Move to TBB2019 and bug fix to capture functor (#1917)

* Move to TBB2019 and bug fix to capture functor

* Change to use TBB release tag

* remove lightweight from codegen

* Enable TBB flow graph tracing

CMakeLists.txt

@@ -68,7 +68,7 @@ else()
 endif()
 
 # Prevent Eigen from using any LGPL3 code
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DEIGEN_MPL2_ONLY")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DEIGEN_MPL2_ONLY -DTBB_USE_THREADING_TOOLS")
 
 if($ENV{NGRAPH_USE_PREBUILT_LLVM})
     set(NGRAPH_USE_PREBUILT_LLVM TRUE)

cmake/external_tbb.cmake

@@ -20,7 +20,7 @@
 
 if(NGRAPH_TBB_ENABLE)
     set(TBB_GIT_REPO_URL https://github.com/01org/tbb)
-    set(TBB_GIT_TAG "tbb_2018")
+    set(TBB_GIT_TAG "2019_U1")
 
     configure_file(${CMAKE_SOURCE_DIR}/cmake/tbb_fetch.cmake.in ${CMAKE_CURRENT_BINARY_DIR}/tbb/CMakeLists.txt)
     execute_process(COMMAND "${CMAKE_COMMAND}" -G "${CMAKE_GENERATOR}" .

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -23,14 +23,7 @@
 #include <typeinfo>
 #include <unordered_map>
 
-// Kill clang diagnostics bug
-#pragma clang diagnostic push
-#pragma clang diagnostic ignored "-Wreserved-id-macro"
-
-#undef __TBB_PREVIEW_LIGHTWEIGHT_POLICY
-#define __TBB_PREVIEW_LIGHTWEIGHT_POLICY 1
-
-#pragma clang diagnostic pop
+#define TBB_PREVIEW_FLOW_GRAPH_TRACE 1
 
 #include <tbb/flow_graph.h>
 
@@ -425,8 +418,6 @@ void runtime::cpu::CPU_ExternalFunction::compile()
                 "CPU Backend: Tracing and performance breakdowns might not be accurate with TBB "
                 "enabled due to concurrent graph execution");
         }
-        writer << "#undef __TBB_PREVIEW_LIGHTWEIGHT_POLICY \n";
-        writer << "#define __TBB_PREVIEW_LIGHTWEIGHT_POLICY 1\n";
         writer << "#include <tbb/flow_graph.h>";
     }
 
@@ -674,10 +665,9 @@ using namespace ngraph::runtime;
             writer << "\n";
             writer << "if (ctx->first_iteration) {\n";
             writer.indent++;
-            writer << "tbb::flow::continue_node<tbb::flow::continue_msg, tbb::flow::lightweight>* "
+            writer << "tbb::flow::continue_node<tbb::flow::continue_msg>* "
                       "flowgraph_node_start"
-                   << " = new tbb::flow::continue_node<tbb::flow::continue_msg, "
-                      "tbb::flow::lightweight>"
+                   << " = new tbb::flow::continue_node<tbb::flow::continue_msg> "
                       "(*(ctx->G), [&](const tbb::flow::continue_msg &msg)\n{});\n";
         }
 
@@ -774,12 +764,10 @@ using namespace ngraph::runtime;
                 }
                 if (m_use_tbb)
                 {
-                    writer << "tbb::flow::continue_node<tbb::flow::continue_msg, "
-                              "tbb::flow::lightweight>* "
+                    writer << "tbb::flow::continue_node<tbb::flow::continue_msg>* "
                               "flowgraph_node_"
                            << node->get_name()
-                           << " = new tbb::flow::continue_node<tbb::flow::continue_msg, "
-                              "tbb::flow::lightweight>"
+                           << " = new tbb::flow::continue_node<tbb::flow::continue_msg> "
                               "(*(ctx->G), [&](const tbb::flow::continue_msg &msg)\n{\n";
                     writer.indent++;
                 }
@@ -941,8 +929,8 @@ using namespace ngraph::runtime;
             writer << "}\n";
 
