Hi All,

this is similar to Naval's question here:
viewtopic.php?f=74&t=120694

My setting: I have a non-quadratic sparse matrix (N+1xN) with rank N and want to solve it.
Now I am left with SparseQR as the other solvers require a quadratic matrix, right?

I use

Code: Select all

   SparseQR<SparseMatrix<double,ColMajor>, COLAMDOrdering<int>  > phi_solver_ ;


and then

Code: Select all

      phi_solver_.analyzePattern(phi_matrix_);

gives no compiler warning, but a very similar Assertion failed during the run:

Code: Select all

 
karman: /home/hecke/turb/channelflow/eigen3/Eigen/src/OrderingMethods/Ordering.h:141: void Eigen::COLAMDOrdering<Index>::operator()(const MatrixType&, Eigen::COLAMDOrdering<Index>::PermutationType&) [with MatrixType = Eigen::SparseMatrix<double>; Index = int; Eigen::COLAMDOrdering<Index>::PermutationType = Eigen::PermutationMatrix<-1>]: Assertion `info && "COLAMD failed "' failed.


If I compile in release mode I get a segfault...

Unfortunately, I cannot find enough info to solve this on my own, so here's my cry for help!

addendum:
If I change the Ordering method to NaturalOdering, (which seems kind of justified as the coefficients in my matrix are all in a tight range and there are similarly many coefficients in one column) analyzePattern() runs fine, but I get a problem from within factorize()

Code: Select all

karman: /home/hecke/turb/channelflow/nsolver/../eigen3/Eigen/src/Core/DenseCoeffsBase.h:394: Eigen::DenseCoeffsBase<Derived, 1>::Scalar& Eigen::DenseCoeffsBase<Derived, 1>::operator()(Eigen::DenseCoeffsBase<Derived, 1>::Index) [with Derived = Eigen::Matrix<int, -1, 1>; Eigen::DenseCoeffsBase<Derived, 1>::Scalar = int; Eigen::DenseCoeffsBase<Derived, 1>::Index = long int]: Assertion `index >= 0 && index < size()' failed.


Still, there seems to be a problem with a matrix having more rows than columns...

thanks a bunch
Hecke

Can you try with the latest "stable" branch (https://bitbucket.org/eigen/eigen/get/3.2.tar.bz2) or even the devel branch to make sure this issue has not been solved already. Otherwise, please save you matrix to a zipped file so that we can check what's going wrong.

Code: Select all

#include <unsupported/Eigen/SparseExtra>
...
Eigen::saveMarket(A, "filename.mtx", Eigen::Symmetric);
Eigen::saveMarketVector(B, "filename_b.mtx");


Hey Ggael,

I was really hoping for you to chime in. Thanks a lot!

I used the latest stable version, and trying the unstable one leads to the same error.

You find the matrix producing the error here:
http://www.nld.ds.mpg.de/~hecke/archiv/phi_matrix.zip

If I can be of any help, please let me know.

cheers
Hecke

hm, it works very well for me.

Perhaps try to call phi_matrix_.makeCompress() before passing it to SparseQR?

Code: Select all

#include <Eigen/SparseQR>
#include <iostream>
#include <unsupported/Eigen/SparseExtra>
using namespace Eigen;
int main(int argc, char** argv)
{
  typedef SparseMatrix<double> SpMat;
  SpMat A;
  loadMarket(A, argv[1]);
 
  SparseQR<SpMat,COLAMDOrdering<int> > qr(A);
  std::cout << qr.info() << " == " << Success << "\n\n";
 
  VectorXd x(A.cols()), b(A.rows());
  x.setRandom();
  b = A * x;
  x.setZero();
 
  x = qr.solve(b);
  std::cout << qr.info() << " == " << Success << "\n\n";
  std::cout << "err: " << (A*x-b).norm() / b.norm() << "\n";
 
  return 0;
}


Great, Ggael,

that did the trick.

you are my hero!

thanks.
Hecke

Now that's clear:

https://bitbucket.org/eigen/eigen/commits/c3858caad887/
https://bitbucket.org/eigen/eigen/commits/9eb4359d389c/

Yep, now it is hard to miss

thanks again
Hecke