CPU: Implement scalar-tensor and vector dot product

Also make Eigen format construction explicit

src/ngraph/runtime/cpu/emitter.cpp

@@ -44,7 +44,7 @@ static unordered_map<type_index, string> element_type_names = {{TI(ngraph::eleme
                                                               };
 
 
-#define EIGEN_VECTOR_FORMAT(x) "{" + to_string(x) + "}"
+#define EIGEN_VECTOR_FORMAT(x) "fmt::V{" + to_string(x) + "}"
 
 static std::string EIGEN_MATRIX_FORMAT(const ngraph::Shape& shape,
                                        const ngraph::Strides& strides)
@@ -54,7 +54,7 @@ static std::string EIGEN_MATRIX_FORMAT(const ngraph::Shape& shape,
     {
         if (!i)
         {
-            I += "{" + to_string(shape[i]);
+            I += "fmt::M{{" + to_string(shape[i]);
         }
         else
         {
@@ -73,7 +73,7 @@ static std::string EIGEN_MATRIX_FORMAT(const ngraph::Shape& shape,
             I += ", " + to_string(strides[i]);
         }
     }
-    I += "}";
+    I += "}}";
     return I;
 }
 
@@ -102,6 +102,60 @@ void Emitter::EMITTER_DECL(EmitAdd)
 
 void Emitter::EMITTER_DECL(EmitDot)
 {
+    auto& arg_nodes = n->get_arguments();
+    assert(arg_nodes.size() == 2);
+
+    auto arg0_tensor_type =
+        dynamic_pointer_cast<const TensorViewType>(arg_nodes.at(0)->get_value_type());
+    assert(nullptr != arg0_tensor_type);
+
+    auto arg1_tensor_type =
+        dynamic_pointer_cast<const TensorViewType>(arg_nodes.at(1)->get_value_type());
+    assert(nullptr != arg1_tensor_type);
+
+    auto arg0_shape = arg0_tensor_type->get_shape();
+    auto arg1_shape = arg1_tensor_type->get_shape();
+    auto& arg0_element_type = arg0_tensor_type->get_element_type();
+
+    if (arg0_shape.empty() || arg1_shape.empty())
+    {
+        auto& first = (arg0_shape.empty() ? inputs[0] : inputs[1]);
+        auto& second = (arg0_shape.empty() ? inputs[1] : inputs[0]);
+
+        TU += "    {\n";
+              "        auto arg1 = call_frame->get_tensor_view_data<" + element_type_names[TI(arg0_element_type)] +
+                       ">(" + to_string(second.get_index()) + ");\n"
+              "        auto out  = call_frame->get_tensor_view_data<" + element_type_names[TI(arg0_element_type)] +
+                       ">(" + to_string(outputs[0].get_index()) + ");\n"
+              "        EigenVector<" + element_type_names[TI(arg0_element_type)] +
+                       ">(out, " EIGEN_VECTOR_FORMAT(outputs[0].get_layout<DenseTensorViewLayout>()->get_size()) ") = "
+              "call_frame->get_tensor_view_data<" + element_type_names[TI(arg0_element_type)] +
+                       ">(" + to_string(first.get_index()) + ")[0] * EigenVector<" +
+                       element_type_names[TI(arg0_element_type)] +
+                       ">(arg1, " EIGEN_VECTOR_FORMAT(second.get_layout<DenseTensorViewLayout>()->get_size()) ");\n"
+              "    }\n";
+    }
+    else if ((arg0_shape.size() == 1) && (arg1_shape.size() == 1))
+    {
+        TU += "    {\n"
+              "        auto arg0 = call_frame->get_tensor_view_data<" + element_type_names[TI(arg0_element_type)] + ">(" +
+                       to_string(inputs[0].get_index()) + ");\n"
+              "        auto arg1 = call_frame->get_tensor_view_data<" + element_type_names[TI(arg0_element_type)] + ">(" +
+                       to_string(inputs[1].get_index()) + ");\n"
+              "        auto out  = call_frame->get_tensor_view_data<" + element_type_names[TI(arg0_element_type)] + ">(" +
+                       to_string(outputs[0].get_index()) + ");\n"
+              "        EigenVector<" + element_type_names[TI(arg0_element_type)] + ">(out, "
+                       EIGEN_VECTOR_FORMAT(outputs[0].get_layout<DenseTensorViewLayout>()->get_size()) ") << \n"
+              "        EigenVector<" + element_type_names[TI(arg0_element_type)] + ">(arg0, "
+                       EIGEN_VECTOR_FORMAT(inputs[0].get_layout<DenseTensorViewLayout>()->get_size()) ").dot("
+              "EigenVector<" + element_type_names[TI(arg0_element_type)] + ">(arg1, "
+                       EIGEN_VECTOR_FORMAT(inputs[1].get_layout<DenseTensorViewLayout>()->get_size()) "));\n"
+              "    }\n";
+    }
+    else
+    {
+        throw ngraph_error("Dot product for given tensors unimplemented");
+    }
 }
 
