Now [DASH/SS]SampleSource instances are creating ChunkTrackStream
instances dynamically, it makes sense to always require that the
min retry count be injected from that level. The SampleSource
implementations should also use the retry count when refreshing
the manifest.
The option to configure the retry count on the SampleSource classes
will arrive in a later CL.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=122424774
- This means DrmInitData is propagated through sample queues (i.e.
is effectively attached to every sample, so we can see when it
changes when reading from the queue).
- It also allows different DrmInitData per track, which is possible
in muxed MKV/WebM, and per Representation for DASH, although we
wont be able to seamlessly adapt in the latter case.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=121821928
I've seen a sample that has a LIST chunk prior to fmt. It
seems best just to skip all chunks until the format is
located. As per the RIFF spec:
"Programs must expect (and ignore) any unknown chunks
encountered, as with all RIFF forms."
https://www.aelius.com/njh/wavemetatools/doc/riffmci.pdf
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=121682470
Dash/SS SampleSources now instantiate ChunkSource and
ChunkSampleSource (renamed to ChunkTrackStream) instances
on demand as tracks are enabled. The TrackGroups exposed
by the DASH/SS SampleSources are now constructed at the
top level.
Note that this change resolves the TODOs at the top of the
ChunkSource classes, allowing multiple adaptation sets of
the same type.
Next steps will include:
- Bring back UTC timing element support for DASH, which
will be an extra request during preparation in the DASH
SampleSource.
- Simplification of manifest fetching to use a Loader directly
in the two top level SampleSource classes. ManifestFetcher
should eventually go away once HLS no longer needs it.
- Eventually, some consolidation between DASH/SS. There's a
lot of common code there now.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=121001777
Parse the duration of the media directly from the manifest
in the DASH/SS SampleSource implementations.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=121001296
This change pulls manifest refresh responsibility up to the
top level Dash/SS SampleSource implementations. In following
steps more of the manifest processing logic will be pulled
up (e.g. extracting track groups from the initial manifest),
which will allow ChunkSampleSource/ChunkSource instances to
be further simplified and created on demand.
I've avoided moving/renaming anything for now, so as to keep
it fairly easy to review.
Note that this change does the TODO related to releasing the
manifest fetchers.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=121001139
Reading the format and/or a sample needs to be done as a
single operation. Else you can have a situation where the
queue is initially empty, and this happens:
1) Read downstream format X
2) Read downstream format X (unchanged)
3) Write format Y
4) Write first sample
5) Read first sample
The first sample then appears to be format X rather than Y.
Note that readData in the SampleSource implementations always
looks roughly the same. readReset is identical in all cases.
isReady is identical in all cases now I've fixed them to be
that way. So it should be pretty easy to get DefaultTrackOutput
to implement TrackStream directly, at which point a whole load
of duplication will disappear.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=120546377
Having moved HLS to use single sample queues per track, these
classes have become relatively similar. This CL aligns the two
to make this more obvious. It remains unclear whether it'll
ever be sensible to merge them; there are still some niggly
complications for HLS.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=120429618
- RollingSampleBuffer -> DefaultTrackOutput
- TsChunk -> HlsMediaChunk
- Established hls.playlist package for HLS playlist things
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=120325049
This also fixes the largest queued timestamp to be the
correct value if upstream samples are discarded.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=120207054
At this point the only reason preventing the chunk package
from using RollingSampleBuffer directly, rather than
DefaultTrackOutput, is that the latter maintains the largest
parsed timestamp. This will be pushed inside the former in
the next CL. Following that, splicing logic will also be
pushed inside of RollingSampleBuffer, and HLS will be moved
over to using a single RollingSampleBuffer per track, with
the splicing done inline.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=120206658
This is the first step toward merging of RollingSampleBuffer
and DefaultTrackOutput, which is a precursor to removing some
indirection for DASH/SS playbacks, and to moving HLS playbacks
to use a single output buffer per track (with inline splicing).
