Taskflow

Taskflow helps you quickly write parallel and heterogeneous task programs in modern C++

Why Taskflow?

Taskflow is faster, more expressive, and easier for drop-in integration than many of existing task programming frameworks in handling complex parallel workloads.

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.

Static Tasking	Dynamic Tasking

Taskflow supports conditional tasking for you to make rapid control-flow decisions across dependent tasks to implement cycles and conditions that were otherwise difficult to do with existing tools.

Conditional Tasking

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

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

Taskflow provides visualization and tooling needed for profiling Taskflow programs.

Taskflow Profiler

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

"Taskflow is the cleanest Task API I've ever seen." Damien Hocking @Corelium Inc
"Taskflow has a very simple and elegant tasking interface. The performance also scales very well." Glen Fraser
"Taskflow lets me handle parallel processing in a smart way." Hayabusa @Learning
"Taskflow improves the throughput of our graph engine in just a few hours of coding." Jean-Michaël @KDAB
"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 Taskflow. Technical details can be referred to our IPDPS paper.

Start Your First Taskflow Program

The following program (simple.cpp) creates four tasks A, B, C, and D, where A runs before B and C, and D runs after B and C. When A finishes, B and C can run in parallel.

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

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

  auto [A, B, C, D] = taskflow.emplace(  // create four tasks
    [] () { std::cout << "TaskA\n"; },
    [] () { std::cout << "TaskB\n"; },
    [] () { std::cout << "TaskC\n"; },
    [] () { std::cout << "TaskD\n"; } 
  );                                  
                                      
  A.precede(B, C);  // A runs before B and C
  D.succeed(B, C);  // D runs after  B and C
                                      
  executor.run(taskflow).wait(); 

  return 0;
}

Taskflow is header-only and there is no wrangle with installation. To compile the program, clone the Taskflow project and tell the compiler to include the headers.

~$ git clone https://github.com/taskflow/taskflow.git  # clone it only once
~$ g++ -std=c++17 simple.cpp -I taskflow/taskflow -O2 -pthread -o simple
~$ ./simple
TaskA
TaskC 
TaskB 
TaskD

Visualize Your First Taskflow Program

Taskflow comes with a built-in profiler, TFProf, for you to profile and visualize taskflow programs in an easy-to-use web-based interface.

# run the program with the environment variable TF_ENABLE_PROFILER enabled
~$ TF_ENABLE_PROFILER=simple.json ./simple
~$ cat simple.json
[
{"executor":"0","data":[{"worker":0,"level":0,"data":[{"span":[172,186],"name":"0_0","type":"static"},{"span":[187,189],"name":"0_1","type":"static"}]},{"worker":2,"level":0,"data":[{"span":[93,164],"name":"2_0","type":"static"},{"span":[170,179],"name":"2_1","type":"static"}]}]}
]
# paste the profiling json data to https://taskflow.github.io/tfprof/

In addition to execution diagram, you can dump the graph to a DOT format and visualize it using a number of free GraphViz tools.

// dump the taskflow graph to a DOT format through std::cout
taskflow.dump(std::cout);

Express Task Graph Parallelism

Taskflow empowers users with both static and dynamic task graph constructions to express end-to-end parallelism in a task graph that embeds in-graph control flow.

Create a Subflow Graph
Integrate Control Flow to a Task Graph
Offload a Task to a GPU
Compose Task Graphs
Launch Asynchronous Tasks
Execute a Taskflow
Leverage Standard Parallel Algorithms

Create a Subflow Graph

Taskflow supports dynamic tasking for you to create a subflow graph from the execution of a task to perform dynamic parallelism. The following program spawns a task dependency graph parented at task B.

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

tf::Task B = taskflow.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");  
  B3.succeed(B1, B2);  // B3 runs after B1 and B2
}).name("B");

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

Integrate Control Flow to a Task Graph

Taskflow supports conditional tasking for you to make rapid control-flow decisions across dependent tasks to implement cycles and conditions in an end-to-end task graph.

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

// creates a condition task that returns a random binary
tf::Task cond = taskflow.emplace(
  [](){ return std::rand() % 2; }
).name("cond");

init.precede(cond);

// creates a feedback loop {0: cond, 1: stop}
cond.precede(cond, stop);

Offload a Task to a GPU

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

__global__ void saxpy(size_t N, float alpha, float* dx, float* dy) {
  int i = blockIdx.x*blockDim.x + threadIdx.x;
  if (i < n) {
    y[i] = a*x[i] + y[i];
  }
}
tf::Task cudaflow = taskflow.emplace([&](tf::cudaFlow& cf) {

  // data copy tasks
  tf::cudaTask h2d_x = cf.copy(dx, hx.data(), N).name("h2d_x");
  tf::cudaTask h2d_y = cf.copy(dy, hy.data(), N).name("h2d_y");
  tf::cudaTask d2h_x = cf.copy(hx.data(), dx, N).name("d2h_x");
  tf::cudaTask d2h_y = cf.copy(hy.data(), dy, N).name("d2h_y");
  
  // kernel task with parameters to launch the saxpy kernel
  tf::cudaTask saxpy = cf.kernel(
    (N+255)/256, 256, 0, saxpy, N, 2.0f, dx, dy
  ).name("saxpy");

  saxpy.succeed(h2d_x, h2d_y)
       .precede(d2h_x, d2h_y);
}).name("cudaFlow");

