Skip to content
Snippets Groups Projects

[bugfix] fix number of columns for non-scalar matrix blocks

Merged Jonathan Youett requested to merge bugfix/colcompmatrix_wrong_nr_cilums into master

Merge request reports

Loading
Loading

Activity

Filter activity
  • Approvals
  • Assignees & reviewers
  • Comments (from bots)
  • Comments (from users)
  • Commits & branches
  • Edits
  • Labels
  • Lock status
  • Mentions
  • Merge request status
  • Tracking
  • What's the status on this? Any objections to this commit? Jonathan, would it be difficult to extend an existing text demonstrating failure / that your patch works.

  • The ColCompMatrix is not supporting block entries. When it is setup from a BCRSMatrix<BlockType> mat, then the resulting number of scalar columns is mat.N()*BlockType::columns. This is also correctly set in the ColCompMatrix class but not in the factory class ColCompMatrixInitializer. The tests in dune-istl never observed this because only scalar problems are tested.

  • Could you extend an existing test to show the failure with the ColCompMatrixInitializer as an additional commit in this branch? Just, if it isn't too much work.

  • Jonathan Youett Added 21 commits:

    Added 21 commits:

    • 97f27f45...f2624739 - 20 commits from branch master
    • f648c23f - [bugfix] fix number of columns for non-scalar matrix blocks
  • I had another look at the problem and I found that my fix was not correct. I ran into this problem when trying to use SuperLU to decompose a matrix with rectangular blocks. I got an assertion failure in ColCompMatrixInitialiser that the number of columns is wrong. The problem was that the number of columns was initialised by the number of rows (which is the same for square matrices that appear in the dune-istl tests). I adjusted the patch to fix this

  • Works for me and nobody objected. Thanks for the patch.

  • mentioned in commit 49e78fa1

  • Christoph Grüninger Status changed to merged

    Status changed to merged

  • Mentioned in commit 49e78fa1

Please register or sign in to reply
Loading