[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Request for Sponsorship: VeryFastTree - Parallelized and Optimized Version of FastTree



> > Thanks Nilesh. I have created a new implementation based on your
> > guidelines.
> >
> > I am still considering whether it would be necessary to include all
> > possible
> > compilations. Perhaps with only SSE2, AVX2, and AVX512F, it would be
> > sufficient,
> > or even just SSE2 and AVX2. What do you think, Andreas?
>
> I admit I need to fully trust your insight and possibly the performance
> tests you did.  I have not inspected your code - not to mention that I
> do not feel competent to judge about hardware optimisation means.

More than hardware optimizations, I mean that it may not make sense to increase the
package size with 5 binaries when SSE (the minimum required by Debian) and
AVX2 (more common in current hardware) could be sufficient. AVX512, in the
tests conducted, doesn't have as much impact as SSE to AVX2, and outside of
servers, it is still not as common. The intermediate extensions are the
same. If current computers have AVX2, the previous ones don't make sense,
and for the rest, SSE is sufficient.

Regarding the performance differences, they are analyzed in the supplementary material
of the VeryFastTree paper, which is open access. You can find it here:
https://academic.oup.com/bioinformatics/article/36/17/4658/5861530

> > The only change would be in the rules. I'm leaving the code here before
> > uploading

> Just for the terminology:  You are not permitted to "upload" - this can
> only be done by a Debian Developer or Maintainer.  What you can do is
> pushing to the Git repository and doing so is fine.  The package was
> uploaded to new queue and any changes you do will go to the next release
> of the package (in case it will be accepted by ftpmaster).

Yes, I was referring to pushing the changes to the repository.
Sorry for the terminology.

> >  it in case you have any suggestions.
>
> I had only a quick view on the code.  If you tested this change in
> d/rules and the resulting package contains the changes you want to
> implement feel free to push.
>
> Kind regards and thank you for your contribution
>     Andreas.


Should I push it as is or should I delete some binaries?
Approximately, the size of the binaries would be:
AVX512: 10MB, AVX2: 5MB, AVX: 5MB, SSE4: 3MB SSE: 2.7MB = 25.7MB
for the package seems like a lot when perhaps AVX2 and SSE with 7.7MB would be sufficient.

César


Reply to: