*** Original commit ***
Handle int instead of byte in SSIM.
The value of pixels are converted to integers at the point of use,
move this logic to the initialisation step.
This is a prerequisite step for testing SSIM calculation, which
will lead on to some SSIM improvements being verifiable.
Tested manually and SSIM values match for the same video
before and after this change.
***
PiperOrigin-RevId: 473259446
The value of pixels are converted to integers at the point of use,
move this logic to the initialisation step.
This is a prerequisite step for testing SSIM calculation, which
will lead on to some SSIM improvements being verifiable.
Tested manually and SSIM values match for the same video
before and after this change.
PiperOrigin-RevId: 473231779
Test that HDR editing succeeds on devices supporting HDR editing, tone maps on
devices supporting tone mapping, and throws exceptions on all other devices.
Also, only restrict HDR editing and tone mapping support to API 31+ only when
transcoding, not for all transformations.
PiperOrigin-RevId: 472958965
`requestCalculateSsim` more clearly represents the intention of the caller.
Also rephrase the javadoc to simplify it and make it more precise.
PiperOrigin-RevId: 465575578
Format.NO_VALUE is a placeholder value for an invalid @C.ColorTransfer, used
for example when the decoder doesn't support this transfer function.
When encountering this invalid value, interpret this as COLOR_TRANSFER_SDR.
Confirmed locally that an exception is thrown when transcoding on p4head, and no exception is thrown when transcoding with this CL.
PiperOrigin-RevId: 464135080
This will allow effects preview in ExoPlayer to use the
Effect and FrameProcessor interface (and the interfaces
they depend on) without depending on transformer or the
future effects module.
PiperOrigin-RevId: 464060047
Extract a FrameProcessor.Factory interface from GlEffectsFrameProcessor
and allow it to be customized using a setter on Transformer.Builder.
PiperOrigin-RevId: 463433438
Size requires API 21. Using Pair instead will allow effects to be
used from API 18 during previewing once they are moved out of
transformer.
PiperOrigin-RevId: 463206474
This is needed for applying effects to a playlist.
The effects are applied based on the presentation time of the
frame in its corresponding media item and the offset is added
back before encoding.
Each time the offset changes, end of input stream is signalled
to the texture processors. This is needed because the texture
processors can expect monotonically increasing timestamp within
the same input stream but when the offset changes, the timstamps
jump back to 0.
PiperOrigin-RevId: 462714966
* Introduced `useHdr` for `GlEffect#toGlTextureProcessor`, so
`TextureProcessor` implementations can decide how to handle HDR.
* Creating FP16 color textures for HDR input.
Tested via manual testing, adding a no-op GlEffectWrapper to the transformation to
force use of intermediate textures, adding a linear ramp to the fragment shader,
and trying to ascertain that there's a real reduction in posterization when
switching from 4-bit to 8-bit unsigned bytes, and again from 8-bit unsigned bytes
to 16-bit floating point.
PiperOrigin-RevId: 461613117
This may happen when a containers' color transfer incorrectly does not match
the video's color transfer.
An example of a file with such a mismatch is the current Transformer demo HDR10
sample file.
Manually tested by confirming that no errors are emitted for SDR and HLG sample
files, and that errors are emitted for our incorrect HDR10 sample file.
PiperOrigin-RevId: 461583532
The GlEffectsFrameProcessor that will be part of the effects module
uses the DebugViewProvider. So it does not make sense for it
to be an inner interface of Transformer.
PiperOrigin-RevId: 458014932
The FinalMatrixTransformationProcessorWrapper ensures that the
surface is only replaced when it is not being rendered to and vice
versa.
PiperOrigin-RevId: 458007639
This will be useful for downgrading to a lower resolution during
a slow preview and for processing slide-shows once sequential
multi-asset editing is supported.
PiperOrigin-RevId: 457017255
videoEncoderFormatUnsupported_completesWithError() has recently
been flaky on API 31 emulators on presubmit because a different
exception than the expected exception is thrown.
