Question

This is really confusing me, I would appreciate if anyone could help me out.

(EDIT: thought it was a templated problem, I was mistaken with this)

I want to add multiple copies of the following class with gnu’s parallelised accumulate algorithm (stored in #include <parallel/numeric> )

The class deliberately doesn’t do much, I don’t think this is a thread collision problem?

template<class T>
class NaturalParameters
{
public:
  typedef typename std::vector<T>::iterator iterator;

  NaturalParameters()
    :    m_data(2) //vector with two zeros
  {  }
      
  typename std::vector<T>::const_iterator
  begin() const
  {
    return m_data.begin();
  }

  typename std::vector<T>::const_iterator
  end() const
  {
    return m_data.end();
  }

  NaturalParameters<T>& 
  operator+=(const NaturalParameters<T>& other)
  {
    //do something
    return *this;
  }
      
private:
  std::vector<T> m_data;
};

template<class T>
inline
NaturalParameters<T>
operator+(const NaturalParameters<T>& a, const NaturalParameters<T>& b)
{
  NaturalParameters<T> tmp = a;
  return tmp+=b;
}  

I then run it

int
main  (int ac, char **av)
{
  std::vector<NaturalParameters<double> > NP(1000);
  NaturalParameters<double> init;
  //the following segfaults
  NaturalParameters<double> NP2 = __gnu_parallel::accumulate(NP.begin(), NP.end(), init ); 
  //The following runs fine
  //NaturalParameters<double> NP2 = std::accumulate(NP.begin(), NP.end(), init ); 
}

This really confuses me – I have no idea what the problem is.
I’m using g++ 4.4.5 and compiling with g++ gnu_parallel.cpp -g -fopenmp

EDIT:

Note that this works: (999 elements rather than 1000)

 for(size_t i=0;i<1000;++i){

  std::vector<NaturalParameters> ChildrenNP(999);
  NaturalParameters<double> init;
  NaturalParameters<double> NP = __gnu_parallel::accumulate(ChildrenNP.begin(), ChildrenNP.end(), init ); 
  //NaturalParameters<double> NP = std::accumulate(ChildrenNP.begin(), ChildrenNP.end(), init ); 
   }

The backtrace is:

    Program received signal SIGSEGV, Segmentation fault.
__libc_free (mem=0x12af1) at malloc.c:3709
3709    malloc.c: No such file or directory.
    in malloc.c
(gdb) backtrace
#0  __libc_free (mem=0x12af1) at malloc.c:3709
#1  0x00000000004024f8 in __gnu_cxx::new_allocator<double>::deallocate (this=0x614518, __p=0x12af1) at /usr/include/c++/4.4/ext/new_allocator.h:95
#2  0x0000000000401f0a in std::_Vector_base<double, std::allocator<double> >::_M_deallocate (this=0x614518, __p=0x12af1, __n=18446744073709542049) at /usr/include/c++/4.4/bits/stl_vector.h:146
#3  0x00000000004017b9 in std::_Vector_base<double, std::allocator<double> >::~_Vector_base (this=0x614518, __in_chrg=<value optimized out>) at /usr/include/c++/4.4/bits/stl_vector.h:132
#4  0x00000000004013b9 in std::vector<double, std::allocator<double> >::~vector (this=0x614518, __in_chrg=<value optimized out>) at /usr/include/c++/4.4/bits/stl_vector.h:313
#5  0x00000000004012b8 in NaturalParameters<double>::~NaturalParameters (this=0x614518, __in_chrg=<value optimized out>) at gnu_parallel.cpp:10
#6  0x00000000004023e7 in __gnu_parallel::for_each_template_random_access_ed<__gnu_cxx::__normal_iterator<NaturalParameters<double>*, std::vector<NaturalParameters<double>, std::allocator<NaturalParameters<double> > > >, __gnu_parallel::nothing, __gnu_parallel::accumulate_selector<__gnu_cxx::__normal_iterator<NaturalParameters<double>*, std::vector<NaturalParameters<double>, std::allocator<NaturalParameters<double> > > > >, __gnu_parallel::accumulate_binop_reduct<__gnu_parallel::plus<NaturalParameters<double>, NaturalParameters<double> > >, NaturalParameters<double> > (begin=..., end=..., o=..., f=..., r=..., 
    base=..., output=..., bound=-1) at /usr/include/c++/4.4/parallel/par_loop.h:127
#7  0x0000000000401d70 in std::__parallel::accumulate_switch<__gnu_cxx::__normal_iterator<NaturalParameters<double>*, std::vector<NaturalParameters<double>, std::allocator<NaturalParameters<double> > > >, NaturalParameters<double>, __gnu_parallel::plus<NaturalParameters<double>, NaturalParameters<double> > > (begin=..., end=..., init=..., binary_op=..., parallelism_tag=__gnu_parallel::parallel_unbalanced)
    at /usr/include/c++/4.4/parallel/numeric:99
#8  0x0000000000401655 in std::__parallel::accumulate<__gnu_cxx::__normal_iterator<NaturalParameters<double>*, std::vector<NaturalParameters<double>, std::allocator<NaturalParameters<double> > > >, NaturalParameters<double> > (begin=..., end=..., init=...) at /usr/include/c++/4.4/parallel/numeric:139
#9  0x0000000000400e2c in main (ac=1, av=0x7fffffffe188) at gnu_parallel.cpp:59

Answer 1

This definitely looks like a bug in libstdc++:

/usr/include/c++/4.4/parallel/par_loop.h:87

#   pragma omp single
      {
        num_threads = omp_get_num_threads();
        thread_results = static_cast<Result*>(
                            ::operator new(num_threads * sizeof(Result)));
        constructed = new bool[num_threads];
      }

But line 127 deletes it with

delete[] thread_results;

_Apparently, the construction of thread_results was optimized at one stage, but the deletion statement was never updated to reflect this. The optimization makes sense over just newing up the array (new Result[num_threads]) because it avoids constructing the elements._

Fixing that to

delete thread_results;

removes the bug. You will want to report this to the gnu devs.

---

You might still have some residu problems with the threadsafety of std::__cxx1998::vector::operator=. You can see what I mean by using valgrind. However, it is entirely possible that valgrind reports a fasle positive there.

I just tested the other way around: when using new Result[num_threads] with delete[] (instead of the optimized version in the GNU source) you’ll get a clean valgrind run all the way. I’m pretty sure this will be a false positive, but I’d sure mention it to the GNU devs while you report the bug.