 void Emitter::EMITTER_DECL(EmitMultiply)
@@ -210,8 +264,8 @@ void Emitter::EMITTER_DECL(EmitConcat)
         TU += "    {\n"
               "        auto out  = call_frame->get_tensor_view_data<" + element_type_names[TI(result_element_type)] +
                                    ">(" + to_string(outputs[0].get_index()) + ");\n"
-              "        EigenMatrix<" + element_type_names[TI(result_element_type)] + "> out_matrix(out, {" +
-                       EIGEN_MATRIX_FORMAT(out_layout->get_shape(), out_layout->get_strides()) + "});\n";
+              "        EigenMatrix<" + element_type_names[TI(result_element_type)] + "> out_matrix(out, " +
+                       EIGEN_MATRIX_FORMAT(out_layout->get_shape(), out_layout->get_strides()) + ");\n";
 
         size_t concat_pos[2]{0, 0};
         for (size_t i = 0; i < inputs.size(); i++)
@@ -224,8 +278,8 @@ void Emitter::EMITTER_DECL(EmitConcat)
                   to_string(arg_shape.at(1)) + ") << "
                   "EigenMatrix<" + element_type_names[TI(result_element_type)] + ">(call_frame->"
                   "get_tensor_view_data<" + element_type_names[TI(result_element_type)] + ">(" +
-                  to_string(inputs[i].get_index()) + "), {" +
-                  EIGEN_MATRIX_FORMAT(arg_layout->get_shape(), arg_layout->get_strides()) + "});\n";
+                  to_string(inputs[i].get_index()) + "), " +
+                  EIGEN_MATRIX_FORMAT(arg_layout->get_shape(), arg_layout->get_strides()) + ");\n";
 
             concat_pos[axis] += arg_shape.at(axis);
         }

test/cpu.cpp

@@ -442,8 +442,7 @@ TEST(cpu, equal)
     ASSERT_EQ((vector<char>{1, 1, 0, 0, 0, 1, 1, 0}), result->get_vector());
 }
 
-/*
-TEST(execute, dot_0_0)
+TEST(cpu, dot_0_0)
 {
     auto shape = Shape{0};
     auto A = make_shared<op::Parameter>(element::Float32::element_type(), shape);
@@ -452,7 +451,7 @@ TEST(execute, dot_0_0)
     auto rt = make_shared<TensorViewType>(element::Float32::element_type(), Shape{});
     auto f = make_shared<Function>(make_shared<op::Dot>(A, B), rt, op::Parameters{A, B});
 
-    auto manager = runtime::Manager::get("NGVM");
+    auto manager = runtime::Manager::get("CPU");
     auto external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
@@ -468,6 +467,7 @@ TEST(execute, dot_0_0)
     ASSERT_EQ((vector<float>{0}), result->get_vector());
 }
 
+/*
 TEST(execute, dot_matrix_2x0_0x2)
 {
     auto shape_a = Shape{2, 0};