             // Execute the flow graph
-            writer << "(static_cast<tbb::flow::continue_node<tbb::flow::continue_msg, "
-                      "tbb::flow::lightweight>*>(&(*(ctx->G->begin()))))"
+            writer << "(static_cast<tbb::flow::continue_node<tbb::flow::continue_msg>*>"
+                      "(&(*(ctx->G->begin()))))"
                    << "->try_put(tbb::flow::continue_msg());\n";
             writer << "try { ctx->G->wait_for_all(); } catch(...) { throw; }\n";
         }
@@ -1631,22 +1619,18 @@ void runtime::cpu::CPU_ExternalFunction::build()
             // Build the flow graph
             if (ctx->first_iteration)
             {
-                std::unordered_map<
-                    std::string,
-                    tbb::flow::continue_node<tbb::flow::continue_msg, tbb::flow::lightweight>*>
+                std::unordered_map<std::string, tbb::flow::continue_node<tbb::flow::continue_msg>*>
                     nodename_tbbnode_map;
-                tbb::flow::continue_node<tbb::flow::continue_msg,
-                                         tbb::flow::lightweight>* flowgraph_node_start =
-                    new tbb::flow::continue_node<tbb::flow::continue_msg, tbb::flow::lightweight>(
+                tbb::flow::continue_node<tbb::flow::continue_msg>* flowgraph_node_start =
+                    new tbb::flow::continue_node<tbb::flow::continue_msg>(
                         *(ctx->G), [&](const tbb::flow::continue_msg& msg) {});
                 auto it = enable_nodename_list.begin();
                 for (const auto& p : enables)
                 {
                     auto index = profiler_count++;
-                    tbb::flow::continue_node<tbb::flow::continue_msg, tbb::flow::lightweight>*
-                        flowgraph_node = new tbb::flow::continue_node<tbb::flow::continue_msg,
-                                                                      tbb::flow::lightweight>(
-                            *(ctx->G), [&, index](const tbb::flow::continue_msg& msg) {
+                    tbb::flow::continue_node<tbb::flow::continue_msg>* flowgraph_node =
+                        new tbb::flow::continue_node<tbb::flow::continue_msg>(
+                            *(ctx->G), [&, functor, index](const tbb::flow::continue_msg& msg) {
                                 if (p(ctx) || ctx->first_iteration)
                                 {
                                     if (runtime::cpu::IsTracingEnabled() || m_emit_timing)
@@ -1686,8 +1670,11 @@ void runtime::cpu::CPU_ExternalFunction::build()
                                         m_perf_counters[index].m_call_count++;
                                     }
                                 }
-                                std::advance(functor, 1);
                             });
+#ifdef TBB_PREVIEW_FLOW_GRAPH_TRACE
+                    flowgraph_node->set_name(it->second.c_str());
+#endif
+                    std::advance(functor, 1);
                     nodename_tbbnode_map.insert({it->second, flowgraph_node});
                     it++;
                 }
@@ -1720,9 +1707,7 @@ void runtime::cpu::CPU_ExternalFunction::build()
                 }
             }
             // Execute the flow graph
-            (static_cast<
-                 tbb::flow::continue_node<tbb::flow::continue_msg, tbb::flow::lightweight>*>(
-                 &(*(ctx->G->begin()))))
+            (static_cast<tbb::flow::continue_node<tbb::flow::continue_msg>*>(&(*(ctx->G->begin()))))
                 ->try_put(tbb::flow::continue_msg());
             try
             {

src/ngraph/runtime/cpu/cpu_runtime_context.hpp

@@ -21,6 +21,7 @@
 #include <set>
 
 #define TBB_PREVIEW_GLOBAL_CONTROL 1
+#define TBB_PREVIEW_FLOW_GRAPH_TRACE 1
 #include <tbb/flow_graph.h>
 #include <tbb/global_control.h>
 #include <tbb/task_scheduler_init.h>