[IE PYTHON] fix gil (#8068)

* remove nogil

* Add test

inference-engine/ie_bridges/python/src/openvino/inference_engine/ie_api.pyx

@@ -544,9 +544,7 @@ cdef class IECore:
     # If there are more than one device of a specific type, they all are listed followed by a dot and a number.
     @property
     def available_devices(self):
-        cdef vector[string] c_devices
+        cdef vector[string] c_devices = self.impl.getAvailableDevices()
-        with nogil:
-            c_devices = self.impl.getAvailableDevices()
         return [d.decode() for d in c_devices]
 
 ## This structure stores info about pre-processing of network inputs (scale, mean image, ...)
@@ -925,14 +923,11 @@ cdef class ExecutableNetwork:
     ## A tuple of `InferRequest` instances
     @property
     def requests(self):
-        cdef size_t c_infer_requests_size
+        cdef size_t c_infer_requests_size = deref(self.impl).infer_requests.size()
-        with nogil:
-            c_infer_requests_size = deref(self.impl).infer_requests.size()
         if len(self._infer_requests) == 0:
             for i in range(c_infer_requests_size):
                 infer_request = InferRequest()
-                with nogil:
+                infer_request.impl = &(deref(self.impl).infer_requests[i])
-                    infer_request.impl = &(deref(self.impl).infer_requests[i])
                 infer_request._inputs_list = list(self.input_info.keys())
                 infer_request._outputs_list = list(self.outputs.keys())
                 for input_name in infer_request._inputs_list:
@@ -1052,10 +1047,7 @@ cdef class ExecutableNetwork:
     ## Get idle request ID
     #  @return Request index
     cpdef get_idle_request_id(self):
-        cdef int request_id
+        return deref(self.impl).getIdleRequestId()
-        with nogil:
-            request_id = deref(self.impl).getIdleRequestId()
-        return request_id
 
 ctypedef extern void (*cb_type)(void*, int) with gil
 
@@ -1197,8 +1189,7 @@ cdef class InferRequest:
     cpdef infer(self, inputs=None):
         if inputs is not None:
             self._fill_inputs(inputs)
-        with nogil:
+        deref(self.impl).infer()
-            deref(self.impl).infer()
 
     ## Starts asynchronous inference of the infer request and fill outputs array
     #
@@ -1215,8 +1206,7 @@ cdef class InferRequest:
     cpdef async_infer(self, inputs=None):
         if inputs is not None:
             self._fill_inputs(inputs)
-        with nogil:
+        deref(self.impl).infer_async()
-            deref(self.impl).infer_async()
 
     ## Waits for the result to become available. Blocks until specified timeout elapses or the result
     #  becomes available, whichever comes first.
@@ -1337,8 +1327,7 @@ cdef class IENetwork:
     def __cinit__(self, model = None):
         # Try to create Inference Engine network from capsule
         if model is not None:
-            with nogil:
+            self.impl = C.IENetwork(model)
-                self.impl = C.IENetwork(model)
         else:
             with nogil:
                 self.impl = C.IENetwork()
@@ -1352,9 +1341,7 @@ cdef class IENetwork:
     ## A dictionary that maps input layer names to InputInfoPtr objects.
     @property
     def input_info(self):
-        cdef map[string, C.InputInfo.Ptr] c_inputs
+        cdef map[string, C.InputInfo.Ptr] c_inputs = self.impl.getInputsInfo()
-        with nogil:
-            c_inputs = self.impl.getInputsInfo()
         inputs = {}
         cdef InputInfoPtr input_info_ptr
         for input in c_inputs:
@@ -1367,9 +1354,7 @@ cdef class IENetwork:
     ## A dictionary that maps output layer names to DataPtr objects
     @property
     def outputs(self):
-        cdef map[string, C.DataPtr] c_outputs
+        cdef map[string, C.DataPtr] c_outputs = self.impl.getOutputs()
-        with nogil:
-            c_outputs = self.impl.getOutputs()
         outputs = {}
         cdef DataPtr data_ptr
         for output in c_outputs:

inference-engine/ie_bridges/python/src/openvino/inference_engine/ie_api_impl_defs.pxd

