[GPU] Fix a bug that occurred when the optimized layer is executed. (#13568)

src/plugins/intel_gpu/src/graph/crop.cpp

@@ -256,4 +256,15 @@ void crop_inst::on_execute() {
 void crop_inst::reuse_input() {
     _outputs[0] = _network.get_engine().reinterpret_buffer(input_memory(), node->get_output_layout());
 }
+
+void crop_inst::update_output_memory() {
+    if (!node->can_be_optimized())
+        return;
+
+    if (_outputs[0] && _network.get_engine().is_the_same_buffer(output_memory(), input_memory()))
+        return;
+
+    _outputs[0] = _network.get_engine().reinterpret_buffer(input_memory(), node->get_output_layout());
+}
+
 }  // namespace cldnn

src/plugins/intel_gpu/src/graph/include/crop_inst.h

@@ -54,6 +54,7 @@ public:
     static layout calc_output_layout(crop_node const& node, kernel_impl_params const& impl_param);
     static std::string to_string(crop_node const& node);
     typed_primitive_inst(network& network, crop_node const& node);
+    void update_output_memory() override;
 
 private:
     void on_execute() override;

src/plugins/intel_gpu/src/graph/include/primitive_inst.h

@@ -197,6 +197,8 @@ public:
 
     layout get_node_output_layout() const { return _node_output_layout; }
 
+    virtual void update_output_memory() {}
+
 protected:
     primitive_inst(network& network, program_node const& node, bool allocate_memory);
 

src/plugins/intel_gpu/src/graph/include/reorder_inst.h

@@ -67,6 +67,8 @@ public:
 
     bool has_mean() const { return !argument->mean.empty(); }
 
+    void update_output_memory() override;
+
 private:
     void on_execute() override;
     void reuse_input();

src/plugins/intel_gpu/src/graph/include/reshape_inst.h

@@ -51,6 +51,8 @@ public:
 
     typed_primitive_inst(network& network, reshape_node const& node);
 
+    void update_output_memory() override;
+
 private:
     void on_execute() override;
 

src/plugins/intel_gpu/src/graph/network.cpp

@@ -643,6 +643,15 @@ void network::allocate_primitives() {
             }
         }
     }
+
+    // Update the output memory address of optimized-out layer if it is not valid.
+    for (auto const& node : po) {
+        if (node->can_be_optimized()) {
+            auto opt_inst = _primitives.at(node->id());
+            opt_inst->update_output_memory();
+        }
+    }
+
     // allocate intermediate buffers
     for (auto const& node : po) {
         auto prim = _primitives[node->id()];

src/plugins/intel_gpu/src/graph/reorder.cpp

@@ -243,6 +243,13 @@ void reorder_inst::on_execute() {
 }
 
 void reorder_inst::reuse_input() {
+    update_output_memory();
+}
+
+void reorder_inst::update_output_memory() {
+    if (!node->can_be_optimized())
+        return;
+
     if (static_cast<bool>(_outputs[0]) && _network.get_engine().is_the_same_buffer(output_memory(), input_memory()))
         return;
 

src/plugins/intel_gpu/src/graph/reshape.cpp

@@ -189,6 +189,16 @@ void reshape_inst::on_execute() {
 }
 
 void reshape_inst::reuse_input() {
+    update_output_memory();
+}
+
+void reshape_inst::update_output_memory() {
+    if (!node->can_be_optimized())
+        return;
+
+    if (_outputs[0] && _network.get_engine().is_the_same_buffer(output_memory(), input_memory()))
+        return;
+
     build_deps();  // reshape need deps
     OPENVINO_ASSERT(input_memory_ptr() != nullptr, "[GPU] Failed to reuse input in ", id(), " primitive: input memory was not allocated");
     _outputs = {_network.get_engine().reinterpret_buffer(input_memory(), _impl_params->output_layout)};