Fix issue with deallocation from different .so on Linux platforms

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

Author: pereanub

src/plugins/intel_npu/src/backend/include/zero_backend.hpp

@@ -15,7 +15,7 @@ namespace intel_npu {
 class ZeroEngineBackend final : public IEngineBackend {
 public:
     ZeroEngineBackend();
-    ~ZeroEngineBackend();
+    ~ZeroEngineBackend() override = default;
     const std::shared_ptr<IDevice> getDevice() const override;
     const std::shared_ptr<IDevice> getDevice(const std::string&) const override;
     const std::string getName() const override {

src/plugins/intel_npu/src/backend/include/zero_tensor.hpp

@@ -69,7 +69,7 @@ class ZeroTensor final : public ov::ITensor {
     void prevent_reuse();
     bool can_be_reused();
 
-    ~ZeroTensor() override;
+    ~ZeroTensor() override = default;
 
 private:
     void update_strides() const;

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

@@ -143,11 +143,4 @@ const std::shared_ptr<ZeroInitStructsHolder> ZeroEngineBackend::getInitStructs()
     return _initStruct;
 }
 
-ZeroEngineBackend::~ZeroEngineBackend() {
-    if (_initStruct != nullptr) {
-        _devices.clear();
-        _initStruct->destroy();
-    }
-}
-
 }  // namespace intel_npu

src/plugins/intel_npu/src/backend/src/zero_tensor.cpp

@@ -204,10 +204,4 @@ bool ZeroTensor::can_be_reused() {
     return _can_be_reused;
 }
 
-ZeroTensor::~ZeroTensor() {
-    // make sure that the mem ref is destroyed before destroying the zero context
-    _mem_ref = {};
-    _init_structs->destroy();
-}
-
 }  // namespace intel_npu

src/plugins/intel_npu/src/plugin/include/remote_context.hpp

@@ -21,7 +21,7 @@ class RemoteContextImpl : public ov::IRemoteContext {
 public:
     RemoteContextImpl(const ov::SoPtr<IEngineBackend>& engineBackend, const ov::AnyMap& remote_properties = {});
 
-    ~RemoteContextImpl();
+    ~RemoteContextImpl() override = default;
 
     /**
      * @brief Returns name of a device on which underlying object is allocated.

src/plugins/intel_npu/src/plugin/src/remote_context.cpp

@@ -117,8 +117,4 @@ std::shared_ptr<ov::IRemoteContext> RemoteContextImpl::get_this_shared_ptr() {
     return shared_from_this();
 }
 
-RemoteContextImpl::~RemoteContextImpl() {
-    _init_structs->destroy();
-}
-
 }  // namespace intel_npu

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

@@ -10,7 +10,6 @@
 #include <ze_intel_npu_uuid.h>
 
 #include <memory>
-#include <mutex>
 
 #include "intel_npu/utils/logger/logger.hpp"
 #include "intel_npu/utils/zero/zero_api.hpp"
@@ -82,15 +81,8 @@ class ZeroInitStructsHolder final {
 
     static const std::shared_ptr<ZeroInitStructsHolder> getInstance();
 
-    static void destroy();
-
 private:
     void initNpuDriver();
-    static std::weak_ptr<ZeroInitStructsHolder>& getInstanceStorage();
-    // Static mutex for synchronization
-    static std::mutex& getMutex();
-
-    void destroy_context();
 
     // keep zero_api alive until context is destroyed
     std::shared_ptr<ZeroApi> zero_api;

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

@@ -71,7 +71,7 @@ class ZeroRemoteTensor final : public ov::IRemoteTensor {
     void* get_original_memory() const;
     ze_context_handle_t get_zero_context_handle() const;
 
-    ~ZeroRemoteTensor() override;
+    ~ZeroRemoteTensor() override = default;
 
 private:
     void allocate(const size_t bytes);

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

@@ -7,6 +7,7 @@
 #include <ze_command_queue_npu_ext.h>
 #include <ze_mem_import_system_memory_ext.h>
 
+#include <mutex>
 #include <regex>
 
 #include "intel_npu/utils/zero/zero_utils.hpp"
@@ -365,40 +366,19 @@ ZeroInitStructsHolder::ZeroInitStructsHolder()
 }
 
 const std::shared_ptr<ZeroInitStructsHolder> ZeroInitStructsHolder::getInstance() {
-    std::lock_guard<std::mutex> lock(getMutex());
-    auto instance = getInstanceStorage().lock();
+    static std::mutex mutex;
+    static std::weak_ptr<ZeroInitStructsHolder> weak_instance;
+
+    std::lock_guard<std::mutex> lock(mutex);
+    auto instance = weak_instance.lock();
     if (!instance) {
         instance = std::make_shared<ZeroInitStructsHolder>();
-        getInstanceStorage() = instance;
+        weak_instance = instance;
     }
     return instance;
 }
 
-void ZeroInitStructsHolder::destroy() {
-    std::lock_guard<std::mutex> lock(getMutex());
-    auto instance = getInstanceStorage().lock();
-    if (instance && instance.use_count() == 2) {
-        instance->destroy_context();
-        // reset weak pointer
-        getInstanceStorage().reset();
-        // reset shared pointer
-        instance.reset();
-    }
-    // don't destroy if ref count is higher than 2.
-    // one is after getInstanceStorage().lock() and another one is hold by the caller
-}
-
-std::weak_ptr<ZeroInitStructsHolder>& ZeroInitStructsHolder::getInstanceStorage() {
-    static std::weak_ptr<ZeroInitStructsHolder> weak_instance;
-    return weak_instance;
-}
-
-std::mutex& ZeroInitStructsHolder::getMutex() {
-    static std::mutex mutex;
-    return mutex;
-}
-
-void ZeroInitStructsHolder::destroy_context() {
+ZeroInitStructsHolder::~ZeroInitStructsHolder() {
     if (context) {
         if (context_npu_dditable_ext_decorator->version() >= ZE_MAKE_VERSION(1, 0)) {
             context_options &= ~(ZE_NPU_CONTEXT_OPTION_ENABLE_IDLE_OPTIMIZATIONS);
@@ -421,8 +401,4 @@ void ZeroInitStructsHolder::destroy_context() {
     }
 }
 
-ZeroInitStructsHolder::~ZeroInitStructsHolder() {
-    destroy_context();
-}
-
 }  // namespace intel_npu

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

@@ -133,8 +133,6 @@ ZeroMem::~ZeroMem() {
                       uint64_t(result),
                       ze_result_to_description(result).c_str());
     }
-
-    _init_structs->destroy();
 }
 
 }  // namespace intel_npu