C++ 中的并行循环

Parallel Loops in C++

提问by Exagon

I wonder if there is a light, straight forward way to compute loops such as for and range based for loops in parallel in C++. How would you implement such a thing? From Scala I know the map, filter and foreach functions maybe it would also possible to perform these parallel? Is there an easy way to achieve this in C++. My primary plattform is Linux but it would be nice if it works cross-plattform.

我想知道是否有一种简单、直接的方法可以在 C++ 中并行计算循环，例如 for 和基于范围的 for 循环。你会如何实施这样的事情？从 Scala 我知道 map、filter 和 foreach 函数也许也可以并行执行这些？有没有一种简单的方法可以在 C++ 中实现这一点。我的主要平台是 Linux，但如果它可以跨平台工作，那就太好了。

回答by Exagon

With the parallel algorithms in C++17 we can now use:

使用 C++17 中的并行算法，我们现在可以使用：

std::vector<std::string> foo;
std::for_each(
    std::execution::par_unseq,
    foo.begin(),
    foo.end(),
    [](auto&& item)
    {
        //do stuff with item
    });

to compute loops in parallel. The first parameter specifies the execution policy

并行计算循环。第一个参数指定执行策略

回答by Daniel Langr

What is your platform? You can look at OpenMP, though it's not a part of C++. But it is widely supported by compilers.

你的平台是什么？您可以查看OpenMP，尽管它不是 C++ 的一部分。但它得到了编译器的广泛支持。

As for range-based for loops, see, e.g., Using OpenMP with C++11 range-based for loops?.

至于基于范围的 for 循环，请参阅例如使用 OpenMP 和 C++11 基于范围的 for 循环？.

I've also seen few documents at http://www.open-std.orgthat indicate some efforts to incorporate parallel constructs/algorithms into future C++, but don't know what's their current status.

我还在http://www.open-std.org上看到了一些表明将并行构造/算法合并到未来 C++ 中的努力的文档，但不知道它们目前的状态。

UPDATE

更新

Just adding some exemplary code:

只需添加一些示例代码：

template <typename RAIter>
void loop_in_parallel(RAIter first, RAIter last) {
   const size_t n = std::distance(first, last);

   #pragma omp parallel for
   for (size_t i = 0; i < n; i++) {
       auto& elem = *(first + i);
       // do whatever you want with elem
    }
}

The number of threads can be set at runtime via the OMP_NUM_THREADSenvironment variable.

线程数可以在运行时通过OMP_NUM_THREADS环境变量设置。

回答by bobah

std::asyncmay be a good fit here, if you are happy to let the C++runtime control the parallelism.

std::async如果您愿意让C++运行时控制并行度，可能很适合这里。

Example from the cppreference.com:

来自 cppreference.com 的示例：

#include <iostream>
#include <vector>
#include <algorithm>
#include <numeric>
#include <future>

template <typename RAIter>
int parallel_sum(RAIter beg, RAIter end)
{
    auto len = end - beg;
    if(len < 1000)
        return std::accumulate(beg, end, 0);

    RAIter mid = beg + len/2;
    auto handle = std::async(std::launch::async,
                              parallel_sum<RAIter>, mid, end);
    int sum = parallel_sum(beg, mid);
    return sum + handle.get();
}

int main()
{
    std::vector<int> v(10000, 1);
    std::cout << "The sum is " << parallel_sum(v.begin(), v.end()) << '\n';
}

回答by arkan

With C++11 you can parallelize a for loop with only a few lines of codes. This splits a for loop into smaller chunks and assign each sub loop to a thread:

使用 C++11，您只需几行代码就可以并行化 for 循环。这将 for 循环拆分为更小的块并将每个子循环分配给一个线程：

/// Basically replacing:
void sequential_for(){
    for(int i = 0; i < nb_elements; ++i)
        computation(i);
}

/// By:
void threaded_for(){
    parallel_for(nb_elements, [&](int start, int end){ 
        for(int i = start; i < end; ++i)
            computation(i); 
    } );
}

Or withing a class:

或与一个类：

struct My_obj {

    /// Replacing:
    void sequential_for(){
        for(int i = 0; i < nb_elements; ++i)
            computation(i);
    }

    /// By:
    void threaded_for(){
        parallel_for(nb_elements, [this](int s, int e){ this->process_chunk(s, e); } );
    }

    void process_chunk(int start, int end)
    {
        for(int i = start; i < end; ++i)
            computation(i);
    }
};

To do this, you only need to put the code below in a header file and use it at will:

为此，您只需要将下面的代码放在一个头文件中并随意使用即可：

#include <algorithm>
#include <thread>
#include <functional>
#include <vector>

/// @param[in] nb_elements : size of your for loop
/// @param[in] functor(start, end) :
/// your function processing a sub chunk of the for loop.
/// "start" is the first index to process (included) until the index "end"
/// (excluded)
/// @code
///     for(int i = start; i < end; ++i)
///         computation(i);
/// @endcode
/// @param use_threads : enable / disable threads.
///
///
static
void parallel_for(unsigned nb_elements,
                  std::function<void (int start, int end)> functor,
                  bool use_threads = true)
{
    // -------
    unsigned nb_threads_hint = std::thread::hardware_concurrency();
    unsigned nb_threads = nb_threads_hint == 0 ? 8 : (nb_threads_hint);

    unsigned batch_size = nb_elements / nb_threads;
    unsigned batch_remainder = nb_elements % nb_threads;

    std::vector< std::thread > my_threads(nb_threads);

    if( use_threads )
    {
        // Multithread execution
        for(unsigned i = 0; i < nb_threads; ++i)
        {
            int start = i * batch_size;
            my_threads[i] = std::thread(functor, start, start+batch_size);
        }
    }
    else
    {
        // Single thread execution (for easy debugging)
        for(unsigned i = 0; i < nb_threads; ++i){
            int start = i * batch_size;
            functor( start, start+batch_size );
        }
    }

    // Deform the elements left
    int start = nb_threads * batch_size;
    functor( start, start+batch_remainder);

    // Wait for the other thread to finish their task
    if( use_threads )
        std::for_each(my_threads.begin(), my_threads.end(), std::mem_fn(&std::thread::join));
}

Lastly you could define a macro to get even more compact expression:

最后，您可以定义一个宏以获得更紧凑的表达式：

#define PARALLEL_FOR_BEGIN(nb_elements) tbx::parallel_for(nb_elements, [&](int start, int end){ for(int i = start; i < end; ++i)
#define PARALLEL_FOR_END()})

PARALLEL_FOR_BEGIN(nb_edges)
{
    computation(i);
}PARALLEL_FOR_END();

回答by uSeemSurprised

This can be done using threadsspecifically pthreadslibrary function that can be used to perform operations concurrently.

这可以使用可用于并发执行操作的threads特定pthreads库函数来完成。

You can read more about them here : http://www.tutorialspoint.com/cplusplus/cpp_multithreading.htm

您可以在此处阅读有关它们的更多信息：http: //www.tutorialspoint.com/cplusplus/cpp_multithreading.htm

std::thread can also be used : http://www.cplusplus.com/reference/thread/thread/

std::thread 也可以使用：http: //www.cplusplus.com/reference/thread/thread/

Below is a code in which i use the thread id of each thread to split the array into two halves :

下面是一个代码，其中我使用每个线程的线程 id 将数组分成两半：

#include <iostream>
#include <cstdlib>
#include <pthread.h>

using namespace std;

#define NUM_THREADS 2

int arr[10] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};

void *splitLoop(void *threadid)
{
   long tid;
   tid = (long)threadid;
   //cout << "Hello World! Thread ID, " << tid << endl;
   int start = (tid * 5);
   int end = start + 5;
   for(int i = start;i < end;i++){
      cout << arr[i] << " ";
   }
   cout << endl;
   pthread_exit(NULL);
}

int main ()
{
   pthread_t threads[NUM_THREADS];
   int rc;
   int i;
   for( i=0; i < NUM_THREADS; i++ ){
      cout << "main() : creating thread, " << i << endl;
      rc = pthread_create(&threads[i], NULL, 
                          splitLoop, (void *)i);
      if (rc){
         cout << "Error:unable to create thread," << rc << endl;
         exit(-1);
      }
   }
   pthread_exit(NULL);
}

Also remember while compiling you have to use the -lpthreadflag.

还要记住，在编译时必须使用-lpthread标志。

Link to solution on Ideone : http://ideone.com/KcsW4P

Ideone 解决方案链接：http://ideone.com/KcsW4P

回答by doctorlai

The Concurrency::parallel_for (PPL) is also one of the nice opions to do task parallelism.

