Add BroadcastLike and ScalarConstantLike to op_tbl (#2213)

* add missing ops to op_tbl

* add two missing ops

src/ngraph/autodiff/adjoints.cpp

@@ -37,7 +37,7 @@ using namespace ngraph;
 
 std::shared_ptr<Node> make_zero(const std::shared_ptr<Node>& node)
 {
-    std::shared_ptr<Node> zero = std::make_shared<op::ScalarConstantLike<double>>(node, 0.0);
+    std::shared_ptr<Node> zero = std::make_shared<op::ScalarConstantLike>(node, 0.0);
     std::shared_ptr<Node> bzero = std::make_shared<op::BroadcastLike>(zero, node, AxisSet{});
     return bzero;
 }

src/ngraph/op/acos.cpp

@@ -50,7 +50,7 @@ void op::Acos::generate_adjoints(autodiff::Adjoints& adjoints, const NodeVector&
 
     auto x = get_inputs().at(0).get_output().get_node();
 
-    auto one = make_shared<op::ScalarConstantLike<double>>(x, 1.0);
+    auto one = make_shared<op::ScalarConstantLike>(x, 1.0);
     auto ones = make_shared<op::BroadcastLike>(one, x, AxisSet());
 
     adjoints.add_delta(x, -delta / make_shared<op::Sqrt>(ones - x * x));

src/ngraph/op/broadcast.cpp

@@ -84,9 +84,9 @@ void op::Broadcast::generate_adjoints(autodiff::Adjoints& adjoints, const NodeVe
 
 op::BroadcastLike::BroadcastLike(const std::shared_ptr<Node>& arg,
                                  const std::shared_ptr<Node>& like_arg,
-                                 const AxisSet& broadcast_axes)
+                                 const AxisSet& initial_broadcast_axes)
     : Broadcast("BroadcastLike", {arg, like_arg}, {}, {})
-    , m_initial_broadcast_axes(broadcast_axes)
+    , m_initial_broadcast_axes(initial_broadcast_axes)
 {
     constructor_validate_and_infer_types();
 }

src/ngraph/op/broadcast.hpp

@@ -70,17 +70,17 @@ namespace ngraph
             ///
             /// \param arg The argument to be broadcast.
             /// \param like_arg Provides the shape for the result.
-            /// \param broadcast_axes indicates which axes will be broadcast. If empty,
-            /// arg must be scalar and all axes are broadcast.
+            /// \param initial_broadcast_axes indicates which axes will be broadcast. If empty,
+            ///        arg must be scalar and all axes are broadcast.
             BroadcastLike(const std::shared_ptr<Node>& arg,
                           const std::shared_ptr<Node>& like_arg,
-                          const AxisSet& broadcast_axes);
+                          const AxisSet& initial_broadcast_axes);
 
             virtual std::shared_ptr<Node>
                 copy_with_new_args(const NodeVector& new_args) const override;
 
             void infer_shape() override;
-
+            const AxisSet& get_initial_broadcast_axes() const { return m_initial_broadcast_axes; }
         protected:
             AxisSet m_initial_broadcast_axes;
         };

src/ngraph/op/constant.cpp

@@ -169,6 +169,21 @@ shared_ptr<op::Constant> op::ScalarConstantLikeBase::as_constant() const
     return std::make_shared<op::Constant>(m_element_type, m_shape, m_data);
 }
 
+std::shared_ptr<Node> op::ScalarConstantLike::copy_with_new_args(const NodeVector& new_args) const
+{
+    return std::make_shared<ScalarConstantLike>(new_args.at(0), m_value);
+}
+
+void op::ScalarConstantLike::infer_element_type()
+{
+    m_element_type = get_input_element_type(0);
+    if (nullptr == m_data)
+    {
+        m_data = ngraph::aligned_alloc(m_element_type.size(), m_element_type.size());
+        write_values(std::vector<double>(1, m_value));
+    }
+}
+
 //
 // We have to open up namespace blocks here to work around a problem with gcc:
 //

src/ngraph/op/constant.hpp

@@ -281,7 +281,6 @@ namespace ngraph
         };
 
