[
https://issues.apache.org/jira/browse/LUCENE-9679?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17268895#comment-17268895
]
Uwe Schindler commented on LUCENE-9679:
---------------------------------------
Math.fma() is especially slow for doublearguments if the CPU does not support
it. In that case the non-intrinsified version is using BigDecimal to work
around the precision loss by the two double operations. A bit better is
Math.fma() with float agruments, because it is casting everything to double
first and rounding at end. Actually without CPU support it would round 3 times:
2 times with doubles (mul, add) and then at end again when converting double
result to float.
So with floats, using Math.fma would be an option and may improve a bit on
Haswell(or alter) CPUs. But with doubles: never-ever!
> Try using Math.fma to speed up vector computations
> --------------------------------------------------
>
> Key: LUCENE-9679
> URL: https://issues.apache.org/jira/browse/LUCENE-9679
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Michael Sokolov
> Priority: Major
>
> On the mailing list, Anton Hägerstrand <[email protected]> posted a link to a
> blog post showing hotspots in our indexing:
> https://blunders.io/posts/lucene-bench-2021-01-10 and identified vecto dot
> product computations as one of them. Not surprising in itself, but he also
> suggested we might want to try using {{Math.fma}} in the computation to see
> if it can optimize the multiply-and-add operation that is at the heart of
> both the dot product and l2-norm distance.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]