Where this is heading is that sample format changes will
eventually be attached to samples in the rolling buffer. This
will eliminate the need for piping sample formats around the
edges of the rolling buffer (e.g. via redirection in
ChunkExtractorWrapper and BaseMediaChunk.getSampleFormat).
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=120123093
- Use same constant for unknown/unset microsecond times/durations.
- Change constant values to be nowhere near the "normal" range.
- Misc cleanup.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=119944019
For DASH + SS, sample data is currently plumbed through:
Extractor
->ChunkExtractorWrapper
->ContainerMediaChunk
->DefaultTrackOutput
->RollingSampleNBuffer
This change bypasses the ContainerMediaChunk layer. It
should be possible to completely delete DefaultTrackOutput
in the future, but such a change may well be tied to
changes in HLS format splicing + how we buffer (or don't
buffer) disabled tracks.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=119853857
Both of these features are being promoted to first class
citizens in V2 (multi-period support will be handled via
playlists, seeking-in-window will be handled by exposing
the window/timeline from the player and via the normal
seek API). For now, it's much easier to continue the
refactoring process with the features removed.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=119518675
The only change that's not re-ordering is to add a Util method
for usToMs to replace the ones in the HLS and Chunk sources.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=118775359
This was made possible by the simplification of how DASH/SS
chunk replacement works. It is also a step towards eliminating
continueBuffering(), since continueBuffering() calls are no
longer relied upon to resume a backed off load.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=118774865
DASH + SS previously had a ridiculously complicated chunk
replacement mechanism in resumeFromBackOff. It also didn't
allow replacement of the first media chunk in the queue,
even though it's possible to remove it in the case that no
corresponding samples have been consumed.
This CL moves DASH + SS to the simpler model used in the
HLS implementation, where the chunk source has a single
opportunity to cancel (and hence later replace) the chunk
when the load error occurs. With this change comes the
ability to replace the first media chunk in the queue in
all cases where it's possible to do so.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=118573418
This fixes some nuances with the initial solution. Mainly,
that the TrackStreams returned by selectTrack could not be
used safely until after endTrackSelection was invoked. It
also reduces the need for member variables to track state
between the track selection methods.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=118556651
This removes the need for each SampleSource implementation to
implement 3x "if(condition) {noop}" tests (ChunkSampleSource
and SingleSampleSource were missing some of these checks).
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=118036233
1. SampleSource now has an explicit track selection state. This state is entered
after the source is prepared, and also by calling startTrackSelection.
2. endTrackSelection commits selection changes, and is responsible for doing the
right thing w.r.t starting/stopping/restarting load operations.
3. All sources now start or restart a load in the case of a new track selection.
This fixes a problem where a source could be advanced by repeatedly disabling
and re-enabling whilst paused. Some sources didn't restart a load in this case,
since the position was unchanged, however the downstream renderer would then
consume media up to the first keyframe in order to render something. Hence
each disable/re-enable would advance by a keyframe.
4. This change will enable a subsequent change where we'll discard media for
non-selected tracks earlier than we do currently (i.e. we'll hook the extractor
to a dummy track output, so the samples will never be written to a rolling
buffer). This will enable a further subsequent change where buffer contributions
are per-renderer rather than per-source.
Issue: #1041
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=118024436
Also use MediaCodec buffer flag constants instead of those on MediaExtractor.
This is in preparation for merging InputBuffer and SampleHolder.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=117810136
- Remove special DefaultTrackOutput.sampleData method, and have
SingleSampleMediaChunk use the regular one instead.
- Make DummyTrackOutput behave correctly is allowEndOfInput is
false.
- Simplify progress tracking in ExtractorSampleSource.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=117808659
This change replaces TrackGroup[] with TrackGroupArray. This is
to allow equality based hashCode and equals implementations.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=115563146
The TrackSelector API will look like:
TrackSelection[] selectTrack(
TrackRenderer[] renderers, TrackGroup[] trackGroups);
In this CL:
- SampleSources return TrackGroup[], so that the result can be easily
passed to the selector.
- TrackStream gets its own file to sit alongside other Track* classes.
- A TrackSelection object is introduced to encapsulate group and track
indices.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=115251331
- Format can represent both container and sample formats.
