[patch] optimization for libtheora

[Andy Wingo]

libtheora is almost useless for encoding without these optimizations. For
example, on my p4 2.4ghz machine, I can't encode and stream in realtime with
flumotion.

A patch for libtheora alpha3 is attached. It came from thomas's fedora package:

http://thomas.apestaart.org/download/pkg/fedora-3-i386-gstreamer/libtheora-1.0alpha3-6.fdr.1.3/libtheora-1.0alpha3-6.fdr.1.3.src.rpm

[Andy Wingo (andywingo)]

Created an attachment (id=1205)
optimizations for libtheora

Thomas vander Stichele claims this patch does runtime cpu detection, so it can
be merged unconditionally. I haven't checked it myself though.

[Jeff Waugh (jdub)]

thanks heaps andy, runtime detection was the crucial problem with the previous
patch - i'll get this tested on our archs and into the archive if it works.

[j^ (j)]

whats the stat with adding mmx support for libtheora?
not that theora-mmx is also released along with alpha5

http://downloads.xiph.org/releases/theora/libtheora-mmx-1.0alpha5.tar.bz2

[Jeff Waugh (jdub)]

Does it do run time mmx detection?

[Andy Wingo (andywingo)]

I think xiph was looking at using liboil for the optimizations. At least that
way I have some hope of encoding on my laptop, which is ppc. Now that liboil
0.3.3 has been released with support for the theora routines, I would punt on
this patch and wait for the theora liboil patch to be included in an upstream
release.

[Michael Smith (msmith-xiph)]

(In reply to comment #4)
> Does it do run time mmx detection?

Theora-mmx? Yes.

From a quick look at the code, it does NOT, however, appear to do compile-time
mmx-detection. This means building it for non-x86 presumably will fall over (but
it's probably easy to fix); I guess that makes it inappropriate for ubuntu to
adopt wholesale - unless you do arch-specific packages?.

There are really three directions xiph is looking at for having a fast theora
that can be widely adopted:

 1) theora-mmx, as it stands currently, presumably with things like 'detecting
what you can compile' added. As well as the actual assembly code, theora-mmx
adds the infrastructure for doing cpu-specific optimisations for multiple
architectures (i.e. it doesn't just hack in the mmx code), so this _could_ be a
basis for doing this optimisation for more architectures.

 2) theora-liboil. This basically uses the same actual assembly code as
theora-mmx, but moves the optimisations to inside liboil. Some people have
expressed a preference for using this as a platform for doing optimisations for
more architectures.

 3) Longer term, derf's theora-exp implementation will, hopefully, take over.
It's higher quality code, implements the full theora spec, is faster, and has a
much nicer API. It's not ready now, though.

At this point in time, for Ubuntu, option 1 (with a few basic fixes, probably),
or option 2 would be sensible.

[Jeff Waugh (jdub)]

Unfortunately, even option 1 is too late for breezy. But with dapper, we have an
abundance of choice. :-) Does theora-exp use liboil?