This fixes an issue where a DownloadService implementation
that allows foreground but doesn't provide a scheduler would
not be restarted in the case that it was still in memory but
classed as idle by the platform.
It also speeds up service restart in the case that a
scheduler is provided.
Issue: #6798
PiperOrigin-RevId: 290068960
This method has two use cases:
1. Seeking. Calls are immediately preceded by a call to rewind(), and
the returned value isn't important unless it's ADVANCED_FAILED (i.e.
the caller is only interested in success and failure).
2. Advancing. The return value is important unless it's ADVANCED_FAILED,
in which case the caller wants to treat it as 0.
This change creates separate methods for each use case. The new seekTo
methods automatically rewind and return a boolean. The updated advanceTo
method returns 0 directly in cases where ADVANCED_FAILED was returned.
Arguments that were always hard-coded to true by callers have also been
removed.
This change is a step toward one possible solution for #6155. How we'll
solve that issue is still up for discussion, but this change seems like
one we should make regardless!
Issue: #6155
PiperOrigin-RevId: 290053743
- DownloadManagerHelper now passes all downloads to the
DownloadService when the service is attached (and once
the downloads are known). The service then starts the
foreground notification updater if necessary. This fixes
the ref'd issue.
- Don't call getScheduler() if the service is background
only. This was already documented to be the case on the
DownloadService constructor.
- If the service is started in the foreground on SDK level
26 and higher, satisfy the condition to move the service
to the foreground in onStartCommand rather than in stop().
It's much more obviously correct, and should produce the
same end result.
Issue: #6798
PiperOrigin-RevId: 290050024
This optimization allows a ChunkSampleStream to output track
formats as soon as they're parsed during an InitializationChunk
load, rather than waiting until after the InitializationChunk
load is completed.
In DASH VOD, a single InitializationChunk typically loads the
moov and sidx atoms in that order. Hence for long form content
where the sidx is a non-trivial size, this may result in the
track formats being output a non-negligible period of time
sooner than was previously the case. This allows downstream
renderers to start codec initialization sooner, potentially
decreasing startup latency.
For a single test stream on a fast & stable network, this
pretty consistently reduced elapsed time until both audio and
video codecs have been initialized from ~0.5s to ~0.3 seconds
on a Galaxy S8. For 5 test runs without and with these patches,
the eventTime logged by EventLogger for the second decoder
init were:
Without (secs): 0.47 0.47 0.45 0.48 0.46
With (secs) : 0.32 0.33 0.34 0.31 0.40
PiperOrigin-RevId: 289845089
The current order of operations means that the Format is only passed
to the chunk's output after the DataSource has been opened. This
means that establishing the network connection and the downstream
renderers initializing their codecs are effectively serialized to
occur one after the other.
In the new order, the Format is passed to the chunk's output before
the DataSource has been opened. This allows the downstream renderers
to initialize their codecs in parallel with the network connection
being established, and hence latency at the start of playback is
reduced.
PiperOrigin-RevId: 289841854
As discovered whilst investigating #6798, there are cases
where these methods are not correctly. They were added as
convenience methods that could be overridden by concrete
DownloadService implementations, but since they don't work
properly it's preferable to require application code to
listen to their DownloadManager directly instead.
Notes:
- The original proposal to fix#6798 stored the state change
events in memory until they could be delivered. This approach
is not ideal because the events still end up being delivered
later than they should be. We also want to fix the root cause
in a different way that does not require doing this.
- This change does not fix#6798. It's a preparatory step.
Issue: #6798
PiperOrigin-RevId: 289418555
Unfortunately devices such as the MI 8 do not provide an alternative
working decoder. Some Vivo devices do provide one though.
PiperOrigin-RevId: 288911897
This improves readability by making clearer that reading no bytes from
the input in readFrames() is an expected case if the buffer is full.
PiperOrigin-RevId: 288711841
- Simulate IO exceptions in the test using FlacBinarySearchSeeker for
seeking in FlacExtractorTests. This makes the test slower but covers
more test cases.
