boost::io_service post using a lambda

Question

This question relates to this previous question.

I've implemented the code published there by Richard Hodges. The code published works for me when I use g++ (Debian 4.8.4-1) 4.8.4.

However, the implementation is part of a CUDA library, and I am stuck with CUDA 6.5 which unofficially supported C++11 features.

When I use the code Richard posted:

template <class F>
void submit( F&& f)
{
    std::unique_lock<std::mutex> lock(_cvm);
    ++ _tasks;
    lock.unlock();
    _io_service.post(
    [this, f = std::forward<F>(f)]
                     {
                         f();
                         reduce();
                     });
}

I get an error: error: expected a "]" referring to the lambda line. This makes me think that the header is not being parsed properly. I tried without the template, just passing a reference to my worker class, and without the forwarding.

void submit( trainer & job)
{
    std::unique_lock<std::mutex> lock(_cvm);
    ++ _tasks;
    lock.unlock();
    _io_service.post([this,&]
                     {
                         job();
                         reduce();
                     });
}

And I got error: an enclosing-function local variable cannot be referenced in a lambda body unless it is in the capture list.

So I explicitly added both this and job:

void submit( trainer & job)
{
    std::unique_lock<std::mutex> lock(_cvm);
    ++ _tasks;
    lock.unlock();
    _io_service.post([this,&job]
                     {
                         job();
                         reduce();
                     });
}

At which point, I got stuck at error:

error: could not convert ‘{{((cuANN::trainer_pool*)this)->cuANN::trainer_pool::_io_service}}’ from ‘<brace-enclosed initializer list>’ to ‘boost::asio::io_service::work’ boost::asio::io_service::work _work { _io_service };

FYI, cuANN::trainer_pool is the worker_pool in Richard's example, and the thread pool implementation, and _io_service is simply a member of the class trainer_pool:

class trainer_pool
{   
public:
    trainer_pool ( unsigned int max_threads );
    void start();
    void wait();
    void stop(); 
    void thread_proc();
    void reduce();
    void submit( trainer & job);

private:
    unsigned int _max_threads_;
    boost::asio::io_service _io_service;
    boost::asio::io_service::work _work { _io_service };
    std::vector<std::thread> _threads;
    std::condition_variable _cv;
    std::mutex _cvm;
    size_t _tasks = 0;
};

1. What am I doing wrong?
2. How can I do it using std::bind or boost::bind instead of using a lambda?

PS: the code on http://ideone.com/g38Z4H is my skeleton for g++ (which works). The http://ideone.com/d7Nkop saved as host.cu also showcases the issue.

nvcc -std=c++11 host.cu -lboost_thread -lboost_system -lpthread -o host

Answer 1

this construct:

[this, f = std::forward<F>(f)]

is c++14 and won't compile in c++11.

using [this, &] would be a mistake since there's no guarantee that the function object will still exist (you captured it by reference)

in c++11 consider [this, f] to take a copy.

edit: just realised that job is mutable, therefore:

void submit(trainer & job)
{
    std::unique_lock<std::mutex> lock(_cvm);
    ++ _tasks;
    lock.unlock();
    _io_service.post([this,job]()->void mutable
                     {
                         job();
                         reduce();
                     });
}

Answer 2

I found the problem. NVCC and CUDA works fine with the lambdas.

The offending line was in the header:

boost::asio::io_service::work _work { _io_service };

CUDA 6.5 does not like the initializer list. Moving it to the constructor:

trainer_pool (unsigned int max_threads)
:_max_threads_(max_threads), _work(_io_service)
{}

Seems to have fixed the error and the code now compiles. Many thanks to all who helped.