Hi, there is a long time I have not written anything, sorry for that! I planned to write several articles during these few next weeks. This article will deal with multithreading.
Introduction
In a lot of articles, we can read that multithreading is not easy, and other things like that. In this article, we will learn how to write multi-threaded application in a simple and a fun way.
The idea will be to write something like that
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 |
int main() { auto test = when_all( [] {return 8; }, // return 8 [] {return 25; } // return 25 ) // tuple(8, 25) .then( [](auto a, auto b) {return std::make_tuple(a, b); }, // return tuple(8, 25) [](auto a, auto b) {std::cout << a << "-" << b << " = "; }, // return nothing [](auto a, auto b) {return a - b; } // return - 17 ) // tuple(tuple(8, 25), -17) .deferredExecutor(); // Launch the chain of function in an asynchronous way test.execute(asynchronous_scheduler{}); // Do some work very expensive auto result = test.waitAndGet(); auto first = std::get<0>(result); // tuple(8, 25) auto second = std::get<1>(result); // -17 std::cout << second << std::endl; return 0; } |
If we write a little scheme, it gives us that :
Obviously, when the function f3 starts, functions f1 and f2 must have complete. The same thing happens for the getResult. It must wait for f3, f4, f5 to complete.
Does it really exist?
Actually, it does, and it will even be in the standard in C++20. However, the feature from the C++20 suffers from different problems that are explained in the Vittorio Romeo article. Actually, this article could be a continuation of the series of articles from Vittorio.
Multithreading with Monadic Expression: Implementation
I hope you really love read ununderstandable C++ code because the code is full of template and metaprogramming. I recommend to have some basis in really modern C++ or read the series from Vittorio I was talking before. We are going to see a lot of things both in multithreading and metaprogramming
Why do we need a continuation of the series from Vittorio?
Firstly, the code provided by Vittorio does not work on Visual Studio, and I really want to have my codes work on all compilers and all operating systems. Second, in its article, he only provides a version using a function waitAndGet that prevents us to insert code into the main thread in an easy way.
The third issue is that the group of functions returns a std::tuple
For example, you must have to write something like that
|
1 2 3 4 5 6 7 8 9 10 |
when_all( []{return 8;}, []{return 25;} ).then([](auto tuple) {auto [a, b] = tuple; return a - b;}); // Instead of when_all( []{return 8;}, []{return 25;} ).then([](auto a, auto b) {return a - b;}); |
But, in reality, some people will prefer the first version, other will prefer the second version. What will happen if a function does not return any arguments? According to me, the second version is more natural.
Forwarding
Because we are going to use a lot of metaprogramming, we may need to perfect forward some arguments. Here are two macro that defines a perfect forwarding. There is one macro for a normal variable and another one for auto. I was not able to use the same one for both cases because sometimes, use the form ::std::forward<decltype(variable)>(variable); is too difficult for Visual Studio, that is why I provide these both declarations.
|
1 2 3 4 5 |
#pragma once #include <utility> #define FWD(T, x) ::std::forward<T>(x) #define FWD_AUTO(x) ::std::forward<decltype(x)>(x) |
This code does not really need any explanations.
A Latch
A latch is useful when you want to know if a job is finished or not. Let’s say you have two thread, which one of them is waiting the result coming from the second one. The first one need to wait the second to finish.
The following code shows how to implement a simple “bool_latch”
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 |
#pragma once #include <mutex> #include <condition_variable> class bool_latch { public: /** * Function to call when the job is finished */ void job_done() { { std::scoped_lock<std::mutex> lock(mMutex); mDone = true; } mConditionVariable.notify_one(); } /** * Function to call when you want to wait for the job to be done */ void wait_for_job_done() { std::unique_lock<std::mutex> lock(mMutex); mConditionVariable.wait(lock, [this] {return mDone; }); } private: std::mutex mMutex; std::condition_variable mConditionVariable; bool mDone{ false }; }; |
The code is based on condition_variable, mutex, and a boolean variable
The return type
Each function may return one value. However, what returns a group of functions? The better choice is to return a std::tuple. Thus, if you have three functions that return an int at the same level (running in 3 parallels threads), the result will be a std::tuple<int, int, int>.
Functions that return void.
The simple way for this case is to create an empty type that we will name nothing. Its declaration is straightforward.
|
1 |
struct nothing{}; |
The function, instead to return nothing, must return the type nothing.