This disables it on those emulators to reduce testing noise
until the underlying problem is investigated and resolved.
PiperOrigin-RevId: 456765512
This change is just renaming. There is no functional change intended.
The FrameProcessor interface will be created in a follow-up.
PiperOrigin-RevId: 456741628
After this change GlEffects can use any GlTextureProcessor not just
SingleFrameGlTextureProcessor.
MediaPipeProcessor now implements GlTextureProcessor directly which
allows it to reuse MediaPipe's output texture for its output texture
and avoids an extra copy shader step.
PiperOrigin-RevId: 456530718
This change adds a SurfaceProvider interface which is necessary to
allow for texture processors whose output size becomes available
asynchronously in follow-ups.
VTSP's implementation of this interface wraps the encoder and provides
its input surface together with the output frame width, height, and
orientation as used for encoder configuration.
The FrameProcessorChain converts the output frames to the provided
orientation and resolution using a ScaleToFitTransformation and
Presentation replacing EncoderCompatibilityTransformation.
PiperOrigin-RevId: 455112598
Based on
https://developer.android.com/reference/android/media/MediaCodec#using-an-output-surface,
frame dropping behaviour depends on the target SDK version.
After this change transformer will only use
MediaFormat#KEY_ALLOW_FRAME_DROP if both the target and system SDK
version are at least 29 and default to its pre 29 behaviour where each
decoder output frame must be processed before a new one is rendered
to prevent frame dropping otherwise.
Also remove deprecated Transformer.Builder constructor without a
context and the context setter.
PiperOrigin-RevId: 453971097
Transformer always enabled glAssertionsEnabled, so there should
be no functional change.
ExoPlayer previously disabled glAssertionsEnabled, so GlUtil logged
GlExceptions instead of throwing them. The GlExceptions are now
caught and logged by the callers so that there should also be no
functional change overall.
This change also replaces EGLSurfaceTexture#GlException with
GlUtil#GlException.
PiperOrigin-RevId: 453963741
This removes the prior restriction of needing to remember not to crop and set aspect ratio in the same Presentation.Builder, and makes each class a bit more targeted.
This is partially made feasible by the past work to merge consecutive
MatrixTransformations into a single MatrixTransformationFrameProcessor, which
ensures that there's no loss in quality between successive MatrixTransformations.
PiperOrigin-RevId: 453660582
SingleFrameGlTextureProcessor is now an abstract class containing a
default implementation of the more flexible GlTextureProcessor interface
while still exposing the same simple abstract methods for single frame
processing it previously did.
FrameProcessorChain and GlEffect will be changed to use
GlTextureProcessor in follow-ups.
PiperOrigin-RevId: 453633000
Once the more advanced GlTextureProcessor interface exists,
it will be possible to change the output size of a GlTextureProcessor
between frames. To keep the re-configuration based on the frame sizes
minimal, things indepedent of the frame size, such as the GlProgram,
can be initialized in the constructor.
PiperOrigin-RevId: 451997584
ExoPlayer applies a large time offset to buffers so that, if the input has negative timestamps, generally buffers seen by the decoders should have positive timestamps. Modify how the offset is handled in `Transformer` so that decoders and encoders generally see positive timestamps, by leaving the offset on samples when reading them in the base renderer (remove the code that removed the offset), and then removing the offset when muxing. Also update the frame processor chain and slow motion flattening code to retain the existing behavior after this change (these both need original media presentation timestamps)
Tested via existing end-to-end tests and manually verified that the overlay frame processor shows the expected original media timestamps.
Aside: we don't need the same logic as ExoPlayer to track stream offsets across the decoder yet, because we don't try to handle stream changes during playback in single asset editing. (There is an edge case of multi-period DASH that may not work but I doubt anyone will use that as input to `Transformer` before we change the code to handle multi-asset properly.) In future we should try to refactor interaction with the decoder to use the same code for Transformer and ExoPlayer.
