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

Posted to jira@arrow.apache.org by "Yibo Cai (Jira)" <ji...@apache.org> on 2022/02/22 04:35:00 UTC

[jira] [Commented] (ARROW-13677) [C++] Improve performance of unpack64

    [ https://issues.apache.org/jira/browse/ARROW-13677?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17495854#comment-17495854 ] 

Yibo Cai commented on ARROW-13677:
----------------------------------

Looks compiler is able to vectorize the naive implementation on {{avx2/avx512/neon}}, both for unpack32 and unpack64.
[https://godbolt.org/z/jrz95vrsY]

Compiler doesn't vectorize for {{sse4.2}}, probably it believes the vectorized code is no better than the scalar one.

> [C++] Improve performance of unpack64
> -------------------------------------
>
>                 Key: ARROW-13677
>                 URL: https://issues.apache.org/jira/browse/ARROW-13677
>             Project: Apache Arrow
>          Issue Type: Improvement
>          Components: C++
>            Reporter: Antoine Pitrou
>            Priority: Major
>
> unpack32 benefits from auto-generated SIMD optimizations, but unpack64 doesn't. The latter is used by Parquet for DELTA_BINARY_PACKED encoding.



--
This message was sent by Atlassian Jira
(v8.20.1#820001)