The implementation of these methods was updated from direct java integer
arithmetic in 885ddb167e.
In this change, `RoundingMode.FLOOR` was used to try and maintain
compatibility with java integer division. This was incorrect, because
java integer division uses `DOWN` (i.e. towards zero), rather than
`FLOOR` (i.e. towards negative infinity) semantics.
This change fixes the compatibility.
The dump file changes in this CL relate to tests that exercise edit
list behaviour. This involves manipulating negative timestamps, which
explains why they are impacted by this change.
PiperOrigin-RevId: 684013175
This CL also fixes EOS handling to account for not-yet-copied samples in
`remainingInputToCopyFrameCount`, which would throw off the final output
sample count calculation.
For testing, we allow a tolerance of 0.000017% drift between expected
and actual number of output samples. The value was obtained from running
100 iterations of `timeStretching_returnsExpectedNumberOfSamples()` and
calculating the average delta percentage between expected and actual
number of output samples. Roughly, this means a tolerance of 40 samples
on a 90 min mono stream @48KHz.
PiperOrigin-RevId: 683133461
The workaround in ExternalTextureManager.forceSignalEndOfStream
was being applied even when the decoder did output all frames, and only
the GL pipeline was slow.
This change ensures that workaround is not applied when the decoder
has already produced all output frames.
PiperOrigin-RevId: 680471587
Some devices can encode portrait 720x1080 but not landscape
1080x720. But VideoSampleExporter always prefers encoding
landscape. Have `assumeFormatsSupported` mirror sample exporter
logic more closely
PiperOrigin-RevId: 679495210
To avoid rounding errors, set the `Rounding mode` of the `uvFromVu` and `vuFromUs` results to `HALF_UP`. This `Rounding mode` rounds numbers towards the "nearest neighbor" unless both neighbors are equidistant, in which case round up.
PiperOrigin-RevId: 679003943
This CL adds `SonicTest` and `RandomParameterizedSonicTest` as
initial basic unit testing for `Sonic.java`. The tested scenarios
do not necessarily verify a correct implementation of Sonic, but rather
hope to catch any behaviour change from the current implementation.
The change includes a small fix for a lossy simplification and also
checks whether the output sample count matches the expected drift from
the truncation accumulation error present in Sonic's resampler. This is
important as pre-work for fixing issues with unexpected durations within
`SonicAudioProcessor` and `SpeedChangingAudioProcessor` that cause AV
sync issues for speed changing effects.
This is a partial roll forward of e88d6fe459, which was rolled back in
873d485056.
PiperOrigin-RevId: 677756854
Last buffer was not flipped, so was writing the garbage data between
limit and capacity, rather than the actual data between position and
limit.
As a result, all PCM audio dump files need updating.
PiperOrigin-RevId: 676452990
This is no longer needed now our `compileSdk` implies a new-enough AGP
which does this out-lining automatically via R8. See also
https://issuetracker.google.com/345472586#comment7
There's no plan to remove the `ApiXXX` classes, but no new ones need
to be added.
PiperOrigin-RevId: 675940634
This is to allow not setting the MediaFormat OPERATING_RATE and PRIORITY
altogether. The current behvaiour, if left the value `UNSET`, it'll apply the
our optimizations, but apps might want to disable this optimization.
PiperOrigin-RevId: 675923909
After the change in a879bc2154, the Sequence won't have repeated
EditedMediaItems. Thus if the sequence is looping, the last EditedMediaItems
in the Sequence object might not corresponds to the last item in the "logical"
sequence.
PiperOrigin-RevId: 675912197
The muxer might not have accepted the first sample, if it
is waiting for audio track.
This bug causes issue when
1. VideoSampleExporter gives first sample (timestamp = 0) to the muxer.
2. Muxer does not write it because its waiting for audio track.
3. The video pipleline has processed all the sample and they are ready
to be consumed.
4. VideoSampleExporter fetches the next available sample from encoder (which is still with timestamp = 0) but it changes its timestamp to last timestamp because VideoSampleExporter thinks it has muxed the sample at timestamp zero, but in reality it hasn't. This is because the flag `hasMuxedTimestampZero` is set when queueing the input, rather than actually muxing the input.
This scenario can happen when video is processed much faster than
the audio.
PiperOrigin-RevId: 675565603
to
LAST_SAMPLE_DURATION_BEHAVIOR_SET_FROM_END_OF_STREAM_BUFFER_OR_DUPLICATE_PREVIOUS
This CL also combines LAST_SAMPLE_DURATION_BEHAVIOR_SET_FROM_END_OF_STREAM_BUFFER
and LAST_SAMPLE_DURATION_BEHAVIOR_DUPLICATE_PREVIOUS.
The reason for combining the two enums is that, when the option
to use END_OF_STREAM_BUFFER is selected and if the EOS buffer is
not provided then the muxer anyways fallbacks to duplicate
duration behavior.
The last sample with 0 durations seems less useful so
change the default behavior to non-zero duration.
This will also match the behavior with MediaMuxer.
PiperOrigin-RevId: 675189932
The tests were flaky because CompositionPlayer registers audio sequences'
formats to the audio pipeline in first-come-first-serve order. With the change
in bc8d82355f, the audio format is deterministic.
The test is turned off in 060356ea00
PiperOrigin-RevId: 674261801
Sonic would accumulate truncation errors on float to int conversions
that caused the final output sample count to drift noticeably, by
hundreds of samples on streams of a few minutes of length. The fix now
keeps track of the truncation error and compensates for it.
Other small fixes include eliminating lossy operations (e.g. int
division) and using doubles instead of floats for resampling where
helpful.
This CL also introduces `SonicParameterizedTest`, which helps test
resampling on an arbitrary number of randomly generated parameters,
with random sample data. `SonicParameterizedTest` uses `BigDecimal`s
for calculating sample count values, as to avoid precision issues with
large sample counts.
PiperOrigin-RevId: 673852768
When processing edit lists in MP4 files, the media start position may be a non-keyframe. To ensure proper playback, the decoder must preroll to the preceding keyframe, as these preroll samples are essential for decoding but are not rendered.
Issue: google/ExoPlayer#1659
#cherrypick
PiperOrigin-RevId: 673457615
In case of negative PTS workaround, instead of disallowing all cases where samples are not in presentation order, we now validate that the first sample's presentation time is the smallest. This adjustment allows for correctly applying an offset to ensure all samples have a presentation time >= 0.
#cherrypick
PiperOrigin-RevId: 673434793
Limit input image size in Transformer to be less than 4096x4096.
For very large images, this can reduce memory usage substantially,
and stays away from `GL_MAX_TEXTURE_SIZE` - often 4096
PiperOrigin-RevId: 670555939
These tests do not test performance. Moving them out of this directory
ensures they are run on emulators and on all physical devices.
PiperOrigin-RevId: 670505992