PiperOrigin-RevId: 451846055
Most devices won't support 8k decoding, so they'll skip this test entirely.
As the video is quite short, this test shouldn't be any longer than the nearby,
long-running 4k60 test.
PiperOrigin-RevId: 451423368
SSIM calculation requires the input and output dimensions to be identical.
For devices that can't encode the input dimensions, skip SSIM calculations and
log the cause. Only apply this on tests where the encoder may not support the
input file dimensions.
PiperOrigin-RevId: 451364904
Decode-only video frames (needed when the frame at / first frame after the
clipping start is not a key frame) need to be decoded but not passed to
the frame processor chain or encoder.
The clipping start offset needs to be removed from the frame timestamps
in the passthrough and video pipelines.
There are no changes needed for this in the audio pipeline, as it doesn't
use the input timestamps -- it uses its own timestamps derived from the
buffer sizes instead.
Also add demo option to try this out.
#minor-release
PiperOrigin-RevId: 451353609
Also update names of implementations to match design doc.
In follow-ups, SingleFrameGlTextureProcessor will become
an abstract implementation of a new GlTextureProcessor
interface.
Texture processor makes sense as it processes OpenGL textures.
The term frame processor will be used for something else in
follow-ups.
PiperOrigin-RevId: 451142085
There is a problem with the ImageReader formats used by the
SSIM helper that only occurs for Nexus 5 API 21, so as a workaround
we can skip the SSIM calculation on Nexus 5 API 21.
This skips just the SSIM calculation (by setting the value to
1.0 instead and logging). The tests still run when SSIM is skipped
so that we can detect other failures.
PiperOrigin-RevId: 450903183
When using a MatrixTransformationFrameProcessor per transformation
matrix, each frame processor's shader applies the matrix to the
vertices and clips the result to the NDC range when drawing the
output frame.
This change combines consecutive MatrixTransformations into a single
MatrixTransformationFrameProcessor by multiplying the individual
matrices while updating and clipping the visible polygon after
each matrix and mapping the resulting visible polygon back to the
input space so that its vertices and the combined transformation
matrix can be used in the shader.
PiperOrigin-RevId: 448521068
This listener replaces
FrameProcessorChain#getAndRethrowBackgroundExceptions.
The listener uses a new exception type FrameProcessingException
separate from TransformationException as the frame processing
components will be made reusable outside of transformer soon.
PiperOrigin-RevId: 447455746
ScaleToFitFrameProcessor, PresentationFrameProcessor,
and EncoderCompatibilityFrameProcessor now each implement
MatrixTransformation instead of wrapping
MatrixTransformationFrameProcessor.
PiperOrigin-RevId: 446480286
This change splits AdvancedFrameProcessor into 4 files:
- MatrixTransformationFrameProcessor for the GlFrameProcessor
implementation
- MatrixTransformation and GlMatrixTransformation for the GlEffect
specification
- MatrixUtils for the static matrix helpers
PiperOrigin-RevId: 446236384
To ensure frame processor operations operate on square pixels,
make the frame taller or wider for non-square input pixels.
In addition to automated tests, this was tested by changing the
inputFormat.pixelWidthHeightRatio in the TransformerVideoRenderer.
PiperOrigin-RevId: 444553517
This allows the actual bitmap to be saved, even if output
dimensions are different than expected. Otherwise, differing
output dimensions would throw an exception, preventing the bitmap
from being saved.
PiperOrigin-RevId: 444512210
This test should run on all devices from API 21 (the media uses Baseline
profile level 3.0 H.264) to give us coverage of the full pipeline (forcing
re-encoding) and SSIM calculation on all devices.
PiperOrigin-RevId: 443650002
Some devices under report their resolution support, like 2144 for 2160 in
H265, 1072 for 1080 in H264. This CL only takes care of these two cases,
- reporting 1920x1080 is supported when the device reports 1920x1072, and
- reporting 3840x2160 is supported when the device reports 3840x2144
PiperOrigin-RevId: 443095042
* Group what's now many related test PNGs by moving them to their own directory.
* Move bitmap references to files where they're used, as each bitmap is only
used once each, except the original bitmap.
PiperOrigin-RevId: 441485489
What a minimal implementation should include is now explained in the
interface javadoc while the method name reflects what the method does.
PiperOrigin-RevId: 441432059
The stream with the mp3 track is added because
- We only encode to AAC
- We only encode when the source track is not AAC
Now that we have a way to force encoding, we no longer need the mp3 track.
The test asset is kept for later parameterized testing.
PiperOrigin-RevId: 440876080
The MIME type was set to H265 to force transcoding. Now that we have an encoder
factory that forces encoding, switching back to H264 ensures the quality test
is conducted on more devices (those don't support H265 can be tested now).
However, H265 should be part of the quality test after we have proper mechanism
to skip test based on device capability.
PiperOrigin-RevId: 440132471
We don't currently have enough understanding of the correlation between a
specific SSIM score and video quality. Dropping to .90 to make most tests pass.
Especially when there's no discernible difference from the videos with .9 and
.95 SSIM.
PiperOrigin-RevId: 440047551
The encoder surface is no longer needed for the OpenGL setup and frame
processor initialization, as a placeholder surface is used instead. So
all of the setup can now be done in the factory method.
PiperOrigin-RevId: 438844450
This requires an additional nanos to micros conversion because
the SurfaceTexture uses nanos. But as the timestamps from the
MediaCodec decoder (propagated in DefaultCodec#releaseOutputBuffer) are
in microseconds no precision is lost here.
Also add test that checks output video duration.
PiperOrigin-RevId: 438010490
MAXIMUM_AVERAGE_PIXEL_ABSOLUTE_DIFFERENCE was copied from a test
class, but BitmapTestUtil isn't a test. So the javadoc needs
rewording to reflect that.
PiperOrigin-RevId: 438001833
* The AdvancedFrameProcessor calls use() in updateProgramAndDraw().
* The AdvancedFrameProcessor has the same input and output dimensions.
PiperOrigin-RevId: 437231350
Configuring the frame sizes between frame processors is now the
FrameProcessorChain's rather than the caller's responsibility.
The caller can getOutputSize() and override it for encoder fallback
in configure().
PiperOrigin-RevId: 437048436
The factory method is replaced by a public constructor and
configure() method which configures the input/output surfaces
and handles the OpenGL setup.
This is a prerequisite for removing the responsibility of the
caller to configureSizes() before creating the chain in a follow-up.
PiperOrigin-RevId: 437028882
This allows us to bypass many device-specific issues, that only occur when
decoding directly to an encoder surface, without OpenGL. This also allows us
to maintain fewer code branches, which require additional testing to verify
correctness.
PiperOrigin-RevId: 437003138
Allow apps to modify how frames are presented, via modifying resolution.
A follow-up CL will provide aspect ratio, cropping, etc.
PiperOrigin-RevId: 436963312
SEF is similar to frame editing, where the input and output videos are
not intended to match, so SSIM will not provide a reliable value. To
check SSIM correctly in future, we would need to provide
golden/expected video files to compare output against.
PiperOrigin-RevId: 436707240
The FrameProcessorChain manages a List<GlFrameProcessor>.
FrameProcessorChainDataProcessingTest now tests chaining ScaleToFit-
and AdvancedFrameProcessors.
PiperOrigin-RevId: 436468037
This currently causes the test to fail on Pixel 6 Pro running a recent S build
SQ1D.220205.004.
There is no need to test audio transcoding while we are measuring video
quality.
PiperOrigin-RevId: 435635314
ExternalCopyFrameProcessor's output dimensions match the input
size not the output size. So the intermediate texture size
should match the input size.
