Move the code in its own class as DefaultAudioTrack
is getting very big. It also help for testability.
The new class is easily configurable and highly tested.
Manual test was used to catch any regression.
https://github.com/google/ExoPlayer/issues/8891
PiperOrigin-RevId: 424602011
AnalyticsCollector can't be null when passed into ExoPlayerImplInternal,
so there is no need to pass it around as nullable.
PiperOrigin-RevId: 424594031
Two of the sessions are finished at the same time in the test
and the order of the corresponding callbacks depends on the randomly
generated session string and the order these strings are stored in a
HashSet.
Update test to assert both callbacks are called and they contain the
right arguments, but don't assert on the order of these two callbacks.
PiperOrigin-RevId: 424548819
Ad playback shouldn't be affected by manual speed adjustments set
by the user. This change enforces unit speed for ad playback.
Issue: google/ExoPlayer#9018
PiperOrigin-RevId: 424546258
To avoid the `MediaPeriodQueue`to discard the reading period, we can set the next ad of an ad group early and then (possibly) only change it's duration once we receive the actual duration. This way we avoid a rebuffering as a result of the reading period being discarded.
The change also takes care to properly set ad break and their durations when we join the live stream at the moment when an ad is playing.
PiperOrigin-RevId: 423163467
We currently only end sessions on Timeline updates if the associated media
is no longer in the playlist. But we should also end all sessions that are
finished as a result of the timeline update (similar to how this is done for
discontinuities). This issue was introduced by 394ab7bcfd
PiperOrigin-RevId: 423075855
Adaptive video and audio selections will be limited to formats with
the same level of DecoderSupport and HardwareAccelatationSupport, unless
specifically allowed by new flags.
If different levels of decoder support are available, prefer primary
over fallback decoders and hardware-accelerated over software decoders
(in this order). For video, also prefer more efficient codecs, if both
are supported by hardware-accelerated primary decoders.
Issue: google/ExoPlayer#4835
Issue: google/ExoPlayer#9565
PiperOrigin-RevId: 422345048
This change moves the video track selection to the generic
selection method introcuced for audio and text. This ensures
we can apply the same criteria for fixed and adaptive video
track selections. Implicitly, this reorders the preferences
for adaptive tracks to give non-quality preferences (like
preferred MIME type or preferred role flags) a higher priority
than number of tracks in the selection.
Issue: google/ExoPlayer#9519
PiperOrigin-RevId: 422310902
Live speed adjustment is used for all live playback at the moment,
but has no user visible effect if the media is not played with low
latency. To avoid unnecessary adjustment during playback without
benefit, this change restricts the live speed adjustment to cases
where either the user requested a speed value in the MediaItem or the
media specifically defined a low-latency stream.
Issue: google/ExoPlayer#9329
PiperOrigin-RevId: 421514283
This ensures we test the API level specific logic, in particular
around 5G-NSA detection.
Robolectric has a remaining bug that it doesn't support listening
to service state changes. Hence, we need to ignore some tests on
these API levels still until this is fixed.
PiperOrigin-RevId: 421505951
The media codec renderers have fallback logic in getDecoderInfos
to assume that E-AC3 decoders can handle the 2D version of E-AC3-JOC and
that H264/H265 decoders can handle some base layer of Dolby Vision
content. Both fallbacks are useful if there is no decoder for the
enhanced Dolby formats.
Both fallbacks are not applied during track selection at the moment
because the separate MediaCodecInfo.isCodecSupported method verifies
that the mime type corresponding to format.codecs is the same as the
decoder mime type (which isn't true for the fallback case).
To fix the fallback logic, we can just completely remove this additional
check because it's not needed in the context of this method that is only
called after we already established that the decoder can handle the
format.sampleMimeType.
In addition, we need to map the Dolby Vision profiles to the equivalent
H264/H265 profile to make the codec profile comparison sensible again.
PiperOrigin-RevId: 420959104
* @Override is not a TYPE_USE annotation, so should appear before any modifiers and after Javadocs. @CryptoType is a TYPE_USE annotation, so should appear after modifiers and directly before the type.
PiperOrigin-RevId: 418811744
This more closely matches the pattern we have for all implementations
except DefaultMediaSourceFactory (e.g. ProgressiveMediaSource.Factory)
and other factory interfaces like (Http)DataSource.Factory.
PiperOrigin-RevId: 417826803
To support multi-period content we need to store AdPlaybackStates and
SharedMediaPeriod by the periodUid as a key. While after this no-op CL,
we still only support single-period content, storing these resources by
periodUid is the ground work for multi-period support being added in an
follow-up CL.
PiperOrigin-RevId: 416836445
We check the fraction of the available duration we have already
buffered for live streams to see if we can increase the quality.
This fraction compares against the overall available media duration
at the time of the track selection, which by definition can't include
one of the availabe chunks (as this is the one we want to load next).
That means, for example, that for a reasonable live offset of 3 segments
we can at most reach a fraction of 0.66, which is less than our default
threshold of 0.75, meaning we can never switch up.
By subtracting one chunk duration from the available duration, we make
this comparison fair again and allow all live streams (regardless of
live offset) to reach up to 100% buffered data (which is above our
default value of 75%), so that they can increase the quality.
Issue: google/ExoPlayer#9784
PiperOrigin-RevId: 416791033
*** Original commit ***
Rollback of 0aa23b08b1
*** Original commit ***
Add capability flags for hardware and decoder support
Issue: google/ExoPlayer#9565
***
***
PiperOrigin-RevId: 416285603
*** Original commit ***
Make audio track min buffer size configurable.
Move the code in its own class as DefaultAudioTrack
is getting very big. It also help for testability.
The new class is easily configurable and highly tested.
Manual test was used to catch any regression.
https://github.com/google/ExoPlayer/issues/8891
***
PiperOrigin-RevId: 415469179
Move the code in its own class as DefaultAudioTrack
is getting very big. It also help for testability.
The new class is easily configurable and highly tested.
Manual test was used to catch any regression.
https://github.com/google/ExoPlayer/issues/8891
PiperOrigin-RevId: 415268938
`DefaultAudioSink` already has 3 telescoping
constructors and an other one would be have been
needed to add a buffer size tuning option.
PiperOrigin-RevId: 414703366
Allowing duplicate groups caused some other code working with the
array to use reference equality comparison. This is error-prone,
easily forgotten (e.g. when using the TrackGroups in a map) and
causes bugs when TrackGroups are serialized to disk or to another
process.
All TrackGroups created by ExoPlayer are already unique and custom
code creating TrackGroupArrays with identical groups can easily
distringuish them by adding an id to each group.
Issue: google/ExoPlayer#9718
PiperOrigin-RevId: 413617005
And also tweak existing role flag logic to strictly prefer perfect
matches over partial matches.
Caveat: Video role flags only supported for fixed track selections
(same issue as Issue: google/ExoPlayer#9519).
Issue: google/ExoPlayer#9402
PiperOrigin-RevId: 412292835
Currently, DrmSessionManager takes player specific values (= the
playback looper) through (pre)acquireSession calls and requires
the caller to pass in the same values every time.
Instead, we can configure the DrmSessionManager for playback with
a player once before it's being used. We can't simply extend the
prepare() method as prepare may be called before the player is
created to prewarm the DrmSessionManager.
The new method also takes a PlayerId which is bound to the lifetime
of the player similar to the playback looper.
To avoid breakage of custom MediaSources with DRM, we can keep the
old the SampleQueue.createWithDrm method as deprecated.
PiperOrigin-RevId: 410998240
Currently, TrackSelectionOverrides are documented as being applied
per track type, meaning that one override for a type disables all
other selections for the same track type. However, the actual
implementation only applies it per track group, relying on the
track selector to never select another renderer of the same type.
