MediaCodecRenderer currently has two independent paths to trigger
events at stream changes:
1. Detection of the last output buffer of the old stream to trigger
onProcessedStreamChange and setting the new output stream offset.
2. Detection of the first input buffer of the new stream to trigger
onOutputFormatChanged.
Both events are identical for most media. However, there are two
problematic cases:
A. (1) happens after (2). This may happen if the declared media
duration is shorter than the actual last sample timestamp.
B. (2) is too late and there are output samples between (1) and (2).
This can happen if the new media outputs samples with a timestamp
less than the first input timestamp.
This can be made more robust by:
- Keeping a separate formatQueue for each stream to avoid case A.
- Force outputting the first format after a stream change to
avoid case B.
Issue: google/ExoPlayer#8594
#minor-release
PiperOrigin-RevId: 512586838
Some devices were reported to have wrong PerformancePoint sets
that cause 60 fps to be marked as unsupported even though they
are supported.
Issue: google/ExoPlayer#10898
#minor-release
PiperOrigin-RevId: 512580395
The output info for a new stream is marked pending until the last
sample of the previous stream has been processed. However, this fails
if the previous stream has already been fully processed. We need to
detect this case explicitly to avoid signalling the output change one
sample too late.
#minor-release
PiperOrigin-RevId: 512572854
This test became flaky after ab7e84fb34 because some of the
unrealistic frame times ended up on the same release time.
Using realistic numbers avoids the flakiness.
PiperOrigin-RevId: 512566469
Protected system broadcasts should not specify the export flag.
Marking them as NOT_EXPORTED breaks sticky broadcasts in some
cases.
Issue: google/ExoPlayer#10970
#minor-release
PiperOrigin-RevId: 512020154
The current logic uses manual array operations to keep track of pending
changes. Modernize this code by using an ArrayDeque and a data class.
This also allows to extend the output stream information in the future.
This also fixes a bug where a position reset accidentally assigns a pending
stream offset instead of keeping the current one.
#minor-release
PiperOrigin-RevId: 511787571
When rendering frames at a rate higher than the screen refresh rate,
e.g. playing at 8x, the player is releasing multiple frames at the same
release time (nanos) which are then dropped by the platform. The output
buffers are available later and as a result MediaCodec cannot keep up
decoding fast enough.
This change skips releasing multiple video frames on the same vsync
period and proactivelly drops the frame. The frame is counted as skipped
rather than dropped to differentiate with frames dropped due to slow
decoding.
PiperOrigin-RevId: 510964976
This call may cause performance overhead in some situations,
for example if the AudioTrack needs to query an offload DSP
for the current position. We don't need to check this multiple
times per doSomeWork iteration as the value is unlikely to
change in any meaningful way.
PiperOrigin-RevId: 510957116
Rename ScaleToFitTransformation to ScaleAndRotateTransformation.
This better represents the operations that can be accomplished using this
effect. The name was originally named ScaleToFit* because it's not obvious how
to scale to fit using OpenGL, and this effect handled the scaling to fit in a way that no other MatrixTransformations did.
However, it's hard to discover how to rotate when skimming names of effects, so
it's probably more useful to convey that this effect rotates, than that it
scales to fit.
PiperOrigin-RevId: 510480078
Previously, any constructors instrumented by Robolectric were made public. This
caused two types of issues:
1) If Android classes had non-public constructors which were made public and
added to the Android API, Robolectric allowed tests to incorrectly use the
constructors on older SDK levels (where they were non-public). This most
commonly occurs for AccessibiltyEvent and AccessibilityNodeInfo.
2) When reflection was used to instantiate classes that were instrumented by
Robolectric, all constructors were accessible, which did not match what
happened when running on an Android test.
Update the instrumentation in Robolectric to prevent making all public
constructors.
PiperOrigin-RevId: 510049123
Tests in `SilenceSkippingAudioProcessorTest` used half as many short integers as needed for channel values when generating alternating silence/noise input. Fix this by passing left and right channel input.
PiperOrigin-RevId: 509188074
The AsynchronousMediaCodecAdapter's queuing thread stores any exceptions
raised by MediaCodec and re-throws them on the next call to
queueInputBuffer()/queueSecureInputBuffer(). However, if MediaCodec
raises and error while queueing, it goes into a failed state and does
not announce available input buffers. If there is no input available
input buffer, the MediaCodecRenderer will never call
queueInputBuffer()/queueSecureInputBuffer(), hence playback is stalled.
This change surfaces the queueing error through the adapter's dequeueing
methods.
PiperOrigin-RevId: 508637346
`TrackSelectorResult.rendererConfigurations` can contain null elements:
> A null entry indicates the corresponding renderer should be disabled.
This wasn't caught by the nullness checker because `ExoPlayerImpl` is
currently excluded from analysis.
#minor-release
Issue: google/ExoPlayer#10977
PiperOrigin-RevId: 508619169
Previously, Robolectric's instrumentation updated all constructors to be
public. This caused two main types of problems:
1) When non-public constructors were made public and added to the Android API,
Robolectric allowed tests to incorrectly use the constructors on older SDK
levels (where they were non-public). This most commonly occurs for
AccessibiltyEvent and AccessibilityNodeInfo.
2) When reflection was used to instantiate classes that were instrumented by
Robolectric, all constructors were accessible.
Fix issues across Google3 Robolectric tests that were affected by this issue.
A forthcoming change will fix the instrumentation in Robolectric to prevent
this type of issue from occurring.
