AsyncTasking Launch Asynchronous Tasks from an Executor AsyncTasking_1LaunchAsynchronousTasksFromAnExecutor Launch Asynchronous Tasks from a Subflow AsyncTasking_1LaunchAsynchronousTasksFromAnSubflow Launch Asynchronous Tasks from a Runtime AsyncTasking_1LaunchAsynchronousTasksFromARuntime This chapters discusses how to launch tasks asynchronously so that you can incorporate independent, dynamic parallelism in your taskflows. Taskflow executor provides an STL-styled method, tf::Executor::async, for you to run a callable object asynchronously. The method returns a std::future that will eventually hold the result of that function call. std::future<int>future=executor.async([](){return1;}); assert(future.get()==1); Unlike std::async, the future object returned from tf::Executor::async does not block on destruction until completing the function. If you do not need the return value or use a future to synchronize the execution, you are encouraged to use tf::Executor::silent_async which returns nothing and thus has less overhead (i.e., no shared state management) compared to tf::Executor::async. executor.silent_async([](){ //dosomeworkwithoutreturninganyresult }); Launching asynchronous tasks from an executor is thread-safe and can be called by multiple threads both inside (i.e., worker) and outside the executor. Our scheduler autonomously detects whether an asynchronous task is submitted from an external thread or a worker thread and schedules its execution using work stealing. tf::Taskmy_task=taskflow.emplace([&](){ //launchanasynchronoustaskfrommy_task executor.async([&](){ //launchanotherasynchronoustaskthatmayberunbyanotherworker executor.async([&](){}); }) }); executor.run(taskflow); executor.wait_for_all();//waitforalltaskstofinish Asynchronous tasks created from an executor does not belong to any taskflows. The lifetime of an asynchronous task is managed automatically by the executor that creates the task. You can name an asynchronous task using the overloads, tf::Executor::async(const std::string& name, F&& f) and tf::Executor::silent_async(const std::string& name, F&& f), that take a string in the first argument. Assigned names will appear in the observers of the executor. std::future<void>fu=executor.async("asynctask",[](){}); executor.silent_async("silentasynctask",[](){}); You can launch asynchronous tasks from tf::Subflow using tf::Subflow::async. Asynchronous tasks are independent tasks spawned during the execution of a subflow. When the subflow joins, all asynchronous tasks are guaranteed to finish. The following code creates 100 asynchronous tasks from a subflow and joins their executions explicitly using tf::Subflow::join. tf::Taskflowtaskflow; tf::Executorexecutor; std::atomic<int>counter{0}; taskflow.emplace([&](tf::Subflow&sf){ std::vector<std::future<void>>futures; for(inti=0;i<100;i++){ futures.emplace_back(sf.async([&](){++counter;})); } sf.join();//allofthe100asynchronoustaskswillfinishbythisjoin assert(counter==100); }); executor.run(taskflow).wait(); If you do not need the return value or the future to synchronize the execution, you can use tf::Subflow::silent_async which has less overhead when creating an asynchronous task compared to tf::Subflow::async. tf::Taskflowtaskflow; tf::Executorexecutor; std::atomic<int>counter{0}; taskflow.emplace([&](tf::Subflow&sf){ for(inti=0;i<100;i++){ sf.silent_async([&](){++counter;}); } sf.join();//allofthe100asynchronoustaskswillfinishbythisjoin assert(counter==100); }); executor.run(taskflow).wait(); You should only create asynchronous tasks from a joinable subflow. Launching asynchronous tasks from a detached subflow results in undefined behavior. You can assign an asynchronous task a name using the two overloads, tf::Subflow::async(const std::string& name, F&& f) and tf::Subflow::silent_async(const std::string& name, F&& f). Both methods take an additional argument of a string. taskflow.emplace([](tf::Subflow&sf){ std::future<void>future=sf.async("nameofthetask",[](){}); sf.silent_async("anothernameofthetask",[](){}); sf.join(); }); The asynchronous tasking feature of tf::Subflow is indeed derived from tf::Runtime. You can launch asynchronous tasks from tf::Runtime using tf::Runtime::async or tf::Runtime::silent_async. The following code creates 100 asynchronous tasks from a runtime and joins their executions explicitly using tf::Runtime::corun_all. tf::Taskflowtaskflow; tf::Executorexecutor; std::atomic<int>counter{0}; taskflow.emplace([&](tf::Runtime&rt){ for(inti=0;i<100;i++){ rt.silent_async([&](){++counter;})); } rt.corun_all();//allofthe100asynchronoustaskswillfinishbythisjoin assert(counter==100); }); executor.run(taskflow).wait(); Unlike tf::Subflow::join, you can call tf::Runtime::corun_all multiple times to synchronize the execution of asynchronous tasks between different runs. For example, the following code spawn 100 asynchronous tasks twice and join each execution to assure the spawned 100 asynchronous tasks have properly completed. tf::Taskflowtaskflow; tf::Executorexecutor; std::atomic<int>counter{0}; taskflow.emplace([&](tf::Runtime&rt){ //spawn100asynchronoustasksandjoin for(inti=0;i<100;i++){ rt.silent_async([&](){++counter;})); } rt.corun_all();//allofthe100asynchronoustaskswillfinishbythisjoin assert(counter==100); //spawnanother100asynchronoustasksandjoin for(inti=0;i<100;i++){ rt.silent_async([&](){++counter;})); } rt.corun_all();//allofthe100asynchronoustaskswillfinishbythisjoin assert(counter==200); }); executor.run(taskflow).wait(); By default, tf::Runtime does not join like tf::Subflow. All pending asynchronous tasks spawned by tf::Runtime are no longer controllable when their parent runtime disappears. It is your responsibility to properly synchronize spawned asynchronous tasks using tf::Runtime::corun_all. Creating asynchronous tasks from a runtime allows users to efficiently implement parallel algorithms using recursion, such as parallel sort (tf::Taskflow::sort), that demands dynamic parallelism at runtime.