You are viewing a plain text version of this content. The canonical link for it is here.

Posted to issues@impala.apache.org by "Daniel Becker (JIRA)" <ji...@apache.org> on 2019/06/28 10:26:00 UTC

[jira] [Created] (IMPALA-8726) Autovectorisation leads to worse performance in bit unpacking

Daniel Becker created IMPALA-8726:
-------------------------------------

             Summary: Autovectorisation leads to worse performance in bit unpacking
                 Key: IMPALA-8726
                 URL: https://issues.apache.org/jira/browse/IMPALA-8726
             Project: IMPALA
          Issue Type: Improvement
            Reporter: Daniel Becker
         Attachments: no_vector.png

The compiler (GCC 4.9.2) autovectorises bit unpacking for bit widths 1, 2, 4 and 8 (function BitPacking::UnpackValues), but this leads to actually worse performance (see the attached graph). We should consider whether it is worth disabling autovectorisation for bit unpacking, but future compiler versions may do a better job.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)