标签c ++ openmp shared_ptr datarace

Question

I have a c++ program where I try to use openmp on a for loop. The for loop works with an shared_ptr to my own class that in turn calls another dll. I get errors:

Table 7. Current pointer and in_use_count are inconsistent.

The code looks something like this..

    int n = 1000;
    std::vector<double> result(n),indata(n);

    // populate indata

    std::shared_ptr<MyNS::MyClass> sp_mycl = std::make_shared < MyNS::MyClass >();
    sp_mycl->var1 = 2;
    // populate sp_mycl->v_var4
    #pragma omp parallel for firstprivate(sp_mycl)
    for (auto ii = 0; ii<n;ii++)
    {
        sp_mycl->var2 = indata[ii];
        sp_mycl->calc();
        result[ii] = sp_mycl->var3;
    }

and MyClass.h looks similar to this

namespace MyNS
{
    extern "C" { double * fortran_dll_calc(int *num, double arrinput[],double arroutput[])} // subroutine in fortran dll
    class MyClass
    {
        double var1, var2, var3;
        std::vector<double> v_var4;
        void calc();
    }
}

and MyClass.cpp

using namespace MyNS;
void MyClass::calc()
{
    int len = v_var4.size();
    double *test = new double[len];
    for (auto is = 0; is<len;is++)
        test[is] = v_var4[is];
    double fortran_result[10]; // output from fortran dll 
    fortran_dll_calc(len,test,fortran_result);
    for (int ir = 0;ir < 10;ir++)
        var3 += fortran_result[ir];
}

I use MSVS with Intel parallel studio c++/fortran compiler 2016.

I want the sp_mycl->var1 to have the same inital value for all threads, hence the firstprivate for sp_mycl .

The openmp for loop seems to be going wrong somewhere, and when debugging it sometimes appear to stop inside MyClass::calc() , and sometimes already at sp_mycl->var2= indata[ii] . This much I have found out using some cout outputs.

Is the firstprivate working with shared_ptr to my own defined objects? I am a beginner thus, there might be many mistakes and errors, any comment on anything in the code is appreciated.

Answer 1

OpenMP is going to initialize the private version of sp_mycl for each of the thread as follows:

   auto priv_sp_mycl = sp_mycl;

This is done in parallel and invokes the copy constructor which needs finally needs to update the reference counting for the shared_ptr . However, shared_ptr implementation in STL is not thread-safe so there is where things starts to goes wrong.

However, I am not sure this is what you want. Why do you want to have every thread pointing to the same object? Wouldn't that introduce an additional race condition into your program? Also each thread is writing the result into result[n] ... did you mean result[ii] ?

Anyway, I think this is what you want:

MyNS::MyClass sp_mycl;
sp_mycl.var1 = 2;
// populate sp_mycl->v_var4
#pragma omp parallel for firstprivate(sp_mycl)
for (auto ii = 0; ii<n;ii++)
{
    sp_mycl.var2 = indata[ii];
    sp_mycl.calc();
    result[ii] = sp_mycl.var3;
}