This changes the logic in the following ways:
- If no preferred language is matched, prefer better scores for the selected
audio language.
- If a preferred language is matched, always prefer the better match
irrespective of default or forced flags.
- If a preferred language score and the isForced flag is the same, prefer
tracks with a better selected audio language match.
PiperOrigin-RevId: 259707430
The fLaC prefix is included in the initialization data output
from the MKV extractor, so this is highly likely ot be the
right thing to do.
Issue: #6397
PiperOrigin-RevId: 268244365
We currently don't test if an ad needs to be played in case we are already
playing content.
This is to prevent recreating the current content period when
an ad is marked as skipped. We prefer playing until the designated ad group
position and appending another piece of content. This is less likely to cause
visible discontinuities in case the ad group position is at a key frame
boundary.
However, this means we currently miss updates that require us to play an ad
after a timeline update.
PiperOrigin-RevId: 267553459
Sniffing is performed in ProgressiveMediaPeriod even if a single
extractor is provided. Skip it in that case to improve performances.
Issue:#6325
PiperOrigin-RevId: 266766373
The last selection criteria is the audio bitrate to prefer higher-quality
streams. We shouldn't apply this criterium though if the languages of the
tracks are different.
Issue:#6335
PiperOrigin-RevId: 265064756
Currently the value of FLAG_ALLOW_CACHE_FRAGMENTATION is defined as "1 << 4" but commented as "8". Either the value of FLAG_ALLOW_CACHE_FRAGMENTATION should be "1 << 3", or the comment should be 16. Here I am modifying the comment since it does not affect any current behavior.
PiperOrigin-RevId: 265011839
The app is able to pass a more specialized array type, so the Arrays.copyOf call
produces an array into which it's not valid to store arbitrary AudioProcessors.
Create a new array and copy into it to avoid this problem.
PiperOrigin-RevId: 264779164
This matches the documentation on MetadataDecoder.decode:
"@return The decoded metadata object, or null if the metadata could not be decoded."
PiperOrigin-RevId: 263767144
entries are used in .equals(), so it's good to have them printed in toString() too (for test failures) and it makes logging easier too.
PiperOrigin-RevId: 263335503
The current max video buffer is 13MB which is too small for high quality
streams and doesn't allow the DefaultLoadControl to buffer up to its default
max buffer time of 50 seconds.
Also move util method and constants only used by DefaultLoadControl into this
class.
PiperOrigin-RevId: 263328088
I think we need to start clearing the holder as part of the
DRM rework. When we do this, it'll only be valid to read
from the holder immediately after it's been populated.
PiperOrigin-RevId: 262362725
We're no longer tied to the emsg spec, so we can skip unused fields
and assume ms for duration.
Also remove @Nullable annotation from EventMessageEncoder#encode, it
seems the current implementation never returns null
PiperOrigin-RevId: 262135009
We already allow mixed mime type and mixed sample rate adaptation on request,
so for completeness, we can also allow mixed channel count adaptation.
Issue:#6257
PiperOrigin-RevId: 261930046
This also decouples EventMessageEncoder's serialization schema from the emesg spec (it happens to still match the emsg-v0 spec, but this is no longer required).
PiperOrigin-RevId: 261877918
If we keep streams in chunk sources after selecting new tracks, we also keep
a reference to a stale disabled TrackSelection object. Fix this by updating
the TrackSelection object when keeping the stream. The static part of the
selection (i.e. the subset of selected tracks) stays the same in all cases.
Issue:#6256
PiperOrigin-RevId: 261696082
- When in STATE_SEEK with targetGranule==0, seeking would exit
without checking that the input was positioned at the correct
place.
- Seeking could fail due to trying to read beyond the end of the
stream.
- Seeking was not robust against IO errors during the skip phase
that occurs after the binary search has sufficiently converged.
PiperOrigin-RevId: 261317035
A previous change switched to calculation of the bitrate based on the
first MPEG audio header in the stream. This had the effect of fixing
seeking to be consistent with playing from the start for streams where
every frame has the same padding value, but broke streams where the
encoder (correctly) modifies the padding value to match the declared
bitrate in the header.
Issue: #6238
PiperOrigin-RevId: 261163904
Android considers ALAC initialization data to consider of the magic
cookie only, where-as FFmpeg requires a full atom. Standardize around
the Android definition, since it makes more sense (the magic cookie
being contained within an atom is container specific, where-as the
decoder shouldn't care what container the media stream is carried in)
Issue: #5938
PiperOrigin-RevId: 261124155
Checking inputPosition == 0 isn't sufficient because the synchronization
at the top of read() may advance the input (i.e. in the case where there's
some garbage prior to the seek frame).
PiperOrigin-RevId: 261086901
When using speed adjustment it was possible for playback to get stuck at a
period transition when the channel count changed: SonicAudioProcessor would be
drained at the point of the period transition in preparation for creating a new
AudioTrack with the new channel count, but during draining the incorrect (new)
channel count was used to calculate the output buffer size for pending data from
Sonic. This meant that, for example, if the channel count changed from stereo to
mono we could have an output buffer size that stored an non-integer number of
audio frames, and in turn this would cause writing to the AudioTrack to get
stuck as the AudioTrack would prevent writing a partial audio frame.
Use Sonic's current channel count when draining output to fix the issue.
