[
https://issues.apache.org/jira/browse/NUMBERS-156?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17351014#comment-17351014
]
Alex Herbert commented on NUMBERS-156:
--------------------------------------
Thanks for the update. It is curious that the sqrt2 cannot reduce the error to
0 ulp. Did you ensure the sum and the roundoff were rebalanced to the
non-overlapping split number (x, xx) before the call to sqrt2?
In this case it appears the extra overhead is not worth it for a marginal gain
in precision (lower standard deviation). The {{extLinearSinglePass}} algorithm
is a good choice. It just requires the scaling thresholds to be updated to
handle extremely long input vectors. I think that optimising it for 2D and 3D
can avoid using 3 individual sums with 3 compensations. It is easy to find the
absolute maximum and minimum and then do the same scaling on all inputs as
appropriate. No need to scale each differently. This is what the {{extLinear}}
algorithm is doing (all input cloned and scaled with the same value) and in the
results it has the same accuracy and precision as {{extLinearSinglePass}}.
Do you propose to add a 2D version for completeness or leave that to Math.hypot?
> SafeNorm 3D overload
> --------------------
>
> Key: NUMBERS-156
> URL: https://issues.apache.org/jira/browse/NUMBERS-156
> Project: Commons Numbers
> Issue Type: Improvement
> Reporter: Matt Juntunen
> Priority: Major
> Attachments: performance-all.png, performance-len-1-5.png,
> performance2-1-5.png, performance2-all.png, performance3-1-5.png,
> performance3-all.png, performance4-1-5.png, performance4-all.png,
> performance5-1-5.png, performance5-all.png
>
>
> We should create an overload of {{SafeNorm.value}} that accepts 3 arguments
> to potentially improve performance for 3D vectors.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
