Hi there,

I have downloaded Eigen 3.3-alpha1 and played with it to solve a sparse matrix (row major) on an 8 cores machine. The test code is quite simple. I compiled it using g++ version 4.7.2 with -fopenmp option. But I didn't see any speedup. The Eigen::nbThreads() did tell me that there were 8 threads. The followings are part of the test code.

Code: Select all

typedef Eigen::SparseMatrix<double, Eigen::RowMajor>  MatrixType;
typedef Eigen::VectorXd  VecType;

void run_CG(const MatrixType& mat, const VecType& rhs,
            int maxIters, double tolerance,
            VecType& rst, int& nIters, double& estimateErr) {

    std::cout << "---------- CG ----------" << std::endl;
    Eigen::ConjugateGradient<MatrixType,
                             Eigen::Lower | Eigen::Upper,
                             Eigen::IdentityPreconditioner> cg;
    cg.setMaxIterations(maxIters);
    cg.setTolerance(tolerance);
    cg.compute(mat);
    rst = cg.solve(rhs);
    nIters = cg.iterations();
    estimateErr = cg.error();
}

void run_BiCGSTAB(const MatrixType& mat, const VecType& rhs,
            int maxIters, double tolerance,
            VecType& rst, int& nIters, double& estimateErr) {

    std::cout << "---------- BiCGStab ----------" << std::endl;
    Eigen::BiCGSTAB<MatrixType, Eigen::IdentityPreconditioner> cg;
    cg.setMaxIterations(maxIters);
    cg.setTolerance(tolerance);
    cg.compute(mat);
    rst = cg.solve(rhs);
    nIters = cg.iterations();
    estimateErr = cg.error();
}


I also tried to

Code: Select all

 grep -R "pragma omp" ./*

in the source directory of Eigen. But I didn't see and OpenMP directives.
So I was wondering if i used the wrong version or it is not supported yet.

Thanks,
LZ

Very strange that you cannot find any pragma omp is Eigen's source, because even the 3.0 version has some omp directives. On 3.3-alpha1 you should see:

Code: Select all

$ grep "pragma omp" Eigen/* -R
Eigen/src/Core/products/GeneralMatrixMatrix.h:      #pragma omp critical
Eigen/src/Core/products/GeneralMatrixMatrix.h:        #pragma omp atomic
Eigen/src/Core/products/Parallelizer.h:  #pragma omp parallel num_threads(threads)
Eigen/src/SparseCore/SparseDenseProduct.h:        #pragma omp parallel for schedule(static) num_threads(threads)


The last one is responsible for multi-threading in sparse iterative solvers.

Thank you! Problem sovled