If there is data after the last samples in the container, we may request
continue loading after the last samples have been read but before the load has
completed. In this situation the buffered position is returned as
Long.MAX_VALUE, which prevents continuing loading, yet the media period is not
treated as fully buffered because its buffered position is not
C.TIME_END_OF_SOURCE.
PiperOrigin-RevId: 231406964
We can keep the reading media period and continue playing if we haven't read
beyond the new duration. Otherwise, we can keep the period, but need to reset
the renderers as we already read too far.
PiperOrigin-RevId: 231406252
It better describes what the class does. More importantly, we've had
inconsistent class names since we added offline support, for which we
added ProgressiveDownloader ("ExtractorDownloader" doesn't make any
sense). We could really do with aligning the names for clarity.
(Sorry)
PiperOrigin-RevId: 231387268
This requires to prepare the media source and the periods in a small helper similar
to the metadata retriever. It also gets rid of the need to have abstract protected
methods to load the manifest, to extract the track groups and to convert to stream keys,
as this can now be handled by the media period.
PiperOrigin-RevId: 231385590
Converting a single track selection to stream keys is only possible if the output
is independent from other track selections being made.
This is not the case for DASH and HLS embedded track groups which should select the
already selected primary track if possible (and thus needs to know whether a primary
track group is selected).
Also, update the test method to take a period index.
PiperOrigin-RevId: 231385490
The way it is currently, it's very unclear that an operation on the
version table will correctly belong to a transaction in code such as
this, taken from DefaultDownloadIndex:
writableDatabase.beginTransaction();
try {
writableDatabase.execSQL(...);
versionTable.setVersion(...);
writableDatabase.setTransactionSuccessful();
} finally {
writableDatabase.endTransaction();
}
This change explicitly passes the database, to make it obvious that
the operation will really go into the same transaction:
writableDatabase.beginTransaction();
try {
writableDatabase.execSQL(....);
VersionTable.setVersion(writableDatabase, ...);
writableDatabase.setTransactionSuccessful();
} finally {
writableDatabase.endTransaction();
}
PiperOrigin-RevId: 231374933
Float output is only possible from API 21, but the high-res int
to float conversion path was checking for 32-bit PCM not float
output capability.
PiperOrigin-RevId: 231172495
This will allow CachedContentIndex to start using database
storage without us having to change the SimpleCache and
CachedContentIndex constructors to require a Context or a
DatabaseProvider.
PiperOrigin-RevId: 230884501
If are allowing changing durations of periods, we shouldn't use the end position
of clipped content as part of the id as it may change. This change moves the end
position back to MediaPeriodInfo and adds the next ad group index to the id
instead to ensure we still have unique ids for all content parts.
PiperOrigin-RevId: 230878389
The combination of pre-16 API levels accounting for ~0.5% of the device
population, and that the most important components in ExoPlayer (e.g.
the MediaCodec renderers) have always required API level 16, mean it's
very unlikely this will negatively impact on anyone.
PiperOrigin-RevId: 230701808
We need to support serialization to/from an SQLite table. The
model of passing something around for each class to write into
doesn't work well for SQL, and it would be messy to have two
different structural designs for serialization. This change
centralizes the logic in CachedContentIndex, where a centralized
SQL based version can more easily sit alongside it.
PiperOrigin-RevId: 230692291
This method is a generalization of the existing hasReadStreamToEnd. It is
useful to determine whether a renderer already read beyond a new duration of
a period.
PiperOrigin-RevId: 230689165
1. The method kept track of the current period index to check if the next
period is still in the correct period. This is unneccessary since we no longer
use the period index but the actual uid in MediaPeriodId and mismatches are
already detected by canKeepMediaPeriodHolder.
2. We updated the MediaPeriodIndfo twice: once in getFollowingMediaPeriodInfo
and once in getUpdatedMediaPeriodInfo. That's confusing and difficult to
follow. The only difference is that getUpdatedMediaPeriodInfo keeps the
content position while getFollowingMediaPeriodInfo resets it. This is made more
explicit for readability.