Let’s say you have three functions, f1, f2, f3. f1 and f3 return an int and f2 returns nothing. You will get a std::tuple<int, nothing, int>
How to return nothing instead of void? This function explains that in a straightforward way!
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 |
namespace detail { template<typename F, typename ...Args> /** This function returns f(args...) when decltype(f(args...)) is not void. Returns nothing{} else. @param f - The function @param args... - Arguments to give to the function @result - returns f(args...) if possible, or nothing{} else. */ inline decltype(auto) callThatReturnsNothingInsteadOfVoid(F &&f, Args &&...args) { if constexpr(std::is_void_v<std::invoke_result_t<F, Args...>>) { f(FWD(Args, args)...); return nothing{}; } else return f(FWD(Args, args)...); } } |
Pass a tuple as an argument
Okay, now say you have a std::tuple<int, double, std::string> and, you want to give it to one function with a prototype like returnType function(int, double, std::string).
One way to do that is to use the function apply.
Here there is no problem, but assume that there is a nothing in your tuple like std::tuple<int, nothing, int>.
When you will apply this tuple to the function, you will also pass the nothing.
To avoid such problem, you must filter all the arguments and store the arguments inside a lambda function.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 |
namespace detail { template<typename F> inline decltype(auto) callWithoutNothingAsArgument_Impl(F &&f) { return callThatReturnsNothingInsteadOfVoid(FWD(F, f)); } template<typename F, typename T, typename ...Args> inline decltype(auto) callWithoutNothingAsArgument_Impl(F &&f, T &&t, Args &&...args) { return callWithoutNothingAsArgument_Impl([&f, &t](auto &&...xs) -> decltype(auto) { if constexpr(std::is_same_v < nothing, std::decay_t<T>>) { return f(FWD_AUTO(xs)...); } else { return f(FWD(T, t), FWD_AUTO(xs)...); } }, FWD(Args, args)...); } } template<typename F, typename ...Args> inline decltype(auto) callWithoutNothingAsArgument(F &&f, std::tuple<Args...> tuple) { return std::apply([&f](auto ...xs)->decltype(auto) { return detail::callWithoutNothingAsArgument_Impl(f, xs...); }, tuple); } |
Architecture
This part will be the most difficult part of the article. We will see how the architecture work. It is an architecture with nodes. We will see three kinds of nodes. The root node that is the beginning of the chain of functions, the when_all node that can own one or several functions which will run in parallel and one node result_getter that represents the end of the chain of functions.
Overview
As you may have noticed, there is no type erasure. The type owned by the caller is the result_getter. However, the result_getter, when it executes all the functions, it must return to the root. And to do that, it must go through the when_all.
Now go to explain whole the code!
root
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 |
#pragma once #include <tuple> #include "fwd.h" namespace detail { class root { template<typename Parent, typename... F> friend class when_all; // A node produce nothing using output_type = std::tuple<>; private: // Once you are at the root, you can call the execute function template<typename Scheduler, typename Child, typename ...GrandChildren> void goToRootAndExecute(Scheduler &&s, Child &directChild, GrandChildren &... grandChildren) & { execute(FWD(Scheduler, s), directChild, grandChildren...); } // You must use the scheduler template<typename Scheduler, typename Child, typename ...GrandChildren> void execute(Scheduler &&s, Child &directChild, GrandChildren &... grandChildren) & { s([&]() -> decltype(auto) {return directChild.execute(FWD(Scheduler, s), output_type{}, grandChildren...); }); } }; } |
There is nothing difficult here. The when_all is a friend. There are two functions, the first one is called by the children and its purpose it’s only to reach the top. After, you execute the child functions through the scheduler.
