Original commit message from CVS:
Implemented the hybrid scheduling system for sources and connections
outside the current Bin. Is a bit hackish in one place, but I'll work out
a way to make that cleaner soon. queue.c in tests now works nicely in all
cases. More to come later.
Original commit message from CVS:
First attempt at rebuilding the type/plugin system
- make sure caps/props are saved in the registry
autoplugging is completely broken.
removed the typefactories and typeids from all the pads in the plugins
XML load/save is ok (be sure to rerun gstreamer-register)
Original commit message from CVS:
Snapshot of work-in-progress do deal with out-of-bin elements. Current
system is very likely going to be ditched due to its complexity and the
fact that it doesn't work right now. More than that, it's generating some
really odd results in my test program, which aren't readily explainable.
If you want to keep a working copy of your working copy (heh), don't update
until this is replaced with a hybrid approach. This approach will be the
beginning of a plan generator that can construct full hybrid schedules
given hints from various places.
Hybrid means that cothreads are used, but there are chunks of the pipeline
that are actually dealt with by chaining. This can improve speed by
reducing cothread switches (which are much cheaper than thread switches,
but still far from free, about 570 cycles on my PIII), but does carry a
complexity burden. Luckily, the structure of GStreamer allows that burden
to live entirely in create_plan. Luck? I think not ;-)
Original commit message from CVS:
Another big set of changes. Connections are now also pullfunc based.
gstqueue has been updated, I don't know of any other connections offhand.
There are still a few things that need doing, specifically the concept
of a source or connection with connections to multiple thread contexts is
not dealt with. This may force us to move the threadstate from the
element to the pad, maybe keeping the element's copy for simple cases.
Then the Bin would create a structure to pass to the cothreaded _wrappers
of any such elements, which would detail the pads that are to be dealt with
by this particular cothread context.
That will speed things up to, since we don't have to look through the list
of all pads for every Src or Connection element for every iteration, we can
simply step through the list provided by the plan. Special case might even
have a single pad pointer sitting there to trump the list, if there's only
one (the common case anyway).
Task 23098 is tracking these changes. The main task 22588 depends on that
subtask, as well as 22240, which is a consistency check on PAD_DISABLED.
Original commit message from CVS:
Changed the way things are scheduled, especially sources. A Src used to
have a push() function, and optionally a pushregion() to deal with async
reads, etc. That whole thing has gone away, in favor of providing a
pull() function for the output (Src) pad instead, ala chain functions.
This makes constructing cothreaded schedules out of non-loop elements
somewhat easier. Basically there was always a question as to which pad
was being dealt with. In the pullregion case, cothread-specific data was
used to try to pass the region struct to the right place, which is a slow
hack. And in general, the push function severely limited the kind of
tricks that could be played when there's more than one output pad, such as
a multi-out file reader with async capabilities on each pad independently.
This changes the way cothread scheduling occurs. Instead of the hack to
deal with Src's by calling their push() function (or optionally the
pushregion(), in certain cases), we now are working towards a general
mechanism where pads are the only thing that are dealt with directly.
An optimization was made in the process of doing this: the loopfunction
actually run as the outer [stack] frame of the cothread is now set more
intelligently in create_plan() based on what kind of element it is. We
now have:
loopfunc_wrapper: used for loop-based elements, it simply calls the
loopfunc in a loop, paying attention to COTHREAD_STOPPING (see
below). It currently does other, soon to be depracated, stuff.
pullsrc_wrapper: wraps a Src that's not loop-based (since your options
are now loop- or pull-based)
There will be a couple more to deal with other cases, such as Connections
and chain-based elements. The general idea is that it's a lot more
efficient to make the decisions once in create_plan than to keep doing
this huge if/else chain in the wrapper. Just choose the right wrapper up
front. It'll be most apparent performance-wise in the case of whichever
element context is switched to first for each iteration, since the whole
wrapper setup is done for every iteration.
The tricky part is that there is now a bit of overloading of the function
pointers in a pad. The current meanings (possibly to change a bit more
soon) are:
chainfunc: as always, chainfunc pointer is mirrored between peer pads
(this may change, and the chain func may end up in pushfunc)
pushfunc: SrcPad: gst_pad_pushfunc_proxy, cothread_switch to peer
SinkPad: none (may take over chainfunc, see below) pullfunc:
SrcPad: Src or Connection's function to construct buffers
SinkPad: gst_pad_pullfunc_proxy, cothread_switch to peer
There are a number of issues remaining with the scheduling, not the least
of which is the fact that Connections are still dealt with the old way,
with _push() functions and such. I'm trying to figure out a way to unify
the system so it makes sense. Following the scheduling system is hard
enough, trying to change it is murder.
