Do not set num_threads_per_pool to 1 in cpu executor. (#2648)

* Do not set num_threads_per_pool to 1 in cpu executor.
Ignore cacheability in memory assignment when not reusing memory.

* Fix debug output.

src/ngraph/runtime/cpu/cpu_executor.cpp

@@ -68,13 +68,11 @@ namespace ngraph
                     for (int i = 0; i < num_thread_pools; i++)
                     {
                         int num_threads_per_pool;
-#if defined(EIGEN_OPENMP)
+
                         // Eigen threadpool will still be used for reductions
                         // and other tensor operations that dont use a parallelFor
-                        num_threads_per_pool = 1;
-#else
                         num_threads_per_pool = GetNumCores();
-#endif
+
                         // User override
                         char* eigen_tp_count = std::getenv("NGRAPH_CPU_EIGEN_THREAD_COUNT");
                         if (eigen_tp_count != nullptr)

src/ngraph/runtime/cpu/pass/cpu_memory_assignment.cpp

@@ -712,21 +712,16 @@ bool runtime::cpu::pass::CPUMemoryAssignment::run_on_function(shared_ptr<ngraph:
                             auto input_op = std::static_pointer_cast<op::Op>(input_node);
                             if (auto input_op_annotations = input_op->get_op_annotations())
                             {
-                                // when input is cacheable, do not allow destructive oi
-                                if (input_op_annotations->is_cacheable())
-                                {
-                                    NGRAPH_DEBUG << "cpu_memory_assignment: cacheable input, no "
-                                                    "destructive oi";
-                                    continue;
-                                }
                                 // when reusing memory, ops with different cacheabilities are using different memory manager
                                 // and should not share the same buffer.
-                                else if (!m_disable_memory_sharing &&
-                                         op_annotations->is_cacheable())
+                                if (!m_disable_memory_sharing &&
+                                    input_op_annotations->is_cacheable() !=
+                                        op_annotations->is_cacheable())
                                 {
-                                    NGRAPH_DEBUG << "cpu_memory_assignment: reusing memory with "
-                                                    "non-cacheable input and cacheable output, no "
-                                                    "destructive oi";
+                                    NGRAPH_DEBUG
+                                        << "cpu_memory_assignment: reusing memory with "
+                                           "input and output have different cacheabilities, no "
+                                           "destructive oi";
                                     continue;
                                 }
                             }