Fixed double-buffering timing (#3309)

API is synchronous per thread and threads are coordinated so that
we know when we hit the last iteration everything is done.
Using join() to gate end of iterations was introducing too much
overhead to timing as verified via checking traces.

src/tools/nbench/benchmark_pipelined.cpp

@@ -96,6 +96,10 @@ static void
                 result->read(data->get_data_ptr(),
                              data->get_element_count() * data->get_element_type().size());
             }
+            if (current_iteration == (s_iterations + s_warmup_iterations - 1))
+            {
+                s_timer.stop();
+            }
         }
     }
 }
@@ -177,7 +181,6 @@ vector<runtime::PerformanceCounter> run_benchmark_pipelined(shared_ptr<Function>
     {
         threads[i].join();
     }
-    s_timer.stop();
     float time = s_timer.get_milliseconds();
     cout << time / iterations << "ms per iteration" << endl;