[GitHub] [systemds] Baunsgaard commented on pull request #1128: [SYSTEMDS-2761] Transpose in place

[GitBox <gi...@apache.org>]

Baunsgaard commented on pull request #1128:
URL: https://github.com/apache/systemds/pull/1128#issuecomment-757452080


   The in place transpose is done, but the performance is worse than our previous, therefore this PR is on standby.
   
   After implementing this in place transpose, it is slower than our default transpose that allocate a full matrix in cases where columns != rows.
   This is because for each cell in the output each value is moved 6 times, while in our allocate a new matrix we move once. The allocation of the new matrix does introduce memory overhead, that this in place operation does not.
   
   The two instances that is much faster is the in place transpose with cols == rows, because each cell only require a single swap and instances where either rows or columns == 1, the latter is because even if the previous is a metadata operation, it still copies the vector.
   


----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org