I have a big vector of items that belong to a certain class.
struct item {
int class_id;
//some other data...
};
The same class_id
can appear multiple times in the vector, and the vector is constructed once and then sorted by class_id
. So all elements of the same class are next to each other in the vector.
I later have to process the items per class, ie. I update all items of the same class but I do not modify any item of a different class. Since I have to do this for all items and the code is trivially parallelizable I wanted to use Microsoft PPL with Concurrency::parallel_for_each()
. Therefore I needed an iterator and came up with a forward iterator that returns the range of all items with a certain class_id
as proxy object. The proxy is simply a std::pair
and the proxy is the iterator's value type.
using item_iterator = std::vector<item>::iterator;
using class_range = std::pair<item_iterator, item_iterator>;
//iterator definition
class per_class_iterator : public std::iterator<std::forward_iterator_tag, class_range> { /* ... */ };
By now I was able to loop over all my classes and update the items like this.
std::vector<item> items;
//per_class_* returns a per_class_iterator
std::for_each(items.per_class_begin(), items.per_class_end(),
[](class_range r)
{
//do something for all items in r
std::for_each(r.first, r.second, /* some work */);
});
When replacing std::for_each
with Concurrency::parallel_for_each
the code crashed. After debugging I found the problem to be the following code in _Parallel_for_each_helper
in ppl.h at line 2772 ff.
// 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++);
}
It uses postincrement (so a temporary iterator is returned), dereferences that temporary iterator and takes the address of the dereferenced item. This only works if the item returned by dereferencing a temporary object survives, ie. basically if it points directly into the container. So fixing this is easy, albeit the per class std::for_each
work loop has to be replaced with a for-loop.
//it := iterator somewhere into the vector of items (item_iterator)
for(const auto cur_class = it->class_id; cur_class == it->class_id; ++it)
{
/* some work */
}
My question is if returning proxy objects the way I did is violating the standard or if the assumption that every iterator dereferences into permanent data has been made by Microsoft for their library, but is not documented. At least I could not find any documentation on the iterator requirements for parallel_for_each()
except that either a random access or a forward iterator are expected. I have seen the question about forward iterators and vector but since my iterator's reference type is const value_type&
I still think my iterator is ok by the standard. So is a forward iterator returning a proxy object still a valid forward iterator? Or put another way, is it ok for an iterator to have a value type different from a type that is actually stored somewhere in a container?
#include <vector>
#include <utility>
#include <cassert>
#include <iterator>
#include <memory>
#include <algorithm>
#include <iostream>
#include <ppl.h>
using identifier = int;
struct item
{
identifier class_id;
// other data members
// ...
bool operator<(const item &rhs) const
{
return class_id < rhs.class_id;
}
bool operator==(const item &rhs) const
{
return class_id == rhs.class_id;
}
//inverse operators omitted
};
using container = std::vector<item>;
using item_iterator = typename container::iterator;
using class_range = std::pair<item_iterator, item_iterator>;
class per_class_iterator : public std::iterator<std::forward_iterator_tag, class_range>
{
public:
per_class_iterator() = default;
per_class_iterator(const per_class_iterator&) = default;
per_class_iterator& operator=(const per_class_iterator&) = default;
explicit per_class_iterator(container &data) :
data_(std::addressof(data)),
class_(equal_range(data_->front())), //this would crash for an empty container. assume it's not.
next_(class_.second)
{
assert(!data_->empty()); //a little late here
assert(std::is_sorted(std::cbegin(*data_), std::cend(*data_)));
}
reference operator*()
{
//if data_ is unset the iterator is an end iterator. dereferencing end iterators is bad.
assert(data_ != nullptr);
return class_;
}
per_class_iterator& operator++()
{
assert(data_ != nullptr);
//if we are at the end of our data
if(next_ == data_->end())
{
//reset the data pointer, ie. make iterator an end iterator
data_ = nullptr;
}
else
{
//set to the class of the next element
class_ = equal_range(*next_);
//and update the next_ iterator
next_ = class_.second;
}
return *this;
}
per_class_iterator operator++(int)
{
per_class_iterator tmp{*this};
++(*this);
return tmp;
}
bool operator!=(const per_class_iterator &rhs) const noexcept
{
return (data_ != rhs.data_) ||
(data_ != nullptr && rhs.data_ != nullptr && next_ != rhs.next_);
}
bool operator==(const per_class_iterator &rhs) const noexcept
{
return !(*this != rhs);
}
private:
class_range equal_range(const item &i) const
{
return std::equal_range(data_->begin(), data_->end(), i);
}
container* data_ = nullptr;
class_range class_;
item_iterator next_;
};
per_class_iterator per_class_begin(container &c)
{
return per_class_iterator{c};
}
per_class_iterator per_class_end()
{
return per_class_iterator{};
}
int main()
{
std::vector<item> items;
items.push_back({1});
items.push_back({1});
items.push_back({3});
items.push_back({3});
items.push_back({3});
items.push_back({5});
//items are already sorted
//#define USE_PPL
#ifdef USE_PPL
Concurrency::parallel_for_each(per_class_begin(items), per_class_end(),
#else
std::for_each(per_class_begin(items), per_class_end(),
#endif
[](class_range r)
{
//this loop *cannot* be parallelized trivially
std::for_each(r.first, r.second,
[](item &i)
{
//update item (by evaluating all other items of the same class) ...
//building big temporary data structure for all items of same class ...
//i.processed = true;
std::cout << "item: " << i.class_id << '\n';
});
});
return 0;
}
When you're writing a proxy iterator, the reference
type should be a class type, precisely because it can outlive the iterator it is derived from. So, for a proxy iterator, when instantiating the std::iterator
base should specify the Reference
template parameter as a class type, typically the same as the value type:
class per_class_iterator : public std::iterator<
std::forward_iterator_tag, class_range, std::ptrdiff_t, class_range*, class_range>
~~~~~~~~~~~
Unfortunately, PPL is not keen on proxy iterators and will break compilation:
ppl.h(2775): error C2338: lvalue required for forward iterator operator *
ppl.h(2772): note: while compiling class template member function 'Concurrency::_Parallel_for_each_helper<_Forward_iterator,_Function,1024>::_Parallel_for_each_helper(_Forward_iterator &,const _Forward_iterator &,const _Function &)'
with
[
_Forward_iterator=per_class_iterator,
_Function=main::<lambda_051d98a8248e9970abb917607d5bafc6>
]
This is actually a static_assert
:
static_assert(std::is_lvalue_reference<decltype(*_First)>::value, "lvalue required for forward iterator operator *");
This is because the enclosing class _Parallel_for_each_helper
stores an array of pointer
s and expects to be able to indirect them later:
typename std::iterator_traits<_Forward_iterator>::pointer _M_element[_Size];
Since PPL doesn't check that pointer
is actually a pointer, we can exploit this by supplying a proxy pointer with an operator*
and overloading class_range::operator&
:
struct class_range_ptr;
struct class_range : std::pair<item_iterator, item_iterator> {
using std::pair<item_iterator, item_iterator>::pair;
class_range_ptr operator&();
};
struct class_range_ptr {
class_range range;
class_range& operator*() { return range; }
class_range const& operator*() const { return range; }
};
inline class_range_ptr class_range::operator&() { return{*this}; }
class per_class_iterator : public std::iterator<
std::forward_iterator_tag, class_range, std::ptrdiff_t, class_range_ptr, class_range&>
{
// ...
This works great:
item: item: 5
1
item: 3item: 1
item: 3
item: 3
Press any key to continue . . .