If a container contains a single track (as is true in
DASH and SmoothStreaming) then the container Format can
also contain sufficient information about the samples
to allow for track selection. This avoids the Format to
MediaFormat conversions that we were previously doing in
ChunkSource implementations.
- One important result of this change is that adaptive
format evaluation and static format selection now use the
same format objects, which is a whole lot less confusing
for someone who wants to implement both initial selection
and subsequent adaptation logic. It's not in the V2 doc,
but it may well make sense if the TrackSelector not only
selects the tracks to enable for an adaptive playback, but
also injects a FormatEvaluator when enabling them that will
control the subsequent adaptive selections. That would make
it so that all format selection logic originates from the
same place.
- As part of this change, the adaptiveX variables are removed
from the format object; they don't really correspond to a
single format. This also saves on having to inject the max
video dimensions through a bunch of classes.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=114546777
Duration was originally included in MediaFormat to match the
framework class, but it actually doesn't make much sense. In
many containers there's no such thing as per-stream duration,
and in any case we don't really care. Setting the duration on
each format required excessive piping.
This change moves duration into SeekMap instead, which seems
to make a lot more sense because it's at the container level,
and because being able to seek is generally couplied with
knowing how long the stream is.
This change is also a step toward merging Format and MediaFormat
into a single class (because Format doesn't have a duration),
which is coming soon.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=114428500
[Step 4 - Partial, of []
- The capabilities checks previously performed in VideoFormatSelectorUtil
are now performed in MediaCodecVideoTrackRenderer. This means they'll be
useful for non-chunk use cases (e.g. when using ExtractorSampleSource).
- Added capabilities checks for audio in MediaCodecAudioTrackRenderer. We
didn't check audio capabilities previously.
- Added functionality to allow a TrackRenderer to indicate the extent of
its adaptive support.
The idea here is that a TrackSelector (to be introduced) will have access to:
(a) TrackGroups from the SampleSource that indicate whether they support
adaptive playbacks and the formats of each individual track.
(b) TrackRenderers that indicate whether they support adaptive playbacks as
well as how capable they are of rendering formats of individual tracks.
This is everything that a TrackSelector needs from the player components in
order to decide how to wire things up. Note that a TrackSelector may opt to
treat FORMAT_EXCEEDS_CAPABILITIES as FORMAT_HANDLED at its own risk, if it
thinks that it (or the user) knows better. This is a request that we've seen
from third parties for better handling cases where capabilities aren't
accurately reported by the underlying platform.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=114427428
Delete SingleSampleChunkSource. I don't think it's really
useful for anything, now we have SingleSampleSource.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=113259289
This change removes the need for SourceBuilders to load their
manifests before building their sources. This is done by pushing
initial manifest loads into the ChunkSource classes. This simplifies
the SourceBuilders a lot, and also DemoPlayer.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=113259259
Notes:
1. The logic in ExoPlayerImplInternal is very temporary, until we
have proper TrackSelector implementations. Ignore the fact that
it's crazy and has loads of nesting.
2. This change removes all capabilities checking. TrackRenderer
implementations will be updated to perform these checks in a
subsequent CL.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=113151233
GitHub note - Apologies for the cryptic change descriptions,
they relate to a design doc that's not externally visible.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=113043764
When a seek is performed, renderers currently perform the
actions that they need to take in two places: Some changes
are performed in seekTo implementations. Other changes are
performed when discontinuities are read from the source.
In HLS we need to perform what is effectively a seek
originating in the source. To support this, this CL allows
discontinuities read from the source to modify the playback
position. All actions that renderers perform as a result
of a seek are moved to be performed when a discontinuity is
received.
Best way to understand CL:
- Look at SampleSource interface change and then at the
concrete implementations, to make sure they've been
changed properly.
- Look at SampleSourceTrackRenderer change.
- Look at concrete renderers. The general pattern is that
code previously performed in seekTo and READ_DISCONTINUITY
is merged into onDiscontinuity().
Note: This will be further untangled in V2.
Issue #676
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=112720746
