Re: [eigen] SparseMatrix StorageIndex issue

[Gerard Bahi <zomeck@xxxxxxxxx>]

Hi Gael,

Thanks for your answer. 

I think an assert is a good idea. Otherwise one gets crashes when using SPQR as it forces long for storage index type.

Kind Regards,

Gerard

On 31 December 2017 at 15:20, Gael Guennebaud <gael.guennebaud@xxxxxxxxx> wrote:

Hi,

On Thu, Dec 21, 2017 at 3:08 PM, Gerard Bahi <zomeck@xxxxxxxxx> wrote:

Hi,

I am using Eigen 3.3.4, on clang 4.0

I was changing the StorageIndex type and I noticed something that might be a problem with my Eigen configuration or a bug.

Given the following code:

#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
#include <Eigen/Eigen>

using StorageIndexType = long;
Eigen::SparseMatrix<double, Eigen::ColMajor, StorageIndexType> A(10,10);

We've never tested such a configuration because it makes no sense. We should probably prohibit it as it is doom to fail. In your case the problem lies in this line:

Index realloc_size = (std::min<Index>)(NumTraits<StorageIndex>::highest(),  size + Index(reserveSizeFactor*double(size)));

where NumTraits<long>::highest() is wrongly converted to int as -1. Of course this line is easy to fix, there will be many other not-fixable issues. So better statically assert IMO.

gael

 

When adding for(int i=0; i<10; ++i) A.insert(i,i) = 1; or A.setIdentity(); I get a bad_alloc exception.

However running:
A.resize(10, 10);
for(int i=0; i<10; ++i) A.insert(i,i) = 1;

Does not cause the throw. It will throw if using setIdentity instead of insert even with the resize though.

I have also noticed that printing the matrix like this:
std::cerr << "A =\n" << A << std::endl;

Causes the bad_alloc throw as well on  (SparseMatrix.h:829) // s << static_cast<const SparseMatrixBase<SparseMatrix>&>(m)

The streaming with Eigen debug ON print shows the following (before crashing):
Nonzero entries:
(1,0) (1,1) (1,2) (1,3) (1,4) (1,5) (1,6) (1,7) (1,8) (1,9) (_,_) (_,_) (_,_) (_,_) (_,_) (_,_) (_,_) (_,_) (_,_) (_,_)

Outer pointers:
0 1 2 3 4 5 6 7 8 9  $
Inner non zeros:
1 1 1 1 1 1 1 1 1 1  $

I have tried making all combinations of int and long of EIGEN_DEFAULT_DENSE_INDEX_TYPE and StorageIndexType, only this combination crashes.

Trying to use the matrix with SPQR also crashes (which was my initial intention):
Eigen::SPQR<Eigen::SparseMatrix<double, Eigen::ColMajor, StorageIndexType> > solver;
solver.compute(A);

Kind regards,

Gerard