Currently the scan uses Boyer-moore method and its performance is good.
but, it can be optimized from an implementation of view.
The original scan code is implemented by byte array and index-based access.
In _scan_for_start_code(), the index is increasing from start to end and the
base address of the byte array is referred to as return value.
In the case, index-based access can be replaced by pointer access, which
improve the performance by removing index-related operations.
Its performace is enhanced by approximately 8% on arm-based embedded devices.
Although it seems trivial, it can affect the overall performance because the
_scan_for_start_code() function is very often called when H.264/H.265 video is
played.
In addition, the technique can apply for all architectures and it is good in
view of readability and maintainability.
https://bugzilla.gnome.org/show_bug.cgi?id=731442
Adds a utility struct that is capable of storing and aggregating flow returns
associated with pads.
This way all demuxers will have a standard function to use and have the
same expected results.
Includes tests.
https://bugzilla.gnome.org/show_bug.cgi?id=709224
Buffer pool set_config() may return FALSE if requested configuration needed
small changes. Reget the config and try setting it again (validating the
changes first). This ensure we have a configured pool if possible.
https://bugzilla.gnome.org/show_bug.cgi?id=727916
exit() will call atexit handlers, which may try to
clean up things or wait for things to get cleaned up,
which we don't want or need. We just want to stop
and let the parent know about the failure as quickly
as possible in case fork() is used.
Fixes timeouts on assert failures in checks where
an exit handler waits for things to stop, but they
don't stop because they haven't been shut down,
and they haven't been shut down because there's no
simple way to do so on failures.
http://sourceforge.net/p/check/patches/50/
Currently, if prepare() takes too much time, we skip the call to render().
The side effect of this, is that we endup starving the render(). The solution
in this patch is to always render frames that are on time before prepare() is
executed. This will maximize the number of frames we display and smoothly
degrade the rendering performance.
https://bugzilla.gnome.org/show_bug.cgi?id=729335
They are very confusing for people, and more often than not
also just not very accurate. Seeing 'last reviewed: 2005' in
your docs is not very confidence-inspiring. Let's just remove
those comments.
Keep it simple. Likely also makes things easier for bindings,
and efficiency clearly has not been a consideration given how
the existing code handled these lists.
In order to be deterministic, multiple waiting GstClockIDs needs to be
released at the same time, or else one can get into the situation that
the one being released first can add itself back again before the next
one waiting is released.
Test added for new API and old tests rewritten to comply.
This reverts commit b9313afc75.
This should be fixed in upstream libcheck instead. We want
to keep diff of our local copy to upstream libcheck
to a minimum.
We iterate the current discont group backwards and push each GOP forwards,
starting from the last one. However if the first buffer in the current
discont group is a keyframe, we will keep it around until next time,
which is far from ideal. Just push it.
This prevents situations where a first branch would get seeked and
receive a buffer before all branches got seeked, and thus collected
would get called based on EOS from the previous segment.
As a consequence, during the process of seeking, don't decrease
the eospads number when a FLUSH_STOP is received.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=724571
GST_CHECKS can be simply "test*" to run run all tests (including those that are
marked broken). Update the sparse comments a bit to tell how this works.
Don't set the size to -1 in automatic_eos mode (which also updates the
duration to -1). We only want automatic_eos mode influence the maxsize
calculations without any side effects.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=724564
This defaults to TRUE and if it is set to FALSE it is the subclasses
responsibility to return GST_FLOW_EOS from the create() vmethod once
the stream is done.
Store the eos event seqnum and use it when creating the
new eos event to be pushed downstream. To know if the eos
was caused by the eos events received on send_event, a
'forced_eos' flag is used to use the correct seqnum on
the event pushed downstream.
Useful if the application wants to check if the EOS message
was generated from its own pushed EOS or from another source
(stream really finished).
Also adds a test for this
https://bugzilla.gnome.org/show_bug.cgi?id=722791
If on passthrough during reverse playback, do not accumulate buffers as
baseparse will never check for DISCONT flag to push those buffers.
So just push buffers downstream as if it was forward playback.
https://bugzilla.gnome.org/show_bug.cgi?id=721941
TIME segments are being ignored and a standard initialized
segment is used instead. This causes issues as not properly detecting
reverse playback or not cliping output based on the segment.
This seems to be a regression from one of the GstSegment/GstEvent
redesigns on the 0.10 -> 1.0 transition
It wasn't required, instead baseparse was using it to check the media
caps to identify if it was handling audio or video.
