How can be OpenMP used with the class and Eigen? - c++

Question

I am trying to make code using OpenMP, some class and Eigen.
The overall code is at the bottom.
The code is a simplified version of what I want to do in my project.
The simplified code is just a matrix inverse and multiplication using class and OpenMP.
If I don't use the OpenMP, I get the results of
solution in for loop at 0th iteration is... 1 1 1
solution in class at 0th iteration is... 1 1 1
solution in for loop at 1th iteration is... 4 4 4
solution in class at 1th iteration is... 4 4 4
solution in for loop at 2th iteration is... 9 9 9
solution in class at 2th iteration is... 9 9 9
solution in for loop at 3th iteration is... 16 16 16
solution in class at 3th iteration is... 16 16 16
However, If I use the OpenMP, I get the results of
9 11 91
1
16 16
solution in class at 3th iteration is... 116 16 161
1 1
solution in class at 2th iteration is... 1
9 9 9
solution in for loop at 1th iteration is... solution in for loop at 9
13th iteration is... 1
solution in class at 0th iteration is... 4 4 4 16 solution in for loop at 1
solution in for loop at 01 11
1 solution in for loop at 1 solution in for loop at 41 3th iteration is... th iteration is...
16 9 solution in for loop at th iteration is... 1 solution in class at 33th iteration is... th iteration is... 161
16 solution in for loop at 016 16 2th iteration is... 1616 1616 16 161616th iteration is...
solution in for loop at 19 solution in class at 3th iteration is... 9 9
16
The results are corrupted.
Why does this happen?
Also, at first, I found that only using "#pragma omp parallel for" results in the error that occurs when LDLT is not initialized.
So, I added the following instead of "#pragma omp parallel for".
#pragma omp parallel for private(j,k) firstprivate(foo)
class Foo {
private:
std::vector<Eigen::Matrix<double, 3, 3>> a;
Eigen::LDLT<Eigen::MatrixXd> Minv; // Minv
std::vector<Eigen::Matrix<double, 3, 1>> b;
std::vector<Eigen::Matrix<double, 3, 1>> solution;
public:
Foo() {};
~Foo() {};
void Initialization() {
a.resize(4);
b.resize(4);
solution.resize(4);
for (int i = 0; i < 4; i++) {
a.at(i).setZero();
b.at(i).setZero();
solution.at(i).setZero();
}
}
void SetInv(int idx) {
a.at(idx) = (1.0/(idx+1.0))*Eigen::Matrix3d::Identity();
b.at(idx) = (idx+1)*Eigen::Vector3d::Ones();
Minv.compute(a.at(idx));
}
void Calculation(int idx) {
solution.at(idx) = Minv.solve(b.at(idx));
}
Eigen::Matrix<double, 3, 1> &ReturnSolution(int idx) {
return solution.at(idx);
}
void ShowSolution(int idx) {
std::cout << " solution in class at " << idx << "th iteration is... "<< solution.at(idx).transpose() << std::endl;
}
};
I hope to know what the corrupted result is made.

Answer 1

First of all, the results are corrupted, because using OpenMP you made a multithreaded code. There are several threads doing cout at the same time - therefore you can see the mess.
if you want to have clear output, you need implement a mutex to make your code thread safe. Simplified example:
#include <mutex>
std::mutex mtx;
#pragma omp parallel for private(j,k) firstprivate(foo)
{
/* Your work happens here */
{ // use brackets to show scope where mutex is locked
std::lock_guard<std::mutex> lock(mtx);
std::cout << "Some result you want to print";
}
}
And link to doc: https://en.cppreference.com/w/cpp/thread/lock_guard
Also, at first, I found that only using "#pragma omp parallel for"
results in the error that occurs when LDLT is not initialized.
Please provide sample code with your implementation of #pramgma omp... so it will be easier to answer your question.