n_frames could end up being quite big (potentially up to G_MAXINT64). Which
would result in overflowing 64bits when multiplying it by GST_SECOND.
Instead move GST_SECOND to the num argument
If we are shutting down, don't spawn a cleanup thread to cleanup old
groups and instead queue them to be cleaned up in the state change
thread.
This avoids (hopefully for good) having a race between the state change
thread and other threads trying to deactivate elements/pads.
Deactivating pads from two threads isn't 100% MT-safe. There is a
slim chance that the GstPadActivateFunc might be called twice with
the same values (in this case from the cleanup thread *and* from
the GstElement change_state function when going from PAUSED to READY).
In order to avoid that, call any existing cleanup function *before*
calling the parent change_state implementation on downwards state
changes.
When deactivating pads, we need to ensure that the streaming threads
going through the pads we wish to deactivate can cleanly return.
Failure to do that would result in the streaming locks of those
pads never being released. The end result would be a deadlock
when stopping decodebin2.
In order to avoid that situation, release the "dyn" lock around
the deactivation code. And refactor the code to cope with the
list of blocked pads having potentially changed when re-acquiring
the lock.
We have a dedicated one-shot thread to handle cleanup of old groups.
While this is a good idea. It's an even better idea to make sure
that thread is *completed* before the parsebin element to which
it is related isn't freed/gone.
* There can only be one cleanup thread happening at any point in time.
If there is already one, we wait for the previous one to finish.
* When shutting down (NULL=>READY) make sure the thread is finished
https://bugzilla.gnome.org/show_bug.cgi?id=790007
We have a dedicated one-shot thread to handle cleanup of old groups.
While this is a good idea. It's an even better idea to make sure
that thread is *completed* before the decodebin2 element to which
it is related isn't freed/gone.
* There can only be one cleanup thread happening at any point in time.
If there is already one, we wait for the previous one to finish.
* When shutting down (NULL=>READY) make sure the thread is finished
https://bugzilla.gnome.org/show_bug.cgi?id=790007
Instead of emitting 'drained' whenever every single chain is drained
(which would result in plenty of signal emission, and would also
occur when switching groups), only emit it when the top-level chain
is drained.
Furthermore, mark unknown (and therefore unexposed) pads as drained
since we'll never get EOS on them.
https://bugzilla.gnome.org/show_bug.cgi?id=787367
If we can expose the main chain, recheck whether we are shutting
down or not.
decodebin2 might have been set to READY/NULL during the attempt
to expose, which would cause it to fail ... but it is not a fatal
issue.
By select-streams event, current implementation of decodebin3
supports deactivate output stream (i.e., decoder element)
in reassign slot(), but cannot activate any slot without track change.
https://bugzilla.gnome.org/show_bug.cgi?id=778015
Application might choose only specific type among all available types
using select-streams event. In this case, it is desired that reconfigure
of playsink to clear unused stream path.
https://bugzilla.gnome.org/show_bug.cgi?id=778015
When an empty mix matrix is passed, audio-channel-mixer
will now generate a (potentially truncated) identity matrix,
this replicates the behaviour of audiomixmatrix in first-channels
mode.
https://bugzilla.gnome.org/show_bug.cgi?id=788833
remove_format_info was a bit confusing to read, this removes
it in favor of standard gst_caps_map_in_place calls.
This no longer simplifies the resulting caps, but I
consider this should be the job of basetransform.
https://bugzilla.gnome.org/show_bug.cgi?id=785471
Use the intended sequence for re-using elements:
* EOS
* STREAM_START if element is to be re-used
This avoids having elements (such as queue/multiqueue/queue2) not
properly resetting themselves.
When delaying EOS propagation (because we want to wait until all
streams of a group are done for example), we re-trigger them by
first sending the cached STREAM_START and then EOS (which will
cause elements to re-set themselves if needed and accept new
buffers/events).
https://bugzilla.gnome.org/show_bug.cgi?id=785951
It is forwarding messages to the playbin bus, thus forwarding messages
that contain a floating reference to the application. This generally
makes bindings unhappy, we must not leak floating references to them.
Crossfading is a bit more complex than just having two pads with the
right keyframes as the blending is not exactly the same.
The difference is in the way we compute the alpha channel, in the case
of crossfading, we have to compute an additive operation between
the destination and the source (factored by the alpha property of both
the input pad alpha property and the crossfading ratio) basically so
that the crossfade result of 2 opaque frames is also fully opaque at any
time in the crossfading process, avoid bleeding through the layer
blending.
Some rationnal can be found in https://phabricator.freedesktop.org/T7773.
https://bugzilla.gnome.org/show_bug.cgi?id=784827
channels=1 is always mono, having it 'unpositioned' does not make
sense.