Tested:
TAP --sample ran all affected tests and none failed
http://test/OCL:507861075:BASE:507805409:1675803313108:f2128fa4
PiperOrigin-RevId: 508087822
If a renderer error happens while processing readahead data for the
next item in the playlist, we currently throw the error immediately
and only set the item id in the error details. This makes it harder
to associate the error to the right item. For example, the user
facing UI is likely not updated to show the failing item when the
error is reported.
This can be improved slighly by force setting the position to the
failing item. The playback still fails immediately, but this can't
be avoided because the renderer itself went into an error state.
PiperOrigin-RevId: 507808635
The AudioTrackPositionTracker needs to correct positions by
the speed set on the AudioTrack itself whenever it makes
estimations based on real-time (=the real-time playout
duration is not equal to the media duration played).
This happens for the main playback path already, but not for
the mode in which the position is estimated from the playback
head position and also not in the phase after the track has
been stopped. Both cases are not very noticeable during
normal playback, but become relevant when playing in offload
mode.
PiperOrigin-RevId: 507736408
Based on [this conversation thread](https://chat.google.com/room/AAAA--f88ao/76Rem_cRCK8), I've opted to update the existing FrameProcessor.create() rather than deprecate it, as it is unlikely to be in use by apps outside google3.
PiperOrigin-RevId: 506920930
Before this CL, the `renderedLastFrame` flag is not set if the last frame is released immediately (force render), or when it's dropped.
PiperOrigin-RevId: 506358626
Flushing resets all the texture processors within the `FrameProcessor`. This
includes:
- At the back, the FinalMatrixTextureProcessorWrapper, and its MatrixTextureProcessor
- At the front, the ExternalTextureManager
- All the texture processors in between
- All the ChainingGlTextureProcessorListeners in between texture processors
- All the internal states in the aforementioned components
The flush process follows the order, from `GlEffectsFrameProcessor.flush()`
1. Flush the `FrameProcessingTaskExecutor`, so that after it returns, all tasks queued before calling `flush()` completes
2. Post to `FrameProcessingTaskExecutor`, to flush the `FinalMatrixTextureProcessorWrapper`
3. Flushing the `FinalMatrixTextureProcessorWrapper` will propagate flushing through, via the `ChainingGlTextureProcessorListener`
Startblock:
has LGTM from christosts
and then
add reviewer andrewlewis
PiperOrigin-RevId: 506296469
Can be used to combine multiple media items into a single timeline window.
Issue: androidx/media#247
Issue: google/ExoPlayer#4868
PiperOrigin-RevId: 506283307
AudioProcessor-based speed changes are not supported with
offload, so we should use the AudioTrack-based speed changes
by default if the user requests a speed change.
This moves the decision of which speed change path is used
into the Configuration, so that it can change for each
AudioTrack in a playlist.
PiperOrigin-RevId: 505621911
This makes it easier to reason about some parts of the code and
will eventually allow to easily switch between AudioProcessor-
based on AudioSink-based speed adjustment.
The current state saves the applicable playback parameters
in separate variables depending on which speed adjustment
path is used. Moreover, the AudioProcessor-based logic keeps
a chain of pending parameter changes and we derive the last
applicable one everytime we need the current parameters.
After this change, this is simplified by
- keeping a common value for playback parameters independent
of the actual path we use for adjustment.
- keeping the final ("current") parameters directly, instead
of deriving it from a chain of yet to be applied parameters.
PiperOrigin-RevId: 505097294
MergingMediaPeriod creates its track groups with ids concatenating position in its periods array and the underlying child track group id. The ids can be used in selectTracks for matching to periods list.
Issue: google/ExoPlayer#10930
PiperOrigin-RevId: 505074653
`FrameProcessorManager` now reports it's not ready when the output surface is set but not the output resolution. `FrameProcessorManager` allows playback without an output surface, like with ExoPlayer.
PiperOrigin-RevId: 504626088
Resampling as a term is overloaded, however it was raised as a
confusing term in a meeting recently, as the 1P team initially thought
this would change the sample rate. The naming of this `AudioProcessor`
now matches `ToInt16PcmAudioProcessor`.
PiperOrigin-RevId: 503441019
This is needed for constrained multi-asset to shift the timestamps of
the media items that are not the first in the sequence.
PiperOrigin-RevId: 502409923
Previously, after calling MCVR.setOutput() with null, `frameProcessorManager`'s output surface is cleared. What was unexpected is `ExoPlayerInternal` notifies a zero output resolution after clearing the output surface. This zero resolution causes FrameProcessor to fail.
PiperOrigin-RevId: 501861993
Starting with Android 13 (API 33) an app needs to request the
permission to post notifications or notifications are suppressed.
This change documents this in the class level JavaDoc of the
`DownloadService`.
Issue: google/ExoPlayer#10884
PiperOrigin-RevId: 501346908
When audio processors are enabled during tunneling, they must produce
output immediately, ensuring that the timestamps of the output samples
correspond to the input and that no additional samples are produced.
This requirement is documented in the Javadoc of DefaultAudioSink.
However, this alone doesn't guarantee all buffers are immediately
written to the AudioTrack, because the AudioTrack writes are
non-blocking and may need multiple attempts.
When draining the audio sink at the end of the stream, we currently
fail in this situation because we assert that the timestamp must be
set (=the drain operation is a no-op). But this may not be true when
the previous non-blocking write wasn't fully handled. We can fix this
by saving the last timestamp and reusing it during draining.
Issue: google/ExoPlayer#10847
PiperOrigin-RevId: 500943891
Initialising the fields as Integer and then getting a String on compute time is slow. Instead we directly initialise these fields as String. Improves the time taken in bundling PlayerInfo further to less than 200ms from ~300ms.
Also modified a test to improve productive coverage.
PiperOrigin-RevId: 500003935
