A header-only C++ library to help you quickly write parallel and heterogeneous programs using task models

Cpp-Taskflow is by far faster, more expressive, and easier for drop-in integration than existing task programming libraries such as OpenMP Tasking and Intel TBB FlowGraph in handling complex parallel workloads.

Cpp-Taskflow lets you quickly implement task decomposition strategies that incorporate both regular and irregular compute patterns, together with an efficient work-stealing scheduler to optimize your multithreaded performance.

Without Cpp-Taskflow	With Cpp-Taskflow

Cpp-Taskflow has a unified interface for both static tasking and dynamic tasking, allowing users to quickly master our parallel task programming model in a natural idiom.

Static Tasking	Dynamic Tasking

Cpp-Taskflow supports conditional tasking for you to implement cyclic and dynamic control flows that are otherwise difficult to do with existing task programming frameworks.

Conditional Tasking

Cpp-Taskflow is composable. You can create large parallel graphs through composition of modular and reusable blocks that are easier to optimize at an individual scope.

Taskflow Composition

Cpp-Taskflow supports heterogeneous tasking for you to accelerate a wide range of scientific computing applications by harnessing the power of CPU-GPU collaborative computing.

Concurrent CPU-GPU Tasking

We are committed to support trustworthy developments for both academic and industrial research projects in parallel computing. Check out Who is Using Cpp-Taskflow and what our users say:

• "Cpp-Taskflow is the cleanest Task API I've ever seen." damienhocking
• "Cpp-Taskflow has a very simple and elegant tasking interface. The performance also scales very well." totalgee
• "Cpp-Taskflow lets me handle parallel processing in a smart way." Hayabusa
• "Best poster award for open-source parallel programming library." Cpp Conference 2018
• "Second Prize of Open-source Software Competition." ACM Multimedia Conference 2019

See a quick presentation and visit the documentation to learn more about Cpp-Taskflow. Technical details can be referred to our IEEE IPDPS19 paper.

• Get Started with Cpp-Taskflow
• Create a Taskflow Application
  • Step 1: Create a Taskflow
  • Step 2: Define Task Dependencies
  • Step 3: Execute a Taskflow
• Dynamic Tasking
• Conditional Tasking
  • Step 1: Create a Condition Task
  • Step 2: Scheduling Rules for Condition Tasks
• Composable Tasking
• Concurrent CPU-GPU Tasking
  • Step 1: Create a cudaFlow
  • Step 2: Compile and Execute a cudaFlow
• Visualize a Taskflow Graph
• Monitor Thread Activities
• API Reference
• System Requirements
• Compile Unit Tests, Examples, and Benchmarks
• Who is Using Cpp-Taskflow?

The following example simple.cpp shows the basic Cpp-Taskflow API you need in most applications.

#include <taskflow/taskflow.hpp>  // Cpp-Taskflow is header-only

int main(){
  
  tf::Executor executor;
  tf::Taskflow taskflow;

  auto [A, B, C, D] = taskflow.emplace(
    [] () { std::cout << "TaskA\n"; },               //  task dependency graph
    [] () { std::cout << "TaskB\n"; },               // 
    [] () { std::cout << "TaskC\n"; },               //          +---+          
    [] () { std::cout << "TaskD\n"; }                //    +---->| B |-----+   
  );                                                 //    |     +---+     |
                                                     //  +---+           +-v-+ 
  A.precede(B);  // A runs before B                  //  | A |           | D | 
  A.precede(C);  // A runs before C                  //  +---+           +-^-+ 
  B.precede(D);  // B runs before D                  //    |     +---+     |    
  C.precede(D);  // C runs before D                  //    +---->| C |-----+    
                                                     //          +---+          
  executor.run(taskflow).wait();

  return 0;
}

Compile and run the code with the following commands:

~$ g++ simple.cpp -I path/to/include/taskflow/ -std=c++17 -O2 -lpthread -o simple
~$ ./simple
TaskA
TaskC  <-- concurrent with TaskB
TaskB  <-- concurrent with TaskC
TaskD

Cpp-Taskflow defines a very expressive API to create task dependency graphs. Most applications are developed through the following three steps:

Create a taskflow object to build a task dependency graph:

tf::Taskflow taskflow;

A task is a callable object for which std::invoke is applicable. Use the method emplace to create a task:

tf::Task A = taskflow.emplace([](){ std::cout << "Task A\n"; });

You can add dependency links between tasks to enforce one task to run before or after another.

A.precede(B);  // A runs before B.

To execute a taskflow, you need to create an executor. An executor manages a set of worker threads to execute a taskflow through an efficient work-stealing algorithm.

tf::Executor executor;

The executor provides a rich set of methods to run a taskflow. You can run a taskflow multiple times, or until a stopping criteria is met. These methods are non-blocking with a std::future return to let you query the execution status. Executor is thread-safe.

executor.run(taskflow);       // runs the taskflow once
executor.run_n(taskflow, 4);  // runs the taskflow four times

// keeps running the taskflow until the predicate becomes true
executor.run_until(taskflow, [counter=4](){ return --counter == 0; } );

You can call wait_for_all to block the executor until all associated taskflows complete.

executor.wait_for_all();  // block until all associated tasks finish

Notice that the executor does not own any taskflow. It is your responsibility to keep a taskflow alive during its execution, or it can result in undefined behavior. In most applications, you need only one executor to run multiple taskflows each representing a specific part of your parallel decomposition.

Another powerful feature of Taskflow is dynamic tasking. Dynamic tasks are those tasks created during the execution of a taskflow. These tasks are spawned by a parent task and are grouped together to a subflow graph. To create a subflow for dynamic tasking, emplace a callable with one argument of type tf::Subflow.

// create three regular tasks
tf::Task A = tf.emplace([](){}).name("A");
tf::Task C = tf.emplace([](){}).name("C");
tf::Task D = tf.emplace([](){}).name("D");

// create a subflow graph (dynamic tasking)
tf::Task B = tf.emplace([] (tf::Subflow& subflow) {
  tf::Task B1 = subflow.emplace([](){}).name("B1");
  tf::Task B2 = subflow.emplace([](){}).name("B2");
  tf::Task B3 = subflow.emplace([](){}).name("B3");
  B1.precede(B3);
  B2.precede(B3);
}).name("B");
            
A.precede(B);  // B runs after A 
A.precede(C);  // C runs after A 
B.precede(D);  // D runs after B 
C.precede(D);  // D runs after C 

By default, a subflow graph joins its parent node. This ensures a subflow graph finishes before the successors of its parent task. You can disable this feature by calling subflow.detach(). For example, detaching the above subflow will result in the following execution flow:

// create a "detached" subflow graph (dynamic tasking)
tf::Task B = tf.emplace([] (tf::Subflow& subflow) {
  tf::Task B1 = subflow.emplace([](){}).name("B1");
  tf::Task B2 = subflow.emplace([](){}).name("B2");
  tf::Task B3 = subflow.emplace([](){}).name("B3");

  B1.precede(B3);
  B2.precede(B3);

  // detach the subflow to form a parallel execution line
  subflow.detach();
}).name("B");

A subflow can be nested or recursive. You can create another subflow from the execution of a subflow and so on.

Taskflow supports conditional tasking for users to implement dynamic and cyclic control flows. You can create highly versatile and efficient parallel patterns through condition tasks.

A condition task evalutes a set of instructions and returns an integer index of the next immediate successor to execute. The index is defined with respect to the order of its successor construction.

tf::Task init = tf.emplace([](){ }).name("init");
tf::Task stop = tf.emplace([](){ }).name("stop");

// creates a condition task that returns 0 or 1
tf::Task cond = tf.emplace([](){
  std::cout << "flipping a coin\n";
  return rand() % 2;
}).name("cond");

// creates a feedback loop
init.precede(cond);
cond.precede(cond, stop);  // cond--0-->cond, cond--1-->stop

executor.run(tf).wait();

If the return value from cond is 0, it loops back to itself, or otherwise to stop. Cpp-Taskflow terms the preceding link from a condition task a weak dependency (dashed lines above). Others are strong depedency (solid lines above).