when_all
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 |
#pragma once #include <atomic> #include "enumerate_args.h" #include "result_getter.h" namespace detail { struct movable_atomic_size_t : std::atomic_size_t { using std::atomic_size_t::atomic; movable_atomic_size_t(movable_atomic_size_t &&v) : std::atomic_size_t(v.load(std::memory_order_acquire)) {} }; template<typename parent,="" typename="" ...fs=""> class when_all : Parent, Fs... { friend class root; template<typename, typename...=""> friend class ::detail::when_all; template<typename> friend class result_getter; using input_type = typename Parent::output_type; using output_type = std::tuple<decltype(callwithoutnothingasargument(std::declval<fs&>(), std::declval<input_type>()))...>; public: /** There is SFINAE here because of something "kind of weird". Let's say you have to build something like Parent(std::move(parent)). Instead to try to reach the move constructor, it will try to instantiate this function with FsFWD = empty. However, Fs = {f1, f2, f3...}; The SFINAE is to forbid this error, and so, Parent(parent) will reach the move constructor **/ template<typename parentfwd,="" typename="" ...fsfwd,=""> 0)>> when_all(ParentFWD &&parent, FsFWD && ...f) : Parent(FWD(ParentFWD, parent)), Fs(FWD(FsFWD, f))... {} template<typename ...ffwd=""> decltype(auto) then(FFwd &&...ffwd) && { static_assert(sizeof...(FFwd) > 0, "Must have several functions objects"); return make_when_all(std::move(*this), FWD(FFwd, ffwd)...); } decltype(auto) deferredExecutor() && { return make_result_getter(std::move(*this)); } template<typename scheduler=""> decltype(auto) executeWaitAndGet(Scheduler &&s) && { auto f = deferredExecutor(); f.execute(FWD(Scheduler, s)); return f.waitAndGet(); } private: Parent &getParent() { return static_cast<parent&>(*this); } template<typename scheduler,="" typename="" ...children=""> void goToRootAndExecute(Scheduler &&s, Children&... children) & { // this is for ADL this->getParent().goToRootAndExecute(FWD(Scheduler, s), *this, children...); } template<typename scheduler,="" typename="" child,="" typename...="" grandchildren=""> void execute(Scheduler &&s, input_type r, Child &directChild, GrandChildren &...grandChildren) & { auto exec = [&, r](auto i, auto &f) { ::std::get < i >(mResult) = callWithoutNothingAsArgument(f, r); if (mLeft.fetch_sub(1, std::memory_order::memory_order_release) == 1) directChild.execute(FWD(Scheduler, s), std::move(mResult), grandChildren...); }; auto executeOneFunction = [&s, &exec](auto i, auto f) { if constexpr(i == sizeof...(Fs)-1) exec(i, f); else { s([exec, i, f]() { exec(i, f); }); } }; enumerate_args(executeOneFunction, static_cast<fs&>(*this)...); } private: output_type mResult; movable_atomic_size_t mLeft{ sizeof...(Fs) }; }; template<typename parent,="" typename="" ...f=""> inline auto make_when_all(Parent &&parent, F&& ...f) { return when_all<std::decay_t<parent>, std::decay_t<f>...>{FWD(Parent, parent), FWD(F, f)...}; } } template<typename ...f=""> auto when_all(F&& ...f) { static_assert(sizeof...(F) > 0, "Must have several functions objects"); return ::detail::make_when_all(detail::root{}, FWD(F, f)...); }</typename></f></std::decay_t<parent></typename></fs&></typename></typename></parent&></typename></typename></typename></input_type></decltype(callwithoutnothingasargument(std::declval<fs&></typename></typename,></typename></atomic> |
This is the most difficult part.
However, everything is not difficult. There are two things that are difficult. The output_type and the execute function.
|
1 2 |
using input_type = typename Parent::output_type; using output_type = std::tuple<decltype(callWithoutNothingAsArgument(std::declval<Fs&>(), std::declval<input_type>()))...>; |
The input_type is straightforward, we take the output from the parent. The output_type is a bit tricky. The idea is to call all the function, and for each of them, add the result to a tuple.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 |
template<typename scheduler,="" typename="" child,="" typename...="" grandchildren=""> void execute(Scheduler &&s, input_type r, Child &directChild, GrandChildren &...grandChildren) & { auto exec = [&, r](auto i, auto &f) { ::std::get < i >(mResult) = callWithoutNothingAsArgument(f, r); if (mLeft.fetch_sub(1, std::memory_order::memory_order_release) == 1) directChild.execute(FWD(Scheduler, s), std::move(mResult), grandChildren...); }; auto executeOneFunction = [&s, &exec](auto i, auto f) { if constexpr(i == sizeof...(Fs)-1) exec(i, f); else { s([exec, i, f]() { exec(i, f); }); } }; enumerate_args(executeOneFunction, static_cast<fs&>(*this)...); } </fs&></typename> |
There are several things to think about.
