/*

2019/02/25 - contributed by Paolo Bolzoni

cpp-taskflow works on directed acyclic graphs.

And here we want to pass information between the flow elements.

To do so, we see the cpp-taskflow arcs as objects memory where the functions on

the nodes read from or write to.

The function on every node will *read from* the objects of memory of its

incoming edges and *write to* the objects of its outcoming edges.

The cpp-taskflow semantics ensures the synchronization.

Nodes without incoming edges will require the input from somewhere else;

instead nodes without outcoming edges have to execute some side effects to be

useful.

In this example we fill up (in parallel) two vectors of the results of a fair

percentile die and we pick up the maximum values from each cell, and output the

result.

.----------------.

| fill in vector |----|

'----------------' |->.-------------. .-----------------.

| pick up max |---->| print in stdout |

.----------------. |->'-------------' '-----------------'

| fill in vector |----|

'----------------'

The output will be twenty random integer between 1 and 100, that are clearly

not uniform distributed as they favor larger numbers.

The code assumes the taskflow is executed once, when using the Framework

feature the programmer needs care to keep the invariants.

It is then suggested to use const references (eg., vector<int> const&) for the

objects related to the incoming arcs and references for outcoming ones.

*/

#include <taskflow/taskflow.hpp>

//All those includes are just to init the mersenne twister

#include <array>

#include <algorithm>

#include <functional>

#include <random>

//until here

#include <vector>

#include <iostream>

std::mt19937 init_mersenne_twister() {

std::array<std::uint32_t, std::mt19937::state_size> seed_bits{};

std::random_device real_random{};

std::generate(seed_bits.begin(), seed_bits.end(), std::ref(real_random));

std::seed_seq wrapped_seed_bits(seed_bits.begin(), seed_bits.end());

return std::mt19937(wrapped_seed_bits);

}

class Fill_in_vector {

public:

Fill_in_vector(std::vector<int>& v, int length)

: v_{v}, length_{length} {}

void operator()() {

auto rng = init_mersenne_twister();

std::uniform_int_distribution<int> percentile_die(1, 100);

//the taskflow is used only once, so we can mess up with length_ value

while (length_ > 0) {

--length_;

v_.push_back( percentile_die(rng) );

}

}

private:

std::vector<int>& v_;

int length_;

};

class Pick_up_max {

public:

Pick_up_max(std::vector<int>& in1, std::vector<int>& in2, std::vector<int>& out)

: in1_{in1}, in2_{in2}, out_{out} {}

void operator()() {

for (std::vector<int>::size_type i{}, e = in1_.size(); i < e; ++i) {

in1_[i] = std::max(in1_[i], in2_[i]);

}

// the taskflow is executed once, so we avoid one copy

out_.swap(in1_);

}

private:

std::vector<int>& in1_;

std::vector<int>& in2_;

std::vector<int>& out_;

};

class Print {

public:

Print(std::vector<int> const& v)

: v_{v} {}

void operator()() {

bool first{ true };

for (auto i : v_) {

if (!first) {

std::cout << ", ";

}

std::cout << i;

first = false;

}

std::cout << "\n";

}

private:

std::vector<int> const& v_;

};

int main() {

// Set up the memory for the arcs

std::vector<int> in1{}, in2{}, out{};

// Prepare the functors for taskflow

tf::Taskflow tf;

auto [

fill_in_vector1,

fill_in_vector2,

pick_up_max,

print

] = tf.emplace(

Fill_in_vector(in1, 20),

Fill_in_vector(in2, 20),

Pick_up_max(in1, in2, out),

Print(out)

);

// Set up dependencies

fill_in_vector1.precede(pick_up_max);

fill_in_vector2.precede(pick_up_max);

pick_up_max.precede(print);

// Execution

tf.wait_for_all();

return 0;

}