Mentions légales du service

Skip to content
Snippets Groups Projects

Diagonal copy support

Merged BOUCHERIE Raphael requested to merge rboucher/chameleon:Diag/copy_support into master
All threads resolved!

Moved the data descriptor for temporary copies of the diagonal to the driver level (Fixes #33 (closed))

Merge request reports

Checking pipeline status.

Merged by avatar (Apr 23, 2025 7:00am UTC)

Loading

Pipeline #3791 passed

Pipeline passed for 253e9ad4 on master

Activity

Filter activity
  • Approvals
  • Assignees & reviewers
  • Comments (from bots)
  • Comments (from users)
  • Commits & branches
  • Edits
  • Labels
  • Lock status
  • Mentions
  • Merge request status
  • Tracking
  • added 1 commit

    Compare with previous version

  • added 1 commit

    • d83922ab - changed ge2gb definition in compute_z.h

    Compare with previous version

  • Mathieu Faverge resolved all discussions

    resolved all discussions

  • @all This pull request starts the process of moving up to the driver level the workspaces. The benefit of this is to allocate the workspace only once and use it through multiple algorithms when they are chained together as it is done in reduction algorithms for example, or in solver drivers. As of today, those workspaces induce a sequence wait at the end of each pz file which is a big slowdown. It is now removed.

    The goal is next to move these workspaces a level higher to enable the user to allocate those workspaces only once for multiple MORSE_***_Async calls.

    If everyone is ok, we'll merge it before the end of the week.

  • Although this API change may look weird at first sight, D & T indeed play a similar role.

  • @agullo I don't understand what you meant.

    I agree with you that D and T are similar, and the next changes should include both, it is related to issue #34 (closed) and the change from allocation of large global matrices as workspaces to allocation on demand of the used tiles. Similar idea will be applied to T's matrices in QR/LQ algorithms, Band matrix in reduction algorithms, A^t in he2hb, and other temporary buffers in algorithm that requires temporary workspace. The idea is really to remove the synchronization that workspaces imply at the end of each parallel algorithm by doing it only when everything is computed.

  • Mathieu Faverge mentioned in commit 253e9ad4

    mentioned in commit 253e9ad4

  • Please register or sign in to reply
    Loading