Question

I was looking for some suitable 2D element container. What I want is the ability to iterate through every element of the container using, for example BOOST_FOREACH and I also would like to have an ability to construct subview (slices / subranges) of my container and, probably iterate through them too.

Right now I am using boost::numeric::ublas::matrix for these purposes, but, well, it doesn’t look as a good solution for me, because, well, it’s a BLAS matrix, although it behaves very well as a plain 2d element container (custom unbounded / bounded storages are also very sweet).

Another boost alternative, boost::multi_array is bad, because you can’t iterate through every element using one BOOST_FOREACH statement and because constructing views has extremely obfuscated syntax.

Any alternatives?

Thank you.

Answer 1

I do the following (array type is container/iterator range concept):

ublas::matrix<douple> A;
foreach (double & element, A.data())
{
}

However, this will not work for slices: your best solution is to write an iterator for them.

Here is an example of using multi_array to provide storage of a custom class.
Perhaps you could do the same:

template<size_t N, typename T>
struct tensor_array : boost::multi_array_ref<T,N> {
    typedef boost::multi_array_ref<T,N> base_type;

    typedef T value_type;
    typedef T& reference;
    typedef const T& const_reference;

    tensor_array() : base_type(NULL, extents())
    {
        // std::cout << "create" << std::endl;
    }
    template<class A>
    tensor_array(const A &dims)
        : base_type(NULL, extents())
    {
        //std::cout << "create" << std::endl;
        resize(dims);
    }

    template<typename U>
    void resize(const U (&dims)[N]) {
        boost::array<U,N> dims_;
        std::copy(dims, dims + N, dims_.begin());
        resize(dims_);
    }

    template<typename U>
    void resize(const boost::array<U,N> &dims) {
        size_t size = 1;
        boost::array<size_t,N> shape;
        for (size_t i = 0; i < N; ++i)  {
            size *= dims[i];
            shape[N-(i+1)] = dims[i];
        }
        data_.clear();
        data_.resize(size, 0);
        // update base_type parent
        set_base_ptr(&data_[0]);
        this->num_elements_ = size;
        reshape(shape);
    }

    size_t size() const { return data_.size(); }
    size_t size(size_t i) const { return this->shape()[N-(i+1)]; }
    tensor_array& fill(const T &value) {
        std::fill(data_.begin(), data_.end(), value);
        return *this;
    }
private:
    typedef boost::detail::multi_array::extent_gen<N> extents;
    std::vector<T> data_;
};