task_group Dynamic Dependencies

task_group Dynamic Dependencies#

Note

To enable this extension, define the TBB_PREVIEW_TASK_GROUP_EXTENSIONS macro with a value of 1. When available and enabled, the feature-test macro TBB_HAS_TASK_GROUP_DEPENDENCIES is defined.

Description#

The oneAPI Threading Building Blocks (oneTBB) implementation extends the tbb::task_group specification with an API for defining predecessor-successor relationships between tasks, such that a successor task can begin execution only after all of its predecessors are completed.

An unsubmitted task is one that has not been submitted for execution.

A submitted task is one that has been submitted for execution, such as by passing a task_handle to task_group::run.

A non-empty task_handle object represents an unsubmitted task, while a task_completion_handle can represent any task.

Both submitted and unsubmitted tasks can serve as predecessors. However, only unsubmitted tasks may be used as successors.

The tbb::task_group::set_task_order(pred, succ) function establishes a dependency such that succ cannot begin execution until pred has completed.

tbb::task_handle predecessor = tg.defer(pred_body);
tbb::task_handle successor = tg.defer(succ_body);

tbb::task_group::set_task_order(predecessor, successor);

The tbb::task_group::transfer_this_task_completion_to function allows transferring the completion of the currently executing task to another task. This function must be invoked from within the task body. All successors of the currently executing task will execute only after the task receiving the completion has finished.

tbb::task_handle t = tg.defer([&tg] {
    tbb::task_handle comp_receiver = tg.defer(receiver_body);
    tbb::task_group::transfer_this_task_completion_to(comp_receiver);
    tg.run(std::move(comp_receiver));
});

tbb::task_handle succ = tg.defer(succ_body);

tbb::task_group::set_task_order(t, succ);
// Since t transfers its completion to comp_receiver,
// succ_body will execute after receiver_body

API#

Synopsis#

// <oneapi/tbb/task_group.h> synopsis
namespace oneapi {
    namespace tbb {
        class task_handle {
        public:
            // Only the requirements for destroyed object are changed
            ~task_handle();
        };

        class task_group {
            // Only the behavior in case of dependent tasks is changed
            void run(task_handle&& handle);

            static void set_task_order(task_handle& pred, task_handle& succ);
            static void set_task_order(task_completion_handle& pred, task_handle& succ);

            static void transfer_this_task_completion_to(task_handle& handle);
        };
    } // namespace tbb
} // namespace oneapi

// // <oneapi/tbb/task_arena.h> synopsis
namespace oneapi {
    namespace tbb {
        class task_arena {
            // Only the behavior in case of dependent tasks is changed
            void enqueue(task_handle&& handle);
        }; // class task_arena

        namespace this_task_arena {
            // Only the behavior in case of dependent tasks is changed
            void enqueue(task_handle&& handle);
        } // namespace this_task_arena
    } // namespace tbb
} // namespace oneapi

Member Functions of task_handle Class#

~task_handle();

Destroys the task_handle object and associated task if it exists.

Extension

If the associated task is involved in a predecessor-successor relationship, the behavior is undefined.

Member Functions of task_group Class#

void run(task_handle&& h);

Schedules the task object pointed by h for the execution.

Extension

If the task associated with h has predecessors, scheduling the task execution is postponed until all of the predecessors have completed, while the function returns immediately.

Note

The failure to satisfy the following conditions leads to undefined behavior:

  • h is not empty.

  • *this is the same task_group that h is created with.

static void set_task_order(task_handle& pred, task_handle& succ);
static void set_task_order(task_completion_handle& pred, task_handle& succ);

Registers the task associated with pred as a predecessor that must complete before the task associated with succ can begin execution.

It is thread-safe to concurrently add multiple predecessors to a single successor and to register the same predecessor with multiple successors.

It is thread-safe to concurrently add successors to both the task transferring its completion and the task receiving the completion.

It is thread-safe to concurrently add a successor to a task_completion_handle while the task_handle associated with the same task is being run.

The behavior is undefined in the following cases:

  • Either pred or succ is empty.

  • The tasks referred by pred and succ belong to different task_group instances.

  • The task referred by task_completion_handle was destroyed without being submitted for execution.

static void transfer_this_task_completion_to(task_handle& handle);

Transfers the completion of the currently executing task to the task associated with handle.

After the transfer, the successors of the currently executing task will be reassigned to the task associated with handle.

It is thread-safe to transfer successors to the task while concurrently adding successors to it or to the currently executing task.

The behavior is undefined in the following cases:

  • handle is empty.

  • The function is called outside the body of a task_group task.

  • The function is called for the task whose completion has already been transferred.

  • The currently executing task and the task associated with handle belong to different task_group instances.

Member Functions of task_arena Class#

void enqueue(task_handle&& h);

Enqueues a task owned by h into the task_arena for processing.

Extension

If the task associated with h has predecessors, scheduling the task execution is postponed until all of the predecessors have completed, while the function returns immediately.

The behavior of this function is identical to the generic version (template<typename F> void task_arena::enqueue(F&& f)), except parameter type.

Note

h should not be empty to avoid an undefined behavior.

this_task_arena Namespace#

void enqueue(task_handle&& h);

Enqueues a task owned by h into the task_arena that is currently used by the calling thread.

Extension

If the task associated with h has predecessors, scheduling the task execution is postponed until all of the predecessors have completed, while the function returns immediately.

The behavior of this function is identical to the generic version (template<typename F> void task_arena::enqueue(F&& f)), except parameter type.

Note

h should not be empty to avoid an undefined behavior.

Example#

The following example demonstrates how to perform parallel reduction over a range using the described API.

#define TBB_PREVIEW_TASK_GROUP_EXTENSIONS 1
#include "oneapi/tbb/task_group.h"

struct reduce_task {

    struct join_task {
        void operator()() const {
            result = *left + *right;
        }

        std::size_t& result;
        std::unique_ptr<std::size_t> left;
        std::unique_ptr<std::size_t> right;
    };

    tbb::task_handle operator()() const {
        tbb::task_handle next_task;

        std::size_t size = end - begin;
        if (size < serial_threshold) {
            // Perform serial reduction
            for (std::size_t i = begin; i < end; ++i) {
                result += i;
            }
        } else {
            // The range is too large to process directly
            // Divide it into smaller segments for parallel execution
            std::size_t middle = begin + size / 2;

            auto left_result = std::make_unique<std::size_t>(0);
            auto right_result = std::make_unique<std::size_t>(0);

            
            tbb::task_handle left_leaf = tg.defer(reduce_task{begin, middle, *left_result, tg});
            tbb::task_handle right_leaf = tg.defer(reduce_task{middle, end, *right_result, tg});

            tbb::task_handle join = tg.defer(join_task{result, std::move(left_result), std::move(right_result)});

            tbb::task_group::set_task_order(left_leaf, join);
            tbb::task_group::set_task_order(right_leaf, join);

            tbb::task_group::transfer_this_task_completion_to(join);

            // Save the left leaf for further bypassing
            next_task = std::move(left_leaf);

            tg.run(std::move(right_leaf));
            tg.run(std::move(join));
        }

        return next_task;
    }

    std::size_t begin;
    std::size_t end;
    std::size_t& result;
    tbb::task_group& tg;
};

std::size_t calculate_parallel_sum(std::size_t begin, std::size_t end) {
    tbb::task_group tg;

    std::size_t reduce_result = 0;
    tg.run_and_wait(reduce_task{begin, end, reduce_result, tg});

    return reduce_result;
}