Also rename configureOutputDimensions to configureOutputSize.
PiperOrigin-RevId: 435058789
Use android.util.Size, whose naming is much easier to understand than Pair<Integer, Integer>, in both FrameProcessor and EncoderUtil.
PiperOrigin-RevId: 434813986
* Move auto-adjustments for transformation matrices from the
VideoTranscodingSamplePipeline constructor to the new
ScaleToFitFrameProcessor.
* Add GlFrameProcessor#getOutputDimensions() to allow for GlFrameProcessors with
different input and output dimensions. This is a prerequisite for
Presentation.
* Tested with unit tests (and manually just in case).
* A follow up CL will implement change the FrameProcessor input to be scale and
rotate values as requested by the user. This was kept out of this CL to
reduce CL review size. Presentation will also be implemented in a follow up
CL.
PiperOrigin-RevId: 434774854
If an OpenGL call blocks because the encoder's input surface is full,
this will now block the background thread while the main thread can
continue querying encoder output and free up encoder capacity until
it accepts more input unblocking the background thread.
PiperOrigin-RevId: 433283287
This will allow for easier customisation of the additional tasks
performed by the test runner, such as calculating metrics like SSIM.
PiperOrigin-RevId: 432434850
This test tests the same cases as the FrameEditorDataProcessingTest
as currently the main FrameEditor functionality is to apply a
transformation matrix using a TransformationFrameProcessor.
PiperOrigin-RevId: 431642066
We use SSIM to measure the transcoding quality between. SSIM is a widely used
tool that compares the luma channel between two images, and generates a score
from 0 to 1 that indicates "how similar" the two images are.
In `SsimHelper`, we decode the two videos, extract matching frames and
calculates the mean SSIM (SSIM averaged all matching frames) for both videos.
Matching frames are referred to as "comparisonFrame" in the CL, which is
selected based on the frame number and a user-set comparison interval.
For instance, if the interval is 7, then every seventh frames are compared.
We use MediaCodec/MediaExtractor to decode the video, and use ImageReader to
extract the decoded frame.
The SSIM calculation logic is a inspired by and modified from the CTS
[MSSIMMatcher](https://cs.android.com/android/platform/superproject/+/master:cts/tests/tests/uirendering/src/android/uirendering/cts/bitmapcomparers/MSSIMComparer.java;l=1?q=mssimcom)
that has some errors and extra features we don't need (like handling RGB
images).
Adds TranscodeQualityTest to ensure high quality transcoding.
PiperOrigin-RevId: 430951206
All (later customizable) GlFrameProcessors after the
ExternalCopyFrameProcessor receive their input from a normal OpenGL
texture not an external texture, so they won't need to worry about
the textureTransformMatrix.
PiperOrigin-RevId: 430165652
There are two major blockers to this test:
- H265 muxing is not available for API<24, so setting video mimeType to H265
will fail on those devices.
- AMR audio encoding is buggy on some device and it's not a widely used format.
The solution: use a video that is encoded with AVC/MP3, to ensure transcoding
to AVC/AAC.
PiperOrigin-RevId: 429648598
Re-enable tests that have no muxer support for timestamps going backwards.
Tests running on the B-frame sample will be added in a future commit.
#mse-bug-week
PiperOrigin-RevId: 429599177
Tested:
Verified that the additional information is available through
instrumentation tests, as well as via manual testing.
#mse-bug-week
PiperOrigin-RevId: 429038695
Fallback is only disabled for robolectric and instrumentation tests.
For MH tests, fallback is not disabled, as it may be needed due to
the broad range of devices available.
PiperOrigin-RevId: 426403167
Tested by confirming transformations still work and write to a output file in a
scoped-storage directory on a:
* Nexus 6P API 23 emulator
* Google Pixel 4 API 31 physical device
PiperOrigin-RevId: 425644266
- Add a checkbox in the demo app to enable experimental HDR editing.
- Add an `experimental_` method to `TransformationRequest` to enable HDR editing.
