Bugzilla – Bug 37
sparse row-major multiplication crashes depending on matrix sizes
Last modified: 2011-03-23 09:47:19 UTC
Created attachment 8 [details]
When multiplying an mxn SparseMatrix by an nxp, iff m > p I get
Eigen3/src/Sparse/SparseMatrix.h:190: void Eigen3::SparseMatrix<_Scalar, _Flags, _Index>::startVec(typename Eigen3::ei_traits<Eigen3::SparseMatrix<_Scalar, _Options, _Index> >::Index) [with _Scalar = double, int _Options = 1, _Index = int]: Assertion `m_outerIndex[outer+1]==0 && "You must call startVec for each inner vector sequentially"' failed.
Column-major matrices don't have this problem. Haven't tried mixed.
I attach a test program:
$ testEigenSparseRowMajorProduct m n n p
Created attachment 130 [details]
I also recently ran into this problem (I am using Eigen 3 Beta 4). It seems that its cause is the res.resize() instruction in function sparse_product_impl, line 136 in file src/Sparse/SparseSparseProduct.h.
I was able to get things running by replacing said line with this ugly hack:
if (res.innerSize() != rows || res.outerSize() != cols)
I've tested this with:
- ColMajor ColMajor ColMajor
- ColMajor ColMajor RowMajor
- RowMajor RowMajor RowMajor
The problem thus seems to be that resizing should happen according to inner/outer size of res, not according to rows/cols of res.
I'm attaching the patched SparseSparseProduct.h and another test case.
Created attachment 131 [details]
Silly me... A clean solution seems to be simply checking for ResultType::IsRowMajor. I'm attaching a patch that should be usable.
Created attachment 133 [details]
Patch for src/Sparse/SparseSparseProduct.h
thanks for the patch. I'm currently extending the unit tests and I'll apply your patch soon...