One of the most powerful features that distinguishes Taskflow from other systems is its support for conditional tasking, also known as the control taskflow programming model (CTFG). CTFG allows you to embed control flow directly within a taskflow graph, enabling tasks to make decisions dynamically during execution. This mechanism supports advanced in-graph control flow patterns, such as dynamic branching, loops, and conditionals—that are typically difficult or impossible to express in traditional task graph models.
A condition task returns an integer index indicating which successor task to execute next. The index corresponds to the position of the successor in the order it was added during task construction. The following example creates an if-else block using a condition task.
The condition task cond is connected to two successor tasks, yes and no, via precede(yes, no). When cond returns 0, the execution moves on to yes. When cond returns 1, the execution moves on to no.
A condition task can form a cycle to express iterative control flow. The example below demonstrates a simple yet commonly used feedback loop implemented using a condition task that returns a random binary value. If the return value from cond is 0, the task loops back to itself; otherwise, it proceeds to stop.
Creating a taskflow with complex control flow often requires only a few lines of code to implement. Different control flow paths can execute in parallel, making it easy to express both logic and concurrency. The code below creates a taskflow with three condition tasks to demonstrate this capability:
The above code creates three condition tasks to implement three different control-flow tasks:
cond_1 that loops back to B on returning 0, or proceeds to E on returning 1,cond_2 that goes to G on returning 0, or H on returning 1,cond_3 that loops back to itself on returning 0, or proceeds to L on returning 1 In this particular example, we can clearly see the advantage of CTFG: the execution of cond_1 can overlap with cond_2 or cond_3, enabling greater concurrency in control-driven workloads. Unlike traditional task graph models that require static structure or external orchestration to handle control flow, CTFG allows tasks to make decisions dynamically and continue execution without global synchronization barriers. This design leads to better parallelism, reduced overhead, and more expressive task graphs, especially in workloads with branching or iterative control flows.
In order to understand how an executor schedules condition tasks, we define two dependency types, strong dependency and weak dependency. A strong dependency is a preceding link from one non-condition task to another task. A weak dependency is a preceding link from one condition task to another task. The number of dependencies of a task is the sum of its strong dependencies and weak dependencies. The table below lists the number of strong dependencies and weak dependencies of each task in the previous example:
| task | strong dependency | weak dependency | dependencies |
|---|---|---|---|
| A | 0 | 0 | 0 |
| B | 1 | 1 | 2 |
| C | 1 | 0 | 1 |
| D | 1 | 0 | 1 |
| E | 0 | 1 | 1 |
| F | 1 | 0 | 1 |
| G | 0 | 1 | 1 |
| H | 0 | 1 | 1 |
| I | 1 | 0 | 1 |
| K | 1 | 0 | 1 |
| L | 0 | 1 | 1 |
| M | 1 | 0 | 1 |
| cond_1 | 1 | 0 | 1 |
| cond_2 | 1 | 0 | 1 |
| cond_3 | 1 | 1 | 2 |
You can query the number of strong dependencies, the number of weak dependencies, and the number of dependencies of a task.
When you submit a task to an executor, the scheduler starts with tasks of zero dependencies (both zero strong and weak dependencies) and continues to execute successive tasks whenever their strong dependencies are met. However, the scheduler skips this rule when executing a condition task and jumps directly to its successors indexed by the return value.
Each task has an atomic join counter to keep track of strong dependencies that are met at runtime. When a task completes, the join counter is restored to the task's strong dependency number in the graph, such that the subsequent execution can reuse the counter again.
Let's take a look at an example to understand how task-level scheduling works. Suppose we have the following taskflow of one condition task cond that forms a loop to itself on returning 0 and moves on to stop on returning 1:
The scheduler starts with init task because it has no dependencies (both strong and weak dependencies). Then, the scheduler moves on to the condition task cond. If cond returns 0, the scheduler enqueues cond and runs it again. If cond returns 1, the scheduler enqueues stop and then moves on.
Condition tasks are powerful but require careful graph construction. The following pitfalls are the most common sources of bugs when using conditional tasking, ranging from silent deadlocks to non-deterministic task races.
Every taskflow must have at least one task with zero dependencies for the scheduler to start with. When a condition task forms a cycle, it is easy to accidentally create a graph where every task has at least one incoming edge, leaving the scheduler with no entry point.
The figure below shows common pitfalls and their remedies.
In the error1 scenario, there is no source task for the scheduler to start with. The simplest fix is to add a task S with no dependencies that precedes the rest of the graph, giving the scheduler an unambiguous entry point.
A task race occurs when the same task can be scheduled more than once simultaneously through different paths. This typically happens when a task has both a strong dependency from a regular task and a weak dependency from a condition task.
In the error2 scenario, task D can be scheduled twice: once by E through its strong dependency, and once by C through its weak dependency (when C returns 1). If both paths activate D at the same time, D runs concurrently with itself, which is undefined behavior. The fix is to insert an auxiliary task D-aux between the two paths so that D always has exactly one active predecessor at a time.
In the risky scenario, task X may be raced by M and P if M returns 0 and P returns 1 simultaneously, triggering X from two different condition tasks at once. Whenever a task sits at the junction of multiple condition task outputs, carefully check whether those condition tasks can fire concurrently.
A deadlock occurs when a condition task loops back to a task via a strong dependency rather than a weak one. Because the scheduler only moves a task to the ready queue when all its strong dependencies are satisfied, a loop that creates a strong back-edge will permanently block: the loop body waits for the condition task to complete, but the condition task can never re-run because the loop body has not been executed yet.
The wrong while-loop implementation at Implement While-Loop Control Flow is a concrete example of this pitfall. When the body task i++ directly precedes the loop condition task cond with body.precede(cond), it creates a strong dependency. After init runs and decrements cond's strong dependency count by one, cond still waits for i++ to complete before it can run. But i++ only runs after cond returns 0, so neither task can proceed and the graph deadlocks. The correct fix is to introduce a dedicated back-edge condition task back that returns 0 unconditionally to cond, creating a weak dependency instead.
You can use conditional tasking to implement if-else control flow. The following example creates a nested if-else control flow diagram that executes three condition tasks to check the range of i.
You can use condition tasks to implement switch-style control flow. The following example demonstrates this by creating a switch structure that randomly selects and executes one of three cases using four condition tasks.
Assuming swcond returns 1, the program outputs:
Keep in mind, both switch and case tasks must be described as condition tasks. The following implementation is a common mistake in which case tasks are not described as condition tasks.
In this faulty implementation, task target has three strong dependencies but only one of them will be met. This is because swcond is a condition task, and only one case task will be executed depending on the return of swcond.
You can use conditional tasking to implement do-while-loop control flow. The following example creates a do-while-loop control flow diagram that repeatedly increments variable i five times using one condition task.
The program outputs:
You can use conditional tasking to implement while-loop control flow. The following example creates a while-loop control flow diagram that repeatedly increments variable i five times using two condition task.
The program outputs:
Notice that, when you implement a while-loop block, you cannot direct a dependency from the body task to the loop condition task. Doing so will introduce a strong dependency between the body task and the loop condition task, and the loop condition task will never be executed. The following code shows a common faulty implementation of while-loop control flow.
In the taskflow diagram above, the scheduler starts with init and then decrements the strong dependency of the loop condition task, while i<5. After this, there remains one strong dependency, i.e., introduced by the loop body task, i++. However, task i++ will not be executed until the loop condition task returns 0, causing a deadlock.
A multi-condition task is a generalized version of conditional tasking. In some cases, applications need to jump to multiple branches from a parent task. This can be done by creating a multi-condition task which allows a task to select one or more successor tasks to execute. Similar to a condition task, a multi-condition task returns a vector of integer indices that indicate the successors to execute when the multi-condition task completes. The index is defined with respect to the order of successors preceded by a multi-condition task. For example, the following code creates a multi-condition task, A, that informs the scheduler to run on its two successors, B and D.
1.13.1