turn on remove_unnecessary_lock

test=develop

paddle/fluid/framework/details/build_strategy.h

@@ -91,7 +91,7 @@ struct BuildStrategy {
   int num_trainers_{1};
   int trainer_id_{0};
   std::vector<std::string> trainers_endpoints_;
-  bool remove_unnecessary_lock_{false};
+  bool remove_unnecessary_lock_{true};
 
   // NOTE:
   // Before you add new options, think if it's a general strategy that works

paddle/fluid/pybind/pybind.cc

@@ -993,7 +993,7 @@ All parameter, weight, gradient are variables in Paddle.
             PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized.");
             self.remove_unnecessary_lock_ = b;
           },
-          R"DOC(The type is BOOL. If set True, some locks in GPU ops would be released and ParallelExecutor would run faster. Default False.)DOC")
+          R"DOC(The type is BOOL. If set True, some locks in GPU ops would be released and ParallelExecutor would run faster. Default True.)DOC")
       .def_property(
           "num_trainers",
           [](const BuildStrategy &self) { return self.num_trainers_; },

python/paddle/fluid/tests/unittests/test_dist_base.py

@@ -122,7 +122,7 @@ class TestDistRunnerBase(object):
         if args.batch_merge_repeat > 1:
             pass_builder = build_stra._finalize_strategy_and_create_passes()
             mypass = pass_builder.insert_pass(
-                len(pass_builder.all_passes()) - 2, "multi_batch_merge_pass")
+                len(pass_builder.all_passes()) - 3, "multi_batch_merge_pass")
             mypass.set_int("num_repeats", args.batch_merge_repeat)
 
         if args.update_method == "nccl2":