Add new debug config disable_memory_reuse

src/plugins/intel_gpu/include/intel_gpu/runtime/debug_configuration.hpp

@@ -120,6 +120,7 @@ class debug_configuration {
     int disable_async_compilation;              // Disable async compilation
     int disable_dynamic_impl;                   // Disable dynamic implementation
     int disable_runtime_buffer_fusing;          // Disable runtime buffer fusing
+    int disable_memory_reuse;                   // Disable memmory reuse among layers
     std::set<int64_t> dump_iteration;           // Dump n-th execution of network.
     static const debug_configuration *get_instance();
     bool is_dumped_layer(const std::string& layerName, bool is_output = false) const;

src/plugins/intel_gpu/src/graph/primitive_inst.cpp

@@ -1180,6 +1180,10 @@ memory::ptr primitive_inst::allocate_output(engine& _engine, memory_pool& pool,
     if (_node.is_in_shape_of_subgraph())
         reusable_across_network = false;
 
+    GPU_DEBUG_GET_INSTANCE(debug_config);
+    GPU_DEBUG_IF(debug_config->disable_memory_reuse) {
+        reusable_across_network = false;
+    }
     // For outputs, cpu prim we want to have lockable alloc type
     // Also if the successor of a node is an cpu, then memory needs to be lockable.
     bool is_cpu = _node.get_selected_impl() ? _node.get_selected_impl()->is_cpu() : false;

src/plugins/intel_gpu/src/runtime/debug_configuration.cpp

@@ -139,6 +139,7 @@ static void print_help_messages() {
                               "the following order: number of iterations for pre-allocation(int), max size of single iteration in bytes(int), "
                               "max per-dim allowed diff(int), unconditional buffers preallocation ratio(float). For example for disabling memory"
                               "preallocation at all, you can use OV_GPU_MemPreallocationOptions='0 0 0 1.0'");
+    message_list.emplace_back("OV_GPU_DisableMemoryReuse", "Disable memory reuse");
 
     auto max_name_length_item = std::max_element(message_list.begin(), message_list.end(),
         [](std::pair<std::string, std::string>& a, std::pair<std::string, std::string>& b){
@@ -177,7 +178,8 @@ debug_configuration::debug_configuration()
         , max_kernels_per_batch(0)
         , disable_async_compilation(0)
         , disable_dynamic_impl(0)
-        , disable_runtime_buffer_fusing(0) {
+        , disable_runtime_buffer_fusing(0)
+        , disable_memory_reuse(0) {
 #ifdef GPU_DEBUG_CONFIG
     get_gpu_debug_env_var("Help", help);
     get_common_debug_env_var("Verbose", verbose);
@@ -208,6 +210,7 @@ debug_configuration::debug_configuration()
     get_gpu_debug_env_var("DisableAsyncCompilation", disable_async_compilation);
     get_gpu_debug_env_var("DisableDynamicImpl", disable_dynamic_impl);
     get_gpu_debug_env_var("DisableRuntimeBufferFusing", disable_runtime_buffer_fusing);
+    get_gpu_debug_env_var("DisableMemoryReuse", disable_memory_reuse);
     std::string dump_iteration_str;
     get_gpu_debug_env_var("DumpIteration", dump_iteration_str);
     std::string mem_preallocation_params_str;