PiperOrigin-RevId: 288285057
This typo was introduced in ddb70d96ad
when migrating a static method with parameter `durationUs` to an
instance method where the correct field to use was `blockDurationUs`
(but `durationUs` also exists).
The test that catches this was only added in 45013ece1e (and
therefore configured with the wrong expected output data).
issue:#6833
PiperOrigin-RevId: 288274197
It's unreliable for at least one IMA ADPCM file I've found.
Calculating the blockIndex to seek to using exact properties
also seems more robust.
Note this doesn't change anything for the existing PCM test,
since averageBytesPerSecond is set correctly. It does make
a difference for an upcoming IMA ADPCM test though.
PiperOrigin-RevId: 287814947
Add support for MP3 as an encoding format for passthrough.
This change does not change the observable
behavior of Exoplayer.
Also name the magics.
#exo-offload
PiperOrigin-RevId: 286146539
- Make extractor output samples that are uniformly distributed
with respect to time, with a target of ~10 samples per second.
The old approach could in theory put every frame into its own
sample, which would be very inefficient downstream because we'd
need to pass them individually to MediaCodec. It could also put
data corresponding to a long duration of time into a single
sample (e.g. if the sample rate of the content is low), which
is bad downstream because we decide whether to set the decodeOnly
flag on a per sample basis. More generally, the new approach
is more predictable :).
- Stop using the WavSeekMap to get sample timestamps, and instead
calculate them directly from the number of frames output. It's
more obviously correct, particularly for data formats like IMA
ADPCM where we'll need to adjust the data prior to output.
PiperOrigin-RevId: 285750010
Also change IcyInfo.rawMetatadata from String to byte[]
ICY doesn't specify the character encoding, and there are streams
not using UTF-8 (issue:#6753). It seems the default of at least one
server is ISO-8859-1 so let's support that as a fallback:
https://github.com/savonet/liquidsoap/issues/411#issuecomment-288759200
Also update IcyDecoder to skip strings it doesn't recognise at all
instead of decoding invalid characters.
The feed from issue:#6753 now decodes accents correctly:
EventLogger: ICY: title="D Pai - Le temps de la rentrée", url="null"
PiperOrigin-RevId: 285388522
We previously used Locale.fromLanguageTag and then toLanguageTag to automatically
filter out invalid tags and to replace deprecated tags. However, this only works
for API 21+.
This change does no longer rely on the platform methods and instead:
1. Keeps invalid tags as they are. This may help if developers rely on
non-spec-complaint language tags (for example, #6681).
2. Adds a list of deprcated tags and their modern replacements to our code
3. Normalizes some short codes that have been superseded by macrolanguage codes by
their macrolanguage equivalent.
Issue:#6681
PiperOrigin-RevId: 285384841
Retrieve the sample number in the extractor instead of passing a holder
to FlacBinarySearchSeeker. This change makes the code easier to
understand and is required to implement the seeking from the seek table.
PiperOrigin-RevId: 285241862
- WavHeader is now immutable and contains only values parsed out
of the WAVE FMT chunk. It no longer contains a C.PcmEncoding
encoding, or mutable data bounds.
- WavHeaderReader now parses the WAVE header chunks without any
additional logic (e.g. validating the block alignment value,
which is format type dependent).
- The SeekMap part of WavHeader is split out into WavSeekMap.
PiperOrigin-RevId: 285232498
This makes the format metadata null (instead of an empty Metadata
object) when it is not provided, and is therefore consistent with the
other extractors behavior.
PiperOrigin-RevId: 284171148
Saves around 200 lines of code. High level overview:
- Rename SampleQueue to SampleDataQueue.
- Rename SampleMetadataQueue to SampleQueue.
This CL should not introduce behavior changes. The only significant
changes in synchronization should be:
+ Add synchronized keyword to isReady.
- Seems to be necessary.
+ Add synchronized keyword to SampleQueue.sampleMetadata.
- Before this change, SampleQueue.sampleMetadata could acquire the
SampleMetadataQueue lock three times in a single method call.