The pending_segment was removed and a checked_media boolean
replaced it for a more accurate naming.
https://bugzilla.gnome.org/show_bug.cgi?id=721350
A GAP event is handled as an empty buffer by sinks and they expect
to receive start up events before GAP events (like a segment).
This is important specially if there is a GAP at the beginning of
a stream (before any buffers) so that the segment event can be
pushed downstream before the GAP
https://bugzilla.gnome.org/show_bug.cgi?id=721350
* fix typo GstBufferFlag -> GstBufferFlags
* fix typo GstFeatures -> GstCapsFeatures
* fix typo GstAllocatorParams -> GstAllocationParams
* fix typo GstContrlSources -> GstControlSource
* do not refer to gstcheck as an object
* make references gtk_init() and tcase_set_timeout() not be references
* gst_element_get_pad() renamed gst_element_get_static_pad()
* gst_clock_id_wait_async_full() renamed gst_clock_id_wait_async()
* _drop_element() is really gst_queue_array_drop_element()
* gst_pad_accept_caps() was removed, do not refer to it
* separate GST_META_TAG_MEMORY_STR declaration from description
* do not describe removed gst_collect_pads_collect()
* correctly link to GstElementClass' virtual set_context()
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=719614
Fix a typo in a doc string - the property is round-trip-limit, not
roundtrip-limit.
Remove a bogus GST_WARNING that can print an uninitialised variable
and is redundant anyway.
Sometimes, packets might take a very long time to return. Such packets
usually are way too late and destabilize the regression with their
obsolete data. On Wi-Fi, round-trips of over 7 seconds have been observed.
If the limit is set to a nonzero value, packets with a round-trip period
larger than the limit are ignored.
Signed-off-by: Carlos Rafael Giani <dv@pseudoterminal.org>
https://bugzilla.gnome.org/show_bug.cgi?id=712385
Keep a rolling average of the round trip time for network clock
observations, favouring shorter round trips as being more accurate.
Don't pass any clock observation to the clock slaving if it has a
round-trip time greater than 2 times the average.
Actual shifts in the network topology will be noticed after some
time, as the rolling average incorporates the new round trip times.
Even though this parameter is not used, it should be const to fit in with the
coding standards for other similar parameters. Client code already passes in
const strings under the expectation that they won’t be modified.
https://bugzilla.gnome.org/show_bug.cgi?id=710442
pads->data is the public list. It is dynamically rebuilt at each call to
check_collected, in check_pads to be specific. When you add a pad and
collectpads have been started, it is not added to the public list.
Thus there exists a possible race where :
1) You would add a pad to collectpads while running.
2) You set collectpads to flushing before check_collected has been called again
-> the pad is not set to flushing
3) the pad starts pushing data as downstream might not be prepared, in the case
of adder it then returns FLOW_FLUSHING.
4) elements like demuxers, when they get a FLOW_FLUSHING, stop their tasks,
never to be seen again.
https://bugzilla.gnome.org/show_bug.cgi?id=708636
The change should have been from PARAM_CONSTRUCT_ONLY to
PARAM_CONSTRUCT, otherwise bindings are affected, since
they look for the CONSTRUCT flag.
See ec55363d
The seqnum of the segment after a seek should be the same of
the seek event. Downstream elements might rely on seqnums to
identify events related to a seek.
This is particularly important when a demuxer maps a TIME seek
into a BYTES seek for upstream and it needs to identify the
corresponding segment event and map it back into TIME to push
downstream, possibly using the values from the original seek
event.
https://bugzilla.gnome.org/show_bug.cgi?id=707530
If a pad is removed while a collectpads element (say adder) is in a chain
function waiting to be collected, there is a possibility that an unref happens
on a NULL pointer.
https://bugzilla.gnome.org/show_bug.cgi?id=707536
This avoids triggering plenty of extra code/methods/overhead downstream when
we can just quickly check whenever we want to set caps whether they are
identical or not
https://bugzilla.gnome.org/show_bug.cgi?id=706600
Use custom code to implement flush-stop, we can't reuse the set_flushing code
because we can't touch the live_playing flag and we need to signal the
streaming thread.
In some specific cases (like transmuxing) we want to force the element
to actually parse all incoming data even if the element deems it is not
necessary.
This property simply ignores requests from the element to enable passthrough
mode which results in processing always being enabled.
https://bugzilla.gnome.org/show_bug.cgi?id=705621
Adds a variant of the _push function that doesn't check the queue limits
before adding the new item. It is useful when pushing an element to the
queue shouldn't lock the thread.
One particular scenario is when the queue is used to serialize buffers
and events that are going to be pushed from another thread. The
dataqueue should have a limit on the amount of buffers to be stored to
avoid large memory consumption, but events can be considered to have
negligible impact on memory compared to buffers. So it is useful to be
used to push items into the queue that contain events, even though the
queue is already full, it shouldn't matter inserting an item that has
no significative size.
This scenario happens on adaptive elements (dashdemux / mssdemux) as
there is a single download thread fetching buffers and putting into the
dataqueues for the streams. This same download thread can als generate
events in some situations as caps changes, eos or a internal control
events. There can be a deadlock at preroll if the first buffer fetched
is large enough to fill the dataqueue and the download thread and the
next iteration of the download thread decides to push an event to this
same dataqueue before fetching buffers to other streams, if this push
locks, the pipeline will be stuck in preroll as no more buffers will be
downloaded.
There is a somewhat common practice in dash streams to have a single
very large buffer for audio and one for video, so this will always
happen as the download thread will have to push an EOS right after
fetching the first buffer for any stream.
API: gst_data_queue_push_force
https://bugzilla.gnome.org/show_bug.cgi?id=705694
When the range for a property is defined as -INT_MAX-1 .. INT_MAX, like
the xpos in a videomixer the following expression in the macro
definitions of convert_g_value_to_##type (and the equivalent in
convert_value_to_##type)
v = pspec->minimum + (g##type) ROUNDING_OP ((pspec->maximum - pspec->minimum) * s);
are converted to:
v = -2147483648 + (g##type) ROUNDING_OP ((2147483647 - -2147483648) * s);
(2147483647 - -2147483648) overflows to -1 and the net result is:
v = -2147483648 + (g##type) ROUNDING_OP (-1 * s);
so v only takes the values -2147483648 for s == 0 and 2147483647
for s == 1.
Rewriting the expression as minimum*(1-s) + maximum*s gives the correct
result in this case.
https://bugzilla.gnome.org//show_bug.cgi?id=705630
Calling gst_buffer_get_size represented 2/3 of the cost of helper_find_peek
which was called whenever a typefindfunction wanted to peek at data.
We already know the size (from the GstMapInfo), so just use that.
Pass the fixed caps we're asked to accept as a filter for the caps
query, so we don't get a fully-expanded set of caps back (which we don't
need and can take a lot of time for intersection).
This reduces the time for camerabin to produce a second frame on a
logitech C910 camera from around 52 seconds to a bit less then 16
seconds on my system.
https://bugzilla.gnome.org/show_bug.cgi?id=702632
When we asynchronously go from READY to PLAYING, also call the
state change function so that subclasses can update their state for PLAYING.
Because the PREROLL lock is not recursive, we can't make this without
races and we must assume for now that the subclass can handle concurrent calls
to PAUSED->PLAYING and PLAYING->PAUSED. We can make this assumption because not
many elements actually do something in those state changes and the ones that
did would be broken even more without this change.
https://bugzilla.gnome.org/show_bug.cgi?id=702282
Doing it after every single create() is not very efficient and not necessary.
Especially on network file systems fstat() is not cached and causes network
traffic, making the source possibly unusable slow.
https://bugzilla.gnome.org/show_bug.cgi?id=652037
This makes sure that at least one buffer per second is rendered if buffers
are dropped before ::prepare. Without this change, at least one buffer per
second wouldn't be too late before ::prepare anymore but would be dropped
before ::render because of last_render_time being set before ::prepare
already.
This function works just like gst_data_queue_pop, but it doesn't
remove the object from the queue.
Useful when inspecting multiple GstDataQueues to decide from which
to pop the element from.
Add: gst_data_queue_peek
Importantly, this patch converts DTS to running time. Less importantly,
and possibly a problem for some muxers, is that it orders buffers by
DTS (if it is valid, otherwise PTS). This is generally correct, but
might be somewhat surprising to muxers.
Also note that once converted to running time, DTS can end up negative.
gst_pad_get_current_caps() on the source pad might yield NULL caps
if we're being shut down and the source pad has already been
deactivated by the other thread that's changing state. Just bail
out in that case, instead of passing NULL caps to the transform_size
function, which it might not expect.
Fixes spurious warnings in audioresample shutdown unit test.
https://bugzilla.gnome.org/show_bug.cgi?id=693996
... and tracking of DTS. Fixes cases where PTS is locked on to the
DTS of an incoming buffer with no PTS with invalid data, leading to
no outgoing PTS (since it is not allowed smaller than DTS).
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=691481
Use GSIZE_TO_POINTER instead. sizeof(GType) may be larger
than sizeof(gulong) and sizeof(int), so the casts may
chop off some bits from the GType value on some architectures.
