You are viewing a plain text version of this content. The canonical link for it is here.
Posted to jira@arrow.apache.org by "Joris Van den Bossche (Jira)" <ji...@apache.org> on 2021/04/19 08:22:00 UTC
[jira] [Updated] (ARROW-12413) [C++][Compute] Improve power kernel
[ https://issues.apache.org/jira/browse/ARROW-12413?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Joris Van den Bossche updated ARROW-12413:
------------------------------------------
Description:
See comments at https://github.com/apache/arrow/pull/9841#issuecomment-818526907
Power kernel is modified from add/sub/mul/div kernels. But power is much more complicated than basic arithmetic operations. Some limitations in current implementation:
- Base and exponent must be the same type, this is not reasonable
- Null handling is delayed to the end, and power is always calculated even for nulls (with whatever base/exponent happen to be in the slots). This is a big waste of time for expensive power calculations.
For checked power, integer overflow is checked for every multiplication. As the integer upper bound is already know, it can be calculate directly if the power will overflow. It will remove overflow checks in each iteration, but will introduce one expensive logarithm operation. Maybe deserve a try.
was:See comments at https://github.com/apache/arrow/pull/9841#issuecomment-818526907
> [C++][Compute] Improve power kernel
> -----------------------------------
>
> Key: ARROW-12413
> URL: https://issues.apache.org/jira/browse/ARROW-12413
> Project: Apache Arrow
> Issue Type: Improvement
> Components: C++
> Reporter: Yibo Cai
> Priority: Major
>
> See comments at https://github.com/apache/arrow/pull/9841#issuecomment-818526907
> Power kernel is modified from add/sub/mul/div kernels. But power is much more complicated than basic arithmetic operations. Some limitations in current implementation:
> - Base and exponent must be the same type, this is not reasonable
> - Null handling is delayed to the end, and power is always calculated even for nulls (with whatever base/exponent happen to be in the slots). This is a big waste of time for expensive power calculations.
> For checked power, integer overflow is checked for every multiplication. As the integer upper bound is already know, it can be calculate directly if the power will overflow. It will remove overflow checks in each iteration, but will introduce one expensive logarithm operation. Maybe deserve a try.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)