Skip to content

C++ implementation of Python-like "zip" parallel iteration.

License

Notifications You must be signed in to change notification settings

CommitThis/zip-iterator

Repository files navigation

C++: Zip Iteration

This document is licensed CC-BY-NC-SA 3.0, everything else is licensed under the MIT terms.

Python is one of my favourite languages. One of the features it has is the zip function, which allows for parallel iteration of arrays and I wondered if something like that was possible in C++. While I haven't seen anything like this before and I didn't really go to the trouble of researching this; I thought it would be an interesting exercise I could try and do myself.

In any case, using some modern C++ techniques we can do this reasonably easily. Being C++, I am not terribly confident I've covered all of the various details properly, and any feedback on this would be appreciated.

While the value categories in C++ aren't too complicated, I have simplified the categories to an lvalue (that which is named, and has a predictable address, e.g. int x = 1), and an rvalue (that which does not, is a temporary; is a new instance returned by a function e.g. f(){ return 1; }.

The Loop

This was the motivating example, use structured bindings to unpack values from two iterables in parallel.

for (auto && [x, y] : zip(a, b)) {
    c.push_back(x + z);
}

I wanted both x and y to be references and therefore modifiable, but only if the respective zipped containers was passed in as non-const, and values otherwise.

A range-based for loop executes functions over a range, using syntactic sugar to do so. It requires that the range expression represents a suitable sequence, e.g. an array or object with begin and end functions available.

https://en.cppreference.com/w/cpp/language/range-for

Therefore the zip function ought to return such an object. It will be a temporary (it is an rvalue), but that is ok as it's lifetime is extended until the end of the loop. The begin and end functions of this object must then return suitable iterators.

The zip call also needs to be able to handle arguments of different types, and not require the type be supplied at use.

Finally, the value_type of the iterated sequence must be de-structurable, which can be an array, tuple-like, or an object with non-static data members.

https://en.cppreference.com/w/cpp/language/structured_binding

So, three components will be needed:

  • A utility function for users to call that handles type deduction;
  • An object that meets the requirements for range-for that returns
  • A suitable iterator that proxies the iterators for the passed in types.

The Utility Function

template <typename T, typename U>
auto zip(T && t, U && u) {
    return zipper<T, U>{std::forward<T>(t), std::forward<U>(u)};
}

Universal references are used to accept arguments of any type. This is not because we want to accept temporaries (which does appear to work), but because we want to accept both const and non-const types. We could be more clever at this point to be more strict with some more template trickery, but I think this is good enough for this example.

In any case, we can see that the types of the arguments are deduced and are then used as template arguments to the object that returns the iterators.

The Zipper

I couldn't think of a great name for this, but basically is a really simple object that contains references (or values, largely depends on what the type arguments are) to the objects to be iterated over, and returns our special zip iterators.

template <typename T, typename U>
class zipper {
public:
    using Iter1 = select_iterator_for<T>;
    using Iter2 = select_iterator_for<U>;

    using zip_type = zip_iterator<Iter1, Iter2>;

    template <typename V, typename W>
    zipper(V && a, W && b)
        : m_a{a}
        , m_b{b}
    {}

    auto begin() -> zip_type {
        return zip_type{std::begin(m_a), std::begin(m_b)};
    }
    auto end() -> zip_type {
        return zip_type{std::end(m_a), std::end(m_b)};
    }

private:
    T m_a;
    U m_b;
};

The only trickery here is making sure we select the create iterator type to be used by the proxy. If we pass in a std::vector<>&, we want to use the std::vector<>::iterator type, however if it's a std::vector<> const & then we need to select std::vector<>::const_iterator. A relatively simple use of std::conditional_t solves this:

template <typename T>
using select_iterator_for = std::conditional_t<
    std::is_const_v<std::remove_reference_t<T>>, 
    typename std::decay_t<T>::const_iterator,
    typename std::decay_t<T>::iterator>;

This selects the correct iterator type. Basically, if what's passed in is const, the const_iterator is selected and otherwise iterator. The reason why we std::decay_t is used is because the types are references, and you can't access their type definitions as it's considered an incomplete type. std::decay_t returns it's fundamental type and solves this issue.

As per https://en.cppreference.com/w/cpp/types/is_const, the reference needs to be removed from the original type, although I'm not sure why!

