mboehm7 commented on pull request #1078:
URL: https://github.com/apache/systemds/pull/1078#issuecomment-706680533
well, for now I would recommend to properly maintain the non-zeros. It can
only become a moderate overhead if the output is large compared to the compute
(cubic compute and
mboehm7 commented on pull request #1078:
URL: https://github.com/apache/systemds/pull/1078#issuecomment-706566376
The current runtime/compiler requires the non-zeros to be exact or unknown
(-1) - so we cannot simply set it to fully dense. This information could be
mistakenly used (e.g.,