PiperOrigin-RevId: 260156541
1. Using the Locale on API<21 doesn't make any sense because it's a no-op
anyway. Slightly restructured the code to avoid that.
2. API<21 often reports languages with non-standard underscores instead of
dashes. Normalize that too.
3. Some invalid language tags on API>21 get normalized to "und". Use original
tag in such a case.
Issue:#6153
PiperOrigin-RevId: 258773463
2-letter codes (ISO 639-1) are the standard Android normalization and thus we
should prefer them to 3-letter codes (although both are technically allowed
according the BCP47).
This helps in two ways:
1. It simplifies app interaction with our normalized language codes as the
Locale class makes it easy to convert a 2-letter to a 3-letter code but
not the other way round.
2. It better normalizes codes on API<21 where we previously had issues with
language+country codes (see tests).
3. It allows us to normalize both ISO 639-2/T and ISO 639-2/B codes to the same
language.
PiperOrigin-RevId: 258729728
If we use the default start position, we currently resolve it immediately
even if we need to play an ad first, and later try to project forward again
if we believe that the default start position should be used.
This causes problems if a specific start position is set and the later
projection after the preroll ad shouldn't take place.
The problem is solved by keeping the content position as TIME_UNSET (= default
position) if an ad needs to be played first. The content after the ad can
then be resolved to its current default position if needed.
PiperOrigin-RevId: 258583948
Sending MESSAGE_PREPARE_SOURCE should happen last in the constructor.
It was previously happening before initialization finished (and in
particular before pendingMediaPeriods was instantiated).
Issue: #6146
PiperOrigin-RevId: 257158275
Currently, we sometimes apply new playback parameters directly and sometimes
through the list of playbackParameterCheckpoints. Only when using the checkpoints,
we also reset the offset and corresponding position for speedup position
calculation. However, these offsets need to be changed in all cases to prevent
calculation errors during speedup calculation[1].
This change channels all playback parameters changes through the checkpoints to
ensure the offsets get updated accordingly. This fixes an issue introduced in
31911ca54a.
[1] - The speed up is calculated using the ratio of input and output bytes in
SonicAudioProcessor.scaleDurationForSpeedUp. Whenever we set new playback
parameters to the audio processor these two counts are reset. If we don't reset
the offsets too, the scaled timestamp can be a large value compared to the input
and output bytes causing massive inaccuracies (like the +20 seconds in the
linked issue).
Issue:#6117
PiperOrigin-RevId: 256533780
Before this change we'd release the audio track and create a new one as soon
as audio processors had drained when reconfiguring.
Fix this behavior by stop()ing the AudioTrack to play out all written data.
Issue: #2446
PiperOrigin-RevId: 244812402
We currently handle most the control code logic after handling special
characters. This includes filtering out repeated control codes and checking
for the correct channel. As the special character sets are control codes as well,
these checks should happen before parsing the characters.
Issue:#6133
PiperOrigin-RevId: 256993672
1. Only output video starting from a keyframe
2. When calculating the timestamp offset to adjust live streams to start
at t=0, use the timestamp of the first tag from which a sample is actually
output, rather than just the first audio/video tag. The test streams in
the referenced GitHub issue start with a video tag whose packet type is
AVC_PACKET_TYPE_SEQUENCE_HEADER (i.e. does not contain a sample) and whose
timestamp is set to 0 (i.e. isn't set). The timestamp is set correctly on
tags that from which a sample is actually output.
Issue: #6111
PiperOrigin-RevId: 256147747
We currently don't display the last frame because the seek time is behind the
last frame's timestamps and it's thus marked as decodeOnly.
This case can be detected by checking whether all data sent to the codec is
marked as decodeOnly at the time we read the end of stream signal. If so, we
can re-enable the last frame. This should work for almost all cases because the
end-of-stream signal is read in the same feedInputBuffer loop as the last
frame and we therefore haven't released the last frame buffer yet.
Issue:#2568
PiperOrigin-RevId: 251425870
We are currently queuing periods in a way such that the new start position
lines up with the end of the previous period (to ensure continuous playback).
However, if the start position of the new period is larger than the total of
all previously played period durations, we may end up with negative renderer
timestamps when seeking back to the beginning of this new period. Negative
timestamps should be avoided as most decoders have problems handling them
correctly.
This change forces a renderer reset if we detect such a seek to a negative
renderer time and also resets the renderer offset to 0 every time all
renderers are disabled, as this is the only time where we can savely change
the offset of an existing media period.
Also, if playback starts with an ad, we choose the content position as
renderer offset to prevent the whole issue from occurring for the seek-behind-
midroll case.
Issue:#6009
Issue:#5323
PiperOrigin-RevId: 253790054
In some edge cases the renderer position may be slightly ahead of the
buffered position and the total buffered duration is thus negative. We already
filter that in ExoPlayerImpl for the publicly accessible value. However, we
forward the unfiltered value to other components like the LoadControl, which
may be confusing.
Issue:#6015
PiperOrigin-RevId: 253780460
This permission has normal access right and can't be revoked by the user.
However, an app can choose to revoke it when using ExoPlayer, e.g. if
no network is required and the app doesn't want to list this permission.
Support this use case by gracefully catching the exception in the relevant
places.
Issue:#6019
PiperOrigin-RevId: 253759332