This fixes pipeline such as:
gst-validate-1.0 audiotestsrc ! audio/x-raw,channels=2,rate=44100,layout=interleaved ! audioconvert ! audioresample ! audio/x-raw, rate=44100, channels=1 ! avenc_mp2 ! fakesink
https://bugzilla.gnome.org/show_bug.cgi?id=785407
Do not remove other parsebin's input streams. It will cause unexpected
removal of any input streams in multi-parsebin use case.
Basically, the purpose of blocking buffers is similar to checking
no-more-pads of chain/group. That is, it gives hint to know the timing
to remove old (EOSed) streams of the parsebin and to add/reuse slots
for new input streams. But, that doesn't mean that we need to remove
other parsebin's EOSed stream. Each parsebin has most likely its
own streaming thread and therefore EOSed time can be much different.
(i.e., much early EOS of subtitle only parsebin)
https://bugzilla.gnome.org/show_bug.cgi?id=785120
Fields related to stream handling (input_streams,
output_streams, slots, guint slot_id) where used totally unprotected
until know.
This lead to several races, especially playing back RTSP streams.
To protect those fields, the OBJECT_LOCK can not be used as we sometimes
need to be able to post message on the bus while holding it.
decodebin3 already has a lock to manage stream selection, and in the end
it makes sense to protect all the stream management fields with the same
lock which is why we reuse the SELECTION_LOCK here.
https://bugzilla.gnome.org/show_bug.cgi?id=784012
decodebin3 checks input streams and pushes EOS if all input streams
are EOSed. If not, fake EOS is pushed to the corresponding slot.
When adaptivedemux is used with multi-track configuration,
adaptivedemux never ever push EOS to non-selected track
because streaming thread for the slot stops with not-linked flow return.
So, decodebin3 should generate EOS itself to finish playback.
https://bugzilla.gnome.org/show_bug.cgi?id=777735
linked input of slot can be old input, so urisourcebin should check
eos state to figure out whether it's new one or not.
If not, urisourcebin never ever forwards EOS to downstream at the end
of presentation, because the old input is still there without removal
https://bugzilla.gnome.org/show_bug.cgi?id=777735
group-id in stream-start event might be updated in
parse_chain_output_probe (). This cause duplicated stream-start
twice with identical stream-id and seq-num, but only group-id is
different. Although there is no change, stream-start event will
be followed by the first buffer.
https://bugzilla.gnome.org/show_bug.cgi?id=771088
This makes it possible for GstDiscoverer to work with sources that
have multiple source pads and hence will trigger the creation of multiple
decodebin instances such as rtspsrc.
Based on the work of Vineeth TM <vineeth.tm@samsung.com>
https://bugzilla.gnome.org/show_bug.cgi?id=754178
The base class is trying to align the processed data, but it endup
removing the GstVideoMeta. That caused wrong result. Instead, just copy
from the process function with the appropriate alignment.
https://bugzilla.gnome.org/show_bug.cgi?id=781204
And only set low-percent/high-percent if not using downloadbuffer, just
like in old uridecodebin. using the watermark based buffering causes
playback to hang never finish buffering with downloadbuffer.
With both audiorate and videorate, it seems more sensible to apply rate
adjustments after the first buffer appears. For example, with v4l2src,
there is often a small delay before the first video buffer turns up, and
this can cause a stuttery start because of videorate trying to ensure a
perfect stream.
Those multiqueue are the ones dealing with adaptive demuxers. They should
have a time limit set so that they don't end up buffering too much data.
They would previously be set with no limits at all, which would cause them
to grow indefinitely until downstream blocks.
gst_video_rate_flush_prev() ensures that the pushed buffer is writable
by calling gst_buffer_make_writable() on videorate->prevbuf.
In drop-only mode we always push buffers directly when they are received
from GstBaseTransform (gst_video_rate_transform_ip()) and do not keep them
around. GstBaseTransform already ensures that those buffers are
writable so there is no need to do it twice.
This change saves us from copying buffers in drop-only mode as we no longer
calls gst_buffer_make_writable() with a buffer having a refcount of 2
(one ref owned by GstBaseTransform and one in videorate->prevbuf).
https://bugzilla.gnome.org/show_bug.cgi?id=780767
When caps changes while streaming, the new caps was getting processed
immediately in videoaggregator, but the next buffer in the queue that
corresponds to this new caps was not necessarily being used immediately,
which resulted sometimes in using an old buffer with new caps. Of course
there used to be a separate buffer_vinfo for mapping the buffer with its
own caps, but in compositor the GstVideoConverter was still using wrong
info and resulted in invalid reads and corrupt output.
This approach here is more safe. We delay using the new caps
until we actually select the next buffer in the queue for use.
