[GPU] Fix network loading time related to onednn engine creation (#12375)

+ Creation of onednn_engine in ocl_engine is changed to on-demand.
+ benchmark cache_dir option takes longer than cl_cache_dir env in loading network.
+ For clDNN execution, benchmark cache_dir created onednn_engine if just ONEDNN_ENABLE config is ON.

Signed-off-by: Min, Byungil <byungil.min@intel.com>

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

@@ -44,7 +44,7 @@ public:
     static queue_types detect_queue_type(engine_types engine_type, void* queue_handle);
 
 #ifdef ENABLE_ONEDNN_FOR_GPU
-    virtual dnnl::stream& get_onednn_stream() = 0;
+    virtual dnnl::stream& get_onednn_stream() const = 0;
 #endif
 
 protected:

src/plugins/intel_gpu/src/runtime/ocl/ocl_engine.cpp

@@ -51,16 +51,20 @@ ocl_engine::ocl_engine(const device::ptr dev, runtime_types runtime_type,
 
     _usm_helper.reset(new cl::UsmHelper(get_cl_context(), get_cl_device(), use_unified_shared_memory()));
 
-#ifdef ENABLE_ONEDNN_FOR_GPU
-    _onednn_engine = std::make_shared<dnnl::engine>(dnnl::ocl_interop::make_engine(casted->get_device().get(), casted->get_context().get()));
-#endif
     _program_stream.reset(new ocl_stream(*this));
 }
 
 #ifdef ENABLE_ONEDNN_FOR_GPU
 dnnl::engine& ocl_engine::get_onednn_engine() const {
-    if (!_onednn_engine)
-        throw ov::Exception("[GPU] onednn engine is nullptr");
+    const std::lock_guard<std::mutex> lock(onednn_mutex);
+    if (!_onednn_engine) {
+        auto casted = std::dynamic_pointer_cast<ocl_device>(_device);
+        if (!casted)
+            throw ov::Exception("[GPU] Invalid device type stored in ocl_engine");
+
+        _onednn_engine = std::make_shared<dnnl::engine>(dnnl::ocl_interop::make_engine(casted->get_device().get(), casted->get_context().get()));
+    }
+
     return *_onednn_engine;
 }
 #endif

src/plugins/intel_gpu/src/runtime/ocl/ocl_engine.hpp

@@ -44,7 +44,8 @@ public:
     stream& get_program_stream() const override;
 
 #ifdef ENABLE_ONEDNN_FOR_GPU
-    /// Returns onednn engine object which shares device and context with current engine
+    // Returns onednn engine object which shares device and context with current engine
+    // If onednn engine has not been created yet, it creates on-demand.
     dnnl::engine& get_onednn_engine() const override;
 #endif
 
@@ -57,7 +58,8 @@ private:
     std::unique_ptr<cl::UsmHelper> _usm_helper;
 
 #ifdef ENABLE_ONEDNN_FOR_GPU
-    std::shared_ptr<dnnl::engine> _onednn_engine;
+    mutable std::mutex onednn_mutex;
+    mutable std::shared_ptr<dnnl::engine> _onednn_engine;
 #endif
 };
 

src/plugins/intel_gpu/src/runtime/ocl/ocl_stream.cpp

@@ -289,6 +289,7 @@ ocl_stream::ocl_stream(const ocl_engine &engine)
     queue_builder.set_supports_queue_families(queue_families_extension);
 
     _command_queue = queue_builder.build(context, device);
+
 #ifdef ENABLE_ONEDNN_FOR_GPU
     if (config.queue_type == queue_types::in_order) {
         auto onednn_engine = engine.get_onednn_engine();
@@ -317,9 +318,10 @@ ocl_stream::ocl_stream(const ocl_engine &engine, void *handle)
 }
 
 #ifdef ENABLE_ONEDNN_FOR_GPU
-dnnl::stream& ocl_stream::get_onednn_stream() {
+dnnl::stream& ocl_stream::get_onednn_stream() const {
     if (!_onednn_stream)
         throw std::runtime_error("[GPU] onednn stream is nullptr");
+
     return *_onednn_stream;
 }
 #endif

src/plugins/intel_gpu/src/runtime/ocl/ocl_stream.hpp

@@ -83,7 +83,7 @@ public:
     static queue_types detect_queue_type(void* queue_handle);
 
 #ifdef ENABLE_ONEDNN_FOR_GPU
-    dnnl::stream& get_onednn_stream() override;
+    dnnl::stream& get_onednn_stream() const override;
 #endif
 
 private: