GPU Common Function Elimination for compile speedup (#1065)

* remove comments within body of function

* change to emit each op exactly once

* update code per review comments

src/ngraph/codegen/code_writer.hpp

@@ -68,16 +68,16 @@ public:
 
     std::string generate_temporary_name(std::string prefix = "tempvar");
 
-    void block_begin(std::string block_prefix = "")
+    void block_begin()
     {
-        *this << "{" << block_prefix << "\n";
+        *this << "{\n";
         indent++;
     }
 
-    void block_end(std::string block_suffix = "")
+    void block_end()
     {
         indent--;
-        *this << "}" << block_suffix << "\n";
+        *this << "}\n";
     }
 
 private:

src/ngraph/codegen/compiler.cpp

@@ -265,7 +265,6 @@ void codegen::StaticCompiler::add_header_search_path(const string& p)
     vector<string> paths = split(p, ';');
     for (const string& path : paths)
     {
-        NGRAPH_INFO << path;
         if (!contains(m_extra_search_path_list, path))
         {
             m_extra_search_path_list.push_back(path);

src/ngraph/runtime/gpu/gpu_emitter.hpp

@@ -77,7 +77,7 @@ namespace ngraph
                     auto& cuda_emitter =
                         external_function->get_primitive_emitter()->get_cuda_emitter();
 
-                    writer.block_begin("  // " + node->get_name());
+                    writer.block_begin();
                     {
                         std::vector<std::string> dtypes;
                         for (auto& arg : args)

src/ngraph/runtime/gpu/gpu_external_function.hpp

@@ -83,6 +83,13 @@ namespace ngraph
                     const Node&,
                     const std::unordered_map<descriptor::TensorView*, std::vector<size_t>>&);
                 void release_function() { m_function = nullptr; }
+                std::string emit_op_as_function(const Node& node, const std::string& function_name);
+                std::string strip_comments(const std::string& s) const;
+                bool is_functionally_identical(
+                    const Node& n1,
+                    const Node& n2,
+                    const std::unordered_map<const Node*, std::string>& node_cache) const;
+
                 std::unique_ptr<codegen::Compiler> m_compiler;
                 std::unique_ptr<codegen::ExecutionEngine> m_execution_engine;
                 bool m_emit_timing;