CODEWITHSUNDEEP
c++sparse-matrixeigenaccess-violationsolver
user16320
user16320

Reputation: 123

Eigen SparseLU solve error read access violation, this->m_sup_to_col was 0x111011101110112

I am using Visual Studio Community 2015 and I'm trying to use the Eigen library in c++ to solve large sparse linear systems. I have everything wrapped in my class Domain2D with defined stiffness matrix and load vector with load vector being my own long double type:

typedef Eigen::Matrix<long double, Eigen::Dynamic, 1> VectorXld; // somewhere way up in the code

class Domain2D {
private:
    ...
    Eigen::SparseMatrix<long double, Eigen::ColMajor> constant_stiffness;
    Eigen::SparseMatrix<long double, Eigen::ColMajor> convection_stiffness;
    ...
    VectorXld load;
    ...
public:
    ...
    void load_constant_stiffness() {
        ...
        resize, reserve, fill up with elements
    }
    ...
    void load_convection_stiffness() {
        ...
    }
    void load_RHS() {
        ...
    }
}

Finally, somewhere in the class I have the following routine:

void advance_step(const int &step) {
        ...
        Eigen::SparseLU<Eigen::SparseMatrix<long double, Eigen::ColMajor>, Eigen::COLAMDOrdering<int>> solver;
        Eigen::SparseMatrix<long double, Eigen::ColMajor> M;
        VectorXld res, res_new;
        res.resize(2 * N + n);
        ...
        while (d > 1e-5) {
            d = 0;
            load_convection_stiffness();
            load_RHS();
            M = constant_stiffness + convection_stiffness;
            M.makeCompressed();
            solver.analyzePattern(M);
            solver.factorize(M);
            res_new = solver.solve(load);
            ...
        }
        Save("output\\results\\", step);
        t += dt;
    }

The program fails on the last mentioned line: res_new = solver.solve(M);

It shows me a file called "SparseLU_SupermodalMatrix.h", the routine it fails in is void MappedSuperNodalMatrix<Scalar,Index_>::solveInPlace( MatrixBase<Dest>&X) const and the line is Index fsupc = supToCol()[k]; and the error it shows is

Unhandled exception thrown: read access violation.

this->m_sup_to_col was 0x111011101110112.

If there is a handler for this exception, the program may be safely continued.

I was searching a bit and I found that if I use eigen objects in my class, I should add line EIGEN_MAKE_ALIGNED_OPERATOR_NEW; in the public section, which I did, to no avail. All matrices are of a proper size, as well as the vector of long doubles. I have been solving these systems for some time and it worked, but suddenly, this happens. What could be the cause?

P.S.: I loaded the matrix and load vector to mathematica and even with the standard precision, I get a solution which looks absolutely fine, so it's not like the system is ill-conditioned.

Upvotes: 0

Views: 1406

Answers (1)

ggael
ggael

Reputation: 29205

Some tips:

  1. You need to check that the factorization went well by checking that solver.info()==Eigen::Success after the call to factorize().
  2. Try with the head of the 3.3 branch here to see whether this issue has already been resolved.
  3. Make sure you tried without NDEBUG defined (aka. debug mode in visual) to see whether Eigen triggers some insightful assertions.
  4. Checking with double instead of long double might help to isolate the issue. To this end, you can cast locally M and load with .cast<double>().
  5. Running your program within a memory debugger might help too.
  6. After trying all of these, share your matrix M with Eigen's team to let them (i.e., myself ;) ) reproduce the issue.

Upvotes: 3

Related Questions