Hi

I may have stumbled across a peculiar bug - *Matrix Multiplication* I don't know if this is linked to the gcc compiler or is some initialisation issue in eigen or to do with the transpose function specifically, but: if,

MatrixXd res = mat1.transpose() * mat2.transpose();
//half of the matrix is empty with zeros and the columns and rows become jumbled up?
if I first declare:

MatrixXd res;
res = mat1.transpose() * mat2.transpose();

//then all is ok - can anyone confirm this?

also:
chain multiplication:
MatrixXd res;
res = mat1 * (mat1.transpose() * mat2.transpose());
//fails with similar issues as above, solved by:

MatrixXd res;
res = (mat1.transpose() * mat2.transpose());
res = mat1 * res;

Thanks

TOAD

Is this Eigen 2.0 or the development branch?

Just to accomodate our laziness, could you post a compilable test case?

sorry - my fault - It's using Eigen 2.0.4 (Aug 1 release) - I can do a test case if that helps.

OK here's a testcase:

Code: Select all

#include <Eigen/Array>
using namespace Eigen;
int main()
{
  int n = XXSIZE;
  MatrixXd mat1 = MatrixXd::Random(n,n);
  MatrixXd mat2 = MatrixXd::Random(n,n);
  MatrixXd res1 = mat1.transpose() * mat2.transpose();
  MatrixXd res2; res2 = mat1.transpose() * mat2.transpose();
  std::cout << res1.isApprox(res2,1e-3) << std::endl;
}


Testing against 2.0:

Code: Select all

=== 14:48:49 ~$ g++ n.cpp -o n -I eigen2.0 -DXXSIZE=10 && ./n
1
=== 14:48:57 ~$ g++ n.cpp -o n -I eigen2.0 -DXXSIZE=20 && ./n
0


So the bug is triggered when the size is large enough that the cache-friendly path is taken.

The bug exists only in the 2.0 branch, not in the devel branch.

Bisected. The bug had always existed in both branches until it was fixed in the default branch by changeset bf5ef41ebba3 by Gael: "more refactoring in the level3 products"

Good news, I just committed a fix to the 2.0 branch.

Can you please try the 2.0 branch and report? At least the above testcase works now.

I can confirm that this now works, v. 2.0.5. Thanks