         /// \brief A scalar constant whose element type is the same as like.
-        template <typename T>
         class ScalarConstantLike : public ScalarConstantLikeBase
         {
         public:
@@ -292,30 +291,20 @@ namespace ngraph
             ///
             /// \param like A tensor that will supply the element type.
             /// \param value The value of the scalar.
+            template <typename T>
             ScalarConstantLike(const std::shared_ptr<Node>& like, T value)
                 : ScalarConstantLikeBase("ScalarConstantLike", {like})
-                , m_value(value)
+                , m_value(static_cast<double>(value))
             {
                 constructor_validate_and_infer_types();
             }
 
-            std::shared_ptr<Node> copy_with_new_args(const NodeVector& new_args) const override
-            {
-                return std::make_shared<ScalarConstantLike<T>>(new_args.at(0), m_value);
-            }
+            std::shared_ptr<Node> copy_with_new_args(const NodeVector& new_args) const override;
 
         protected:
-            void infer_element_type() override
-            {
-                m_element_type = get_input_element_type(0);
-                if (nullptr == m_data)
-                {
-                    m_data = ngraph::aligned_alloc(m_element_type.size(), m_element_type.size());
-                    write_values(std::vector<T>(1, m_value));
-                }
-            }
+            void infer_element_type() override;
 
-            T m_value;
+            double m_value;
         };
     }
 }

src/ngraph/op/op_tbl.hpp

@@ -62,6 +62,7 @@ NGRAPH_OP(BatchNormInference, ngraph::op)
 NGRAPH_OP(BatchNormTraining, ngraph::op)
 NGRAPH_OP(BatchNormTrainingBackprop, ngraph::op)
 NGRAPH_OP(Broadcast, ngraph::op)
+NGRAPH_OP(BroadcastLike, ngraph::op)
 NGRAPH_OP(Ceiling, ngraph::op)
 NGRAPH_OP(Concat, ngraph::op)
 NGRAPH_OP(Constant, ngraph::op)
@@ -112,6 +113,7 @@ NGRAPH_OP(Reshape, ngraph::op)
 NGRAPH_OP(Result, ngraph::op)
 NGRAPH_OP(Reverse, ngraph::op)
 NGRAPH_OP(ReverseSequence, ngraph::op)
+NGRAPH_OP(ScalarConstantLike, ngraph::op)
 NGRAPH_OP(Select, ngraph::op)
 NGRAPH_OP(SelectAndScatter, ngraph::op)
 NGRAPH_OP(ShapeOf, ngraph::op)

src/ngraph/runtime/gpu/gpu_emitter.cpp

@@ -460,6 +460,11 @@ void runtime::gpu::GPU_Emitter::emit_Broadcast(EMIT_ARGS)
     writer.block_end();
 }
 
+void runtime::gpu::GPU_Emitter::emit_BroadcastLike(EMIT_ARGS)
+{
+    throw unsupported_op("Unsupported op '" + node->description() + "'");
+}
+
 void runtime::gpu::GPU_Emitter::emit_Ceiling(EMIT_ARGS)
 {
     emit_elementwise<ngraph::op::Ceiling>(external_function, writer, node, args, out);
@@ -1415,6 +1420,11 @@ void runtime::gpu::GPU_Emitter::emit_Rnn(EMIT_ARGS)
 }
 #endif
 
+void runtime::gpu::GPU_Emitter::emit_ScalarConstantLike(EMIT_ARGS)
+{
+    throw unsupported_op("Unsupported op '" + node->description() + "'");
+}
+
 void runtime::gpu::GPU_Emitter::emit_Select(EMIT_ARGS)
 {
     emit_elementwise<ngraph::op::Select>(external_function, writer, node, args, out);

src/ngraph/runtime/intelgpu/intelgpu_backend.cpp

@@ -1716,6 +1716,7 @@ runtime::Handle runtime::intelgpu::IntelGPUBackend::compile(shared_ptr<Function>
         case OP_TYPEID::All:
         case OP_TYPEID::AllReduce:
         case OP_TYPEID::Any:
+        case OP_TYPEID::BroadcastLike:
         case OP_TYPEID::FunctionCall:
         case OP_TYPEID::Dequantize:
         case OP_TYPEID::Quantize:
@@ -1723,6 +1724,7 @@ runtime::Handle runtime::intelgpu::IntelGPUBackend::compile(shared_ptr<Function>
         case OP_TYPEID::ReplaceSlice:
         case OP_TYPEID::GenerateMask:
         case OP_TYPEID::ReverseSequence:
+        case OP_TYPEID::ScalarConstantLike:
         case OP_TYPEID::SelectAndScatter:
         case OP_TYPEID::ShapeOf:
         case OP_TYPEID::StopGradient:

src/ngraph/runtime/interpreter/int_backend.hpp

@@ -459,6 +459,7 @@ private:
                                     broadcast_axes);
             break;
         }
+        case OP_TYPEID::BroadcastLike: break;
         case OP_TYPEID::Ceiling:
         {
             size_t element_count = shape_size(node.get_output_shape(0));
@@ -488,6 +489,7 @@ private:
             // Constant is handled in the main loop
             break;
         }
+        case OP_TYPEID::ScalarConstantLike: break;
         case OP_TYPEID::Convert:
         {
             // const op::Convert* c = static_cast<const op::Convert*>(&node);

src/ngraph/serializer.cpp

@@ -573,6 +573,12 @@ static shared_ptr<ngraph::Function>
                 node = make_shared<op::Broadcast>(args[0], shape, axes);
                 break;
             }
+            case OP_TYPEID::BroadcastLike:
+            {
+                auto initial_axes = node_js.at("initial_axes").get<set<size_t>>();
+                node = make_shared<op::BroadcastLike>(args[0], args[1], initial_axes);
+                break;
+            }
             case OP_TYPEID::Ceiling:
             {
                 node = make_shared<op::Ceiling>(args[0]);
@@ -1026,6 +1032,12 @@ static shared_ptr<ngraph::Function>
                     make_shared<op::ReverseSequence>(args[0], args[1], batch_axis, sequence_axis);
                 break;
             }
+            case OP_TYPEID::ScalarConstantLike:
+            {
+                double value = node_js.at("value").get<double>();
+                node = make_shared<op::ScalarConstantLike>(args[0], value);
+                break;
+            }
             case OP_TYPEID::Select:
             {
                 node = make_shared<op::Select>(args[0], args[1], args[2]);
@@ -1338,6 +1350,12 @@ static json write(const Node& n, bool binary_constant_data)
         node["shape"] = tmp->get_broadcast_shape();
         break;
     }
+    case OP_TYPEID::BroadcastLike:
+    {
+        auto tmp = dynamic_cast<const op::BroadcastLike*>(&n);
+        node["initial_axes"] = tmp->get_initial_broadcast_axes();
+        break;
+    }
     case OP_TYPEID::Ceiling: { break;
     }
     case OP_TYPEID::Concat:
@@ -1593,6 +1611,14 @@ static json write(const Node& n, bool binary_constant_data)
         node["sequence_axis"] = tmp->get_sequence_axis();
         break;
     }
+    case OP_TYPEID::ScalarConstantLike:
+    {
+        auto tmp = dynamic_cast<const op::ScalarConstantLikeBase*>(&n);
+        auto constant = tmp->as_constant();
+        node["value"] = constant->get_value_strings()[0];
+        node["element_type"] = write_element_type(constant->get_element_type());
+        break;
+    }
     case OP_TYPEID::Select: { break;
     }
     case OP_TYPEID::SelectAndScatter: