App users can choose arbitrary data that might not be
anticipated by developers. Transformer shouldn't `checkState` based on
media data or file type -- report an error for unsupported data instead.
Public API change `ImageAssetLoader` needs to parse MIME type and now accepts
`Context` as parameter.
PiperOrigin-RevId: 660762459
Package-private until API is more useable.
Similar to frame analyzer mode: uses ImageReader instead of an encoder,
and no muxer.
PiperOrigin-RevId: 658446675
Transformer.experimentalSetMaxFramesInEncoder controls max number
of frames in encoder.
VideoFrameProcessor now allows delayed releasing of frames to Surface,
while still using the original presentation time.
VideoSampleExporter can now configure video graphs to not render frames
automatically to output surface.
VideoSampleExporter.VideoGraphWrapper tracks how many frames are ready
to be rendered to Surface, and how many frames are already in-use by encoder.
PiperOrigin-RevId: 658429969
Add parameterized test for codecs supported by InAppMuxer.
Split TransformerWithInAppMuxerEndToEndParameterizedTest tests
between parameterized and non parameterized
This will be helpful for adding more tests which need not to
be parameterized.
PiperOrigin-RevId: 658353532
This is following a renaming of registerInputFrame to handleInputFrame.
- queueInputBitmap is renamed to handleInputBitmap for consistency with
handleInputFrame.
- registerInputStream is renamed to onInputStreamChanged for consistency
with media3 method names.
PiperOrigin-RevId: 655529699
Build upon Transformer.videoFrameProcessorFactory in MultipleInputVideoGraph
Check that Transformer.videoFrameProcessorFactory is DefaultVideoFrameProcessor
for multi-input video
Fixes https://github.com/androidx/media/issues/1509
PiperOrigin-RevId: 655232381
This test is flaky at p4head, because CompositionPlayer has no logic to
ensure a specific input is used to configure the audio graph. Depending
on which sequence registers input first, it changes the output format
of the media.
By setting effects on the 2nd sequence, both inputs are the same format
and this flakiness is avoided in the test.
PiperOrigin-RevId: 653582441
-----
Context:
* Each sequence is wrapped as a single MediaSource, each being played
by an underlying ExoPlayer.
* Repeat mode is typically implemented in Players as a seek to the next
item in the playlist.
-----
This CL:
Repeat mode is triggered by listening for when the main input player
sees a play when ready change due to the end of the media item.
There is a slight delay at the end of the playback, before it repeats.
Setting repeat mode on the underlying players addresses this, but means
that the players will seek without waiting for the CompositionPlayer,
and as such previewAudioPipeline does not get the correct signals
around blocking/flushing.
PreviewAudioPipeline - The seek position can validly be C.TIME_UNSET,
however preview pipeline did not handle this case.
CompositionPlayer getContentPosition is given (through a lambda) as a
supplier to the State object, which means any comparisons between
previous/new state for this value does not work. In SimpleBasePlayer,
there is logic to use the positionDiscontinuityPositionUs for the
position change (see getPositionInfo called from
updateStateAndInformListeners), however this logic is not considered in
getMediaItemTransitionReason, so a condition needed to be added for
this case.
-----
Tests:
* Dump files clearly show the position and data is repeated.
* Assertions on the reasons for transitions or position
discontinuities.
PiperOrigin-RevId: 653210278
Sources (for example media projection) can populate the `Surface` from
`SurfaceAssetLoader` with timestamps that don't start from zero. But
`MuxerWrapper` assumes the latest sample timestamp can be used as the duration
when it calculates average bitrate and notifies its listener.
This can cause a crash because the calculated average bitrate can be zero if
the denominator duration is large enough.
Use the max minus first sample timestamp across tracks instead to get the
duration.
Side note: the large timestamps from the surface texture when using media
projection arrive unchanged (apart from conversion from ns to us) in effect
implementations and in the muxer wrapper (and are passed to the underlying
muxer). The outputs of media3 muxer and the framework muxer are similar.
PiperOrigin-RevId: 652422674
This should have no influence on app behavior and other policies
and just allows code to depend on new API 35 platform symbols.
PiperOrigin-RevId: 652414026
Also promote all H.265 constants to be public in `NalUnitUtil` with
`H265_` prefixes, for consistency.
A lot of these names are used in h.265 too (and `NalUnitUtil` handles
both), but with different values, so this rename aims to avoid
accidentally using an h.264 value in an h.265 context.
PiperOrigin-RevId: 651774188
This is because the force EOS workaround for videos is longer than the original
muxer timeout. This way the apps depending on Transformer won't have to manually
set the muxer timeout on emulators.
PiperOrigin-RevId: 651355755
VideoSink.registerInputFrame is now called for every input frame (not
only the ones that should be rendered to the input surface) because it's
the VideoSink that decides whether it wants the frame to be rendered.
PiperOrigin-RevId: 651049851
MCVR crashed because MCVR registers a new input stream to VideoSink on every
`onOutputFormatChanged()`, assuming that `onOutputFormatChanged()` is only
invoked on media item transition. However, it can be called multiple times for
one media item.
PiperOrigin-RevId: 649050576
Hardware decoder on some devices fails to write 1920x1080 YUV_420_888 buffers into an ImageReader. This change allows us to remove skipCalculateSsim device workaround in ExportTest.java.
VideoDecodingWrapper now uses media3.MediaExtractorCompat: necessary for parsing of MediaFormat#KEY_CODECS_STRING and decoder capabilities check
PiperOrigin-RevId: 648726721
The Javadoc was indicating that the duration should always be set,
but it doesn't need to be set in most cases for export.
PiperOrigin-RevId: 644685827
Includes adjusting how effects are defined, to make it clearer in a
test that a given ItemConfig has effects associated with it.
PiperOrigin-RevId: 644684308
These calls have no effect because the VideoFrameReleaseControl of
CompositionPlayer is created with allowedJoiningTimeMs set to 0.
PiperOrigin-RevId: 644274524
This is an internal refactor with no logic changed.
There is a duplication in information and a lack of consistency around
the use of test assets in transformer androidTest. This change
refactors each asset to be defined as its own AssetInfo, with the other
relevant parts stored alongside the uri.
Once this is in place, we can consider other useful functionality, such
as having boolean flags for what tracks an asset has, helper methods
for whether an asset is local or remote, and more. This will reduce the
manual overhead necessary to use more assets in tests, and in
particular leads towards easily using new & existing assets in
parameterized tests.
PiperOrigin-RevId: 644040595
This CL is a pure refactor of the existing tests, iterative changes
will be done in follow-ups.
The only logic change is having all MP4 assets have the presentaton of
height=360. The old code had some with height 480.
PiperOrigin-RevId: 643020773
- Video release should check for buffer timestamp (which is renderer-offsetted), rather than the frame timestamp
- ImageRenderer should report ended after all of it's outputs are released, rather than when finished consuming its input.
Add tests for timestamp handling
PiperOrigin-RevId: 642587290
In order to do that, make the VideoSink nullable in MCVR.
We want to avoid calling VideoFrameReleaseControl.setClock directly
from MCVR when the sink is enabled. The goal is to handle all the
communication with the release control from the sink/sink provider.
PiperOrigin-RevId: 642542063
Rational for keeping field package private:
Since the overall export already very complex, by looking at the output file
its hard to know if the desired processing happened or not. Since we now
support sequences with "n" media items, its even more important to know if all
the media items were processed or not.
Although the field exposes implementation details, it seems ok as we get benefit of detailed testing.
PiperOrigin-RevId: 642337888
Before this change, the timestamps output from composition playback is offset
with the renderer offset. After this change, the offset is removed and the
timestamp behaviour converges with Transformer, that is, the timestamps of
video/images frames will follow that of the composition. For example, with a
composition of two 10-s items, clipping the first with 2s at the start, the
timestamp of the first frame in the second item, will be 8s.
PiperOrigin-RevId: 641121358
Ensures valid progress state is returned. Should not return NOT_STARTED
once transformer.start has been called, until export ends.
PiperOrigin-RevId: 640533805
Usages removed in this CL are:
- onProcessedStreamChange, which was already called from the VideoSink
(via VideoFrameRenderControl)
- setOutputSurface, which was also already called from the VideoSink
- setFrameRate, which this CL now sets in the VideoSink
PiperOrigin-RevId: 640530903
Removes the flakiness of
MediaItemExportTest.getProgress_unknownDuration_returnsConsistentStates
by using a longer input asset, such that ExoPlayer does not determine
the duration of the media.
PiperOrigin-RevId: 640502470
Add SeparableConvolution.configure(inputSize) to allow effect configuration
depending on input dimensions.
Add LanczosResample.scaleToFit method to scale input images to fit inside
given dimensions.
PiperOrigin-RevId: 640498008
These changes are possible because getProgress is no longer a blocking
operation on transformer.
* Tests call getProgress after every looper message executed.
* Use longer media assets for getProgress tests to give more progress
intervals.
* Remove conditional assertions.
PiperOrigin-RevId: 639734368
Switch from 4-channel RGBA_16F lookup texture to 1-channel R_16F.
Do not use a bitmap when creating the lookup table texture.
Instead, fill the texture directly.
Do not manually convert 32-bit float to 16-bit. Instead, let OpenGL
libraries do this for us.
PiperOrigin-RevId: 639717235
AudioGraphInput.onMediaItemChanged is called on input thread. Pending
media item changes are processed on processing thread, inside calls to
getOutput().
This change allows multiple pending media item changes to be enqueued,
and processed in sequence.
PiperOrigin-RevId: 638995291
Degammaing has been removed in cb4b2ea55c. The goldens for
TransformerHdrTest (previously TransformerSequenceEffectTestWithHdr)
were not regenerated because the test wasn't running due to its name
(fixed in e41a966237).
PiperOrigin-RevId: 638645635
Add DefaultVideosFrameProcessor experimental flag that controls
whether input Bitmaps are sampled once for a repeating sequence of
output frames with the same contents, or once for each output frame.
PiperOrigin-RevId: 637921350
Only sample from input bitmap when the input image has changed.
Introduce GainmapShaderProgram.newImmutableBitmap API that signals
input bitmap changes to GainmapShaderProgram (DefaultShaderProgram).
PiperOrigin-RevId: 637920207
Add a wait in DefaultCodec.signalEndOfInputStream when no
video encoder output has been seen. This avoids a thread synchronization problem
between writing frames to video surface, and signaling end of stream,
which was hit for video input of only one frame on some devices.
PiperOrigin-RevId: 637844690
PQ and HLG have different luminance ranges (max 10k nits and max 1k nits resp). In GL, colors work in a normalised 0 to 1 scale, so for PQ content, 1=10k nits and and for HLG content, 1=1k nits.
This cl scales and normalises PQ content appropriately so that all HDR content works in the HLG luminance range. This fixes two things
1. Conversions between HLG and PQ are "fixed" (before the output colors looked too bright or too dark depending on which way you are converting)
2. color-altering effects will be able to work consistently across HLG and PQ content
1 is tested in this cl. 2 will be tested when ultra HDR overlays are implemented, both cases have been manually tested to ensure the output looks correct on a screen.
PiperOrigin-RevId: 636851701
Remove redundant test logic to add file size to ExportResult because
the file size is already added to export result as part of an export
finishing.
PiperOrigin-RevId: 636499236
This change avoids a muxer deadlock when:
1. Sequence of items
2. First item has audio track that is shorter than video
3. Audio finishes, and muxer refuses to write more than 500ms of video
consecutively.
SequenceAssetLoader fails to progress to the second item. A muxer
deadlock is possible when the audio of the first item finishes,
audio end-of-stream is not propagated through AudioGraph, and muxer blocks
video, preventing SequenceAssetLoader to move to the next item in sequence.
By triggering silence generation early as soon as audio EOS is
encountered, we ensure SequenceAssetLoader can progress to the next item.
PiperOrigin-RevId: 636179966
This was previously only set on images because it was not ignored on
other media types. This parameter was made no-op for non-images in
7b2a1b4443.