Don't retry to negotiate when we fail to negotiate but instead produce a
NOT_NEGOTIATED error. We only want to retry negotiation if the result from
gst_pad_push() returned NOT_NEGOTIATED.
When negotiation fails, mark the pad as needing a reconfigure again so
that it gets picked up again next time.
Signed-off-by: Niv Sardi <xaiki@evilgiggle.com>
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=691986
The _1_0 suffixed environment variables override the
non-suffixed ones, so if we're in an environment that
sets the _1_0 suffixed ones, such as jhbuild, we need
to set those to make sure ours actually always get
used.
Useful for video parses that want to attach matter or
find out if downstream supports certain metas.
API: GstBaseParseClass::src_query()
API: GstBaseParseClass::sink_query()
https://bugzilla.gnome.org/show_bug.cgi?id=691475
Add a max-bitrate property that will slightly delay rendering of buffers if it
would exceed the maximum defined bitrate. This can be used to do
rate control on network sinks, for example.
API: GstBaseSink::max-bitrate
API: gst_base_sink_set_max_bitrate()
API: gst_base_sink_get_max_bitrate()
Large streams would index one frame every second, which can get quite
large with multi-hour streams, so add an additional byte-based
minimum distance as well, which will kick in for long streams
and make sure we never have more than a couple of thousand index
entries.
https://bugzilla.gnome.org/show_bug.cgi?id=666053
Using multiple libraries causes problems for the C# bindings and
will for similiar languages such as Java when there are bindings
for them.
Also change --library=libgstfoo-X.la to --library=gstfoo-X as
the man page suggests it should be done.
https://bugzilla.gnome.org/show_bug.cgi?id=679315
Use a new GCond, protected with the object lock, to signal completion
of the async state change. We can't reuse the live lock because that
one can be locked when the create function blocks.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=686723
gst_base_src_start_complete() can fail when the thread could not be
started, for example. Make sure it causes the state change to fail by
retrieving the result from _start_complete().
Basetransform attempts to do passthrough mode regardless of the order of
the transform_caps method. Add a method to disable this.
This is needed for elements like capsfilter that want to transform caps
based on the order of the caps property.
The 3rd parameter of gst_base_src_new_seamless_segment in
0.10 is the time associated with the start of the new segment,
not the position in the new segment. Fix the name of the parameter,
the docs, and the implementation to match the needs of the only
extant consumer: DVD playback.
It's not right, and we don't know what extra properties
that event might have set in future (e.g. sparseness).
This change means collectpad users need to create their
own stream-start event now. We could add a utility
function that creates a stream-start event based on
the input stream-start events.
Hacky, because the still-frame code all lives in -base, where we
can't use it - so this is a hacky duplication of -base code. Not
sure which way to fix this: Move baseparse to -base, or move still-frame
events to core?
Make the event handling more like what videodecoder does,
to ensure that all events are passed to child classes before being
placed on the pending queue or pushed onward.
We only deal in TIME format ourselves, but if the subclass can handle
converting other formats into TIME format, we can support that too.
Fixes seeking in DEFAULT (sample) format with flacparse,
and the flacdec unit test.
Sometimes a transform filter would need the buffer pool or the memory
allocator negotiated by the base class, for example, for querying different
parameters, such as a bigger number of buffers to allocate by the buffer pool.
This patch expose a two getters accessors: one for the buffer pool and the
other for the memory allocator.
Sometimes the sources would use the buffer pool or the memory allocator for
something else than just allocating output buffers; for example, querying for
different parameters, such as a bigger number of buffers to allocate by the
pool.
This patch expose a two getters accessors: one for the buffer pool and the
other for the memory allocator.
Don't just return FALSE for seek events with negative rates when
operating in push mode. An upstream demuxer may support this just
fine, so if we're not operating in pull mode always check upstream
first if it can handle the seek event. This fixes reverse playback
where the upstream demuxer supports it (e.g. with qtdemux). The
same code would work fine in 0.10, because baseparse will just
call the default pad event handler if FALSE was returned from the
baseparse event handler, and the pad event handler will just
forward it upstream. In 0.11 the baseclass or subclass is
responsible for chaining up to the parent class or forwarding the
event upstream in any case.
Disable reverse playback in pull mode for now, there seems to
be something going wrong with the segment configuration in that
case.
No need to use a custom main context and custom timeout sources,
just use g_socket_condition_timed_wait() instead, which was added
for exactly this case.