This way we also eliminate the need for buffer_vinfo, since the
pad->info is always in sync with the format of the selected buffer.
https://bugzilla.gnome.org/show_bug.cgi?id=780682
Instead go backwards before segment.stop based on the framerate or the
next buffers end timestamp. Otherwise the first buffer will usually be
dropped because outside the segment.
https://bugzilla.gnome.org/show_bug.cgi?id=781899
When there are more than 64 channels, we don't want to exceed the
bounds of the ordering_map buffer, and in these cases we don't want to
remap at all. Here we avoid doing that.
Based on a patch originally for plugins-good/interleave in
https://bugzilla.gnome.org/show_bug.cgi?id=780331
HLS files can have arbitrary extra tags in them, and
those can be quite long lines. We need to search
further than 256 bytes sometimes just to get past the
first few lines of the file. Make the limit 4KB,
which matches a typical input block size and should
hopefully cover every crazy input.
https://bugzilla.gnome.org/show_bug.cgi?id=780559
The term stride is confusing here, since the stride is always use
to signal the pixel row size of an image (including padding). Also
a frame may have a single stride, which adds to the confusion. This
patch uses frame-size, which simply indicate the frame size in the
case the images have some padding in between.
https://bugzilla.gnome.org/show_bug.cgi?id=780053
This allow using those property through gst-launch-1.0. This type
gained a deserilizer recently. The syntax is: <val1, val2, ...>.
Note that we also use the type int instead of uint to avoid having
to cast when specifying the values. The deserilizers assume
int by default.
https://bugzilla.gnome.org/show_bug.cgi?id=780053
When a clip has video audio and subtitle, if need send gap event
to audio and subtitle, we should make sure all has been sent, so
need every stream keep one send_gap_event.
https://bugzilla.gnome.org/show_bug.cgi?id=780429
When posting 100% buffering due to removing the last
buffering element, we still need to hold the posting
lock as well, to avoid any race with other elements
that might post a buffering message at that exact
moment
Add locking, and handle EOS properly now that urisourcebin
uses custom events in place of real EOS events, so we
need to manually remove buffering messages and potentially
post 100% in that situation
The expanded 4 second buffering was making radio streams that are
being delivered at real-time speeds too slow. We might need
a better plan for matching the queue2 size to incoming bitrate
in the absence of tag information or timestamping.
In uridecodebin, it used tags on the output of decodebin to
adjust the queue2 buffering, but urisourcebin doesn't have that
view - decodebin is downstream from us.
This adds a property to select the maximum number of threads to use for
conversion and scaling. During processing, each plane is split into
an equal number of consecutive lines that are then processed by each
thread.
During tests, this gave up to 1.8x speedup with 2 threads and up to 3.2x
speedup with 4 threads when converting e.g. 1080p to 4k in v210.
https://bugzilla.gnome.org/show_bug.cgi?id=778974
See https://bugzilla.gnome.org/show_bug.cgi?id=773666
This would ideally be solved in baseparse but that requires further
thought at this point, and in the meantime it would be good to have
rawbaseparse not assert on this but handle it gracefully instead.
Probe for MultiQueue source pad might receive EOS twice,
the first is fake-eos and the other is actual EOS.
And the slot can be freed with fake-eos/EOS if the slot has no input.
Since slot freeing is async, double free can be possible.
So, decodebin3 needs to remove the probe also with slot freeing.
https://bugzilla.gnome.org/show_bug.cgi?id=777530
"requested_selection" list might be generated by select-streams event.
And memory of stream-id(s) in select-streams is independent from that of stream-collection.
https://bugzilla.gnome.org/show_bug.cgi?id=775553
The latency query originally had a fallthrough to the default
label at the end as fallback, but that got messed up when the
DURATION and POSITION queries were added, so it then fell through
to the duration query handler instead. Restore original behaviour.
https://bugzilla.gnome.org/show_bug.cgi?id=699077
Duration query would return TRUE and duration=-1. This
worked in the unit test because the unit test implementation
was a bit broken.
Both queries need to access rate with a lock.
Fix broken duration query test as well. It relied on broken
behaviour by the videorate query handler, and also it was
implemented as a downstream query rather than an upstream
query. And we must return HANDLED from the probe so that the
query we intercept actually returns TRUE.
https://bugzilla.gnome.org/show_bug.cgi?id=699077
When the decodebin state change fails because of an error
message, we might not go through PAUSED->READY. Don't leak
a ref to decodebin pads due to pad blocking in that case.
This is because we return ASYNC going to PAUSED, and if
we fail before reaching PAUSED the only transition we'll
see is READY->NULL.
https://bugzilla.gnome.org/show_bug.cgi?id=775893
This adds some extra options that affect pattern=ball mode, allowing the
animation to be synced to running time or wall-time clock for comparing
sync across different instances / pipelines / machines.
Also added is the ability to invert the rendering colours every second,
and some different ball motion patterns.
https://bugzilla.gnome.org/show_bug.cgi?id=740557
The state of urisourcebin (and all elements contained within) can
change at any point in time, including when setting up the typefind
element.
