[DO NOT MERGE] [NPU] Using global command queue

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

@@ -40,6 +40,9 @@ struct Pipeline {
     void closeCommandListIndex(size_t command_list_index);
 
 protected:
+    void getCommandQueue();
+
+    std::shared_ptr<ZeroInitStructsHolder> _init_structs;
     std::shared_ptr<IGraph> _graph;
     const Config _config;
     const uint32_t _id;
@@ -59,9 +62,11 @@ struct Pipeline {
     std::vector<std::unique_ptr<Fence>> _fences;
     std::shared_ptr<EventPool> _event_pool;
     std::vector<std::shared_ptr<Event>> _events;
-    bool sync_output_with_fences_ = true;
+    bool _sync_output_with_fences = true;
     std::shared_ptr<zeroProfiling::NpuInferProfiling> _npu_profiling;
     Logger _logger;
+
+    std::mutex _mutex;
 };
 
 }  // namespace intel_npu

src/plugins/intel_npu/src/backend/src/zero_pipeline.cpp

@@ -15,6 +15,21 @@
 #include "intel_npu/utils/zero/zero_types.hpp"
 #include "zero_remote_tensor.hpp"
 
+namespace {
+
+template <class t>
+bool compare_shared_ptr(const std::shared_ptr<t>& a, const std::shared_ptr<t>& b) {
+    if (a == b) {
+        return true;
+    }
+    if (a && b) {
+        return a.get() == b.get();
+    }
+    return false;
+}
+
+}  // namespace
+
 namespace intel_npu {
 
 Pipeline::Pipeline(const Config& config,
@@ -26,14 +41,11 @@ Pipeline::Pipeline(const Config& config,
                    const std::vector<std::vector<std::shared_ptr<ov::ITensor>>>& input_tensors,
                    const std::vector<std::shared_ptr<ov::ITensor>>& output_tensors,
                    uint32_t group_ordinal)
-    : _graph(graph),
+    : _init_structs(init_structs),
+      _graph(graph),
       _config(config),
       _id(_graph->get_unique_id()),
       _number_of_command_lists(_graph->get_batch_size().has_value() ? *_graph->get_batch_size() : 1),
-      _event_pool{
-          std::make_shared<EventPool>(init_structs->getDevice(),
-                                      init_structs->getContext(),
-                                      _number_of_command_lists ? static_cast<uint32_t>(_number_of_command_lists) : 1)},
       _npu_profiling(npu_profiling),
       _logger("Pipeline", _config.get<LOG_LEVEL>()) {
     OV_ITT_SCOPED_TASK(itt::domains::LevelZeroBackend, "Zero_infer_request::Pipeline::Pipeline");
@@ -43,22 +55,43 @@ Pipeline::Pipeline(const Config& config,
         profiling_query.create(profiling_pool._handle);
     }
 
+    OPENVINO_ASSERT(_sync_output_with_fences || !_config.get<RUN_INFERENCES_SEQUENTIALLY>(),
+                    "In-order execution doesn't work in case synchronization of the inferences is done using events");
+
+    if (!_sync_output_with_fences || _config.get<RUN_INFERENCES_SEQUENTIALLY>()) {
+        _event_pool =
+            std::make_shared<EventPool>(_init_structs->getDevice(),
+                                        _init_structs->getContext(),
+                                        _number_of_command_lists ? static_cast<uint32_t>(_number_of_command_lists) : 1);
+
+        _events.reserve(_number_of_command_lists);
+        for (size_t i = 0; i < _number_of_command_lists; i++) {
+            _events.emplace_back(std::make_shared<Event>(_event_pool, static_cast<uint32_t>(i)));
+        }
+    }
+
     _command_lists.reserve(_number_of_command_lists);
-    _events.reserve(_number_of_command_lists);
-    _fences.reserve(_number_of_command_lists);
-    _logger.debug("Pipeline - emplace_back _event_pool and _command_queue");
     for (size_t i = 0; i < _number_of_command_lists; i++) {
         _command_lists.emplace_back(
-            std::make_unique<CommandList>(init_structs,
+            std::make_unique<CommandList>(_init_structs,
                                           group_ordinal,
-                                          init_structs->getMutableCommandListVersion() ? true : false));
-        _events.emplace_back(std::make_shared<Event>(_event_pool, static_cast<uint32_t>(i)));
-        _fences.emplace_back(std::make_unique<Fence>(*_graph->get_command_queue()));
+                                          _init_structs->getMutableCommandListVersion() ? true : false));
+    }
+
+    _command_queue = _graph->get_command_queue();
+
+    if (_sync_output_with_fences) {
+        _fences.resize(_number_of_command_lists);
+
+        for (size_t i = 0; i < _number_of_command_lists; i++) {
+            _logger.debug("Pipeline - getCommandQueue() - create new fence");
+            _fences[i] = std::make_unique<Fence>(_command_queue);
+        }
     }
 