Also seems to help with the unit test deadlocking with glib 2.33.x
https://bugzilla.gnome.org/show_bug.cgi?id=681575
The fail() definition was changed to not fail with non-GCC compilers,
unfortunately the change was incorrect and appended the first argument
of fail to the expression string instead of making it the message.
This change does mean that fail() now requires a message to be passed
along.
https://bugzilla.gnome.org/show_bug.cgi?id=680755
This specifies if a given taglist applies to the complete
medium or only this specific stream. By default a taglist
has a stream scope.
Fixes bug #677619.
Define a 0 and -1 step amount. They used to almost do the same thing but now, 0
cancels/stops the current step and -1 keeps on stepping until the end of the
segment.
See https://bugzilla.gnome.org/show_bug.cgi?id=679378
Move code that checks for upstream seekability and all that to
the right place, otherwise it will never be done for formats
that have headers such as FLAC, as handle_and_push frame will
be called the first time only after headers have been processed
(and framecount is > 0). This then makes us report that we
can't seek, which disables the seek bar in totem.
when we have a new step event with a -1 amount, make sure that we follow the
regular code path so that the stop_end handler is called as usual. This takes
care of flushing the buffer in case of a flushing step and also posts a step end
message.
See https://bugzilla.gnome.org/show_bug.cgi?id=679378
In 0.11 the caller may provide a buffer to be filled by the source to
pull_range/get_range/create, but it's easy to miss this new case when
porting code from 0.10. Provide fallback that copies the created data
into the provided buffer for now.
This makes oggdemux in pull-mode work with dataurisrc.
Make gst_query_add_allocation_meta() take a copy of the passed caps instead of
taking ownership. This makes it easier for the caller in most cases because it
doesn't have to make a copy and deal with NULL values.
Make GstAllocator a GstObject instead of a GstMiniObject, like bufferpool.
Make a new gstallocator.c file. Make a GstAllocator subclass for the default
allocator.
Make it possible to add API specific flags to the ALLOCATION query. This makes
it possible to also check what kinds of subfeatures of the metadata API are
supported.
This is a queue which has the same API as GQueue, except that:
* It uses an array, instead of a doubled-linked-list
* The array can only grow.
This code is not-threadsafe. It is up to the owner to make sure the
proper locking is taken before calling this API.
Rename gst_base_sink_wait_eos() to gst_base_sink_wait() to avoid confusion and
introspection problems with the ::wait_eos vmethod. Also this method can be used
to wait for other things than EOS. Update the docs a little.
Add the running-time of the buffer that caused the async operation to complete
to the async-done message.
Update bin to handle the new async-done message.
Use the new RESET_TIME message to reset the start-time of the pipeline to the
requested time.
Make basesink request a new running-time when the flush-stop message tells it to
insteasd of waiting for preroll.
Use the pad methods to set and check the reconfigure flags
Clear the reconfigure flag before we negotiate so that we don't miss any
reconfigure events while negotiating
Release the object lock before we get the time of the clock because that code
might take other locks.
Fix potential clock refcount error because we released the object lock but
didn't ref the clock.
... since processing might still continue (if e.g. NOT_LINKED)
and then proper state (e.g. offset) needs to be maintained
(e.g. to arrange for a new frame setup).
The proto for helper_find_suggest has a different argument than the actual
function in the same file has. This causes the Sun Studio compiler to fail.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=676624
Conflicts:
libs/gst/base/gsttypefindhelper.c
Add a prepare method that is called before sync happens. The purpose of this
method is to prepare the rendering of the giving buffer so that the following
render() call after sync is a quick as possible.
In gst_base_src_start_complete() we do a perform_seek() that will eventually
start the streaming thread which acquires the live lock and then goes to sleep
in the case of appsrc. Right after we perform seek we also try to acquire the
live lock which might then deadlock.
fix this by taking the stream lock before performing the seek. This makes sure
that the streaming thread cannot start and grab the live lock until we are done
and release the stream lock again.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=676048
This allows subclasses to override it, as is necessary for e.g. the
video-crop meta. It is now necessary that after decide_allocation()
there is always a allocator and a configured buffer pool inside the
query.
It causes the timestamp to go wrong, should not cause much of a performance
increase and in the cases where it is faster, it is broken in 0.10 as well.
We should try to review this when rewriting the adapter for 0.11 memory
features.
Fixes https://bugzilla.gnome.org/show_bug.cgi?id=674791
Rename the _get_value_array() functions to _get_g_value_array() and reintroduce
the former to operate on plain unboxed c datatypes (like in 0.10). The _g_value
variants are for bindings while the _value ones are more suited to processing
in elements.
