How to properly initialize a boost multi_array of objects?

Question

I have been surprised to find that boost::multi_array seems to allocate its initial elements differently from, say, std::vector. It does not seem to fill each element with a unique element (using its default value or default constructor). I'm having trouble finding more information about this.

Is there a way to make the multi_array fill itself with a unique object at each element?

For example, consider the following:

static int num = 0;

struct A {
   int n;
   A() : n((::num)++) {
      std::cout << "A()" << std::endl;
   }
   virtual ~A() {}

   void print() {
      std::cout << "n=" << n << std::endl;
   }
};

int main() {
   std::cout << "vector:" << std::endl;
   std::vector<A> v(3);
   for (auto x : v) {
      x.print();
   }

   std::cout << "multi:" << std::endl;
   boost::multi_array<A, 2> m(boost::extents[2][2]);
   for (auto x : m) {
      for (auto y : x) {
         y.print();
      }
   }
}

This results in the output:

vector:
A()
A()
A()
n=0
n=1
n=2
multi:
A()
n=3
n=3
n=3
n=3

Why is the constructor called only once for the multi_array? How can the multi_array be "filled out" with unique objects (using A's default constructor)?

Answer 1

To quickly fill the whole array do something like fill_n¹:

std::fill_n(a.data(), a.num_elements(), 0);

With boost multi_array you can use a view to your own memory buffer to get the same performance (std::uninitialized_copy is your friend). (actually, you could even map an array view on existing memory, and you want to keep the existing values).

I've written a comparative demo about this here: pointers to a class in dynamically allocated boost multi_array, not compiling

Live On Coliru

#include <boost/multi_array.hpp>
#include <type_traits>
#include <memory>

struct octreenode { int a; int b; };

class world {
public:
    world(double x, double y, double z, int widtheast, int widthnorth, int height)
            : 
                originx(x), originy(y), originz(z), 
                chunkseast(widtheast), chunksnorth(widthnorth), chunksup(height)
    {
#define OPTION 4

#if OPTION == 1
        static_assert(std::is_trivially_destructible<octreenode>::value, "assumption made");
        //std::uninitialized_fill_n(chunk.data(), chunk.num_elements(), octreenode {1, 72});
        std::fill_n(chunk.data(), chunk.num_elements(), octreenode {1, 72});
#elif OPTION == 2
        for(auto a:chunk) for(auto b:a) for(auto&c:b) c = octreenode{1, 72};
#elif OPTION == 3
        for (index cz = 0; cz < chunksnorth; ++cz) {
            for (index cx = 0; cx < chunkseast; ++cx) {
                for (index cy = 0; cy < chunksup; ++cy) {
                    chunk[cz][cx][cy] = octreenode{1, 72};
                }
            }
        }
#elif OPTION == 4
        static_assert(std::is_trivially_destructible<octreenode>::value, "assumption made");
        for (index cz = 0; cz < chunksnorth; ++cz) {
            for (index cx = 0; cx < chunkseast; ++cx) {
                for (index cy = 0; cy < chunksup; ++cy) {
                    new (&chunk[cz][cx][cy]) octreenode{1, 72};
                }
            }
        }
#endif
        (void) originx, (void) originy, (void) originz, (void) chunksup, (void) chunkseast, (void) chunksnorth;
    }

private:
    double originx, originy, originz;
    int chunkseast, chunksnorth, chunksup;

#if 1
    typedef boost::multi_array<octreenode, 3> planetchunkarray; // a boost_multi for chunks
    typedef planetchunkarray::index index;
    planetchunkarray chunk{boost::extents[chunksnorth][chunkseast][chunksup]};
#else
    static_assert(boost::is_trivially_destructible<octreenode>::value, "assumption made");

    std::unique_ptr<octreenode[]> raw { new octreenode[chunksnorth*chunkseast*chunksup] };
    typedef boost::multi_array_ref<octreenode, 3> planetchunkarray;
    typedef planetchunkarray::index index;
    planetchunkarray chunk{raw.get(), boost::extents[chunksnorth][chunkseast][chunksup]};
#endif
};

int main() {
    world w(1,2,3,4,5,6);
}

The variant using multi_array_ref is an example of how to avoid copy-constructing the elements (it's akin to the optimization used by std::vector when it uses uninitialized memory for reserved but unused elements).

---

¹ Of course for unique values, use std::iota or std::generate

• http://en.cppreference.com/w/cpp/algorithm/iota
• http://en.cppreference.com/w/cpp/algorithm/generate_n

Answer 2

So on further study, I learned two things:

1. boost::multi_array uses the copy constructor to initialize objects into the container, not the default constructor.

2. The for (auto x : container) way of looping in C++11 seems (at least with clang++ 3.5) to loop over copies of the container elements, rather than iterators (or references).

Modifying the original question's example to demonstrate point 1.

Adding a copy constructor (and corresponding counter), and using auto& x for the object loops rather than auto x:

 static int num = 0;
 static int cpy = 0;
 struct A {
    int n;
    int c;
    A() : n((::num)++), c(0) {
       std::cout << "A_def()" << std::endl;
    }
    A(const A& o) : n(0), c((::cpy)++) {
       std::cout << "A_cpy()" << std::endl;
    }
    virtual ~A() {}

    void print() {
       std::cout << "n=" << n << ",c=" << c << std::endl;
    }
 };

 int main() {
    std::cout << "vector:" << std::endl;
    std::vector<A> v(3);
    for (auto& x : v) {
       x.print();
    }

    std::cout << "multi:" << std::endl;
    boost::multi_array<A, 2> m(boost::extents[2][2]);

    for (auto x : m) {
       for (auto& y : x) {
          y.print();
       }
    }
 }

Produces the output

 vector:
 A_def()  // <- vector allocation starts
 A_def()
 A_def()
 n=0,c=0  // <- vector printing starts, using "for (auto& x)"
 n=1,c=0
 n=2,c=0
 multi:
 A_def()  // <- a temporary object for multi_array allocation
 A_cpy()  // <- multi_array allocation starts
 A_cpy()
 A_cpy()
 A_cpy()
 n=0,c=0  // <- multi_array prints starts, using "for (auto& y)"
 n=0,c=1
 n=0,c=2
 n=0,c=3

Modifying the example above to demonstrate point 2.

Same class definition as above in this answer, but removing the auto& x from the object loops, and going back to using auto x as done in the original question.

    std::cout << "vector:" << std::endl;
    std::vector<A> v(3);
    for (auto x : v) {
       x.print();
    }

    std::cout << "multi:" << std::endl;
    boost::multi_array<A, 2> m(boost::extents[2][2]);

    for (auto x : m) {
       for (auto y : x) {
          y.print();
       }
    }

Produces output that shows the copy constructor gets called during the print loops, even for elements in the vector.

 vector:
 A_def()  // <- vector allocation starts
 A_def()
 A_def()
 A_cpy()  // <- vector printing starts, using "for (auto x)"
 n=0,c=0
 A_cpy()
 n=0,c=1
 A_cpy()
 n=0,c=2
 multi:
 A_def()  // <- a temporary object for multi_array allocation
 A_cpy()  // <- multi_array allocation starts
 A_cpy()
 A_cpy()
 A_cpy()
 A_cpy()  // <- multi_array printing starts, using "for (auto y)"
 n=0,c=7
 A_cpy()
 n=0,c=8
 A_cpy()
 n=0,c=9
 A_cpy()
 n=0,c=10