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
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
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
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
We currently check certain properties about inputColorInfo twice: once in checkColors in the DVFP and once when creating the samplers in the defaultshaderprogam. The inputcolorinfo shouldn't change between these two components so removing the checks in the shader program to have one source of truth for inputColorInfo checks, which will make the code simpler when it comes to adding to color-related features (like the ones related to removing de-gamma-ing)
PiperOrigin-RevId: 626400960
Makes OverlayFrameAnchor works as described in the OverlaySettings documentation. Currently the code does the opposite e.g setting the anchor to (+1,-1) makes the code anchor to the top left rather than the bottom right.
PiperOrigin-RevId: 621585558
Otherwise, apps overriding BitmapOverlay.getVertexTransformation may not
realize this is being done, and may ask why bitmaps are flipped vertically from
what they expect
Reference: https://github.com/androidx/media/issues/1128
PiperOrigin-RevId: 621544348
nullness checks were removed a while ago from editing, so no need for them now.
Bit neater to have less code to parse too.
Also, removes RequiresNonNull
PiperOrigin-RevId: 616887502
This also makes it more consistent with newer testId tests that don't have nullness, so tests don't look inconsistent, and was pretty easy to do :P
PiperOrigin-RevId: 616856888
Confirmed that the HLG extension displays properly on Pixel 7 Pro, API 34 and Samsung S24, API 34. On these devices, this fixes the washed out HLG preview issue on API 34, where the PQ ext had a washed out look, and where the HLG ext is supported (this bug didn't occur on API 33, only 34).
More info on manual tests done to sanity-check:
* Test cases: Transformer debug SurfaceView and ExoPlayer.setVideoEffects
* Test inputs: HLG and PQ
* Test devices: Pixel 7 Pro (API 33 & 34), Samsung Galaxy S24 (API 34)
* Added debugging: Logging colorInfo used in GlUtil.createEglSurface
No regressions were seen. HLG extension is used more in API 34, and behavior stays the same on API 33. Only human-visible change without logging is that HLG content looks better on API 34, for Samsung S24 and Pixel 7 Pro.
PiperOrigin-RevId: 616131192
This new test is for `ExoPlayer.setVideoEffects()`. It plays the
one-second-long video, applies an overlay that prints the video frame timestamp
onto the frame, captures the output frame and compares the captured output
frame with golden.
PiperOrigin-RevId: 610781590
VideoFrameProcessor requires every input frame to be registered before they are
rendered to its input Surface. This CL adds the option to register the input
frame only once. This is useful for camera inputs where not all input frames
come with changing FrameInfo.
PiperOrigin-RevId: 606294894
The public APIs of these modules reference symbols in some of their
dependencies, so these should be API dependencies, not implementation:
> An API dependency is one that contains at least one type that is
> exposed in the library binary interface, often referred to as its ABI
> (Application Binary Interface).
https://docs.gradle.org/current/userguide/java_library_plugin.html#sec:java_library_recognizing_dependencies
Transformer also uses symbols from `lib-common`, but these are already
an API dep of `lib-exoplayer` so no need to duplicate that here.
PiperOrigin-RevId: 605660621
Previously, we missed the BT2020->BT709 color-space conversion.
A user-visible impact of this is that red and green channels used to be
undersaturated, but now are more correctly saturated.
PiperOrigin-RevId: 605411926
removing the useHdr parameter from queueInputBitmap() it suggests we support changing between HDR and SDR within a stream, which we don't support. instead, identifying whether to use HDR from the shaderprogram which is informed by the inputColorInfo when the stream is registered.
PiperOrigin-RevId: 603681736
Transformer export and ExoPlayer previewing both read inputColorInfo from
registerInputStream now, instead of maintaining a consistent input color
throughout multiple streams in a sequence.
Therefore, we can remove inputColor-related arguments and methods now.
PiperOrigin-RevId: 603423509
If nullness should result in an exception, we should throw as soon as possible,
so that stack traces are easier to follow.
Did a quick scan of media3.effect+transformer and this covers all uses of
peek that immediately throw there.
PiperOrigin-RevId: 603393366
VideoFrameProcessor treats BT601 and BT709 as roughly equivalent now, so we
shouldn't be making checks that assume BT709 <-> requires tone-mapping.
Also, the color transfer is a better determinant for tone-mapping than color range, so use just the transfer to determine if tone-mapping is required.
PiperOrigin-RevId: 603083100
This happens when using `ExoPlayer.setVideoEffects()`.
This CL also fixes the first frame not rendered problem, originally solved in
7e65cce967, but rolled back in 5056dfaa2b because the solution introduces
the flash that is observed in b/292111083.
Before media3 1.1 release, the output size of `VideoFrameProcessor` is not
reported to the app. This was changed later after introducing
`CompositingVideoSinkProvider`, where the video size after processing **is**
reported to the app. After this CL, the size is again, not reported.
PiperOrigin-RevId: 602345087
Many usages are needed to support other deprecations and some
can be replaced by the recommended direct alternative.
Also replace links to deprecated/redirected dev site
PiperOrigin-RevId: 601795998
since we now can support taking in the inputColor upon registering the stream, there is no need to hardcode the image input color anymore. We should now be able to support switching between texture and image input which we couldn't before, but this is untested and not necessary.
PiperOrigin-RevId: 601784149
This also fixes issue introduced by frames being released from a prior version of a
GlShaderProgram
Tested by seeking within a playlist with one SDR then one HDR video.
PiperOrigin-RevId: 599475959
Previously, 8f69bb0d9d updated external input (video input)
but not internal input (image/texture input). Update internal input as
well to match.
PiperOrigin-RevId: 599235813
7e65cce967 introduced a regression on ExoPlayer.setVideoEffects()
where there is flash on the screen after the first few frames are shown.
Before 7e65cce967, the first frames of the content are missed, until
MediaCodecVideoRenderer sends the onVideoSizeChanged() callback. The
first frame is processed but not shown, the onVideoSizeChanged() is
triggered and the renderer receives a video output resolution message as
a response from the UI.
7e65cce967 fixed the missed first frames by setting a surface on the
CompositingVideoSinkProvider before the provider is initialized, and as
as result:
- the first frames are rendered
- the MediaCodecVideoRenderer sends the the onVideoSizeChanged() after
frames are shown
- the UI sends a video output resolution change
- the MediaCodecVideoRenderer updates the CompositingVideoSinkProvider
which causes the flash.
The underlying problem is with onVideoSizeChanged() not being
triggered early enough, before the first frame is shown.
Because the flashing is a regression, this commit is reverting
7e65cce967 but keeps the test that was added as ignored, so that the
test is re-enabled as soon as we address the underlying issue.
PiperOrigin-RevId: 597814013
Partially addresses the following TODO, by simplifying the DefaultShaderProgram
API surface.
```
// TODO(b/274109008): Refactor DefaultShaderProgram to create a class just for sampling.
```
PiperOrigin-RevId: 597575575
Whenever the inputColorInfo updates, update the samplingGlShaderProgram.
Also, allow either SDR or gamma2.2 to be used for HDR->SDR tone-mapping
`outputColorInfo` request. This is required because we can't update the
`outputColorInfo`, but plan to always use gamma2.2 for `outputColorInfo` in the
future.
This allows VideoFrameProcessor to work as is for exoplayer previewing, but
only when not seeking. As we haven't plumbed the per-stream inputColorInfo from
ExoPlayer down to VFP.registerInputStream, follow-up CLs will be needed to
properly support previewing with changing inputColorInfo.
PiperOrigin-RevId: 596627890
Instead, for input videos, use the colorInfo provided by the extractor. Similarly, for input images, use sRGB, the only color currently in use.
Textures do still need the input ColorInfo provided though.
PiperOrigin-RevId: 592875967
Move `inputColorInfo` from `VideoFrameProcessor`'s `Factory.create` to `FrameInfo`,
input via `registerInputStream`.
Also, manually tested on exoplayer demo that setVideoEffects still works.
PiperOrigin-RevId: 591273545
This is because currently
1. Player sets a surfaceView to render to
2. Player intializes the renderer
3. MCVR initializes the VideoSinkProvider, by extension VideoGraph
But when 1 happens, MCVR doesn't set the surfaceView on the VideoGraph because
it's not initialized. Consequently after VideoGraph is initialized, it doesn't
have a surface to render to, and thus dropping the first a few frames.
Also adds a test for first frame to verify the correct first frame is rendered.
PiperOrigin-RevId: 591228174
Otherwise, there's a memory leak of ~30MB, as this is never released.
This likely used to be considered released as part of what now became
`intermediateGlShaderPrograms`, but its release was missed after we split
`finalShaderProgramWrapper` out from the larger glShaderProgram list.