The Zip Iterator

The zip iterator holds two iterators for the actual type being iterated over, and proxies their usual functions in tandem.

template <typename Iter1, typename Iter2>
class zip_iterator
{
public:
    using value_type = std::pair<
        select_access_type_for<Iter1>,
        select_access_type_for<Iter2>
    >;

    zip_iterator() = delete;

    zip_iterator(Iter1 iter_1_begin, Iter2 iter_2_begin)
        : m_iter_1_begin {iter_1_begin}
        , m_iter_2_begin {iter_2_begin}
    {}

    auto operator++() -> zip_iterator& {
        ++m_iter_1_begin;
        ++m_iter_2_begin;
        return *this;
    }

    auto operator++(int) -> zip_iterator {
        auto tmp = *this;
        ++*this;
        return tmp;
    }

    auto operator!=(zip_iterator const & other) {
        return !(*this == other);
    }

    auto operator==(zip_iterator const & other) {
        return 
            m_iter_1_begin == other.m_iter_1_begin ||
            m_iter_2_begin == other.m_iter_2_begin;
    }

    auto operator*() -> value_type {
        return value_type{*m_iter_1_begin, *m_iter_2_begin};
    }

private:
    Iter1 m_iter_1_begin;
    Iter2 m_iter_2_begin;
};

Again, fairly straight forward. The iterator is constructed from two iterators, which are incremented whenever the iterator itself is incremented. It's not really a proper iterator in the sense that it doesn't meet the minimum requirements for such (e.g. doesn't have a category), but it is a minimal example that works.

The comparison we need to make in order to detect when iteration should end is when any of the iterators equal their end position. This isn't strictly correct from a semantic point of view -- it returns true as soon as any of the contained iterators equal their respective pair. This is necessary as the distance between the starting points of either may be different, and then they would never be equal at the same time leading to dereferencing past the end, and the loop never terminating.

De-referencing the iterator returns a std::pair containing either references or values corresponding to the internal de-referenced iterators.

The consideration here is what the access type is when the internal iterators are is de-referenced. In most cases this should be a reference, but this is not always possible. For example, std::vector<bool> is a special case in that it doesn't actually store bools as you might think. It does some special bit twiddling to save on space (although arguably this is implementation defined). A normal iterator of such is incapable of returning a reference, and de-referencing returns a value.

template <typename Iter>
using select_access_type_for = std::conditional_t<
    std::is_same_v<Iter, std::vector<bool>::iterator> ||
    std::is_same_v<Iter, std::vector<bool>::const_iterator>,
    typename Iter::value_type,
    typename Iter::reference
>;

The access type selection is there because I was originally zipping a vector of ints and a vector of bools. std::vector has a specialisation for bool that means that the value can't be accessed by reference, but instead by value. This is obviously very specific for a std::vector, but this could either be extended or replaced and I think works quite well to demonstrate how the access type could be changed. This could, however, be surprising for users.

What Is The Loop Actually Doing?

  • First, the zip function is called that deduces the passed in typed which are then used to create:
  • The zipper object. This contains either a reference or value to what's been passed in and has begin and end functions which the range-for uses to generate a loop;
  • On each pass of the loop, the iterator from begin is dereferenced and assigned to the value on the left hand side (the "range declaration") and finally incremented;
  • The range declaration is a structured binding, and the temporary std::pair from the dereferenced iterator is unpacked into it.

There is also some strangeness going on with the structured bindings. If one of the destructured variables is a reference, then a const qualification will have effect. If it is a value (e.g. in the case of std::vector<bool>), then the const qualifier is applied.

Further, the binding will fail if it is qualified as a non-const lvalue reference.

The Story So Far...

In any case, at this point, we should be able to iterate in parallel over two containers:

auto a = std::vector<int>{1, 2, 3, 4, 5, 6};
auto const & b = a;
auto c = std::vector<int>{};

for (auto && [x, y] : zip(a, b)) {
    c.push_back(x + y);
}

This works as you might expect, x and y are either const or non-const references. What's curious here is how the structured binding is used. Ideally, if you wanted to iteration and wanted to enforce const on any of the values you might write something like:

for (auto const & [x, y] : zip(a, b)) ...

The const in this case refers to the type being unpacked, which is a temporary std::pair. This will have no affect whatsoever on the actual types it contains. If you want to enforce this, you would need to make sure that the passed in type (a or b) are actually const references.

Additionally, because the returned type is a temporary, you will never be able to bind it to an lvalue reference.

