[
https://issues.apache.org/jira/browse/ARROW-8901?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17114518#comment-17114518
]
Wes McKinney commented on ARROW-8901:
-------------------------------------
We probably need at least int8 through int64 (so we can use take to unpack
dictionaries). A different code path will probably be used for running "take"
in a selection vector context (per ARROW-8903)
> [C++] Reduce number of take kernels
> -----------------------------------
>
> Key: ARROW-8901
> URL: https://issues.apache.org/jira/browse/ARROW-8901
> Project: Apache Arrow
> Issue Type: Improvement
> Components: C++
> Reporter: Wes McKinney
> Priority: Major
>
> After ARROW-8792 we can observe that we are generating 312 take kernels
> {code}
> In [1]: import pyarrow.compute as pc
>
> In [2]: reg = pc.function_registry()
>
> In [3]: reg.get_function('take')
>
> Out[3]:
> arrow.compute.Function
> kind: vector
> num_kernels: 312
> {code}
> You can see them all here:
> https://gist.github.com/wesm/c3085bf40fa2ee5e555204f8c65b4ad5
> It's probably going to be sufficient to only support int16, int32, and int64
> index types for almost all types and insert implicit casts (once we implement
> implicit-cast-insertion into the execution code) for other index types. If we
> determine that there is some performance hot path where we need to specialize
> for other index types, then we can always do that.
> Additionally, we should be able to collapse the date/time kernels since we're
> just moving memory.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
