MJPEG Tools Svn

This is what I'm calling "version 3 beta" of y4mdenoise! I've finally
found a practical implementation of my original idea.

Now, each matched pixel-group is flood-filled, and the resulting region
is added to the search-border. Matched pixel-groups that already exist
in the search-border are not flood-filled again, saving lots of time.
The match-throttles still exist to keep pathological cases from
happening, but now regions can be removed from the search-border,
pruned of resolved pixels, and re-added to the search-border if it's no
longer the biggest match. y4mdenoise is finally able to apply the
largest matches and work its way down to the smaller matches without
losing any potentially useful information.

This is "beta" because I haven't finished all the performance
improvements I want to do, and I haven't rewritten the manual page yet,
but it's working well enough to unleash on the other mjpegtools
developers for their comments.

Fixed more multithreaded problems in y4mdenoise. I wasn't using enough
private allocators, and so threads could potentially stomp each other
when allocating memory.

Another fix to y4mdenoise's threading, this time to get rid of a
potential double condition-signal. (One condition-signal is enough.)

If zero is specified for -z/-Z, skip the zero-motion pass entirely.
(I thought I had done this already...oops!)

y4mdenoise would sometimes hang if an earlier process in the pipeline
quit unexpectedly, preventing the pipeline from cleaning up properly.
The read-ahead thread just has to wake up its client before it exits
with an error, and now it works.

Fixed the installation directory of the mjpeg-howto file.
Line 103 creates a ${_prefix}/share/info directory, but the %post
and %postun sections referred to ${_prefix}/info . Now those
sections properly refer to ${_prefix}/share/info.

Lots of changes to y4mdenoise that I've been sitting on for a while.

1) Updated the copyright year.
2) Fixed C++ usage for GCC 4.3.0, mostly related to templates.
3) Diked out usage of SearchBorder from MotionSearcher, since it's
not currently in use.
4) -p option is now a bitfield. No more forcing read/write threads
on someone just because they want color to be denoised in a
separate thread!
5) Reference pixels now calculate the pixel value in AddSample(), not
GetValue(). This should perform much better.
6) Threading rewritten so that it doesn't suck royally.
7) Better denoising algorithm: if match-size-throttling is in use
(which it usually is), it picks the largest such match, not merely
the first one bigger than the match-size-throttle. This leads to a
pretty serious quality increase!
8) Specified GPL v2 explicitly. (No accidental GPL v3!)
9) Added comments, debugging code, etc.

I deeply apologize if I broke the build. I tried to integrate in the
changes others made over the years, but most of my changes
conflicted (which isn't surprising).

Sigh, some distributions don't like Quicktime. Avoid crashing the program due
to uninitialized pointer if HAVE_LIBQUICKTIME is not defined.

Use memmove instead of bcopy. Seems some linux systems have issues with bcopy on 64bit AMD systems

Added a patch that tests for a broken sdl.m4
Found in the tracker: Remove AC_CANONICAL_TARGET call
http://sourceforge.net/tracker/?func=detail&atid=305776&aid=2491094&group_id=5776

sdl is fixed and that test is not needed any more.

Posted by: dgp85