[GPU] Fix incorrect output result in primitives with two outputs (#8902)

inference-engine/thirdparty/clDNN/api/cldnn/graph/network.hpp

@@ -168,6 +168,7 @@ public:
     void execute_primitive(const std::shared_ptr<primitive_inst>& primitive,
                            const std::vector<event::ptr>& events);
     void allocate_primitives();
+    void configure_primitives_second_output();
     void build_insts_deps();
     uint32_t get_id() const { return net_id; }
     stream& get_stream() const { return *_stream; }

inference-engine/thirdparty/clDNN/src/network.cpp

@@ -214,6 +214,7 @@ network::network(program::ptr program, stream::ptr stream, bool is_internal, boo
     }
 
     allocate_primitives();
+    configure_primitives_second_output();
     check_names();
     build_insts_deps();
     build_exec_order();
@@ -552,6 +553,41 @@ void network::allocate_primitives() {
     }
 }
 
+void network::configure_primitives_second_output() {
+    std::map<cldnn::memory::ptr, std::vector<const cldnn::program_node*>> mutable_datas_ptrs;
+    for (auto& inst : _primitives) {
+        auto& node = inst.second->get_node();
+
+        if (!node.is_type<mutable_data>())
+            continue;
+
+        mutable_datas_ptrs[node.as<mutable_data>().get_attached_memory_ptr()].push_back(&node);
+    }
+
+    for (auto item : mutable_datas_ptrs) {
+        if (item.second.size() != 2)
+            continue;
+
+        auto is_first_node_input_md = [&](const cldnn::program_node* first,
+                                          const cldnn::program_node* second) {
+            for (auto user : first->get_users()) {
+                for (auto next_user : user->get_users()) {
+                    if (next_user == second)
+                        return true;
+                }
+            }
+            return false;
+        };
+
+        auto is_first_node_input = is_first_node_input_md(item.second[0], item.second[1]);
+
+        auto input_md_inst = is_first_node_input ? _primitives[item.second[0]->id()] : _primitives[item.second[1]->id()];
+        auto output_md_inst = is_first_node_input ? _primitives[item.second[1]->id()] : _primitives[item.second[0]->id()];
+
+        output_md_inst->set_output_memory(input_md_inst->output_memory_ptr(), false);
+    }
+}
+
 void network::build_insts_deps() {
     for (auto& inst : _primitives) {
         inst.second->build_deps();

inference-engine/thirdparty/clDNN/tests/test_cases/set_output_memory_gpu_test.cpp

@@ -327,7 +327,7 @@ TEST(set_output_memory_gpu, basic_opt) {
     network.set_input_data("input1", input1);
     network.set_input_data("input2", input2);
     network.set_output_memory(outputID, output_mem);
- 
+
     auto outputs = network.execute();
     auto output = outputs.at(outputID).get_memory();
     //  check for correct output memory setting
@@ -349,14 +349,14 @@ TEST(set_output_memory_gpu, mutable_output_data) {
     auto& engine = get_test_engine();
     const int top_k = 2;
     auto input = engine.allocate_memory({ data_types::f32, format::bfyx,{ batch_num, feature_num, x_size , y_size } });
-    auto second_input = engine.allocate_memory({ data_types::f32, format::bfyx, { top_k, feature_num, x_size , y_size } });
-    auto final_output = engine.allocate_memory({ data_types::f32, format::bfyx,{ 1, 1, 1 , 1 } });
+    auto final_output = engine.allocate_memory({ data_types::f32, format::bfyx, { top_k, feature_num, x_size , y_size } });
+    auto second_input = engine.allocate_memory({ data_types::f32, format::bfyx,{ 1, 1, 1 , 1 } });
 
     topology topology;
     topology.add(input_layout("Add_1396", input->get_layout()));
     topology.add(cldnn::mutable_data("second_input", second_input));
     topology.add(cldnn::mutable_data("12220_md_write", final_output));
-    topology.add(arg_max_min("arg_max", { "Add_1396", "12220_md_write", "second_input" }, arg_max_min::min, top_k, arg_max_min::batch));
+    topology.add(arg_max_min("arg_max", { "Add_1396", "second_input", "12220_md_write" }, arg_max_min::min, top_k, arg_max_min::batch));
     topology.add(cldnn::mutable_data("pred/sink_port_0", {"arg_max"},final_output) );
 
     std::vector<float> input_vec = {