[
https://issues.apache.org/jira/browse/MATH-732?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13202115#comment-13202115
]
Sébastien Brisard commented on MATH-732:
----------------------------------------
Hi Kurt,
your implementation is amazingly faster. I think we definitely should go ahead
and replace the previous impl. We require all pieces of code (including private
methods) to be fully documented. I'll do that, no worries, but I might get back
to you (off-list) to get some more information, if you do not mind.
Unfortunately, I haven't had time to try and spot the weaknesses of the
previous implementation. I suspect this has to do with the use of {{Complex}}
arrays, instead of plain {{double}} arrays. Indeed, since {{Complex}} is
immutable, we end up creating *a lot* of small objects.
In fact, looking at your implementation, I'm wondering whether we should not
give up {{Complex[]}} arrays altogether. I would favour a more standard data
layout, in a {{double}} array, where {{data[2 * i]}} would be the real part,
and {{data[2 * i + 1]}} would be the imaginary part of the specified data.
Presently, your implementation takes the real and imaginary parts as two
separate arrays. What do you think?
I will post this question on the mailing list.
Best regards,
Sébastien
> Major Speed Improvement to 1D Discrete Fourier Transform (approximately 5x-9x
> improvement). Preserved public API 100%. Removed unnecessary use of instance
> variables and instance state.
> ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: MATH-732
> URL: https://issues.apache.org/jira/browse/MATH-732
> Project: Commons Math
> Issue Type: Improvement
> Affects Versions: 3.0
> Reporter: Kurt Ostfeld
> Assignee: Sébastien Brisard
> Labels: api, perfomance
> Fix For: 3.0
>
> Attachments: DFTPerformanceWithPatch.png,
> DFTPerformanceWithoutPatch.png, FastFourierTransformer.java,
> FastFourierTransformerTest.java, Main.java
>
>
> I wrote my own Discrete Fourier Transform function in Java and ran some
> benchmarks and found that it ran dramatically faster than the Apache library
> implementation. This is a pretty straight forward implementation of the
> standard Cooley Tukey algorithm that I read out of a textbook. This passes
> all the Apache library test cases plus several I had written on my own. I
> created a source code library patch that preserves the public API completely
> while changing the internal implementation to achieve the performance
> improvement.
> In addition to the performance issue, I suggest that Discrete Fourier
> Transform functionality be provided as stateless pure functions (in Java this
> would be implemented with static methods) just like most other math
> functions. As-is, the library requires the user to instantiate a Transformer
> instance which maintains instance state, which is an unecessary complexity
> for a pure math function. I held off on this change since it would change the
> public API and affect existing code. However, I see similar backward
> compatability breaking API changes are already in the FastFourierTransformer
> class in the 3.0 code base.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
