[
https://issues.apache.org/jira/browse/ARROW-14481?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17434613#comment-17434613
]
Eduardo Ponce commented on ARROW-14481:
---------------------------------------
Looking more carefully ARROW-13879 does modifies code used by unary string
transform and predicate functions:
1. [Kernel registration was slightly modified here for string UTF8
transforms|https://github.com/apache/arrow/pull/11233/files#diff-eb8300bc4dea7d1c46b2576b7dbd8e42b927ab7d42c031f4aecae892a72ee244L4162-R4229]
which converts 2 explicit statements into a for-loop along with a function
call to a generator dispatcher.
2. [A similar change is here for the unary string
predicates.|https://github.com/apache/arrow/pull/11233/files#diff-eb8300bc4dea7d1c46b2576b7dbd8e42b927ab7d42c031f4aecae892a72ee244L4210-R4273]
3. Also, [string predicates exec functor was modified
here|https://github.com/apache/arrow/pull/11233/files#diff-eb8300bc4dea7d1c46b2576b7dbd8e42b927ab7d42c031f4aecae892a72ee244L4171-R4259]
to provide the predicate as a template parameter instead of as a function
parameter.
Now, #3 should not have a noticeable impact, but #1 and #2 can because of the
loop iteration and extra function call.
Nevertheless, these are code paths that are not critical and should only be
called once.
> [C++] Investigate recent regressions in some utf8 kernel benchmarks
> -------------------------------------------------------------------
>
> Key: ARROW-14481
> URL: https://issues.apache.org/jira/browse/ARROW-14481
> Project: Apache Arrow
> Issue Type: Improvement
> Components: C++
> Reporter: David Li
> Priority: Major
> Labels: kernel
>
> See [https://conbench.ursa.dev/benchmarks/6ccff6887e7c47148a09fe46f18c8688/]
> Some (on the surface) unrelated commits have caused performance for a few
> string kernels to plummet. We should try to replicate locally.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