The exec lambda is the function that will be executed in a parallel thread. We store the value inside the good position of the tuple. If all functions have finished, we call the execute function for the directChild. This condition is checked by mLeft.
The executeOneFunction lambda is the function that computes if the function must be launched through the scheduler or not. (If there is only one function, there is no need to launch it through the scheduler because the root already did that).
The enumerate_args execute the function with each arguments, and give the index as an argument as well.
The enumerate_args is following :
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 |
#pragma once #include <type_traits> #include "fwd.h" namespace detail { template<typename F, std::size_t ...Is, typename ...Args> void enumerate_args_impl(F &&f, std::index_sequence<Is...>, Args && ...args) { using expander = int[]; expander{ 0, ((void)f(std::integral_constant<std::size_t, Is>{}, FWD(Args, args)), 0)... }; } } template<typename F, typename ...Args, typename Indices = std::index_sequence_for<Args...>> void enumerate_args(F &&f, Args &&...args) { detail::enumerate_args_impl(FWD(F, f), Indices{}, FWD(Args, args)...); } |
result_getter
Once you are here, you may have already build your own result_getter
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 |
#pragma once #include "bool_latch.h" #include "nothing.h" namespace detail { template<typename Parent> class result_getter : Parent { using input_type = typename Parent::output_type; using result_type = epurated_tuple_without_nothing_t<input_type>; template<typename, typename...> friend class when_all; public: template<typename ParentFWD> result_getter(ParentFWD &&parent) : Parent(std::move(parent)) {} template<typename Scheduler> void execute(Scheduler &&s) & { // this is for ADL this->getParent().goToRootAndExecute(FWD(Scheduler, s), *this); } result_type waitAndGet() & { mLatch.wait_for_job_done(); return mResult; } private: Parent &getParent() { return static_cast<Parent&>(*this); } template<typename Scheduler> void execute(Scheduler &&, input_type r) & { auto setResult = [this](auto ...xs) { if constexpr (sizeof...(xs) == 1) mResult = result_type{ xs... }; else mResult = std::make_tuple(xs...); mLatch.job_done(); }; callWithoutNothingAsArgument(setResult, r); } private: result_type mResult; bool_latch mLatch; }; // Ensure move semantic template<typename Parent, typename = std::enable_if_t<std::is_rvalue_reference_v<Parent&&>>> inline auto make_result_getter(Parent &&parent) { return result_getter<std::decay_t<Parent>>{std::move(parent)}; } } |
The idea here is to execute the function and wait for the latch to return the result. All the part for epurated tuple is done here :
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 |
namespace detail { template<typename...> struct tuple_without_nothing_impl; template<typename Tuple> struct tuple_without_nothing_impl<Tuple> { using type = Tuple; }; template<typename ...PreviousArgs, typename T, typename ...NextArgs> struct tuple_without_nothing_impl<std::tuple<PreviousArgs...>, T, NextArgs...> { using type = std::conditional_t< std::is_same_v<T, nothing>, typename tuple_without_nothing_impl<std::tuple<PreviousArgs...>, NextArgs...>::type, typename tuple_without_nothing_impl<std::tuple<PreviousArgs..., T>, NextArgs...>::type>; }; } template<typename Tuple> struct tuple_without_nothing; template<typename ...Args> struct tuple_without_nothing<std::tuple<Args...>> { using type = typename detail::tuple_without_nothing_impl<std::tuple<>, Args...>::type; }; template<typename Tuple> using tuple_without_nothing_t = typename tuple_without_nothing<Tuple>::type; template<typename Tuple> struct epurated_tuple { using type = std::conditional_t< std::tuple_size_v<Tuple> == 1, std::tuple_element_t<0, Tuple>, tuple_without_nothing_t<Tuple> >; }; template<> struct epurated_tuple<std::tuple<>> { using type = std::tuple<>; }; template<typename Tuple> using epurated_tuple_without_nothing_t = typename epurated_tuple<tuple_without_nothing_t<Tuple>>::type; |
The idea is to remove the nothings from the tuple, and if the tuple owns only one type, the tuple is removed and we get only the type.
Conclusion
In this article, you learned how to write safe and easy to read multi-threaded functions. There is still a lot of things to do, but it is really usable and easy to use. And if someone else has to read your code, he should be able to do so.
If you want to test it online, you can go on wandbox