In order to avoid ending up with typefind starting without being fully
connected, lock the state and connect to the 'have-type' signal.
Due to the special nature of adaptivedemux, reconfigure happens
frequently with seek/track-change.
In very exceptional cases, the following sequence is possible:
* EOS event is pushed to queue element and still buffers are queued
* During draining remaining buffers, reconfiguration downstream
happens due to track switch.
* The queue gets a not-linked flow return from downstream
* Because the sinkpad is EOS, the queue registers an
error on the bus, causing the pipeline to fail.
Avoid the sinkpad getting marked EOS in the first place, by using a
custom event in place of EOS.
https://bugzilla.gnome.org/show_bug.cgi?id=777009
When shutting down decodebin2 and parsebin, they set their
output pads to flushing, and there is a very small window
where elements might send a sticky event such as a tag event
(which silently fails due to flushing) and then sends a buffer,
and the buffer will return GST_FLOW_ERROR because it can't
forward sticky events. The element will then send an error
message on the bus. This can also happen when elements send EOS
just as shutdown is happening. Since we're about to destroy all
the elements inside parsebin and decodebin anyway, just discard
error messages from them.
A nicer but more difficult fix for GStreamer 2.0 is to make
all event pushing / handling in core return a GstFlowReturn
like buffers do, so we can report a FLUSHING state cleanly.
Make sure ticks start with an accumulator value of 0 by incrementing it
after filling in samples instead of before and by resetting the accumulator
every time a tick begins. This prevents it from being discontinuous at the
beginning of the tick.
https://bugzilla.gnome.org/show_bug.cgi?id=774050
When plugging and then exposing a parser, don't fail
if it fails to send sticky events. The most likely
reason is that things were flushed due to the app
immediately doing a seek, but we can't detect flushing
separately to other error conditions without a
gst_pad_send_event_full() core function that returns
a GstFlowReturn.
In some case we might have EncodingProfile that will be defined
in a way that, for example if a Preset is not present, another
profile for that stream should be used.
A test is added showing the feature.
https://bugzilla.gnome.org/show_bug.cgi?id=776188
There are cases when there is no demuxer involved that could do the
buffering, e.g. HLS with raw MP3 or AAC. In this case we want to place
the buffering multiqueue after the parser.
Before this change, we've considered the first element after the
adaptive streaming demuxer as a parser. This is not always true, e.g.
id3demux. Instead we now wait until we actually have a parser (or
decoder).
Fixes playback on such HLS streams.
Compositor does not support it currently and it needs special support
for handling this correctly, and is rather non-trivial to implement for
all formats.
Playbin3 takes lock when querying duration and handling
stream-collection message. So,to post stream-collection message,
duration query should be dropped when input pad is being unlinked.
https://bugzilla.gnome.org/show_bug.cgi?id=773341
max_buffer_usage is the index of the oldest buffer in the queue,
starting at zero, not the number of buffers queued.
find_limits returns the index of the oldest buffer that satisfies the
limits in its min_idx parameter, not the number of buffers needed. Fix
this use too in order to keep passing the tests that read
buffers-queued.
https://bugzilla.gnome.org/show_bug.cgi?id=775351
If a client gets dropped and the iteration gets restarted, bufpos is
incremented again for all clients that preceded the dropped one, causing
havoc.
Adjust the bufpos for all clients first before trying to drop any.
https://bugzilla.gnome.org/show_bug.cgi?id=774908
Optimize LE<->BE conversion by adding a dedicated fast path instead of
using the generic converter. Implement transform_ip function in order to do the
endian swap in place.
This saves buffer allocation for the intermediate format, can be done in place
and also performs the conversion in one step instead of unpack-convert-pack.
For all bit widths the naive algorithm is implemented, which provides the best
performance when compiled with -O3. ORC was considered but eventually removed
as it requires a dedicated function for in-place conversion (due to the
"restrict" parameters).
A more complex algorithm for the 24-bit conversion with unrolled loop and
32-bit processing is implemented in the #if 0 section. It performs better if
compiled with -O2. With -O3 however the naive algorithm performs better.
https://bugzilla.gnome.org/show_bug.cgi?id=773073
For frame->buffer, baseparse is doing that automatically for us. For
frame->output_buffer it doesn't and assumes that the subclass is already
doing that. Consistency!
Deterministic generation of snow and smpte is important for tests so
that it's not affected by other videotestsrc elements in current or
possibly previous tests.
https://bugzilla.gnome.org/show_bug.cgi?id=773102
find_suitable_mask() had complexity O(n^2) on the number of bits.
For common case like 2-channel audio the mask was calculated in about 4k loop
cycles.
Optimize both n_bits_set() and find_suitable_mask() to O(n) where n is the
number of bits set in the mask.
https://bugzilla.gnome.org/show_bug.cgi?id=772864