     for (size_t i = 0; i < _number_of_command_lists; i++) {
         size_t io_index = 0;
-        for (const auto& desc : graph->get_input_descriptors()) {
+        for (const auto& desc : _graph->get_input_descriptors()) {
             if (input_tensors.at(io_index).size() > 1) {
                 void* data = nullptr;
                 auto remote_tensor = std::dynamic_pointer_cast<ZeroRemoteTensor>(input_tensors.at(io_index).at(i));
@@ -68,7 +101,7 @@ Pipeline::Pipeline(const Config& config,
                     data = remote_tensor->get_original_memory();
                 }
 
-                graph->set_argument_value(desc.idx, data);
+                _graph->set_argument_value(desc.idx, data);
 
                 ++io_index;
                 continue;
@@ -82,7 +115,7 @@ Pipeline::Pipeline(const Config& config,
                 data = remote_tensor->get_original_memory();
             }
 
-            graph->set_argument_value(
+            _graph->set_argument_value(
                 desc.idx,
                 static_cast<unsigned char*>(data) +
                     (i * input_tensors.at(io_index).at(0)->get_byte_size()) / _number_of_command_lists);
@@ -91,7 +124,7 @@ Pipeline::Pipeline(const Config& config,
         }
 
         io_index = 0;
-        for (const auto& desc : graph->get_output_descriptors()) {
+        for (const auto& desc : _graph->get_output_descriptors()) {
             void* data = nullptr;
             auto remote_tensor = std::dynamic_pointer_cast<ZeroRemoteTensor>(output_tensors.at(io_index));
             if (remote_tensor == nullptr) {
@@ -100,7 +133,7 @@ Pipeline::Pipeline(const Config& config,
                 data = remote_tensor->get_original_memory();
             }
 
-            graph->set_argument_value(
+            _graph->set_argument_value(
                 desc.idx,
                 static_cast<unsigned char*>(data) +
                     (i * output_tensors.at(io_index)->get_byte_size()) / _number_of_command_lists);
@@ -119,7 +152,7 @@ Pipeline::Pipeline(const Config& config,
             _command_lists.at(i)->appendNpuTimestamp(reinterpret_cast<uint64_t*>(_npu_profiling->npu_ts_infer_start));
         }
 
-        _command_lists.at(i)->appendGraphExecute(static_cast<ze_graph_handle_t>(graph->get_handle()),
+        _command_lists.at(i)->appendGraphExecute(static_cast<ze_graph_handle_t>(_graph->get_handle()),
                                                  profiling_query.getHandle());
 
         /// append timestamp command if feature was activated
@@ -138,7 +171,7 @@ Pipeline::Pipeline(const Config& config,
         }
 
         // appendBarrier used in L0 as well
-        if (!sync_output_with_fences_) {
+        if (!_sync_output_with_fences) {
             _command_lists.at(i)->appendBarrier();
             _events.at(i)->AppendSignalEvent(*_command_lists.at(i));
         }
@@ -147,9 +180,30 @@ Pipeline::Pipeline(const Config& config,
     _logger.debug("Pipeline - initialize completed");
 }
 
+void Pipeline::getCommandQueue() {
+    _logger.debug("Pipeline - getCommandQueue() started");
+
+    std::lock_guard<std::mutex> lock(_mutex);
+
+    if (!compare_shared_ptr(_command_queue, _graph->get_command_queue())) {
+        _command_queue = _graph->get_command_queue();
+
+        if (_sync_output_with_fences) {
+            for (size_t i = 0; i < _number_of_command_lists; i++) {
+                _logger.debug("Pipeline - getCommandQueue() - create new fence");
+                _fences[i] = std::make_unique<Fence>(_command_queue);
+            }
+        }
+    }
+
+    _logger.debug("Pipeline - getCommandQueue() completed");
+}
+
 void Pipeline::push() {
     _logger.debug("Pipeline - push() started");
 
