CPU: Use transpose kernel from MKL in 2D Reshape

Also, add more codegen options

src/ngraph/codegen/compiler.cpp

@@ -143,10 +143,17 @@ std::unique_ptr<llvm::Module> execution_state::compile(const string& source, con
     LO->OpenMP = 1;
     LO->OpenMPUseTLS = 1;
 
-    if (debuginfo_enabled)
-    {
     // CodeGen options
     auto& CGO = Clang->getInvocation().getCodeGenOpts();
+    CGO.OptimizationLevel = 3;
+    CGO.RelocationModel = "static";
+    CGO.ThreadModel = "posix";
+    CGO.OmitLeafFramePointer = 1;
+    CGO.VectorizeLoop = 1;
+    CGO.VectorizeSLP = 1;
+
+    if (debuginfo_enabled)
+    {
         CGO.setDebugInfo(codegenoptions::FullDebugInfo);
     }
 
@@ -161,6 +168,8 @@ std::unique_ptr<llvm::Module> execution_state::compile(const string& source, con
     // Enable various target features
     // Most of these are for Eigen
     auto &TO = Clang->getInvocation().getTargetOpts();
+    // TODO: This needs to be configurable and selected carefully
+    TO.CPU = "broadwell";
     TO.FeaturesAsWritten.emplace_back("+sse4.1");
     TO.FeaturesAsWritten.emplace_back("+sse4.2");
     TO.FeaturesAsWritten.emplace_back("+avx");

src/ngraph/runtime/cpu/cpu_kernels.hpp

@@ -86,3 +86,15 @@ namespace cblas
                      const ngraph::element::Int64::type ldc);
     }
 }
+
+namespace mkl
+{
+    extern "C" {
+    void MKL_Somatcopy(char ordering,
+                       char trans,
+                       size_t rows, size_t cols,
+                       const ngraph::element::Float32::type alpha,
+                       const ngraph::element::Float32::type* A, size_t lda,
+                       ngraph::element::Float32::type* B, size_t ldb);
+    }
+}

src/ngraph/runtime/cpu/emitter.cpp

@@ -979,7 +979,25 @@ void Emitter::EMITTER_DECL(EmitReshape)
         auto arg0_layout = inputs[0].get_layout<DenseTensorViewLayout>();
         auto out_layout = outputs[0].get_layout<DenseTensorViewLayout>();
 
-        TU += "    {\n"
+        // Emit an MKL transpose call if possible
+        if (result_element_type == ngraph::element::Float32::element_type())
+        {
+            TU +=
+                "    {\n"
+                "        auto arg0 = call_frame->get_tensor_view_data<" + element_type_names[TI(result_element_type)] +
+                ">(" + to_string(inputs[0].get_index()) + ");\n"
+                "        auto out  = call_frame->get_tensor_view_data<" + element_type_names[TI(result_element_type)] +
+                ">(" + to_string(outputs[0].get_index()) + ");\n"
+                "        mkl::MKL_Somatcopy('R', 'T', " + to_string(arg_shape[0]) + ",\n"
+                "                          " + to_string(arg_shape[1]) + ", 1.0f,\n"
+                "                           arg0, " + to_string(arg_shape[1]) + ",\n"
+                "                           out, " + to_string(arg_shape[0]) + ");\n"
+                "    }\n";
+        }
+        else
+        {
+            TU +=
+                "    {\n"
                 "        auto arg0 = call_frame->get_tensor_view_data<" + element_type_names[TI(result_element_type)] +
                 ">(" + to_string(inputs[0].get_index()) + ");\n"
                 "        auto out  = call_frame->get_tensor_view_data<" + element_type_names[TI(result_element_type)] +
@@ -990,6 +1008,7 @@ void Emitter::EMITTER_DECL(EmitReshape)
                 EIGEN_MATRIX_FORMAT(arg0_layout->get_shape(), arg0_layout->get_strides()) + ").transpose();\n"
                 "    }\n";
         }
+    }
     // Other cases (reordering of axes for tensors with rank>2) are not handled yet.
     else
     {