But what about.....

Iterating Over More Than Two Types

This is slightly more complicated than the previous example. I won't go to the trouble of repeating what i've said earlier, however the full source can be found on ########

The utility function is modified to take a variable amount of arguments:

template <typename ... T>
auto zip(T && ... t) {
    return zipper<T ...>{std::forward<T>(t)...};
}

In the zipper class, instead of storing the container/references in named members, we can use a tuple instead.

template <typename ... Args>
class zipper {
public:
    using zip_type = zip_iterator<select_iterator_for<T> ...>;

// ... snip ...

private:
    std::tuple<T...> m_args;
};

As for creating the iterator through begin and end, we have to find some way of unpacking the tuple in order to call these functions on the underlying types.

C++17 has a utility function that can do this: std::apply. It takes a lambda expression as the first argument (which in itself takes as arguments the values of the tuple) and the tuple as the second:

std::apply([](auto && ... args){ /* more magic here */ }, my_tuple);

The arguments can then be expanded as an ordinary parameter pack's using std::begin(args)...:

auto begin() -> zip_type {
    return std::apply([](auto && ... args){ 
            return zip_type(std::begin(args)...); 
        }, m_args);
}

So, basically what's going on is that the tuple is being turned into a parameter pack that's expanded in the constructor for our custom iterator.

With respect to the zip iterator we use a tuple member to store the iterators, and use the same technique to provide operator* and operator++:

template <typename ... Iters>
class zip_iterator
{
public:

    // ... snip ...

    auto operator++() -> zip_iterator& {
        std::apply([](auto && ... args){ ((args += 1), ...); }, m_iters);
        return *this;
    }

    auto operator*() -> value_type {
        return std::apply([](auto && ... args){ 
                return value_type(*args...); 
            }, m_iters);
    }
};

The use of this technique with operator* is exactly the same, only we're de-referencing the internal iterators and unpacking the results into yet another tuple. With operator++, we're not returning anything, and just applying the same operator to the contained iterators.

Finally, the equivalence operator is a special case. In a twist of fate, we also need to do parallel iteration on all of the iterators! We can't use std::apply as, as far as I can tell, only works on one tuple at a time. With a somewhat terrifying set of template functions:

template <typename ... Args, std::size_t ... Index>
auto any_match_impl(
    std::tuple<Args...> const & lhs,
    std::tuple<Args...> const & rhs,
    std::index_sequence<Index...>) -> bool
{
    auto result = false;
    result = (... | (std::get<Index>(lhs) == std::get<Index>(rhs)));
    return result;
}

template <typename ... Args>
auto any_match(
    std::tuple<Args...> const & lhs, 
    std::tuple<Args...> const & rhs) -> bool
{
    return any_match_impl(lhs, rhs, std::index_sequence_for<Args...>{});
}

The second function is a helper to call the actual implementation, so that a sequence of integers is generated with respect to the size of the template arguments.

At this point, both the tuples are the same type and therefore have exactly the same size.

In the implementation function, a fold expression is used to expand the index sequence; comparing the type instances of each tuple at the same index, which is aggregated into the result. This was probably the most difficult thing I had to deal with throughout; it wasn't an expansion I was at all familiar with, and I apologise in advance if I haven't explained it properly!

Conclusion

Hopefully, I have demonstrated how parallel iteration over multiple types is possible. We can now quite happily do something like this (observe the varying sizes!):

auto a = std::vector<int>{1, 2, 3, 4, 5, 6};
auto b = std::vector<int>{1, 2, 3, 4, 5, 6, 7};
auto c = std::vector<int>{0, 0, 0, 0, 0};
auto const & d = b;

for (auto && [x, y, z] : c9::zip(a, d, c)) {
    z = x + y;
}

auto expected = std::vector<int>{2, 4, 6, 8, 10};
assert(c == expected);

There are two ways I can currently see that could be improved:

  • Making the iterators have knowledge of their end position. This could be used to make the equality operation semantically correct in cases (which I can't see as being useful) and also help with runtime safety;
  • One of the concepts of C++ is flexible operations over data; you may want to "zip" across ranges within containers, that aren't either at the begin or end positions. In which case you might want a zip signature that looks like zip(begin_a, end_a, begin_b, end_b, ...). This would be easy for two sets of data, but might be more complicated for the variadic case.

About

C++ implementation of Python-like "zip" parallel iteration.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •