You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@systemml.apache.org by "Matthias Boehm (JIRA)" <ji...@apache.org> on 2017/03/03 05:06:45 UTC

[jira] [Created] (SYSTEMML-1369) Loop vectorization rewrite for indexed copies

Matthias Boehm created SYSTEMML-1369:
----------------------------------------

             Summary: Loop vectorization rewrite for indexed copies
                 Key: SYSTEMML-1369
                 URL: https://issues.apache.org/jira/browse/SYSTEMML-1369
             Project: SystemML
          Issue Type: Sub-task
            Reporter: Matthias Boehm


For complex models, users often have to resort to a task-parallel parfor implementation because it's not easily expressed in vectorized form. In this context, we often encounter indexed copies that could have been easily vectorized. For example, the below snippet stems from a recent benchmark comparison against SystemML, which implemented Hidden Markov Model learning:
{code}
parfor (i in 1:ncol(labels)) {
   topics[id, i] = labels[1, i]
}
{code}

This task aims to automatically rewrite such loops into vectorized form - for above example into {{topics[id, 1:ncol(labels)] = labels[1, 1:ncol(labels)]}}, where the unnecessary right indexing is subsequently remove by following HOP rewrites.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)