PiperOrigin-RevId: 636078142
Fragment shaders in OpenGL ES shader language aren't guaranteed
to support highp, required to correctly represent pixel coordinates
inside large images (e.g. 1920x1080).
This change moves coordinate mirroring for images out of fragment shader.
Fixes http://Issue: androidx/media#1331
PiperOrigin-RevId: 635732208
If MediaCodec allocates passes an image buffer with a cropped region,
SurfaceTexture.getTransformMatrix will cut off 2 pixels from each dimensions.
The resulting videos will appear a little stretched.
This patch inspects the SurfaceTexture transform matrix, and guesses what the
unscaled transform matrix should be.
Behind experimentalAdjustSurfaceTextureTransformationMatrix flag
PiperOrigin-RevId: 635721267
For each item, AudioGraphInput now pads the input audio track with silence
to the duration given in onMediaItemChanged.
Possibly resolves Issue: androidx/media#921 .
PiperOrigin-RevId: 634753721
ExoPlayer sometimes calls AudioSink.playToEndOfStream before configuring
the sink. Before this CL, the composition player was failing if this
happened.
PiperOrigin-RevId: 634306592
This class is not ready for production app usage yet, so it is still
marked `@RestrictTo(LIBRARY_GROUP)` for now. Apps can experiment with it
in a non-prod context by suppressing the associated lint error.
* Issue: androidx/media#1014
* Issue: androidx/media#1185
* Issue: androidx/media#816
PiperOrigin-RevId: 633921353
This annotation is only needed on public classes.
This change also removes the `/* package */` comment from some `public`
classes.
PiperOrigin-RevId: 633864544
This treats heic as a separate mimetype to heif (even though heic files are a subset of heif files). This is in line with other platform classes like android.content.ContentResolver
https://developer.android.com/media/platform/supported-formats#image-formats was updated to include avif support or API level 34, so added this MimeType as well and updated our associated util.
solves Issue: androidx/media#1373
PiperOrigin-RevId: 633616214
Add ctts box implementation to handle muxing B-frame videos.
Add method convertPresentationTimestampsToCompositionOffset to
provide sample offsets. Return empty ctts box in case of video
does not contain B-frame. Add ctts box to MoovStructure to handle
muxing the video containing B-frames.
PiperOrigin-RevId: 633537106
Also add documentation that suggests to use them in
PriorityTaskManager and adjust codec priorities in
Transformer's DefaultDe/EncoderFactory accordingly.
PiperOrigin-RevId: 633272667
The process crashes unexpectedly on these devices.
The new changes skips running these tests instead of marking them pass, when
the device needs to be skipped.
PiperOrigin-RevId: 633183638
Two devices were producing B-frames earlier and were causing
frame mismatch. So we had added a workaround for them.
Those devices does not produce B-frames now
after disabling high profile on them, so we don't need workaround now.
PiperOrigin-RevId: 633127755
Use different textures in calls to queueInputTexture(). Allows the texture to be deleted one it is used in transformer and effect.
PiperOrigin-RevId: 632430866
allows apps to use the logic statically before running transformer.
Also ensure the mime type reported is the mime type outputted
PiperOrigin-RevId: 631811763
The Javadoc is stating that the image duration is ignored for non-image
input but this is incorrect (for example, it affects seeking
performance). The Javadoc will be updated in another CL.
PiperOrigin-RevId: 631730980
After this CL, DVFP waits for flushing until all frames registered previously
arrives.
Previously, ETM records the difference between the number of registered frames,
and the number of frames arrivd on the SurfaceTexture, when flushing. (Note
that ETM is flushed the last in the chain, as flushing is done backwards from
FinalShaderProgramWrapper). ETM then waits until the number of frames arrive
after flush.
The normal flow is, MediaCodecVideoRenderer (MCVR) registers a new decoded
frame, in `processOutputBuffer()` to DVFP, MCVR call `codec.releaseOutputBuffer()`
to have MediaCodec render the frame, and then the frame arrives in DVFP's ETM.
