[
https://issues.apache.org/jira/browse/ARROW-5760?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17115287#comment-17115287
]
Wes McKinney commented on ARROW-5760:
-
Another problem I noticed with the current implementation of Take and Filter:
different x86 is generated for applying these operations on arrays with the
same underlying C type. For example, instructions for moving {{int64_t}} values
are being generated for Int64Type, Date64Type, Time64Type, and TimestampType,
when only one underlying "data movement function" is needed. As part of
improving the performance of Take and Filter we should also ensure that we
eliminate this unneeded binary bloat in the shared library
> [C++] Optimize Take and Filter
> --
>
> Key: ARROW-5760
> URL: https://issues.apache.org/jira/browse/ARROW-5760
> Project: Apache Arrow
> Issue Type: New Feature
> Components: C++
>Reporter: Ben Kietzman
>Assignee: Ben Kietzman
>Priority: Major
> Fix For: 2.0.0
>
>
> There is some question of whether these kernels allocate optimally- for
> example when Filtering or Taking strings it might be more efficient to pass
> over the filter/indices twice, first to determine how much character storage
> will be needed then again into allocated memory:
> https://github.com/apache/arrow/pull/4531#discussion_r297160457
> Additionally, these kernels could probably make good use of scatter/gather
> SIMD instructions.
> Furthermore, Filter's bitmap is currently lazily expanded into the indices of
> elements to be appended to the output array. It would probably be more
> efficient to expand to indices in batches, then gather using an index batch.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)