Taskflow also supports SYCL, a general-purpose heterogeneous programming model, to program GPU tasks in a single-source C++ environment using the task graph-based approach.

tf::Task syclflow = taskflow.emplace_on([&](tf::syclFlow& sf){
  tf::syclTask h2d_x = cf.copy(dx, hx.data(), N).name("h2d_x");
  tf::syclTask h2d_y = cf.copy(dy, hy.data(), N).name("h2d_y");
  tf::syclTask d2h_x = cf.copy(hx.data(), dx, N).name("d2h_x");
  tf::syclTask d2h_y = cf.copy(hy.data(), dy, N).name("d2h_y");
  tf::syclTask saxpy = sf.parallel_for(sycl::range<1>(N), 
    [=] (sycl::id<1> id) {
      dx[id] = 2.0f * dx[id] + dy[id];
    }
  ).name("saxpy");
  saxpy.succeed(h2d_x, h2d_y)
       .precede(d2h_x, d2h_y);
}, sycl_queue).name("syclFlow");

Compose Task Graphs

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.

tf::Taskflow f1, f2;

// create taskflow f1 of two tasks
tf::Task f1A = f1.emplace([]() { std::cout << "Task f1A\n"; })
                 .name("f1A");
tf::Task f1B = f1.emplace([]() { std::cout << "Task f1B\n"; })
                 .name("f1B");

// create taskflow f2 with one module task composed of f1
tf::Task f2A = f2.emplace([]() { std::cout << "Task f2A\n"; })
                 .name("f2A");
tf::Task f2B = f2.emplace([]() { std::cout << "Task f2B\n"; })
                 .name("f2B");
tf::Task f2C = f2.emplace([]() { std::cout << "Task f2C\n"; })
                 .name("f2C");

tf::Task f1_module_task = f2.composed_of(f1)
                            .name("module");

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

Launch Asynchronous Tasks

Taskflow supports asynchronous tasking. You can launch tasks asynchronously to incorporate independent, dynamic parallelism in your taskflows.

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

// create asynchronous tasks directly from an executor
tf::future<std::optional<int>> future = executor.async([](){ 
  std::cout << "async task returns 1\n";
  return 1;
}); 
executor.silent_async([](){ std::cout << "async task of no return\n"; });

// launch an asynchronous task from a running task
taskflow.emplace([&](){
  executor.async([](){ std::cout << "async task within a task\n"; });
});

executor.run(taskflow).wait();

Execute a Taskflow

The executor provides several thread-safe methods to run a taskflow. You can run a taskflow once, multiple times, or until a stopping criteria is met. These methods are non-blocking with a tf::future<void> return to let you query the execution status.

// runs the taskflow once
tf::Future<void> run_once = executor.run(taskflow); 

// wait on this run to finish
run_once.get();

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

// runs the taskflow five times
executor.run_until(taskflow, [counter=5](){ return --counter == 0; });

// block the executor until all submitted taskflows complete
executor.wait_for_all();

Leverage Standard Parallel Algorithms

Taskflow defines algorithms for you to quickly express common parallel patterns using standard C++ syntaxes, such as parallel iterations, parallel reductions, and parallel sort.

// standard parallel CPU algorithms
tf::Task task1 = taskflow.for_each( // assign each element to 100 in parallel
  first, last, [] (auto& i) { i = 100; }    
);
tf::Task task2 = taskflow.reduce(   // reduce a range of items in parallel
  first, last, init, [] (auto a, auto b) { return a + b; }
);
tf::Task task3 = taskflow.sort(     // sort a range of items in parallel
  first, last, [] (auto a, auto b) { return a < b; }
);

// standard parallel GPU algorithms
tf::cudaTask cuda1 = cudaflow.for_each( // assign each element to 100 on GPU
  dfirst, dlast, [] __device__ (auto i) { i = 100; }
);
tf::cudaTask cuda2 = cudaflow.reduce(   // reduce a range of items on GPU
  dfirst, dlast, init, [] __device__ (auto a, auto b) { return a + b; }
);
tf::cudaTask cuda3 = cudaflow.sort(     // sort a range of items on GPU
  dfirst, dlast, [] __device__ (auto a, auto b) { return a < b; }
);

Supported Compilers

To use Taskflow, you only need a compiler that supports C++17:

GNU C++ Compiler at least v8.4 with -std=c++17
Clang C++ Compiler at least v6.0 with -std=c++17
Microsoft Visual Studio at least v19.27 with /std:c++17
AppleClang Xode Version at least v12.0 with -std=c++17
Nvidia CUDA Toolkit and Compiler (nvcc) at least v11.1 with -std=c++17
Intel C++ Compiler at least v19.0.1 with -std=c++17
Intel DPC++ Clang Compiler at least v13.0.0 with -std=c++17 and SYCL20

Taskflow works on Linux, Windows, and Mac OS X.

Learn More about Taskflow

Visit our project website and documentation to learn more about Taskflow. To get involved:

See release notes to stay up-to-date with newest versions
Read the step-by-step tutorial at cookbook
Submit an issue at GitHub issues
Find out our technical details at references
Watch our technical talks at YouTube

CppCon20 Tech Talk	MUC++ Tech Talk

We are committed to support trustworthy developments for both academic and industrial research projects in parallel and heterogeneous computing. At the same time, we appreciate all Taskflow contributors!

License

Taskflow is licensed with the MIT License. You are completely free to re-distribute your work derived from Taskflow.

ujfj1986/taskflow