PiperOrigin-RevId: 590954785
To prepare to move `inputColorInfo` from `VFP.Factory.create` to
`VFP.registerInputStream`, move all usage of `inputColorInfo` to be *after*
`registerInputStream`.
To do this, defer creation of `externalShaderProgram` instances, which require
`inputColorInfo`. However, we must still initialize `InputSwitcher` and OpenGL
ES 3.0 contexts in the VFP create, to create and request the input surface from
ExternalTextureManager.
PiperOrigin-RevId: 590937251
Despite GL 3.0 not being required on API 29+, it is experimentally
determined to always be supported on our testing devices, on API 29+.
That said, still fall back to OpenGL 2.0 if 3.0 is not supported,
just in case.
PiperOrigin-RevId: 590569772
Before, a translucent overlay over an opaque video would result in a
translucent output. This is not consistent with physical properties of light
(if putting a translucent object in front of an opaque object, you can't see
behind the opaque object).
Using the mixing properties from DefaultVideoCompositor.
PiperOrigin-RevId: 586636275
In `TextOverlay` and `DrawableOverlay`, treat `Bitmap` as a buffer, where we
allocate it rarely and reuse it as long as possible before making a new one.
In `BitmapOverlay`, avoid allocating GL textures too often as well.
Strongly reduces allocations and memory usage growth (saving ~100-150 MB on 4k60fps
at high end), at the cost of more code complexity and low-end using 70MB more, on
1/1 comparisons.
PiperOrigin-RevId: 585990602
In `ExternalTextureManager` fields are accessed from the GL thread and the class needs to be constructed on the GL thread.
Also visibly document threading requirement in the parent class.
PiperOrigin-RevId: 585941284
The issue that motivated adding this (frames unexpectedly being dropped by the
decoder) has been addressed, so we can turn off the logging to reduce
unnecessary allocations during transformation. We can easily turn on debug
logging in future as needed by setting `DebugTraceUtil.DEBUG = true`.
Also avoid allocations for string building at logging call sites by passing a
format string for extra info. Formatting the string now only happens when
debugging is turned on.
Tested manually by running transformations in the new state (DEBUG = false) and
with debugging turned on.
PiperOrigin-RevId: 585666349
Move the reflective loading of VideoFrameProcessor from
MediaCodecVideoRenderer to the CompositingVideoSinkProvider. This is so
that all reflective code lives in one place. The
CompositingVideoSinkProvider already has reflective code to load the
default PreviewingVideoGraph.Factory.
PiperOrigin-RevId: 584852547
In OverlayShaderProgram, this method is used quite a lot, and is the only method from Util.java in this file. Marginally reduce complexity by using a static import instead.
PiperOrigin-RevId: 584828455
Reduce short-lived allocations of potentially large objects, like Bitmap.
Unfortunately, this does make the TextureOverlay interface more messy though, requiring a way to signal whether the texture should be flipped vertically.
PiperOrigin-RevId: 584661400
Split CompositingVideoSinkProvider.VideoSinkImpl in two classes:
- VideoSinkImpl now only receives input from MediaCodecVideoRenderer and
forwards frames to its connected VideoFrameProcessor
- VideoFrameRenderControl takes composited frames out of the VideoGraph
and schedules the rendering of those.
- CompositingVideoSinkProvider connects VideoSinkImpl with
VideoFramesRenderer.
PiperOrigin-RevId: 584605078
Use renderengine's PQ to SDR tone-mapping implementation instead
of naive implementation from before.
This improves luminance on highlights, as seen in the test image.
PiperOrigin-RevId: 582318045
All methods in VideoFrameProcessor are expected to be called by the owning thread,
as far as I understand (vs. 10 threads each queuing frames/textures/streams, which
invalidates blocking done by registerInputStream and flush)
PiperOrigin-RevId: 582295240
Throws when calling flush when there's no active input, for example
before an input stream is registered or after all output streams have
ended.
PiperOrigin-RevId: 577165419
Much simpler than ExternalTextureManager flushing, because
ExternalTextureManager complexity is due to decoder race condition behavior,
which we won't see for Bitmaps due to a more well-defined interface.
This is needed to test texture output flushing.
PiperOrigin-RevId: 570896363
When running on emulators, ExternalTextureManager needs a longer timeout for
forcing EOS, but we didn't catch a device name running on blaze: `generic_x86`
PiperOrigin-RevId: 565513152
Organize logic a bit by splitting logic about how we draw using OpenGL onto a
texture, out from the larger class, which has lots of logic discussing how we select
frames and streams.
