A proper way to create a matrix in c++

Question

I want to create an adjacency matrix for a graph. Since I read it is not safe to use arrays of the form matrix[x][y] because they don't check for range, I decided to use the vector template class of the stl. All I need to store in the matrix are boolean values. So my question is, if using std::vector<std::vector<bool>* >* produces too much overhead or if there is a more simple way for a matrix and how I can properly initialize it.

EDIT: Thanks a lot for the quick answers. I just realized, that of course I don't need any pointers. The size of the matrix will be initialized right in the beginning and won't change until the end of the program. It is for a school project, so it would be good if I write "nice" code, although technically performance isn't too important. Using the STL is fine. Using something like boost, is probably not appreciated.

Solution 1

Note that also you can use boost.ublas for matrix creation and manipulation and also boost.graph to represent and manipulate graphs in a number of ways, as well as using algorithms on them, etc.

Edit: Anyway, doing a range-check version of a vector for your purposes is not a hard thing:

template <typename T>
class BoundsMatrix
{
        std::vector<T> inner_;
        unsigned int dimx_, dimy_;

public:
        BoundsMatrix (unsigned int dimx, unsigned int dimy)
                : dimx_ (dimx), dimy_ (dimy)
        {
                inner_.resize (dimx_*dimy_);
        }

        T& operator()(unsigned int x, unsigned int y)
        {
                if (x >= dimx_ || y>= dimy_)
                        throw std::out_of_range("matrix indices out of range"); // ouch
                return inner_[dimx_*y + x];
        }
};

Note that you would also need to add the const version of the operators, and/or iterators, and the strange use of exceptions, but you get the idea.

Solution 2

Best way:

Make your own matrix class, that way you control every last aspect of it, including range checking.

eg. If you like the "[x][y]" notation, do this:

class my_matrix {
  std::vector<std::vector<bool> >m;
public:
  my_matrix(unsigned int x, unsigned int y) {
    m.resize(x, std::vector<bool>(y,false));
  }
  class matrix_row {
    std::vector<bool>& row;
  public:
    matrix_row(std::vector<bool>& r) : row(r) {
    }
    bool& operator[](unsigned int y) {
      return row.at(y);
    }
  };
  matrix_row& operator[](unsigned int x) {
    return matrix_row(m.at(x));
  }
};
// Example usage
my_matrix mm(100,100);
mm[10][10] = true;

nb. If you program like this then C++ is just as safe as all those other "safe" languages.

Solution 3

The standard vector does NOT do range checking by default.

i.e. The operator[] does not do a range check.

The method at() is similar to [] but does do a range check.
It will throw an exception on out of range.

std::vector::at()
std::vector::operator[]()

Other notes: Why a vector<Pointers> ?
You can quite easily have a vector<Object>. Now there is no need to worry about memory management (i.e. leaks).

std::vector<std::vector<bool> >   m;

Note: vector<bool> is overloaded and not very efficient (i.e. this structure was optimized for size not speed) (It is something that is now recognized as probably a mistake by the standards committee).

If you know the size of the matrix at compile time you could use std::bitset?

std::vector<std::bitset<5> >    m;

or if it is runtime defined use boost::dynamic_bitset

std::vector<boost::dynamic_bitset>  m;

All of the above will allow you to do:

m[6][3] = true;

Solution 4

If you want 'C' array performance, but with added safety and STL-like semantics (iterators, begin() & end() etc), use boost::array.

Basically it's a templated wrapper for 'C'-arrays with some NDEBUG-disable-able range checking asserts (and also some std::range_error exception-throwing accessors).

I use stuff like

boost::array<boost::array<float,4>,4> m;

instead of

float m[4][4];

all the time and it works great (with appropriate typedefs to keep the verbosity down, anyway).

---

UPDATE: Following some discussion in the comments here of the relative performance of boost::array vs boost::multi_array, I'd point out that this code, compiled with g++ -O3 -DNDEBUG on Debian/Lenny amd64 on a Q9450 with 1333MHz DDR3 RAM takes 3.3s for boost::multi_array vs 0.6s for boost::array.

#include <iostream>
#include <time.h>
#include "boost/array.hpp"
#include "boost/multi_array.hpp"

using namespace boost;

enum {N=1024};

typedef multi_array<char,3> M;
typedef array<array<array<char,N>,N>,N> C;

// Forward declare to avoid being optimised away
static void clear(M& m);
static void clear(C& c);

int main(int,char**)
{
  const clock_t t0=clock();

  {
    M m(extents[N][N][N]);
    clear(m);
  }

  const clock_t t1=clock();

  {
    std::auto_ptr<C> c(new C);
    clear(*c);
  }

  const clock_t t2=clock();

  std::cout 
    << "multi_array: " << (t1-t0)/static_cast<float>(CLOCKS_PER_SEC) << "s\n"
    << "array      : " << (t2-t1)/static_cast<float>(CLOCKS_PER_SEC) << "s\n";

  return 0;
}

void clear(M& m)
{
  for (M::index i=0;i<N;i++)
    for (M::index j=0;j<N;j++)
      for (M::index k=0;k<N;k++)
    m[i][j][k]=1;
}


void clear(C& c)
{
  for (int i=0;i<N;i++)
    for (int j=0;j<N;j++)
      for (int k=0;k<N;k++)
    c[i][j][k]=1;
}

Solution 5

What I would do is create my own class for dealing with matrices (probably as an array[x*y] because I'm more used to C (and I'd have my own bounds checking), but you could use vectors or any other sub-structure in that class).

Get your stuff functional first then worry about how fast it runs. If you design the class properly, you can pull out your array[x*y] implementation and replace it with vectors or bitmasks or whatever you want without changing the rest of the code.

I'm not totally sure, but I thing that's what classes were meant for, the ability to abstract the implementation well out of sight and provide only the interface :-)

View more solutions

Comments

[Admin]

unless you use a debugging version of stdlib, or you're calling vector::at

[janneb]

For multidimensional arrays, boost::multiarray ( boost.org/doc/libs/1_38_0/libs/multi_array/doc/index.html ) is probably a better choice.

[timday]

No way; I've tried that in the past and it's performance was horrific compared with the C-array equivalent. And its storage is always heap allocated c.f boost::array which can use the stack.

[janneb]

Fascinating; in my benchmarks there is essentially no performance difference between a static C style array vs. boost::multiarray. Then again, I tested element access of large arrays so heap allocation performance is not a problem.

[timday]

See update for basis of my belief it's slow. I'm using g++ 4.3.2, boost 1.35.

[janneb]

double or float, and don't include initialization, then multi_array is slightly faster. Some experimenting leads me to believe that multi_array initialization zeroes the array; it is responsible for half the time (testing with different sized arrays). However, for char and int, array (CONTINUED...)

[janneb]

consistently outperforms multi_array. This is with g++ 4.1.2, boost 1.33.

[peedurrr]

That's really nice.

[Alexis Wilke]

throw 0?! Why not std::out_of_range("matrix indices out of range")

[D Drmmr]

BoundsMatrix<bool> doesn't work. Anyway, this is hardly a "range-check version of a vector". You don't even provide iterators, for which it is much harder to provide range checks.

[Diego Sevilla]

@DDrmmr, it was an indication on how to do it, not a complete implementation. Do you think including those special cases as type checks or template specializations would have led to a better answer?

[RoG]

I couldn't get this to compile: error: cannot bind non-const lvalue reference of type ‘bool&’ to an rvalue of type ‘bool’ at return row.at(y);.