CPU: Always create an MKLDNN CPU engine

src/ngraph/codegen/code_writer.cpp

@@ -19,7 +19,6 @@ using namespace ngraph;
 
 codegen::CodeWriter::CodeWriter()
     : indent(0)
-    , emitted_mkldnn_preamble(false)
     , m_pending_indent(true)
     , m_temporary_name_count(0)
 {

src/ngraph/codegen/code_writer.hpp

@@ -36,7 +36,6 @@ public:
     void operator+=(const std::string&);
 
     size_t indent;
-    bool emitted_mkldnn_preamble;
 
     template <typename T>
     friend CodeWriter& operator<<(CodeWriter& out, const T& obj)

src/ngraph/runtime/cpu/cpu_emitter.cpp

@@ -91,7 +91,6 @@ void runtime::cpu::CPU_Emitter::EmitMKLDNNPreamble(codegen::CodeWriter& writer)
 {
     writer << "using namespace mkldnn;\n";
     writer << "auto cpu_engine = engine(engine::cpu, 0);\n";
-    writer.emitted_mkldnn_preamble = true;
 }
 
 void runtime::cpu::CPU_Emitter::EmitNop(codegen::CodeWriter& writer,
@@ -1854,11 +1853,6 @@ void runtime::cpu::CPU_Emitter::EmitConvolution(codegen::CodeWriter& writer,
         images_dilated = images_dilated || (s != 1);
     }
 
-    if (!writer.emitted_mkldnn_preamble)
-    {
-        EmitMKLDNNPreamble(writer);
-    }
-
     // TODO: MKLDNN streams should be static so we need to either implement
     // codegen for statics or move primitive and stream construction out
     // of the generated function and only generate code to run/rerun the stream
@@ -1983,11 +1977,6 @@ void runtime::cpu::CPU_Emitter::EmitMaxPool(codegen::CodeWriter& writer,
 
     // TODO: Optimize for 1D
 
-    if (!writer.emitted_mkldnn_preamble)
-    {
-        EmitMKLDNNPreamble(writer);
-    }
-
     // TODO: Remove element type restriction
     if (arg_rank == 4 && max_pool->get_window_shape().size() == 2 &&
         args[0].get_element_type() == element::f32)

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -481,6 +481,8 @@ using namespace ngraph::runtime;
             writer << "tbb::flow::graph G;\n\n";
         }
 
+        runtime::cpu::CPU_Emitter::EmitMKLDNNPreamble(writer);
+
         bool temporaries_used = false;
         size_t worst_case_tmp_size = 0;
         for (shared_ptr<Node> node : current_function->get_ordered_ops())