@@ -162,18 +162,18 @@ cdef extern from "ie_api_impl.hpp" namespace "InferenceEnginePython":
         object getMetric(const string & metric_name) except +
         object getConfig(const string & metric_name) except +
         int wait(int num_requests, int64_t timeout) nogil
-        int getIdleRequestId() nogil
+        int getIdleRequestId()
         shared_ptr[CExecutableNetwork] getPluginLink() except +
 
     cdef cppclass IENetwork:
         IENetwork() nogil except +
-        IENetwork(object) nogil except +
+        IENetwork(object) except +
         string name
         size_t batch_size
         string precision
         map[string, vector[size_t]] inputs
-        const map[string, InputInfo.Ptr] getInputsInfo() nogil except +
+        const map[string, InputInfo.Ptr] getInputsInfo() except +
-        map[string, DataPtr] getOutputs() nogil except +
+        map[string, DataPtr] getOutputs() except +
         void addOutput(string &, size_t) except +
         void setAffinity(map[string, string] & types_affinity_map, map[string, string] & layers_affinity_map) except +
         void setBatch(size_t size) except +
@@ -193,8 +193,8 @@ cdef extern from "ie_api_impl.hpp" namespace "InferenceEnginePython":
         void setBlob(const string &blob_name, const CBlob.Ptr &blob_ptr, CPreProcessInfo& info) except +
         const CPreProcessInfo& getPreProcess(const string& blob_name) except +
         map[string, ProfileInfo] getPerformanceCounts() except +
-        void infer() nogil except +
+        void infer() except +
-        void infer_async() nogil except +
+        void infer_async() except +
         int wait(int64_t timeout) nogil except +
         void setBatch(int size) except +
         void setCyCallback(void (*)(void*, int), void *) except +
@@ -219,7 +219,7 @@ cdef extern from "ie_api_impl.hpp" namespace "InferenceEnginePython":
         void unregisterPlugin(const string & deviceName) except +
         void registerPlugins(const string & xmlConfigFile) except +
         void addExtension(const string & ext_lib_path, const string & deviceName) except +
-        vector[string] getAvailableDevices() nogil except +
+        vector[string] getAvailableDevices() except +
         object getMetric(const string & deviceName, const string & name) except +
         object getConfig(const string & deviceName, const string & name) except +
 

inference-engine/ie_bridges/python/tests/test_IECore.py

@@ -5,7 +5,7 @@ import os
 import pytest
 from sys import platform
 from pathlib import Path
-from threading import Thread
+from threading import Event, Thread
 from time import sleep, time
 from queue import Queue
 
@@ -274,3 +274,52 @@ def test_load_network_release_gil(device):
     # Assert there were never any long gil locks
     assert message_queue.qsize() == 0, \
         f"More than 0 GIL locks occured! Latency: {message_queue.get()})"
+
+
+def test_nogil_safe(device):
+    call_thread_func = Event()
+    core = IECore()
+    net = core.read_network(model=test_net_xml, weights=test_net_bin)
+
+    def thread_target(thread_func, thread_args):
+        call_thread_func.wait()
+        call_thread_func.clear()
+        thread_func(*thread_args)
+
+    def main_thread_target(gil_release_func, args):
+        call_thread_func.set()
+        gil_release_func(*args)
+        assert not call_thread_func.is_set()
+
+    def test_run_parallel(gil_release_func, args, thread_func, thread_args):
+        thread = Thread(target=thread_target, args=[thread_func, thread_args])
+        thread.start()
+        main_thread_target(gil_release_func, args)
+        thread.join()
+
+    main_targets = [{
+                     core.read_network: [test_net_xml, test_net_bin],
+                     core.load_network: [net, device],
+                    },
+                    {
+                     core.load_network: [net, device],
+                    }]
+
+    thread_targets = [{
+                       core.get_versions: [device,],
+                       core.read_network: [test_net_xml, test_net_bin],
+                       core.load_network: [net, device],
+                       core.query_network: [net, device],
+                       getattr: [core, "available_devices"],
+                      },
+                      {
+                       getattr: [net, "name"],
+                       getattr: [net, "input_info"],
+                       getattr: [net, "outputs"],
+                       getattr: [net, "batch_size"],
+                      }]
+
+    for main_target, custom_target in zip(main_targets, thread_targets):
+        for nogil_func, args in main_target.items():
+            for thread_func, thread_args in custom_target.items():
+                test_run_parallel(nogil_func, args, thread_func, thread_args)