No functional change intended, but a few method calls are shuffled around to
simplify things.
PiperOrigin-RevId: 565426225
As discussed offline, this is important for users of the class, and not all users may choose to read method javadoc, so best to make sure it's visible by leaving it at the class-level.
PiperOrigin-RevId: 565057677
That said, only SDR is supported for now, so this will always throw if
HDR is input. This will also throw if different ColorInfo values are input
because color SDR mixing (ex. between sRGB and BT709) is not yet supported.
PiperOrigin-RevId: 563457729
Instead of invoking "undefined behavior" may occur when a class is used
incorrectly, which is clear but not very helpful to callers, specify what Apps
or users of a class should do when issues occur.
Apply this to Compositor as well, where this otherwise might seem unclear.
PiperOrigin-RevId: 563406219
Overlays may be overlaid over:
* In VideoFrameProcessor, videos or images (or texture input).
* In Compositor, other videos.
In Compositor, Overlays may consist of video, so it could be confusing
for videoFrameAnchor to contrast with overlayAnchor.
Also, rename overlayAnchor to overlayFrameAnchor, since it's modifying
the anchor in the overlay's frame, so this name seems slightly more precise.
PiperOrigin-RevId: 562004292
Implement VideoCompositor support of:
* Different input and output sizes
* CompositorSettings, to customize output size based on input texture sizes
* OverlaySettings, to place an input frame in an arbitrary position on
the output frame.
Also, refactor Overlay's matrix logic to make it more reusable between
Compositor and Overlays
PiperOrigin-RevId: 561931854
Split this logic out from `OverlayShaderProgram`, to share logic like
this with `VideoCompositor`.
Also, remove the unused `context` parameter from `OverlayShaderProgram`
`VideoCompositor` will reuse `OverlayMatrixProvider`'s storing of the matrices,
though it may generate the matrices differently, since `OverlayShaderProgram`
applies this in the fragment shader, whereas `VideoCompositor` applies this in the
vertex shader.
PiperOrigin-RevId: 560142164
The actual logic scales the alpha, instead of setting it, so rename
this to what it does.
We would also prefer alpha to be scaled here, to not lose alpha information
(for example for PNGs with transparent regions).
PiperOrigin-RevId: 560121708
The first call to method `registerInputStream` doesn't block.
Later successive calls to the method blocks until the previous register call
finishes.
PiperOrigin-RevId: 559694490
Replace Queue<Long> with LongArrayQueue which provides queue semantics
for long primitives. LongArrayQueue is forked from IntArrayQueue which
in turn was forked from Androidx CircularIntArray.
IntArrayQueue is deleted and we now use CircularIntArray directly from
Androidx Collection.
PiperOrigin-RevId: 559129744
In case BaseShaderProgram implementations have an expensive
configure() call, we can only call configure() when absolutely necessary,
aka when the input size changes.
To reduce scope, this doesn't reduce the amount of
configure() calls that may be made outside this class, in
DefaultFrameDroppingShaderProgram and
FinalShaderProgramWrapper.
PiperOrigin-RevId: 558754314
In addition to this being how Alpha should be handled, tests do fail under
API 29 without this CL. Unfortunately while these tests do fail under API 29
without this CL, we currently only run this test on API 33 emulators, so we
didn't catch this failure earlier (until compositor tests on mh failed on all
devices under API 29 and succeeded on all at or over 29).
PiperOrigin-RevId: 557781757
`BlankFrameProducer` was originally internal to
`DefaultVideoFrameProcessorVideoFrameRenderingTest`. We later copied it in
c221958889. This CL converges the usages.
PiperOrigin-RevId: 555953620
This effect can scale Alpha values.
This is needed for testing Compositor, and coincidentally was requested
by a partner in the past. This also will generally be more useful in
full multi-asset, where we may want to have one input have different alpha
values than another
Also, update fragment_shader_copy_es2.glsl to not throw away alpha values.
This means ThumbnailStripShaderProgram (the only place this glsl is used) will
consider input alpha and output alpha appropriately
PiperOrigin-RevId: 555915092
This adds an effect to change the speed of the video track. The speed
of the audio track can already be changed using a SonicAudioProcessor.
Issue: androidx/media#559
PiperOrigin-RevId: 555155418
When implementing pause and resume, overlay appears as black frame on resume, suggesting the texture isn't being remade properly due to lack of proper deletion.
PiperOrigin-RevId: 554832580