+    getCommandQueue();
+
     if (_config.get<RUN_INFERENCES_SEQUENTIALLY>()) {
         if (_id) {
             auto previousIndex = _graph->get_last_submitted_id();
@@ -164,10 +218,10 @@ void Pipeline::push() {
 
     for (size_t i = 0; i < _command_lists.size(); ++i) {
         OV_ITT_TASK_CHAIN(ZERO_PIPELINE_IP_PUSH, itt::domains::LevelZeroBackend, "Pipeline", "push");
-        if (sync_output_with_fences_) {
-            _graph->get_command_queue()->executeCommandList(*_command_lists.at(i), *_fences.at(i));
+        if (_sync_output_with_fences) {
+            _command_queue->executeCommandList(*_command_lists.at(i), *_fences.at(i));
         } else {
-            _graph->get_command_queue()->executeCommandList(*_command_lists.at(i));
+            _command_queue->executeCommandList(*_command_lists.at(i));
         }
     }
 
@@ -179,7 +233,7 @@ void Pipeline::pull() {
     OV_ITT_TASK_CHAIN(ZERO_PIPELINE_IP_PULL, itt::domains::LevelZeroBackend, "Pipeline", "pull");
 
     for (size_t i = 0; i < _command_lists.size(); ++i) {
-        if (sync_output_with_fences_) {
+        if (_sync_output_with_fences) {
             _fences.at(i)->hostSynchronize();
         } else {
             _events.at(i)->hostSynchronize();
@@ -194,17 +248,17 @@ void Pipeline::pull() {
 };
 
 void Pipeline::reset() const {
-    _logger.debug("Pipeline - rest() started");
+    _logger.debug("Pipeline - reset() started");
 
     for (size_t i = 0; i < _command_lists.size(); ++i) {
-        if (sync_output_with_fences_) {
+        if (_sync_output_with_fences) {
             _fences.at(i)->reset();
         } else {
             _events.at(i)->reset();
         }
     }
 
-    _logger.debug("Pipeline - rest() completed");
+    _logger.debug("Pipeline - reset() completed");
 };
 
 void Pipeline::updateCommandList(uint32_t arg_index, const void* arg_data, size_t byte_size) {

src/plugins/intel_npu/src/common/include/intel_npu/common/igraph.hpp

@@ -42,9 +42,10 @@ class IGraph : public std::enable_shared_from_this<IGraph> {
 
     const std::vector<ArgumentDescriptor>& get_input_descriptors() const;
     const std::vector<ArgumentDescriptor>& get_output_descriptors() const;
+
     const std::shared_ptr<CommandQueue>& get_command_queue() const;
 
-    void set_workload_type(const ov::WorkloadType workloadType) const;
+    virtual void set_workload_type(const ov::WorkloadType workloadType) = 0;
 
     std::mutex& get_mutex();
 
@@ -59,8 +60,8 @@ class IGraph : public std::enable_shared_from_this<IGraph> {
 
 protected:
     /**
-     * @brief Determines if batching can be addressed inside the plugin. In the positive case, the batch size used by
-     * the model will also be deduced and returned.
+     * @brief Determines if batching can be addressed inside the plugin. In the positive case, the batch size used
+     * by the model will also be deduced and returned.
      * @details Batching can be handled by the plugin only if:
      *  - The batch axis is the first axis.
      *  - The batch size received by the compiler takes the default value of 1.
@@ -72,22 +73,23 @@ class IGraph : public std::enable_shared_from_this<IGraph> {
      *
      * @param metadata Metadata containing the shape values as seen by both the compiler and IR model. These will
      * ultimately be used for determining the batch size.
-     * @returns The batch size deduced by the algorithm or the default value of 1 if batching cannot be performed inside
-     * the plugin.
+     * @returns The batch size deduced by the algorithm or the default value of 1 if batching cannot be performed
+     * inside the plugin.
      */
     std::optional<size_t> get_batch_size(const NetworkMetadata& metadata);
 
+    virtual void create_new_command_queue() = 0;
+
     ze_graph_handle_t _handle = nullptr;
     NetworkMetadata _metadata;
 
     std::vector<ArgumentDescriptor> _input_descriptors;
     std::vector<ArgumentDescriptor> _output_descriptors;
 
