Re: [eigen] Sparse Matrix with block sparsity pattern

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

Hi,

the implementation in g2o does not seems to be very efficient memory-wise unless blocks are statically allocated, and even in this case there is a lot of redundancy. Some years ago we started an implementation where you define the block sizes along the row/column dimensions and then the non-zeros are packed as in a SparseMatrix:

https://bitbucket.org/eigen/eigen/src/5c43a259f0b51963a00dd85e941a1fd044ac1cf4/unsupported/Eigen/src/SparseExtra/BlockSparseMatrix.h

IIRC, even though it is not completely finished, it supports matrix-vector products and thus iterative solvers, but I did not checked for a while....

gael

On Fri, Dec 15, 2017 at 11:29 AM, Tim Rakowski <tim.rakowski@xxxxxxxxx> wrote:

Hi Justin,

g2o has an implementation: https://github.com/RainerKuemmerle/g2o

Take a look at class BlockSolver in https://github.com/RainerKuemmerle/g2o/blob/master/g2o/core/block_solver.h.

Regards,

Tim

On Fri, Dec 15, 2017 at 11:14 AM Justin Carpentier <jcarpent@xxxxxxx> wrote:

Hi Eigen's users,

I'm currently looking for a C++ implementation of Sparse Matrices with a
generic Block Sparse pattern, compatible or inspired from the Eigen library..
What I mean by Block Sparse pattern is that each non-zero element
belongs to a non-zero block of a given dimension. In other words, the
sparsity is not element-wise but much more block-wise.

Is someone has any idea of such existing implementation?

Best,

Justin