[NPU] Get same zero context for different CORE (#29917)

### Details:
 - *Do not reinitialize driver and device for each ov::Core object*

---------

Signed-off-by: Bogdan Pereanu <[email protected]>

Author: pereanub

src/plugins/intel_npu/src/backend/src/zero_backend.cpp

@@ -14,7 +14,7 @@ namespace intel_npu {
 ZeroEngineBackend::ZeroEngineBackend() : _logger("ZeroEngineBackend", Logger::global().level()) {
     _logger.debug("ZeroEngineBackend - initialize started");
 
-    _initStruct = std::make_shared<ZeroInitStructsHolder>();
+    _initStruct = ZeroInitStructsHolder::getInstance();
 
     auto device = std::make_shared<ZeroDevice>(_initStruct);
     _devices.emplace(std::make_pair(device->getName(), device));

src/plugins/intel_npu/src/utils/include/intel_npu/utils/zero/zero_init.hpp

@@ -73,6 +73,8 @@ class ZeroInitStructsHolder final {
         return false;
     }
 
+    static const std::shared_ptr<ZeroInitStructsHolder>& getInstance();
+
 private:
     void initNpuDriver();
 

src/plugins/intel_npu/src/utils/src/zero/zero_init.cpp

@@ -309,6 +309,11 @@ ZeroInitStructsHolder::ZeroInitStructsHolder()
     THROW_ON_FAIL_FOR_LEVELZERO("pfnDeviceGetGraphProperties", result);
 }
 
+const std::shared_ptr<ZeroInitStructsHolder>& ZeroInitStructsHolder::getInstance() {
+    static std::shared_ptr<ZeroInitStructsHolder> instance = std::make_shared<ZeroInitStructsHolder>();
+    return instance;
+}
+
 ZeroInitStructsHolder::~ZeroInitStructsHolder() {
     if (context) {
         log.debug("ZeroInitStructsHolder - performing zeContextDestroy");

src/plugins/intel_npu/tests/functional/behavior/remote_tensor_tests/remote_run.hpp

@@ -753,7 +753,6 @@ TEST_P(RemoteRunTests, checkResultsAfterChangingStateTensorsWithRemoteTensors) {
     auto model = createModelWithStates(element::f32, original_shape);
 
     auto context = core->get_default_context(target_device).as<ov::intel_npu::level_zero::ZeroContext>();
-    ;
 
     compiled_model = core->compile_model(model, target_device, configuration);
     ov::InferRequest inference_request;
@@ -851,7 +850,6 @@ TEST_P(RemoteRunTests, checkResultsAfterChangingStateDataWithRemoteAndRandomTens
     auto model = createModelWithStates(element::f32, original_shape);
 
     auto context = core->get_default_context(target_device).as<ov::intel_npu::level_zero::ZeroContext>();
-    ;
 
     compiled_model = core->compile_model(model, target_device, configuration);
     ov::InferRequest inference_request;
@@ -940,7 +938,6 @@ TEST_P(RemoteRunTests, checkResultsAfterChangingStateDataWithRemoteAndRandomTens
     auto model = createModelWithStates(element::f32, original_shape);
 
     auto context = core->get_default_context(target_device).as<ov::intel_npu::level_zero::ZeroContext>();
-    ;
 
     compiled_model = core->compile_model(model, target_device, configuration);
     ov::InferRequest inference_request;
@@ -1018,6 +1015,42 @@ TEST_P(RemoteRunTests, checkResultsAfterChangingStateDataWithRemoteAndRandomTens
     }
 }
 
+TEST_P(RemoteRunTests, CheckContextFromDifferentOvCores) {
+    // Skip test according to plugin specific disabledTestPatterns() (if any)
+    SKIP_IF_CURRENT_TEST_IS_DISABLED()
+    ov::Core core0;
+    ov::Core core1;
+
+    auto context0 = core->get_default_context(target_device).as<ov::intel_npu::level_zero::ZeroContext>();
+    auto context1 = core0.get_default_context(target_device).as<ov::intel_npu::level_zero::ZeroContext>();
+    auto context2 = core1.get_default_context(target_device).as<ov::intel_npu::level_zero::ZeroContext>();
+
+    EXPECT_EQ(context0.get(), context1.get());
+    EXPECT_EQ(context1.get(), context2.get());
+}
+
+TEST_P(RemoteRunTests, CheckContextFromDifferentDestroyedOvCores) {
+    // Skip test according to plugin specific disabledTestPatterns() (if any)
+    SKIP_IF_CURRENT_TEST_IS_DISABLED()
+
+    ov::RemoteContext context1, context2;
+    auto context0 = core->get_default_context(target_device).as<ov::intel_npu::level_zero::ZeroContext>();
+
+    {
+        ov::Core internal_core;
+        context1 = internal_core.get_default_context(target_device);
+    }
+
+    {
+        ov::Core internal_core;
+        context2 = internal_core.get_default_context(target_device);
+    }
+
+    EXPECT_EQ(context0.get(), context1.get_params().at(ov::intel_npu::l0_context.name()).as<void*>());
+    EXPECT_EQ(context1.get_params().at(ov::intel_npu::l0_context.name()).as<void*>(),
+              context2.get_params().at(ov::intel_npu::l0_context.name()).as<void*>());
+}
+
 }  // namespace behavior
 }  // namespace test
 }  // namespace ov