This change fixes DefaultTrackSelector to fully adhere to the
TrackSelectionsOverride definition. This solves problems when
overriding tracks for extension renderers (see Issue: google/ExoPlayer#9675)
and also simplifies a workaround added to StyledPlayerView.
#minor-release
PiperOrigin-RevId: 409121711
The asynchronous MediaCodec adapter queues input buffers in a
background thread. If a codec queueuing operation throws an exception,
the buffer enqueuer will store it as a pending exception and re-throw it
the next time the adapter will attempt to queue another input buffer.
The buffer enqueuer's flush() and shutdown() may throw an exception if
the pending error is set. This is subject to a race-condition in which
the pending error can be set while the adapter is flushing/shutting down
the enqueuer, e.g., if an input buffer is still being queued and the
codec throws an exception. As a result, the adapter cannot flush or
shutdown gracefully.
This change makes the buffer enqueuer to ignore any pending error
when flushing/shuttinf down so that the adapter can flush/release
gracefully even if a queueing error was detected.
PiperOrigin-RevId: 409113054
MediaSource can be reused with other Player instances after they
have been released, so we need to set the PlayerId when preparing
the source. Access can mostly be handled by the implementation in
BaseMediaSource.
PiperOrigin-RevId: 408878824
And in a couple of related places.
This is for consistency with the rest of the codebase where
we exclusively use indices.
#minor-release
PiperOrigin-RevId: 408273372
When operating the MediaCodec in asynchronous mode, after a
MediaCodec.flush(), we start MediaCodec in the callback thread,
which might trigger errors in some platforms. This change adds an
experimental flag to move the call to MediaCodec.start() back to the
playback thread.
PiperOrigin-RevId: 407801013
When we have multiple overrides for TrackGroups associated with
one renderer, we need to look at all of them to find the non-empty
one. Empty ones should only be used to remove previously selected
tracks for this group and otherwise be ignored.
Currently this is broken because the first override (no matter if
it's empty or not) is used as the final selection for this renderer.
Issue: google/ExoPlayer#9649
#minor-release
PiperOrigin-RevId: 407792330
Where this introduced an inconsistency (e.g. assigning to something
called `windowIndex`), I generally renamed the transitive closure of
identifiers to maintain consistency (meaning this change is quite
large). The exception is code that interacts with Timeline and Window
directly, where sometimes I kept the 'window' nomenclature.
#minor-release
PiperOrigin-RevId: 407040052
This helps to prevent issues where decoders can't handle negative
timestamps. In particular it avoids issues when the media accidentally
or intentionally starts with small negative timestamps. But it also
helps to prevent other renderer resets at a later point, for example
if a live stream with a large start offset is enqueued in the playlist.
#minor-release
PiperOrigin-RevId: 406786977
The static and dynamic metadata now build up in a list, such that when
the MediaMetadata is built, they are applied in an event order. This
means that newer/fresher values will overwrite older ones. The MediaItem
values are then applied at the end, as they take priority over any other.
#minor-release
PiperOrigin-RevId: 405383177
The current API exposes an `ImmutableMap` of
`TrackGroup` -> `TrackSelectionOverride`.
This has several disadvantages:
- A difficult to use API for mutation
(`ImmutableMap.Builder` doesn't support key removal).
- There is no track selection specific methods,
how the generic map API mapps to the selection override is not complex
but to obvious for a casual reader.
- The internal data type is exposed, making internal refactor difficult.
This was done to have the API ready as quick as possible.
When transitioning the clients to the map API in <unknown commit>,
it became clear that the map API was too verbose and not mapping
to the clients needs, so utility methods
were added to make operations clearer and more concise.
Nevertheless, having to use utility method to use easily and correctly
an API is not the sign of a good API.
This cl refactors the track selection API for several improvements:
- Add a type `TrackSelectionParameters` that encapsulate the internal
data structure (map currently).
- For iteration, expose as a list.
- Add a `Builder` for easy mutable operations.
- Add track selection specific methods to avoid having utilities functions.
- Those operations are the same as `DefaultTrackSelector.Parameters`
for easier migration. (`setOverride` was renamed to `addOverride`)
- Move `TrackSelection` classes outside of `TrackSelectionParameters`
as their own top level classes.
The migration of the client code is straightforward as most of it
were already using the previously mentioned utility functions
that are now native methods.
The full migration has not been done yet, and is pending on this cl approval.
PiperOrigin-RevId: 405362719
Common houses DataSource as an interface for reading data,
but most of the concrete implementations are in ExoPlayer.
This means that in practice, if an app wants to use a module
that reads using DataSource (e.g. extractor), they may be
forced to depend on ExoPlayer as well to get a concrete
implementation (e.g. FileDataSource). This change moves the
DataSource implementations into common to resolve this.
PiperOrigin-RevId: 403222081
An upcoming change will modify ExoPlayer.Builder#build() to return
ExoPlayer, so any places that explicitly need a SimpleExoPlayer
instance should be using SimpleExoPlayer.Builder.
PiperOrigin-RevId: 403028312
Update the UI module, the demos and most other users
to make use of the new player TracksInfo and track
selection override APIs.
PiperOrigin-RevId: 402817857
Enable subtitle output in the PlaybackOutput and disable the text
renderer in the MkvPlaybackTest. Add WebvttPlaybackTest to test the
output of side-loaded WebVTT subtitles.
PiperOrigin-RevId: 402526588
*** Original commit ***
Migrate callers of ExoPlayer.Builder#build() to buildExoPlayer()
An upcoming change will update build() to return Player.
PiperOrigin-RevId: 401468532
TracksInfo is very similar to
`MappingTrackSelector.MappedTracksInfo` with some
fields removed to simplify the Player API,
notably it doesn't expose the renderer concept.
A significant difference is the addition of a `selected` boolean
field which avoids having a separate `getCurrentTrackSelection`
API.
This cl is a part of the bigger track selection change,
splitted for ease of review.
In particular, the MediaSession implementation and UI usage
have been slitted in child cls.
Find all cls with the tag:
#player-track-selection
PiperOrigin-RevId: 400937124
This change makes asynchronous queueing non-experimental, it enables the
feature by default on devices with API level >= 31 (Android 12+) and
exposes APIs for apps to either fully opt-in or opt-out from the
feature.
The choice to use or not asynchronous queueing is moved out of
MediaCodecRenderer to a new MediaCodecAdapter factory, the
DefaultMediaCodecAdapterFactory. This is because, at the moment,
if an app passes a custom adapter factory to a MediaCodecRenderer and
then enables asynchronous queueing on it, the custom
adapter factory is not used but this is not visible to the user.
The default behavior of DefaultMediaCodecAdapterFactory is to create
asynchronous MediaCodec adapters for devices with API level >= 31
(Android 12+), and synchronous MediaCodec adapters on devices with older
API versions.
DefaultMediaCodecAdapterFactory exposes methods to force enable or force
disable the use of asynchronous adapters so that applications can enable
asynchronous queueing on devices with API versions before 31 (but not
before 23), or fully disable the feature. For applications that build
MediaCodecRenderers directly, they will need to create a
DefaultMediaCodecAdapterFactory and pass it to the renderer constructor.
For applications that rely on the DefaultRenderersFactory, additional
methods have been added on the DefaultRenderersFactory to control
enabling/disabling asynchronous queueing.
Issue: #6348
PiperOrigin-RevId: 400733506
This moves `SelectionOverride` from `DefaultTrackSelector`
to `TrackSelectionParameters`.
It is then use to allow track selection override per
track selection array.
Note that contrary to
`DefaultTrackSelector.Parameters.selectionOverride`, the renderer
concept is not exposed.
This cl is a part of the bigger track selection change,
splitted for ease of review.
Find all cls with the tag:
#player-track-selection
PiperOrigin-RevId: 399933612
There are two very similar tests checking for release events,
one running with Robolectric and one instrumentation test.
The instrumentation test only adds the interaction with the player
to release it while the renderers are active. This same interaction
can be added to the Robolectric test as well.
This arguably improves the realism of the Robolectric test too
as we listen for real player events instead of simulating the same
events.
PiperOrigin-RevId: 399694869
Before releasing r2.15.0, we had a regression that crashed
PlaybackStatsListener. A change in the AnalyticsCollector made
it to send an additional AnalyticsListener.onEvents() callback after
calling Player.release() and AnalyticsListener.onEvents() appeared
to arrive with event times that were not monotonically increasing.
The AnalyticsListener.onEvents() callback that contained
AnalyticsListener.EVENT_PLAYER_RELEASED was called with a timestamp that
was smaller than event times of previously AnalyticsListener.onEvents()
calls.
A first fix changed the order of events being forwarded to
AnalyticsListener. Upon calling Player.release(), the AnalyticsCollector
would call AnalyticsListener.onPlayerReleased() and its associated
AnalyticsListener.onEvents() on the same stack call. This fix maintained
that event times are monotonically increasing, but made
AnalyticsListener.onPlayerReleased() be called earlier.
This change:
- Further changes AnalyticsCollector to ensure that
AnalyticsListener.onPlayerReleased() and its related
AnalyticsListener.onEvents() are the last callbacks to be called,
and the associated timestamp is bigger than previously reported
event times.
- Adds an instrumentation test to guard against the regression.
PiperOrigin-RevId: 399437724
Previously the released preacquired sessions would start their keepalive
timeout, and so no additional resources would be freed in time for the
manager to retry the session acquisition.
This change adds an additional purge of keepalive sessions *after* the
preacquired sessions are released, which fixes the problem.
#exofixit
#minor-release
PiperOrigin-RevId: 396613352
This is a candidate fix for #8906. As mentioned in that issue,
negative positions within windows might be (kind of) valid in
live streaming scenarios, where the window starts at some
non-zero position within the period. However, negative positions
within periods are definitely not valid. Neither are positions
that exceed the period duration.
There was already logic in ExoPlayerImplInternal to prevent a
resolved seek position from exceeding the period duration. This
fix adds the equivalent constraint for the start of the period.
It also moves the application of the constraints into Timeline.
This has the advantage that the constraints are applied as part
of state masking in ExoPlayerImpl.seekTo, removing any UI flicker
where the invalid seek position is temporarily visible.
Issue: #8906
PiperOrigin-RevId: 395917413
The C2 MP3 decoder produces an extra output buffer when draining after
end-of-stream is queued. This output buffer has a later timestamp than the last
queued input buffer so we need to calculate its timestamp to detect a stream
change in the correct position.
Before this CL we used the original input buffer timestamp as the largest
queued timestamp, which caused the stream change to be detected at the correct
position because the original input buffer timestamp was slightly larger than
the actual last output buffer timestamp. After this change we use exact
calculated timestamp as the largest queued timestamp. I manually verified
gapless continues to work on a device using the C2 MP3 decoder by comparing
output of the MP3 gapless and MP3 gapless stripped playlists in the demo app,
and that the last buffer timestamp now matches.
#exofixit
PiperOrigin-RevId: 395428928
Empty buffer with flag C.BUFFER_FLAG_END_OF_STREAM is send at the end
of the stream. Handling that flag properly is necessary to make the
ExoplayerCuesDecoder work properly with components like TextRenderer.
PiperOrigin-RevId: 394472642
Make this behaviour optional, so it can be disabled for
AnalyticsCollectorTest where we don't use
FakeExoMediaDrm.LicenseServer.
PiperOrigin-RevId: 393133721
This cl doesn't implement completely the API for
`ExoPlayerImpl` as
`onTrackSelectionParametersChanged` is not called.
The follow up cl adds `TrackSelectionParameters` in PlaybackInfo
to correctly propagate the change event and mask it.
Additionally `TrackSelectionParameters` is serialized as a Parcelable
for now. It is transitioned to bundleable in a follow up cl.
PiperOrigin-RevId: 392899918