Another useful scheduling addition, mentioned above, is COTHREAD_STOPPING.
It's an element flag that's used to signal whatever code is running in
cothread context that it should be finishing up and exiting soon. An
example of this is in plugins/cobin/spindentity.c. All the loops should
now be composed of do/while loops, rather than while(1) loops:
do {
buf = gst_pad_pull(spindentity->sinkpad);
gst_pad_push(spindentity->srcpad,buf);
} while (!GST_ELEMENT_IS_COTHREAD_STOPPING(element));
The reason for this is that COTHREAD_STOPPING may be set before the above
loop ever gets started. It wouldn't do for the body of the loop to never
once get called, that would simply stall the pipeline. Note that only the
core library code is ever responsible for setting and unsetting this flag.
All elements have to do is respond to it by cleanly exiting the loop and
the function holding it.
This is needed primarily to allow iterations to occur properly.
Basically, there's a single entry point in the cothread scheduling loop,
gst_bin_iterate_func() simply switches to this cothread. If the element
in this context is allowed to loop infinitely, nothing would even switch
back to the context from which the iterate() was originally called. This
is a bit of a problem. The solution is for there to be an implicit switch
back to the originating context. Now, even I'm not sure exactly how this
works, but if the cothread that's switched to actually returns, execution
returns back to the calling context, i.e. iterate_func().
COTHREAD_STOPPING is therefore set just before switching into this
(currently randomly chosen) context, on the assumption that it will return
promptly after finishing its duties. The burden of clearing the flag
falls to the various wrapper functions provided by the Bin code, thus
element writers don't have to worry about doing that at all (and simply
shouldn't).
Related changes:
All the sources in elements/ have been changed to reflect the new system.
FIXMEs:
1) gstpipeline.c calls gst_src_push at some point, dunno why, it's
commented out now.
2) any other sources, including vcdsrc, dvdsrc, and v4lsrc will break
badly and need to be modified to work as pull-based sources.
Original commit message from CVS:
Changes made to the DEBUG system. New header file gstdebug.h holds the
stuff to keep it out of gst.h's hair. DEBUG prints out the process id,
cothread id, source filename and line number. Two new macros DEBUG_ENTER
and DEBUG_LEAVE are used to show the entry and exit of a given function.
This eventually might be used to construct call trace graphs, even taking
cothreads into account. This would be quite useful in visualizing the
scheduling mechanism.
Minor changes to various debug messages.
Also sitting in gstdebug.h is a prototypical DEBUG_ENTER that's capable of
performing DEBUG_LEAVE automatically. It does this by utilizing a
little-known GCC extension that allows one to call a function with the
same parameters as the current function. The macro uses this to basically
call itself. A boolean is used to ensure that when it calls itself it
actually runs the body of the function. In the meantime it prints stuff
out before and after the real function, as well as constructing a
debugging string. This can be used eventually to provide call-wide data
on the DEBUG lines, instead of having to replicate data on each call to
DEBUG. More research is needed into how this would most cleanly be fit
into some other chunk of code, like GStreamer (I think of this DEBUG trick
as a separate project, sorta).
Unfortunately, the aforementioned DEBUG trick interacts quite poorly with
cothreads. Almost any time it's used in a function that has anything
remotely to do with a cothread context (as in, it runs in one), a segfault
results from the __builtin_apply call, which is the heart of the whole
thing. If someone who really knows assembly could analyze the resulting
code to see what's really going on, we might find a way to fix either the
macro or the cothreads (I'm thinking that there's something we missed in
constructing the cothreads themselves) so this works in all cases.
In the meantime, please insert both DEBUG_ENTER and DEBUG_LEAVE in your
functions. Be sure to put DEBUG_ENTER after your variable declarations
and before any functional code, not to put the function name in any DEBUG
strings (it's already there, trust me), and put a DEBUG_LEAVE if you care
enough.
Changes are going to happen in the way DEBUGs and other printouts occur,
so stay tuned.
