Remove heap allocations during execution (#1583)

* elementwise updated

* add support for constructing any order arg list

* add comment for using node_names

* use array instead of vector for args/output

* fix per review comment

* remove dead code

src/ngraph/runtime/gpu/gpu_emitter.hpp

@@ -87,18 +87,22 @@ namespace ngraph
                         dtypes.push_back(out[0].get_type());
                         auto ew_index =
                             cuda_emitter->build_elementwise<T>(dtypes, out[0].get_shape());
-                        writer << "gpu::invoke_primitive(ctx, " << ew_index << ", ";
-                        writer << "std::vector<void*>{" << args.front().get_name();
-                        for (size_t i = 1; i < args.size(); i++)
-                        {
-                            writer << ", " << args[i].get_name();
-                        }
-                        writer << "}.data(), ";
-                        writer << "std::vector<void*>{" << out[0].get_name() << "}.data()";
-                        writer << ");\n";
+                        writer << "void* input[] = {" << node_names(args) << "};\n";
+                        writer << "void* output[] = {" << node_names(out) << "};\n";
+                        writer << "gpu::invoke_primitive(ctx, " << ew_index
+                               << ", input, output);\n";
                     }
                     writer.block_end();
                 }
+
+            private:
+                /// \brief Create a list of node names for each arg in args
+                /// \param args list of tensor arguments
+                /// \param arg_indexes a list of indexes into args for which args to include in
+                ///    the output list, so {1, 2} will include args 1 and 2 and skip 0.
+                /// \ return returns a string containing "arg0_name, arg1_name, etc."
+                static std::string node_names(const std::vector<GPU_TensorViewWrapper>& args,
+                                              std::initializer_list<int> arg_indexes = {});
             };
             Shape get_padded_shape(const Shape& input_shape,
                                    const Shape& padding_below,

src/ngraph/runtime/gpu/gpu_external_function.cpp

@@ -455,8 +455,9 @@ void runtime::gpu::GPU_ExternalFunction::emit_temp_mem_pool_allocation(
     {
         m_writer << "// Allocate the memory pool\n";
         // TODO memory pool malloc.
-        m_writer << "void* pool_base_ptr = ngraph::runtime::gpu::invoke_memory_primitive(ctx, "
-                 << m_tensor_memory_buffers->at(current_function->get_name()) << ");\n";
+        m_writer
+            << "char* pool_base_ptr = (char*)ngraph::runtime::gpu::invoke_memory_primitive(ctx, "
+            << m_tensor_memory_buffers->at(current_function->get_name()) << ");\n";
 
         // Add temporaries to the variable name map
         for (shared_ptr<Node> node : m_function_ordered_ops.at(current_function))
@@ -464,8 +465,8 @@ void runtime::gpu::GPU_ExternalFunction::emit_temp_mem_pool_allocation(
             for (descriptor::Tensor* tensor : node->liveness_new_list)
             {
                 stringstream ss;
-                ss << "((" << tensor->get_element_type().c_type_string()
-                   << "*)((char *)pool_base_ptr + " << tensor->get_pool_offset() << "))";
+                ss << "((" << tensor->get_element_type().c_type_string() << "*)(pool_base_ptr + "
+                   << tensor->get_pool_offset() << "))";
                 m_variable_name_map[tensor->get_name()] = ss.str();
             }
         }