Other miscellaneous improvements:
+ Put all private methods at the bottom.
+ Move release() to the right category.
PiperOrigin-RevId: 283795844
Where media segments are specified using byte ranges, it makes
sense that a server might return 416 (which we don't consider
for blacklisting) if the segment is unavailable, rather than
the 404 (which we do consider for blacklisting) that we expect
when media segments are only specified using a URL.
Issue: #6775
PiperOrigin-RevId: 286620698
The Download constructor considers it invalid to have a failure
reason if the download isn't in the failed state. Unfortunately,
calling DefaultDownloadIndex.removeAllDownloads when there's a
failed download will change the state without clearing the reason.
If the downloads are then read back from the DefaultDownloadIndex
we end up violating the Download constructor assertion.
This change clears the failed reason for any existing rows in the
invalid state, and also fixes the root cause that allows invalid
rows to enter the table in the first place.
Issue: #6785
PiperOrigin-RevId: 286576242
Without this @Nullable, potential subclasses can't override the
method to return null if they don't want to use the renderer as a
media clock.
Issue:#6792
PiperOrigin-RevId: 286545736
In MatroskaExtractor, if the last cue time exceeds the duration
specified in the header, then we end up generating a negative
duration chunk as the last item in the SeekMap. We should probably
not do this, so drop it instead.
Note: Matroska does have a CueDuration element, but it's not used
in the one problematic file I've found.
PiperOrigin-RevId: 285738418
This has been broken since
c3d6be3afd
and broken for ICY (where I noticed the problem) since
5695bae9d8.
ICY symptom is that we see no repeated metadata, because the
Icy-MetaData:1 header doesn't make it to the server so we never get back
icy-metaint.
PiperOrigin-RevId: 285379234
Keyframe was rendered rather than skipped when performing
an exact seek to a non-zero position close to the start of
the stream.
PiperOrigin-RevId: 284798460
Caveats:
- Block additional data is ignored if the block is laced
and contains multiple samples. Note that this is not
a loss of functionality (SimpleBlock cannot have block
additional data, and lacing was previously completely
unsupported for Block)
- Subrip and ASS samples are dropped if they're in laced
blocks with multiple samples (I don't think this is
valid anyway)
Issue: #3026
PiperOrigin-RevId: 284545197
This change constrains the use of sample state member variables to
writeSampleData, finishWriteSampleData and resetWriteSampleData.
Using them elsewhere gets increasingly confusing when considering
features like lacing in full blocks. For example sampleBytesWritten
cannot be used when calling commitSampleToOutput in this case
because we need to write the sample data for multiple samples
before we commit any of them.
Issue: #3026
PiperOrigin-RevId: 284541942
- Remove "lacing" from member variables. They're used even if
there is no lacing (and the fact that lacing is the way of
getting multiple samples into a block isn't important).
Issue: #3026
PiperOrigin-RevId: 284152447
- Change sampleHasReferenceBlock to a block reading variable, which is
what it is (the distinction didn't matter previously, but will do so
when we add lacing support in full blocks because there wont be a 1:1
relationship any more.
- Move sampleRead to be a reading state variable.
- Stop abbreviating "additional"
Issue: #3026
PiperOrigin-RevId: 284000937
An I/O error could occur while handling the start of an mdta box, in
which case retrying would cause another ContainerAtom to be added. Fix
this by skipping the mdta header before updating container atoms.
PiperOrigin-RevId: 284000715
This parameter is a little confusing, especially as the behaviour
can be surprising if the intended use-case isn't clear. This change
moves the description of the parameter into the class javadoc,
adds context/justification and slims down each method's
javadoc to refer to the class-level.
Related to investigating/fixing issue:#6700
PiperOrigin-RevId: 283724826
Fixes issue:#6700
sample_cbs_truncated.adts test file produced using
`$ split -b 31795 sample_truncated.adts` to remove the last 10 bytes
PiperOrigin-RevId: 283530136
When transitioning to a new stream in a different format, the audio
processors are reconfigured. After this, they are drained and then
flushed so that they are ready to handle data in updated formats for the
new stream.
Before this change, some audio processors made the assumption that after
reconfiguration no more input would be queued in their old input format,
but this assumption is not correct: during draining more input may be
queued. Fix this behavior so that the new configuration is not referred
to while draining and only becomes active once flushed.
Issue: #6601
PiperOrigin-RevId: 282515359
flush() is guaranteed to be called in all these cases anyway.
Also clarify documentation for AudioProcessor-specific methods that can
change the 'active' flag.
Issue: #6601
PiperOrigin-RevId: 282515255
This simplifies the contract of configure and is in preparation for
fixing a bug where more input can't be queued when draining audio
processors for a configuration change.
Issue: #6601
PiperOrigin-RevId: 282514367
- Move property to DrmSession; it feels like a more natural place
for it to go (and provides greater flexibility).
- Change flags to a boolean.
PiperOrigin-RevId: 281758729
They are all marked with a JavaDoc comment and the @WorkerThread annotation
which is useful if apps are using threading annotations. All other public
methods in the same classes are marked with @AnyThread to avoid the impression
we forgot to annotate them.
PiperOrigin-RevId: 281490301
The current workaround seems to cause compilation errors inside the
testutils module in Android Studio. This seems to fix them.
This doesn't introduce a circular dependency because it's only
the tests in library-core depending on testutils.
PiperOrigin-RevId: 281318192
Note:
- Fixing this uncovers another bug in how audio processor draining
works, so the test playlist still doesn't play correctly after this
change.
- Once we reconfigure the audio sink based on the ExoPlayer Format
rather than the codec MediaFormat in a later change, this change can
be reverted.
Issue: #6601
PiperOrigin-RevId: 281264149
- Make NonNull, which is already the case when using the manager builder.
- Better document PLAYREADY_CUSTOM_DATA_KEY, now that newPlayReadyInstance
is no more.
PiperOrigin-RevId: 281079288
The implementation of writing HDR10+ static metadata assumed that the
application would use default (big endian) byte order for this metadata but
MediaCodec expects the order to match the specification CTA-861.3.
PiperOrigin-RevId: 281050806
Also remove the "do we really need to do this" comment for AAC.
Parsing from codec specific data is likely to be more robust, so
I think we should continue to do it for formats where we've seen
this problem.
Issue: #6648
PiperOrigin-RevId: 280575466
Previously the renderer EOS (aka last frame rendered), was reported as soon
as the last encoded frame was queued in the codec renderer.
This leaded to EOS reported too early.
PiperOrigin-RevId: 280456277
This speeds up downloads where segments have the same URL with different
byte ranges. We limit the merged segments to 20 seconds to ensure the download
progress of demuxed streams is roughly in line with the playable media duration.
Issue:#5978
PiperOrigin-RevId: 280410761
Clear state for one mode when entering the other in both SimpleExoPlayer
and SimpleDecoderVideoRenderer. The latter is redundant for the case of
renderers that are used inside SimpleExoPlayer, but seems nice to have.
- Entering Surface mode means receiving a non-null Surface, SurfaceHolder
or TextureView in SimpleExoPlayer, or a non-null Surface in
SimpleDecoderVideoRenderer.
- Entering VideoDecoderOutputBufferRenderer means receiving a non-null
VideoDecoderOutputBufferRenderer in SimpleExoPlayer and
SimpleDecoderVideoRenderer.
PiperOrigin-RevId: 280155151
This aligns the method naming and Javadoc. The only remaining
inconsistency I can see is that the initial reference count for
DrmSession is 0 rather than 1. Unfortunately I think it's not
trivial to get these aligned, because DefaultDrmSessionManager
relies on being able to do something between instantiation and
the DrmSession starting to open the session. In practice this
doesn't really matter, since DrmSessions will be obtained via the
manager, which does increment thee reference count to 1 to be
consistent with how ExoMediaDrm acquisition works.
PiperOrigin-RevId: 280136574