Concurrency::parallel_for (PPL) 也是实现任务并行的不错选择之一。

Taken from C++ Coding Exercise – Parallel For – Monte Carlo PI Calculation

取自C++ 编码练习 – Parallel For – Monte Carlo PI Calculation

int main() {
    srand(time(NULL)); // seed
    const int N1 = 1000;
    const int N2 = 100000;
    int n = 0;
    int c = 0;
    Concurrency::critical_section cs;
    // it is better that N2 >> N1 for better performance
    Concurrency::parallel_for(0, N1, [&](int i) {
        int t = monte_carlo_count_pi(N2);
        cs.lock(); // race condition
        n += N2;   // total sampling points
        c += t;    // points fall in the circle
        cs.unlock();
    });
    cout < < "pi ~= " << setprecision(9) << (double)c / n * 4.0 << endl;
    return 0;
}

回答by Adam

As this thread has been my answer almost everytime i've looked for a method to paralleize something, i've decided to add a bit to it. based on the method by arkan(see above).

由于几乎每次我都在寻找一种方法来并行化某些东西时，这个线程一直是我的答案，因此我决定为其添加一些内容。基于 arkan 的方法（见上文）。

The two next methods are almost the same and allow a simple syntax. Simply include the header file in your project and call one of the parallel version:

接下来的两个方法几乎相同，并且允许使用简单的语法。只需在您的项目中包含头文件并调用并行版本之一：

example:

例子：

#include "par_for.h"

int main() {
//replace - 
for(unsigned i = 0; i < 10; ++i){
    std::cout << i << std::endl;
}

//with -
//method 1:
pl::thread_par_for(0, 10, [&](unsigned i){
            std::cout << i << std::endl;   //do something here with the index i
        });   //changing the end to },false); will make the loop sequantial

//or method 2:
pl::async_par_for(0, 10, [&](unsigned i){
            std::cout << i << std::endl;   //do something here with the index i
        });   //changing the end to },false); will make the loop sequantial

return 0;
}

header file - par_for.h:

头文件 - par_for.h：

#include <thread>
#include <vector>
#include <functional>
#include <future>

using namespace std;

namespace pl{

    void thread_par_for(unsigned start, unsigned end, function<void(unsigned i)> fn, bool par = true){

        //internal loop
        auto int_fn = [&fn](unsigned int_start, unsigned seg_size){
            for (unsigned j = int_start; j < int_start+seg_size; j++){
                fn(j);
            }
        };

        //sequenced for
        if(!par){
            return int_fn(start, end);
        }

        //get number of threads
        unsigned nb_threads_hint = thread::hardware_concurrency();
        unsigned nb_threads = nb_threads_hint == 0 ? 8 : (nb_threads_hint);

        //calculate segments
        unsigned total_length = end - start;
        unsigned seg = total_length/nb_threads;
        unsigned last_seg = seg + total_length%nb_threads;

        //launch threads - parallel for
        auto threads_vec = vector<thread>();
        threads_vec.reserve(nb_threads);
        for(int k = 0; k < nb_threads-1; ++k){
            unsigned current_start = seg*k;
            threads_vec.emplace_back(thread(int_fn, current_start, seg));
        }
        {
            unsigned current_start = seg*(nb_threads-1);
            threads_vec.emplace_back(thread(int_fn, current_start, last_seg));
        }
        for (auto& th : threads_vec){
            th.join();
        }
    }




    void async_par_for(unsigned start, unsigned end, function<void(unsigned i)> fn, bool par = true){

        //internal loop
        auto int_fn = [&fn](unsigned int_start, unsigned seg_size){
            for (unsigned j = int_start; j < int_start+seg_size; j++){
                fn(j);
            }
        };

        //sequenced for
        if(!par){
            return int_fn(start, end);
        }

        //get number of threads
        unsigned nb_threads_hint = thread::hardware_concurrency();
        unsigned nb_threads = nb_threads_hint == 0 ? 8 : (nb_threads_hint);

        //calculate segments
        unsigned total_length = end - start;
        unsigned seg = total_length/nb_threads;
        unsigned last_seg = seg + total_length%nb_threads;

        //launch threads - parallel for
        auto fut_vec = vector<future<void>>();
        fut_vec.reserve(nb_threads);
        for(int k = 0; k < nb_threads-1; ++k){
            unsigned current_start = seg*k;
            fut_vec.emplace_back(async(int_fn, current_start, seg));
        }
        {
            unsigned current_start = seg*(nb_threads-1);
            fut_vec.emplace_back(async(launch::async, int_fn, current_start, last_seg));
        }
        for (auto& th : fut_vec){
            th.get();
        }
    }
}

Some simple tests suggest the method with async is faster, probably because the standrad library controlls whether to actually launch a new thread or not.

一些简单的测试表明使用 async 的方法更快，可能是因为标准库控制是否实际启动一个新线程。