-    std::shared_ptr<CommandQueue> _command_queue;
     std::vector<std::shared_ptr<Event>> _last_submitted_event;
 
-    // Used to protect zero pipeline creation in the graph. The pipeline should be created only once per graph when the
-    // first inference starts running
+    // Used to protect zero pipeline creation in the graph. The pipeline should be created only once per graph when
+    // the first inference starts running
     std::mutex _mutex;
 
     std::unique_ptr<BlobContainer> _blobPtr;
@@ -101,6 +103,12 @@ class IGraph : public std::enable_shared_from_this<IGraph> {
      */
     std::optional<std::size_t> _batch_size = std::nullopt;
 
+    std::shared_ptr<CommandQueue> _command_queue;
+    uint32_t _group_ordinal;
+    ze_command_queue_workload_type_t _ze_workload_type;
+    bool _turbo = false;
+    ze_command_queue_priority_t _ze_queue_priority;
+
     Logger _logger;
 };
 

src/plugins/intel_npu/src/common/src/igraph.cpp

@@ -47,26 +47,6 @@ const std::shared_ptr<CommandQueue>& IGraph::get_command_queue() const {
     return _command_queue;
 }
 
-void IGraph::set_workload_type(const ov::WorkloadType workloadType) const {
-    if (_command_queue == nullptr) {
-        return;
-    }
-
-    ze_command_queue_workload_type_t zeWorkloadType;
-    switch (workloadType) {
-    case ov::WorkloadType::DEFAULT:
-        zeWorkloadType = ze_command_queue_workload_type_t::ZE_WORKLOAD_TYPE_DEFAULT;
-        break;
-    case ov::WorkloadType::EFFICIENT:
-        zeWorkloadType = ze_command_queue_workload_type_t::ZE_WORKLOAD_TYPE_BACKGROUND;
-        break;
-    default:
-        OPENVINO_THROW("Unknown value for WorkloadType!");
-    }
-
-    _command_queue->setWorkloadType(zeWorkloadType);
-}
-
 std::mutex& IGraph::get_mutex() {
     return _mutex;
 }

src/plugins/intel_npu/src/compiler_adapter/include/driver_graph.hpp

@@ -32,11 +32,15 @@ class DriverGraph final : public IGraph {
 
     void initialize(const Config& config) override;
 
+    void set_workload_type(const ov::WorkloadType workloadType) override;
+
     ~DriverGraph() override;
 
 private:
     bool release_blob(const Config& config);
 
+    void create_new_command_queue() override;
+
     std::shared_ptr<ZeGraphExtWrappers> _zeGraphExt;
     std::shared_ptr<ZeroInitStructsHolder> _zeroInitStruct;
 

src/plugins/intel_npu/src/compiler_adapter/include/plugin_graph.hpp

@@ -35,9 +35,13 @@ class PluginGraph final : public IGraph {
 
     void initialize(const Config& config) override;
 
+    void set_workload_type(const ov::WorkloadType workloadType) override;
+
     ~PluginGraph() override;
 
 private:
+    void create_new_command_queue() override;
+
     std::shared_ptr<ZeGraphExtWrappers> _zeGraphExt;
     std::shared_ptr<ZeroInitStructsHolder> _zeroInitStruct;
 

src/plugins/intel_npu/src/compiler_adapter/include/ze_graph_ext_wrappers.hpp

@@ -48,7 +48,7 @@ class ZeGraphExtWrappers {
 
     void setGraphArgumentValue(ze_graph_handle_t graphHandle, uint32_t argi_, const void* argv) const;
 
-    void initializeGraph(ze_graph_handle_t graphHandle, const Config& config) const;
+    void initializeGraph(ze_graph_handle_t graphHandle) const;
 
 private:
     std::unordered_set<std::string> getQueryResultFromSupportedLayers(
@@ -60,7 +60,7 @@ class ZeGraphExtWrappers {
                      std::vector<IODescriptor>& inputs,
                      std::vector<IODescriptor>& outputs) const;
 
-    void initialize_graph_through_command_list(ze_graph_handle_t graphHandle, const Config& config) const;
+    void initialize_graph_through_command_list(ze_graph_handle_t graphHandle) const;
 
     std::shared_ptr<ZeroInitStructsHolder> _zeroInitStruct;
     uint32_t _graphExtVersion;