However there might be a discrepancy. When registering the frame, ETM records
the frame on the calling thread, ~instantly. Later when the rendered frame
arrive, ETM records a frame is available on the task executor thread (or
commonly known as the GL thread). More specifically, when a frame arrives
in `onFrameAvailableListener`, ETM posts all subsequent processing to
the task executor. When seeking, the task executor is flushed as the first
step. It might be a frame that has already arrived on ETM, and the processing
of such frame has already been queued into the task executor; only to be
flushed as a result of flushing the task executor. If this happens, the frame
is considered to be never have arrived. This causes a freeze on the app,
because ETM'll wait until this frame arrives to declare flushing has completed.
PiperOrigin-RevId: 631524332
AudioGraphInput now accepts a range of inputs, as long as the effects
provided modify the audio to be int 16.
As part of this, add the workaround to DefaultCodec to ensure pcm
encoding is correct, and remove parameterized tests that are not valid.
PiperOrigin-RevId: 631404152
The second stage of the changes remove the conversion to linear colors in the SDR effects pipeline by default.
also resolves Issue: androidx/media#1050
PiperOrigin-RevId: 630108296
Also adds first frame rendered test for playing back compositions.
- This test checks the output pixels using an `ImageReader` to retrieve the
output bitmap
PiperOrigin-RevId: 630100817
Part of a two stage change to remove the conversion to linear colors in the SDR effects pipeline by default. Changes the boolean to an intdef, introducing a third option that gets all sdr input into the same colorspace.
This is a planned API breaking change, but this change should not change the behavior of the pipeline.
PiperOrigin-RevId: 629013747
This ensures that `C.TIME_UNSET` is more clear in dump files. Some of
these call-sites will **never** pass `C.TIME_UNSET`, but it seems
clearest to always use `addTime` and maybe it will ensure when these
sites are copied in future, `addTime` will be used in the new location
too.
PiperOrigin-RevId: 628363183
Add ctts box implementation to handle muxing B-frame videos.
Add method convertPresentationTimestampsToCompositionOffset to
provide sample offsets. Return empty ctts box in case of video
does not contain B-frame. Add ctts box to MoovStructure to handle
muxing the video containing B-frames.
PiperOrigin-RevId: 628346257
also makes the setter more flexible by ignoring the value of the setter when the output is hdr rather than throwing (since all HDR content must be have a linear color space)
PiperOrigin-RevId: 627388436
Instead of initializing the video sink outside the renderer with an
empty format for composition preview, we initialize it in the renderer
with the input format for video.
PiperOrigin-RevId: 627313708
Before this CL, if all the video (resp. audio) samples were fed before
the audio (resp. video) output format was reported, the
SequenceAssetLoader was incorrectly reporting a MediaItem change with a
null format.
PiperOrigin-RevId: 625978268
Transformer's input shouldn't be constrained to the number of playable audio channels on the current device because the media may be edited (to mix channels for example) or encoded for playback on another device (a server for example).
PiperOrigin-RevId: 625604243
The raw audio decoder's output audio format is stereo when the number of input
channels is (for example) 10 channels. Add a temporary workaround that uses the
input channel count for raw audio. This code should be removed in future when
we bypass the decoder entirely for raw audio.
Tested manually on a WAVE file with 18 audio channels.
PiperOrigin-RevId: 625307624
Before this CL, currentInputBufferBeingOutput was set to null without
adding the buffer to the queue of available buffers, which was making
this buffer unusable. After multiple seeks, playback was stuck because
the AudioGraphInput had no input buffer left.
PiperOrigin-RevId: 624943271
After this CL, Transformer will throw if the clipping start and end
positions are the same because MediaMuxer doesn't support writing a
file with no samples. This should work once we default to the in-app
muxer.
Issue: androidx/media#1242
PiperOrigin-RevId: 624861950
Reduces flakiness of tests that assert on PCM audio. Tests now have to
clearly choose how they want the capturing muxer to handle pcm audio.
Note that the only dump files that have changed are those that deal
with PCM audio (.wav, sowt, twos, silence). Because of the continuous
nature of PCM, timestamps are not part of the dump.
PiperOrigin-RevId: 623155302