Original commit message from CVS:
Added a check to see if the plugin actually exists before even bothering
to try to load it. Saves a fair amount of debugging spew with
_gst_plugin_spew enabled.
Original commit message from CVS:
Added multiple types to the pads.
Added first preview of the capabilities system.
Autoplugging is seriously broken with these (and other) changes.
Original commit message from CVS:
Remove cothreads.[ch] from tests/cothreads/ and replace then with a rule
in the Makefile.am to generate symlinks to the live versions of the code
in gst/ : fixes a compile error, and should keep things cleaner.
Original commit message from CVS:
Enhanced debugging by making DEBUG() print out the cothread ID as well as
the process ID. cothread_getcurrent returns the current cothread ID, or
-1 if cothreads aren't set up in this process context.
Original commit message from CVS:
First pass at an outline for the Filter Writer's Guide. Mostly complete,
though the structure needs some work. Just need someone to write it ;-)
Original commit message from CVS:
Rearranged cothread_switch a bit to move all the code for error conditions
to the end, using goto's to get there. This has the presumed advantage of
consolidating all the normally run code into one chunk, reducing jumps
(and the associated penalties in any modern processor) and limiting cache-
line usage. It may be instructive to look at the generated assembly for
this revision and the previous, to see if gcc is smart enough to do this
for us anyway.
If you want to turn off some of the checks (they are all checks for NULL
pointers, if you're curious) for a speed gain, disable the #define of
COTHREAD_PARANOID at the top.
Original commit message from CVS:
Added code to force the gsttypes plugin to load before gstelements, by
simply having gstelements.c request the plugin. Solves some dependency
problems. This is the correct method of doing this for now, though I had
a thought:
Have a static list of hard dependencies that the plugin system is responsible
for satisfying before even trying to load the plugin. Makes plugin design
easier.
Original commit message from CVS:
Fixed a dumb mistake: in gst_element_set_loop_function, it was forcing
the cothread's entry function to change. This is totally wrong, the
create_plan function in GstBin is responsible for setting that to its
own internal helper function. Code removed outright...
Original commit message from CVS:
Closed 22145 - gstelement.c: support hotswapping of loopfunc
Added a flag GST_ELEMENT_NEW_LOOPFUNC that's set when a new loopfunc is set,
which needs to be obeyed by the loopfuncs themselves. The assumption is
that only a loopfunc or an element that knows what it's doing will ever
change the loopfunc pointer, so they can be responsible for exitting
the loopfunc itself appropriately.
The loopfunc_wrapper function has been changed to call the loopfunc
pointer itself in a loop. If the loopfunc changes, the current loop is
supposed to exit, which causes the loopfunc_wrapper to start up the new
loopfunc.
Original commit message from CVS:
Closed task 22143 - gstelemnt.c: complete gst_element_connect
SImple matter of cleaning up the conflicting checks and actually
connecting the pads in question.
Original commit message from CVS:
gstbin.c: modified so create_plan occurs after the state change of all the
child elements.
gstelement.c: set_state now loops such that each element only deals with
one state change at a time, i.e. NULL->READY,READY->PLAYING,
instead of a single NULL->PLAYING.
Original commit message from CVS:
Reworked AC3 decoder. No seeking yet but at least we do not need ac3parse
anymore.
Reworked dvdsrc to read scrambled data (DeCSS not included). I have
modified DeCSS a bit to work in GStreamer. Can I release the code or is
there some lawyer that's going to sue me?
MPEG2 SSE motion compensation.
Tried to add PTS to the MPEG decoder but failed.
Original commit message from CVS:
The bonobo component now works on audio only (mp3/vorbis). Video
is broken because bonobo does not handle multithreading yet.
Added Bonobo toolbar and menu.
Original commit message from CVS:
Docs updates.
Added LICENSE info to headers/code where missing in gst directory
Added a bonobo wrapper for the media player (it shows up in gshell but
locks up when activating the component, anyone?)
Fixed some XML save/load problems with arguments.
Original commit message from CVS:
Added seeking to the avi decoder by implementing pull_region.
Fixes to the asyncdisksrc.
Added thread specific data to the cothreads.
Original commit message from CVS:
Make sure the Gdk thread lock is not held inside the gstreamer pipeline.
Fix the videosink to not hold the Gdk lock when emitting a signal.
All Gtk GUI apps using GStreamer should now handle the Gdk locks when
handling signals from the pipeline (as it should be)
