[opensource-dev] J2C fast decoder

[Philippe (Merov) Bossut]

Hi Nicky,

As it happens, I've been working on instrumenting the code to add metric
gathering for image decompression as part of the Snowstorm sprint.

You may want to use my branch (
https://bitbucket.org/merov_linden/viewer-development-vwr-22761) and create
a baseline for openjpeg then run a test for Jasper. You'll have to sort out
the failing cases certainly and just throw them so we compare what gets
truly decompressed (though, clearly, working in all cases is pretty critical
if we look at Jasper as an alternative).

Here's what I got comparing KDU and OpenJpeg:
Label     Metric                              KDU(B)     OJ2C(T)
 Diff(T-B)     Percentage(100*T/B)
ImageCompressionTester-1
     TotalBytesInDecompression    5048643    5003370    -45273        99.1
     TotalBytesOutDecompression 40415336  46592896    6177560    115.29
     TimeTimeDecompression        3.74           17.04          13.3
        455.39
ImageCompressionTester-2
     TotalBytesInDecompression    5000744    5000144     -600
99.99
     TotalBytesOutDecompression 46440040  44248324   -2191716    95.28
     TimeTimeDecompression        3.64           15.02           11.37
     412.02

For that test, I output data every time 5MB of compressed data have been
processed. It's partial but shows that OpenJpeg is roughly 4 times slower
than KDU (at least, the version we're using in the official viewer
currently). Would be nice to have a similar set of numbers for Jasper before
going too far down the implementation path.

I wrote a short (and still incompleted) wiki to explain a bit how the metric
gathering system works:
- https://wiki.secondlife.com/wiki/Performance_Testers

BTW, that's something we should be using more generally for other perf
sensitive areas, especially when starting a perf improvement project.

See http://jira.secondlife.com/browse/VWR-22761 for details.

Cheers,
- Merov

On Fri, Sep 3, 2010 at 9:05 AM, Nicky Fullton <nickyd_sl at yahoo.com> wrote:

> Hello,
>
> >> i'm testing in RL office (not or a viewer) JasPer decoder for JPG2000
> >> images, after a short test with openjpeg2000 from EPFL we have tested
> >> last 3 days JasPer (only a POC apps to do some bench), we must do a lot
> >> of work too, but this is a lil question... anybody here around never
> >> tried it as alternative to OpenJPEG/KDU in a viewer?
>
> >I'm not aware of anyone publishing results for such a test, but if you
> >have the time it would be interesting reading.
>
> You might be interested in:
> http://bitbucket.org/NickyD/viewer-development/changeset/027bf44c5582
>
> I made a rather quick hack to try Jasper instead of OpenJpeg to decode
> images.
>
> The patch has some very rough edges. In fact is the decoding into the
> LLImageRaw buffer not correct.
>
> I did not fix this (yet) because the results so far are not very promising.
> Jasper can only decode around 20% of the jpeg, for the other 80% it will
> create an error and then my code falls back to OpenJpeg.
> This fallback makes the whole decoding rather slow, so it is hard to say
> if Jasper would really be any faster.
>
> Right now I am not sure if it would be reasonable to invest more time
> looking at Jasper. First the code would need to fixed upstream, so all
> images can be properly decoded. As this project looks rather dead, one
> with JPEG2000 knowledge might have to step up for this.
>
> On another note, you might like to try:
> http://bitbucket.org/NickyD/viewer-development/changeset/e4eff3e2af39
>
> This will at least skip the step of calling OpenJpeg in
>  LImageJ2COJ::getMetadata (if possible, it will do sanity checks first).
>
> >Some things to keep in
> >mind. OpenJpeg has patches floating around on its ML against 1.3 that
> >reports have claimed up to 40% speed increase in places due to
> >unrolling the inner loops so finding them and testing would be good.
>
> I did not find any of those, but then again maybe I did not look hard
> enough.
> There is certainly some potential in OpenJpeg.
> There are some loops in t1_dec_sigpass and t1_dec_refpass that can be
> easily rewritten. But there is some pretty tricky stuff in t1_dec_clnpass
> that would need some cleaning and mqc decoder (mqc_decode) burns a lot
> of time. But that one is especially hairy as it has side effects on its
> input parameter.
>
> I am not sure if anyone without enough deep knowledge of OpenJpeg (and
> the dedication to recode a good part of it) would be able to improve
> much of it.
>
> Cheers,
>   Nicky
>
>
>
>
> _______________________________________________
> Policies and (un)subscribe information available here:
> http://wiki.secondlife.com/wiki/OpenSource-Dev
> Please read the policies before posting to keep unmoderated posting
> privileges
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.secondlife.com/pipermail/opensource-dev/attachments/20100909/36816076/attachment.htm