When you submit a taskflow to an executor, the scheduler starts with tasks of zero dependency (both weak and strong dependencies) and continues to execute successive tasks whenever strong dependencies are met. However, the scheduler skips this rule for a condition task and jumps directly to its successor indexed by the return value.

It is users' responsibility to ensure a taskflow is properly conditioned. Top things to avoid include no source tasks to start with and task race. The figure shows common pitfalls and their remedies. In the risky scenario, task X may not be raced if P and M is exclusively branching to X.

A good practice for avoiding mistakes of conditional tasking is to infer the execution flow of your graphs based on our scheduling rules. Make sure there is no task race.

A powerful feature of tf::Taskflow is composability. You can create multiple task graphs from different parts of your workload and use them to compose a large graph through the composed_of method.

tf::Taskflow f1, f2;

auto [f1A, f1B] = f1.emplace( 
  []() { std::cout << "Task f1A\n"; },
  []() { std::cout << "Task f1B\n"; }
);
auto [f2A, f2B, f2C] = f2.emplace( 
  []() { std::cout << "Task f2A\n"; },
  []() { std::cout << "Task f2B\n"; },
  []() { std::cout << "Task f2C\n"; }
);
auto f1_module_task = f2.composed_of(f1);

f1_module_task.succeed(f2A, f2B)
              .precede(f2C);

Similarly, composed_of returns a task handle and you can use precede to create dependencies. You can compose a taskflow from multiple taskflows and use the result to compose a larger taskflow and so on.

Cpp-Taskflow enables concurrent CPU-GPU tasking by leveraging Nvidia CUDA Toolkit. You can harness the power of CPU-GPU collaborative computing to implement heterogeneous decomposition algorithms.

A tf::cudaFlow is a graph object created at runtime similar to dynamic tasking. It manages a task node in a taskflow and associates it with a CUDA Graph. To create a cudaFlow, emplace a callable with an argument of type tf::cudaFlow.

tf::Taskflow taskflow;
tf::Executor executor;

const unsigned N = 1<<20;                            // size of the vector
std::vector<float> hx(N, 1.0f), hy(N, 2.0f);         // x and y vectors at host
float *dx{nullptr}, *dy{nullptr};                    // x and y vectors at device

tf::Task allocate_x = taskflow.emplace([&](){ cudaMalloc(&dx, N*sizeof(float));});
tf::Task allocate_y = taskflow.emplace([&](){ cudaMalloc(&dy, N*sizeof(float));});
tf::Task cudaflow = taskflow.emplace([&](tf::cudaFlow& cf) {
  tf::cudaTask h2d_x = cf.copy(dx, hx.data(), N);    // host-to-device x data transfer
  tf::cudaTask h2d_y = cf.copy(dy, hy.data(), N);    // host-to-device y data transfer
  tf::cudaTask d2h_x = cf.copy(hx.data(), dx, N);    // device-to-host x data transfer
  tf::cudaTask d2h_y = cf.copy(hy.data(), dy, N);    // device-to-host y data transfer
  // launch saxpy<<<(N+255)/256, 256, 0>>>(N, 2.0f, dx, dy)
  tf::cudaTask kernel = cf.kernel((N+255)/256, 256, 0, saxpy, N, 2.0f, dx, dy);
  kernel.succeed(h2d_x, h2d_y)
        .precede(d2h_x, d2h_y);
});
cudaflow.succeed(allocate_x, allocate_y);            // overlap data allocations

executor.run(taskflow).wait();

Assume our kernel implements the canonical saxpy operation (single-precision A·X Plus Y) using the CUDA syntax.

