Fix the cpu thread_safe_calls test on Windows. (#2861)

* Fix the cpu thread_safe_calls test on Windows.

* Use notify_one.

src/ngraph/runtime/cpu/cpu_call_frame.cpp

@@ -112,23 +112,21 @@ void runtime::cpu::CPU_CallFrame::inner_call(
                          m_ctx_vec[id]->op_durations,
                          m_external_function->get_function_name() + ".timeline.json");
     }
-    std::unique_lock<std::mutex> lck(m_mutex);
-    m_id_pool[id] = true;
-    m_num_ctx_available++;
-    m_cv.notify_all();
 }
 
 void runtime::cpu::CPU_CallFrame::call(
     const std::vector<std::shared_ptr<runtime::Tensor>>& output_tvs,
     const std::vector<std::shared_ptr<runtime::Tensor>>& input_tvs)
 {
+    auto id = 0;
+    auto disable_caching = false;
+    {
         std::unique_lock<std::mutex> lck(m_mutex);
         while (m_num_ctx_available == 0)
         {
             m_cv.wait(lck);
         }
 
-    auto id = 0;
         for (auto i = 0; i < m_num_ctx; i++)
         {
             if (m_id_pool[i])
@@ -139,7 +137,6 @@ void runtime::cpu::CPU_CallFrame::call(
         }
         NGRAPH_CHECK(id != m_num_ctx);
         m_id_pool[id] = false;
-    auto disable_caching = false;
         if (id != m_prev_ctx)
         {
             // Disable caching since staleness hints are no longer
@@ -148,11 +145,17 @@ void runtime::cpu::CPU_CallFrame::call(
         }
         m_prev_ctx = id;
         m_num_ctx_available--;
-    m_mutex.unlock();
+    }
 
     m_ctx_vec[id]->pc = 0;
     propagate_layouts(output_tvs, m_external_function->get_result_layout_descriptors());
     inner_call(output_tvs, input_tvs, id, disable_caching);
+
+    m_mutex.lock();
+    m_id_pool[id] = true;
+    m_num_ctx_available++;
+    m_mutex.unlock();
+    m_cv.notify_one();
 }
 
 void runtime::cpu::CPU_CallFrame::propagate_layouts(

test/cpu_test.cpp

@@ -951,8 +951,6 @@ TEST(cpu_test, rotated_pooling)
         make_f(false, false), make_f(false, false), "INTERPRETER", "CPU"); // 5D MaxPool
 }
 
-// TODO enable test for Windows
-#ifndef _WIN32
 // for float this will be 18 bits matching
 // for bfloat this will be 6 bits matching
 constexpr int three_quarters_of_available_bits = (MAX_FLOAT_BITS * 3) / 4;
@@ -1049,7 +1047,6 @@ TEST(cpu_test, thread_safe_calls_convolution_2d_2items)
 
     unset_environment("NGRAPH_CPU_CONCURRENCY");
 }
-#endif
 
 TEST(cpu_test, constant_reshape)
 {