 void MKLDNNTensorIteratorNode::getSupportedDescriptors() {
     auto *ti = dynamic_cast<class InferenceEngine::TensorIterator*>(getCnnLayer().get());