template <typename _Forward_iterator, typename _Function, unsigned int _Chunk_size>
class _Parallel_for_each_helper
{
public:
    typedef typename std::iterator_traits<_Forward_iterator>::value_type _Value_type;
    static const unsigned int _Size = _Chunk_size;

    _Parallel_for_each_helper(_Forward_iterator& _First, const _Forward_iterator& _Last, const _Function& _Func) :
        _M_function(_Func), _M_len(0)
    {
        // Add a batch of work items to this functor's array
        for (unsigned int _Index=0; (_Index < _Size) && (_First != _Last); _Index++)
        {
            _M_element[_M_len++] = &(*_First++);
        }
    }

    void operator()() const
    {
        // Invoke parallel_for on the batched up array of elements
        _Parallel_for_impl(0U, _M_len, 1U,
            [this] (unsigned int _Index)
            {
                _M_function(*(_M_element[_Index]));
            }
        );
    }

private:

    const _Function& _M_function;
    _Value_type *    _M_element[_Size];
    unsigned int     _M_len;

    _Parallel_for_each_helper const & operator=(_Parallel_for_each_helper const&);    // no assignment operator
};