Ayzhuang/fix tf maxpool segfault migrate (#2909)

* remove throw from dtor (#2854)

* Migrate from master (#2858)

* Migrate from master. (#2866)

* Add build option NGRAPH_ENABLE_CPU_CONV_AUTO to enable/disable mkldnn convolution_auto for CPU backend. (#2868)

* Fix a bug in mkldnn emitter. (#2878)

* Fix data type for conv builder (#2885)

* Fix data type for conv builder

* Added support for other data types

* Migrate #2889 from master (#2891)

* Migrate from #2884 (#2897)

* Migrate #2864 (#2901)

* Migrate #2864

* Remove extra changes

* Copy PR #2894 to r0.19 branch (#2903)

src/ngraph/runtime/cpu/mkldnn_emitter.cpp

@@ -788,8 +788,8 @@ void MKLDNNEmitter::build_max_pooling_backward(std::vector<mkldnn::primitive*>&
 
     mkldnn::pooling_forward::primitive_desc fwd_pd{fwd_pool_desc, executor::global_cpu_engine};
 
-    size_t ws_index = m_primitive_deps[fwd_pool_index][2];
-    build_memory_primitive(fwd_pd.workspace_primitive_desc().desc(), ws_index);
+    size_t ws_index = fdeps[2];
+    build_memory_primitive(mkldnn_primitives, fwd_pd.workspace_primitive_desc().desc(), ws_index);
     bdeps[1] = ws_index;
 
     // Allocate workspace