3. The durations compatibility check for all following periods was broken as
it compared the same durations (partly due to the confusion caused by 2.)
PiperOrigin-RevId: 230519295
ExoPlaybackExceptions of type SOURCE are always associated with the loading
period and thus we can use the event time for the loading period in
onPlayerError. Renderer and unexpected exceptions are still associated with the
currently playing period.
Issue:#5407
PiperOrigin-RevId: 230216253
The Amlogic awesome decoder reduces the video size of interlaced videos by half
if the internal configuration isn't force reset with new maximum input size
values. The product of these new values must exceed 1920x1088 to force the
reset.
Issue:#5003
PiperOrigin-RevId: 230206675
That's the same position set in MediaPeriod.prepare (where it may be removed
in the future).
Having the position at an earlier point is necessary to fix an
issue with lazy preparation in ConcatenatingMediaSource where the prepare
position was assumed to be known but MediaPeriod.prepare hasn't been called
yet.
Issue:#5350
PiperOrigin-RevId: 229756637
The buffered position is currently based on the mimimum queued timestamp of
all AV tracks. If the tracks have unequal lengths, one track continues loading
without bounds as the "buffered position" will always stay at the shorter
track's duration.
This change adds an optional buffer flag to mark the last sample of the
stream. This is set in the Mp4Extractor only so far. ExtractorMediaSource
uses this flag to ignore AV streams in the buffered duration calculation if
they already finished loading.
Issue:#3670
PiperOrigin-RevId: 229359899
That was previously handled by the player. But since we switched to Handler
messages instead of player messages, we should do that manually.
PiperOrigin-RevId: 229341747
This is not necessary as the track selection needs to be updated with
updateSelectedTrack anyway. It's also error-prone as the selection code
calls into a protected method of a not fully initialized class.
PiperOrigin-RevId: 229331669
DataSpec.FLAG_ALLOW_CACHE_FRAGMENTATION is added to indicate to the
cache when fragmentation is allowed. This flag is set for progressive
requests only.
To avoid breaking changes, CacheDataSink defaults to ignoring the flag
(and enabling fragmentation) for now. Respecting the flag can be
enabled manually. DownloaderConstructorHelper enables respecting of
the flag.
Issue: #4253
PiperOrigin-RevId: 229176835
We currently forget whether a source is seekable at re-preparation. This was
implemented intentionally this way under the assumption that we really can't seek
until we have loaded the seek map again. However, seek operations are only
allowed after a media period is prepared. So there is no harm in remembering
whether a source is seekable.
This problem currently prevents reusing ClippingMediaSources with
ExtractorMediaSource and a non-zero start clip position.
Issue: #5351
PiperOrigin-RevId: 229169441
ExoMediaCrypto.requiresSecureDecoderComponent() is removed, and
FrameworkMediaCrypto.forceAllowInsecureDecoderComponents is made
public to allow determining whether a secure decoder is required
to be implemented in MediaCodecRenderer.
PiperOrigin-RevId: 228909771
There were some edge cases in which we'd forget to release DRM
sessions. For example if we read a format and acquired a
pendingDrmSession (in onInputFormatChanged), then immediately
read another format and overwrote pendingDrmSession, we'd
forget to release the one that's been overwritten.
This change hopefully makes release much clearer. We keep a list
of all drm sessions we're currently holding. Whenever we update
either drmSession or pendingDrmSession, we release any other
sessions that are in the list.
PiperOrigin-RevId: 228905465
The window object is used without being filled with data. This used to work
well for most cases as the same live stream is sending regular updates and the
first update is almost never used if it's not the first item in a playlist.
It causes problems when the first timeline update of a live stream is actually
used for playback (e.g. when the live stream is lazily prepared in a playlist
and played first).
PiperOrigin-RevId: 228530232
Reported in https://github.com/google/ExoPlayer/issues/3860
For failing examples see the github link above.
[Problem]
We drop matching control codes even if they are not received on
consecutive frames.
The specification says
"(4) If the first transmission of a control code pair passes parity,
it is acted upon within one video frame. If the NEXT frame contains
a perfect repeat of the same pair, the redundant code is ignored."