- Add fragment/vertex shaders for the experimental HDR pipeline. The main difference compared to the existing shaders is that we sample from the decoder in YUV rather than RGB (because the YUV -> RGB conversion in the graphics driver is not precisely defined, so we need to do this to get consistent results), which requires the use of ES 3, and then do a crude YUV -> RGB conversion in the shader (ignoring the input color primaries for now).
- When HDR editing is enabled, we force using `FrameEditor` (no passthrough) to avoid the need to select another edit operation, and use the new shaders. The `EGLContext` and `EGLSurface` also need to be set up differently for this path.
PiperOrigin-RevId: 425570639
TransformerTest sounds like a unit test for Transformer but these
tests test behaviour that involves multiple stages of the pipeline.
PiperOrigin-RevId: 425378369
If
a) the end of stream buffer arrives with a frame rather than an
empty buffer or
b) processDataV29() renders several decoder output buffers to the
FrameEditor's input Surface immediately before encountering the
EOS flag
these frames were previously stuck in the FrameEditor's input Surface
and never fed to the encoder.
PiperOrigin-RevId: 424898820
This test tests that all frames are processed when transcoding
video to a different sample MIME type (and that the transformation
completes successfully).
PiperOrigin-RevId: 424896014
Expected images are taken on emulators, so a larger acceptable
difference from expected images must be accepted on physical devices.
PiperOrigin-RevId: 421543441
Being able to see the output of the GL pipeline is useful for debugging. For
example, when we previously saw flakiness it would have been useful to be able
to tell quickly whether the output looked wrong without needing to run a
transformation to the end then inspect the output file, and when working on
support for HDR editing it's useful to be able to do manual testing on devices
that don't support HDR encoding (but do support decoding/processing it with
GL).
Also change the progress indicator to be linear as this looks better in the
demo app when shown next to the debug preview.
PiperOrigin-RevId: 414999491
Increase timeout for dequeueing a frame from the codec to reduce
flakiness. At a timeout of 2 seconds there was a 2/1000 flake rate and
at 3 seconds 0/1000. Set the timeout to 5 seconds to give plenty of
leeway.
PiperOrigin-RevId: 413946915
Allows a transformation matrix to be input into Transformer,
to apply vertex transformations like cropping, rotation,
and other transformations built into android.graphics.Matrix.
Not building out into a VertexTransformation class yet, as
that class structure wouldn't make sense until we can modify
resolution, per TODOs.
PiperOrigin-RevId: 413384409
The test extracts and decodes the first video frame in the test media, renders it to the frame editor's input surface and then processes data. It then reads back the output from the frame editor, converts it to a bitmap and then compares that with a 'golden' bitmap (which is just the same as the test media's first video frame).
PiperOrigin-RevId: 413131811
We need the filename of the output videos to be predictable, because
MobileHarness requires the exact filename to pull the file.
PiperOrigin-RevId: 412092347
The features supported by `TranscodingTransformer` are a
superset of those supported by `Transformer` after merging
the video renderers in
35c891a7bf.
This change removes `TranscodingTransformer` and adds its features to
`Transformer`.
PiperOrigin-RevId: 411072392
In the near future, we will be able to specify which tests run on this
target explicitly, at which point this ignore will be removed.
PiperOrigin-RevId: 410758785
The `GlUtil` wrappers for attributes/uniforms allocate buffers that are passed into OpenGL, and it seems that the distorted output was caused by these buffers being garbage collected.
The issue was difficult to reproduce manually. Add a test that does repeated transcodes in a loop (marked `@Ignore` for now because it runs for a long time and requires network access so may be flaky) to make it easier to run many transcodes and look for inconsistent output. For now the consistency check is just based on the file size which is very likely to change if the actual video frames change.
Verified that the test fails before, and passes once the attributes and uniforms are stored in fields as in this CL.
PiperOrigin-RevId: 410581706