// saxpy (single-precision A·X Plus Y) kernel
__global__ void saxpy(
  int n, float a, float *x, float *y
) {
  // get the thread index
  int i = blockIdx.x*blockDim.x + threadIdx.x;

  if (i < n) {
    y[i] = a*x[i] + y[i];
  }
}

Name you source with the extension .cu, let's say saxpy.cu, and compile it through nvcc:

~$ nvcc saxpy.cu -I path/to/include/taskflow -O2 -o saxpy
~$ ./saxpy

Our source autonomously enables cudaFlow for compilers that support CUDA.

You can dump a taskflow through a std::ostream in GraphViz format using the method dump. There are a number of free GraphViz tools you could find online to visualize your Taskflow graph.

tf::Taskflow taskflow;
tf::Task A = taskflow.emplace([] () {}).name("A");
tf::Task B = taskflow.emplace([] () {}).name("B");
tf::Task C = taskflow.emplace([] () {}).name("C");
tf::Task D = taskflow.emplace([] () {}).name("D");
tf::Task E = taskflow.emplace([] () {}).name("E");
A.precede(B, C, E); 
C.precede(D);
B.precede(D, E); 

taskflow.dump(std::cout);  // dump the graph in DOT to std::cout

When you have tasks that are created at runtime (e.g., subflow, cudaFlow), you need to execute the graph first to spawn these tasks and dump the entire graph.

tf::Executor executor;
tf::Taskflow taskflow;

tf::Task A = taskflow.emplace([](){}).name("A");

// create a subflow of two tasks B1->B2
tf::Task B = taskflow.emplace([] (tf::Subflow& subflow) {
  tf::Task B1 = subflow.emplace([](){}).name("B1");
  tf::Task B2 = subflow.emplace([](){}).name("B2");
  B1.precede(B2);
}).name("B");

A.precede(B);

executor.run(tf).wait();  // run the taskflow to spawn subflows
tf.dump(std::cout);       // dump the graph including dynamic tasks

Understanding thread activities is very important for performance analysis. Cpp-Taskflow provides a default observer of type tf::ExecutorObserver that lets users observe when a thread starts or stops participating in task scheduling.

auto observer = executor.make_observer<tf::ExecutorObserver>();

When you are running a task dependency graph, the observer will automatically record the start and end timestamps of each executed task. You can dump the entire execution timelines into a JSON file.

executor.run(taskflow1);               // run a task dependency graph 1
executor.run(taskflow2);               // run a task dependency graph 2
executor.wait_for_all();               // block until all tasks finish

std::ofstream ofs("timestamps.json");
observer->dump(ofs);                   // dump the timeline to a JSON file

You can open the chrome browser to visualize the execution timelines through the chrome://tracing developer tool. In the tracing view, click the Load button to read the JSON file. You shall see the tracing graph.

Each task is given a name of i_j where i is the thread id and j is the task number. You can pan or zoom in/out the timeline to get into a detailed view.

The official documentation explains a complete list of Cpp-Taskflow API. Here, we highlight commonly used methods.

The class tf::Taskflow is the main place to create a task dependency graph. The table below summarizes a list of commonly used methods.

You can use emplace to create a task from a target callable.

tf::Task task = tf.emplace([] () { std::cout << "my task\n"; });

When a task cannot be determined beforehand, you can create a placeholder and assign the callable later.

tf::Task A = tf.emplace([](){});
tf::Task B = tf.placeholder();
A.precede(B);
B.work([](){ /* do something */ });

The method parallel_for creates a subgraph that applies the callable to each item in the given range of a container.

auto v = {'A', 'B', 'C', 'D'};
auto [S, T] = tf.parallel_for(
  v.begin(),    // iterator to the beginning
  v.end(),      // iterator to the end
  [] (int i) { 
    std::cout << "parallel " << i << '\n';
  }
);
// add dependencies via S and T.

You can specify a chunk size (default one) in the last argument to force a task to include a certain number of items.

auto v = {'A', 'B', 'C', 'D'};
auto [S, T] = tf.parallel_for(
  v.begin(),    // iterator to the beginning
  v.end(),      // iterator to the end
  [] (int i) { 
    std::cout << "AB and CD run in parallel" << '\n';
  },
  2  // at least two items at a time
);

In addition to iterator-based construction, parallel_for has another overload of index-based loop. The first three argument of this overload indicates starting index, ending index (exclusive), and step size.

