Taskflow

Modern C++ Parallel Task Programming

Project GitHub

API Documentation

Profiler

Presented by Tsung-Wei Huang

How can we make it easier for C++ developers to quickly write parallel programs?

Want Parallel Code

• Simple
• Expressive
• Transparent
• Performant
• Productive

It's NOT Easy Though ...

1. Task Dependencies
2. Data Race
3. Concurrency Controls
4. Thread Contention
5. Debugging
6. ...

A Homegrown Example

C++ Thread

std::atomic_int garnish_ready{0}; // dependency variable
std::atomic_int meat_ready{0};    // dependency variable
std::atomic_int plating_ready{0}; // dependency variable
std::thread cook1 ([&] () { 
  garnish = cook_garnish(); garnish_ready = 1; 
});
std::thread cook2 ([&] () { 
  meat = cook_meat(); meat_ready = 1; 
});
std::thread chief ([&] () {
  while(!garnish_ready && !meat_ready);    // spinning
  plates = plating(garnish, meat); plating_ready = 1;
}); 
std::thread waiter1 ([&] () {
  while(!plating_ready); serve(plates[0]); // spinning
});
std::thread waiter2 ([&] () {
  while(!plating_ready); serve(plates[1]); // spinning
});

OpenMP

#pragma omp parallel 
{
  #pragma omp single 
  {
    int c1c, c2c, cw1, cw2;
    #pragma omp task depend(out:c1c)
    garnish = cook_garnish();
    #pragma omp task depend(out:c2c)
    meat = cook_meat();
    #pragma omp task depend(in:c1c,c2c) depend(out:cw1,cw2)
    plates = plating(garnish, meat);
    #pragma omp task depend(in:cw1)
    serve(plates[0]);
    #pragma omp task depend(in:cw2)
    serve(plates[1]);
  }
}

Intel TBB

using namespace tbb::flow;
task_scheduler_init init(default_num_threads());
graph g;

continue_node<continue_msg> cook1(g, [&] (auto) { garnish = cook_garnish(); });
continue_node<continue_msg> cook2(g, [&] (auto) { meat = cook_meat(); });
continue_node<continue_msg> chief(g, [&] (auto) { plates = plating(garnish, meat); });
continue_node<continue_msg> waiter1(g, [&] (auto) { serve(plates[0]); });
continue_node<continue_msg> waiter2(g, [&] (auto) { serve(plates[1]); });
make_edge(cook1, chief); 
make_edge(cook2, chief); 
make_edge(chief, waiter1);
make_edge(chief, waiter2);

cook1.try_put(continue_msg());    // start at cook1
cook2.try_put(continue_msg());    // start at cook2
g.wait_for_all();

Taskflow

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

auto [cook1, cook2, chief, waiter1, waiter2] = 
  taskflow.emplace(
  [&] () { garnish = cook_garnish();         },
  [&] () { meat    = cook_meat();            },
  [&] () { plates  = plating(garnish, meat); },
  [&] () { serve(plates[0]);                 },
  [&] () { serve(plates[1]);                 }
);

cook1.precede(chief);        // cook1 runs before chief
cook2.precede(chief);        // cook2 runs before chief
chief.precede(waiter1);      // chief runs before waiter1
chief.precede(waiter2);      // chief runs before waiter2

executor.run(tf);

Taskflow is FREE

• from explicit thread management
• from difficult lock mechanism
• from daunting class declaration

Development Cost

What about Performance?

Graph Algorithm

Matrix Operation

Real Gain is Tremendous

VLSI Timing Analysis

Runtime Performance

Parallel Scaling Performance

Dynamic Tasking

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

auto B = tf.emplace([] (auto& subflow) {
  auto B0 = subflow.emplace([](){}).name("B0");
  auto B1 = subflow.emplace([](){}).name("B1");
  auto B2 = subflow.emplace([](){}).name("B2");
  auto B3 = subflow.emplace([](){}).name("B3");
  auto B4 = subflow.emplace([](){}).name("B4");
  B0.precede(B1, B2, B3);
  B4.succeed(B1, B2, 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 

Conditional Tasking

tf::Taskflow tf;
auto A = tf.emplace([](){}).name("A");
auto B = tf.emplace([](){}).name("B");
auto C = tf.emplace([](){}).name("C");
auto D = tf.emplace([](){}).name("D");
auto E = tf.emplace([](){ return rand()%3; }).name("E");
auto F = tf.emplace([](){}).name("F");

A.precede(B, C);  // A runs before B and C
B.precede(D);     // B runs before D
C.precede(F);     // C runs before F
D.precede(B);     // D conditions B on 0 (feedback)
D.precede(D);     // D conditions D on 1 (self-loop)
D.precede(E);     // D conditions E on 2

Composability

tf::Taskflow A, B;
auto [taskA1, taskA2, taskA3] = A.emplace(
  []() { std::cout << "Task A1\n"; },
  []() { std::cout << "Task A2\n"; },
  []() { std::cout << "Task A3\n"; }
);
taskA1.precede(taskA2, taskA3);

auto [taskB1, taskB2, taskB3] = B.emplace(
  []() { std::cout << "Task B1\n"; },
  []() { std::cout << "Task B2\n"; },
  []() { std::cout << "Task B3\n"; }
);
// Compose taskflow B
auto module_A = B.composed_of(A); 
taskB1.precede(module_A);
module_A.precede(taskB2, taskB3);

Concurrent CPU-GPU Tasking

tf::Taskflow taskflow;
tf::Executor executor;
auto allocate_x = taskflow.emplace(
  [&](){ cudaMalloc(&dx, N*sizeof(float));}
);
auto allocate_y = taskflow.emplace(
  [&](){ cudaMalloc(&dy, N*sizeof(float));}
);
auto cudaflow = taskflow.emplace([&](tf::cudaFlow& cf) {
  auto h2d_x = cf.copy(dx, hx.data(), N);
  auto h2d_y = cf.copy(dy, hy.data(), N);
  auto d2h_x = cf.copy(hx.data(), dx, N);
  auto d2h_y = cf.copy(hy.data(), dy, N);
  auto op = cf.kernel(N/256, 256, 0, saxpy, N, 2, dx, dy);
  op.succeed(h2d_x, h2d_y).precede(d2h_x, d2h_y);
});
cudaflow.succeed(allocate_x, allocate_y);
executor.run(taskflow).wait();

Monitor Thread Activities

# visit https://github.com/taskflow/tfprof
~$ python3 tfprof.py -o output.tfp taskflow-program args
~$ cat output.tfp
[ 
  ...
]
# paste the content to https://taskflow.github.io/tfprof/

Taskflow Profiler

https://taskflow.github.io/tfprof/

Drop-in Integration

# clone the newest Taskflow
~$ git clone https://github.com/taskflow/taskflow.git

# Taskflow is header-only
~$ cp -r taskflow/taskflow my_project/

# compile you code with g++, clang++, or msvc
~$ g++ -std=c++17 my_project/test.cpp -pthread

We ♥ Feedback

Taskflow is the cleanest Task API I've ever seen.

Taskflow has a very simple and elegant tasking interface. The performance also scales very well.

Best Poster Award in the official CPP Conference, 2018

Second Prize of Open Source Software Competition in ACM Multimedia Conference, 2019

Acknowledgment

• Development Team
• Contributors
• Sponsors (NSF, DARPA)
• ... and all users!