Keyword is the NEXT. The frames must arrive immediately after
each other.
See https://www.law.cornell.edu/cfr/text/47/79.101
[Solution]
Set an additional flag when any data is processed. Control code
duplication checks should be limited only for the first control
byte pairs processed after any control code.
[Test]
Sarnoff tests have equivalent CEA708 and CEA608 Streams.
This is the initialization part of mitigating issue #4253. The
remaining work is on the writing side, and is simply a case of
having startFile return File instances that are sharded into
sub-directories. We still need to decide what scheme we want
to use for doing that.
Issue: #4253
PiperOrigin-RevId: 228306327
Calls to File.length() can be O(N) where N is the number of files
in the containing folder. This is believed to be true for at least
FAT32. Repeated calls for the same file tend to be faster,
presumably due to caching in the file system, however are still
surprisingly expensive. Hence minimizing the number of calls is
preferable.
Issue: #4253
PiperOrigin-RevId: 228179921
They are not longer needed anywhere, are error-prone (because of threading
requirements), and complicate testing and using MediaSources without a player.
PiperOrigin-RevId: 227871157
We currently either use the app thread returned by the player or the thread
the commands are called on depending on whether the media source is already
prepared or not.
This change lets the application decide which callback thread to use. As a
side effect, we also don't longer need access the player instance passed to
MediaSource.prepare.
PiperOrigin-RevId: 227871111
Also configure the FFmpeg context to ignore errors as far as possible (this
appears to have an effect only for certain decoders).
Issue: #5293
PiperOrigin-RevId: 227851397
The top level requirement only tried to ensure that the entire timeline only
has one period. This is already asserted by ImaAdsLoader. AdsMediaSource
itself works fine as long as the wrapped timeline has one period only. This
is now asserted instead.
PiperOrigin-RevId: 227682141
Passing the player through MediaSource.prepare is only needed for the AdsLoader
and complicates other usages of MediaSource. Providing the player directly to
the AdsLoader is also in line with the usage pattern of PlayerView and other
components.
Also rename methods to start/stop to better reflect their usage.
PiperOrigin-RevId: 227682112
Made DownloadState top level class.
Replaced action field DownloadAction fields.
Added removing, removed and restarting states.
Renamed started state to downloading.
PiperOrigin-RevId: 227664735
ExoPlayer methods must not be called from any thread besides the specified
app thread. Therefore we shouldn't use them here. Using a regular Handler
instead is fully equivalent.
Issue:#5240
PiperOrigin-RevId: 227650489
- Increase the default cache file size to 5MB
- Recommended a minimum cache file size of 2MB to discourage
applications from specifying values small enough such that
unreasonably large numbers of cache files are generated
- Allow maxCacheFileSize=C.LENGTH_UNSET, equivalent to setting it
to MAX_VALUE. This is just for API consistency with other APIs
we have that accept LENGTH_UNSET
Issue: #4253
PiperOrigin-RevId: 227524233
CacheEvictor.onStartFile currently receives a length, which is the
maximum size of the content that might be written. The LRU cache
evictor will make sure there's sufficient space for the specified
size. If the actual size of the content is unknown, CacheDataSink
passes maxCacheFileSize.
The current behavior isn't ideal. It seems valid for a developer to
specify maxCacheFileSize=Long.MAX_VALUE if they don't want any
file fragmentation in the cache. However if they then attempt to
cache something with unknown length, LRU cache will try and make
room for content of length Long.MAX_VALUE, and end up evicting the
entire cache to do so.
This change alters the logic so a length is only passed if the
actual size of the content is known. In other cases C.LENGTH_UNSET
is now passed. The LRU evictor will not evict until after the file
is committed. Note a custom LRU evictor could still opt to evict to
ensure some application specified amount of space, if that's the
desired behavior.
PiperOrigin-RevId: 227509525
When a Download is created it's set to queued state but doesn't notify
listeners about this state. DownloadManager checks if it can start the
download. After this checks it notifies the listeners.
With this change Download can immediately check if it can be started and
sends correct notification.
PiperOrigin-RevId: 225967129