// [0, 11) with a step size of 2
auto [S, T] = tf.parallel_for(
  0, 11, 2, 
  [] (int i) {
    std::cout << "parallel_for on index " << i << std::endl;
  }, 
  2  // at least two items at a time
);
// will print 0, 2, 4, 6, 8, 10 (three partitions, {0, 2}, {4, 6}, {8, 10})

Each time you create a task, the taskflow object adds a node to the present task dependency graph and return a task handle to you. A task handle is a lightweight object that defines a set of methods for users to access and modify the attributes of the associated task. The table below summarizes a list of commonly used methods.

The method name lets you assign a human-readable string to a task.

A.name("my name is A");

The method work lets you assign a callable to a task.

A.work([] () { std::cout << "hello world!"; });

The method precede lets you add a preceding link from self to other tasks.

// A runs before B, C, D, and E
A.precede(B, C, D, E);

The method succeed is similar to precede but operates in the opposite direction.

A task is empty is it is not associated with any graph node.

tf::Task task;  // assert(task.empty());

A placeholder task is associated with a graph node but has no work assigned yet.

tf::Task task = taskflow.placeholder();  // assert(!task.has_work());

The class tf::Executor is used for execution of one or multiple taskflow objects. The table below summarizes a list of commonly used methods.

The run series are non-blocking call to execute a taskflow graph. Issuing multiple runs on the same taskflow will automatically synchronize to a sequential chain of executions.

executor.run(taskflow);                 // runs a graph once
executor.run_n(taskflow, 5);            // runs a graph five times
executor.run_until(taskflow, my_pred);  // keeps running until the my_pred becomes true
executor.wait_for_all();                // blocks until all tasks finish

The first run finishes before the second run, and the second run finishes before the third run.

To use the latest Cpp-Taskflow, you only need a C++14 compiler.

• GNU C++ Compiler at least v5.0 with -std=c++14
• Clang C++ Compiler at least v4.0 with -std=c++14
• Microsoft Visual Studio at least v15.7 (MSVC++ 19.14); see vcpkg guide
• AppleClang Xode Version at least v8
• Nvidia CUDA Toolkit and Compiler (nvcc) at least v10.0 with -std=c++14

See the C++ compiler support status.

Cpp-Taskflow uses CMake to build examples and unit tests. We recommend using out-of-source build.

~$ cmake --version   # must be at least 3.9 or higher
~$ mkdir build
~$ cd build
~$ cmake ../ 
~$ make & make test  # run all unit tests

The folder examples/ contains several examples and is a great place to learn to use Cpp-Taskflow.

Please visit benchmarks to learn to compile the benchmarks.

Cpp-Taskflow is being used in both industry and academic projects to scale up existing workloads that incorporate complex task dependencies.

• OpenTimer: A High-performance Timing Analysis Tool for Very Large Scale Integration (VLSI) Systems
• DtCraft: A General-purpose Distributed Programming Systems using Data-parallel Streams
• Firestorm: Fighting Game Engine with Asynchronous Resource Loaders (developed by ForgeMistress)
• Shiva: An extensible engine via an entity component system through scripts, DLLs, and header-only (C++)
• PID Framework: A Global Development Methodology Supported by a CMake API and Dedicated C++ Projects
• NovusCore: An emulating project for World of Warraft (Wrath of the Lich King 3.3.5a 12340 client build)
• SA-PCB: Annealing-based Printed Circuit Board (PCB) Placement Tool
• LPMP: A C++ framework for developing scalable Lagrangian decomposition solvers for discrete optimization problems
• Heteroflow: A Modern C++ Parallel CPU-GPU Task Programming Library
• OpenPhySyn: A plugin-based physical synthesis optimization kit as part of the OpenRoad flow

More...

Cpp-Taskflow is being actively developed and contributed by the these people. Meanwhile, we appreciate the support from many organizations for our developments.

Cpp-Taskflow is licensed under the MIT License.