Fix perf regression in some models (#3260)

* Fix perf regression in vgg16

* Make switch generic

* Remove unused variables

* Review comments

* Remove unused function parameters

* trivial commit to restart CI

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -1194,6 +1194,7 @@ void runtime::cpu::CPU_ExternalFunction::register_common_passes(
     REGISTER_KNOBBED_PASS(RecurrentReshapeElimination, false, ngraph::pass);
     REGISTER_KNOBBED_PASS_WITH_ARGS(
         CoreFusion, true, ngraph::pass, ngraph::pass::FusionType::ALL_FUSIONS);
+    REGISTER_KNOBBED_PASS(CPUPreFusion, true, runtime::cpu::pass);
 
     // Disable CPUFusion if MLIR is enabled to preserve core ops.
     if (std::getenv("NGRAPH_MLIR") == nullptr)

src/ngraph/runtime/cpu/pass/cpu_fusion.cpp

@@ -520,6 +520,70 @@ void ngraph::runtime::cpu::pass::CPUFusion::construct_conv_bias_bprop()
     this->add_matcher(m, callback);
 }
 
+static bool switch_nodes(std::shared_ptr<ngraph::Node> node1,
+                         std::shared_ptr<ngraph::Node> node2,
+                         size_t source_input_index = 0)
+{
+    // check if node1 has only 1 argument, not sure how it will work with >1 args
+    if (node1->inputs().size() > 1)
+    {
+        NGRAPH_DEBUG << "Cannot switch. More than 1 inputs to this node\n";
+        return false;
+    }
+    if (node1->get_users().size() > 1)
+    {
+        NGRAPH_DEBUG << "Cannot switch. More than 1 user of this node\n";
+        return false;
+    }
+    if (node1->outputs().size() > 1)
+    {
+        NGRAPH_DEBUG << "Cannot switch. More than 1 output of this node\n";
+        return false;
+    }
+    if (node2->outputs().size() > 1)
+    {
+        NGRAPH_DEBUG << "Cannot switch. More than 1 output of this node\n";
+        return false;
+    }
+
+    auto target_inputs = node2->output(0).get_target_inputs();
+    // Remove the control_dependency, which shouldn't be there, but in case
+    // Other control_dependencies will work out fine even after switch.
+    node2->remove_control_dependency(node1);
+
+    // actual switch happening after this
+    auto arg = node1->get_argument(source_input_index);
+    node2->input(0).replace_source_output(arg);
+
+    node1->input(0).replace_source_output(node2->output(0));
+
+    // used implementation ref from replace_node
+    for (auto& input : target_inputs)
+    {
+        input.replace_source_output(node1->output(0));
+    }
+    return true;
+}
+
+void ngraph::runtime::cpu::pass::CPUPreFusion::construct_maxpool_relu_switch()
+{
+    auto input_shape = Shape{1, 2, 2, 2};
+    auto input = std::make_shared<pattern::op::Label>(element::f32, input_shape);
+    Shape window_shape{2, 2};
+    auto max_pool = std::make_shared<ngraph::op::MaxPool>(input, window_shape);
+    auto prelu = std::make_shared<ngraph::op::Relu>(max_pool);
+
+    auto callback = [input](pattern::Matcher& m) {
+        NGRAPH_DEBUG << "In callback for construct_maxpool_relu_switch against node = "
+                     << m.get_match_root()->get_name();
+
+        return switch_nodes(m.get_match_root()->get_argument(0), m.get_match_root());
+    };
+
+    auto m = std::make_shared<ngraph::pattern::Matcher>(prelu, "CPUPreFusion.MaxpoolReluSwitch");
+    this->add_matcher(m, callback);
+}
+
 void ngraph::runtime::cpu::pass::CPUFusion::construct_batch_norm_relu()
 {
     auto input_shape = Shape{1, 2, 2, 2};

src/ngraph/runtime/cpu/pass/cpu_fusion.hpp

@@ -27,6 +27,7 @@ namespace ngraph
         {
             namespace pass
             {
+                class CPUPreFusion;
                 class CPUFusion;
                 class CPUQuantFusion;
             }
@@ -34,6 +35,19 @@ namespace ngraph
     }
 }
 
+class CPU_BACKEND_API ngraph::runtime::cpu::pass::CPUPreFusion : public ngraph::pass::GraphRewrite
+{
+public:
+    CPUPreFusion()
+        : GraphRewrite()
+    {
+        construct_maxpool_relu_switch();
+    }
+
+private:
+    void construct_maxpool_relu_switch();
+};
+
 class CPU_BACKEND_API ngraph::runtime::cpu::pass::CPUFusion : public ngraph::pass::GraphRewrite
 {
 public: