Reason for revert:
Reverting again. The perf regression does not seem to be related to dropping frames.
Original issue's description:
> Reland of Split IncomingVideoStream into two implementations, with smoothing and without.
>
> Original issue's description:
>
> Split IncomingVideoStream into two implementations, with smoothing and without.
>
> This CL fixes an issue with the non-smoothing implementation where frames were delivered on the decoder thread. No-smoothing is now done in a separate class that uses a TaskQueue. The implementation may drop frames if the renderer doesn't keep up and it doesn't block the decoder thread.
>
> Further work done:
>
> * I added TODOs and documentation for VideoReceiveStream::OnFrame, where we today grab 5 locks.
>
> * I removed the Start/Stop methods from the IncomingVideoStream implementations. Now, when an instance is created, it should be considered to be "running" and when it is deleted, it's "not running". This saves on resources and also reduces the amount of locking required and I could remove one critical section altogether.
>
> * I changed the VideoStreamDecoder class to not depend on IncomingVideoStream but rather use the generic rtc::VideoSinkInterface<VideoFrame> interface. This means that any implementation of that interface can be used and the decoder can be made to just use the 'renderer' from the config. Once we do that, we can decouple the IncomingVideoStream implementations from the decoder and VideoReceiveStream implementations and leave it up to the application for how to do smoothing. The app can choose to use the Incoming* classes or roll its own (which may be preferable since applications often have their own scheduling mechanisms).
>
> * The non-smoothing IncomingVideoStream implementation currently allows only 1 outstanding pending frame. If we exceed that, the current frame won't be delivered to the renderer and instead we deliver the next one (since when this happens, the renderer is falling behind).
>
> * The lifetime of the VideoStreamDecoder instance is now bound to Start/Stop in VideoReceiveStream and not all of the lifetime of VideoReceiveStream.
>
> * Fixed VideoStreamDecoder to unregister callbacks in the dtor that were registered in the ctor. (this was open to a use-after-free regression)
>
> * Delay and callback pointers are now passed via the ctors to the IncomingVideoStream classes. The thread primitives in the IncomingVideoStream classes are also constructed/destructed at the same time as the owning object, which allowed me to remove one more lock.
>
> * Removed code in the VideoStreamDecoder that could overwrite the VideoReceiveStream render delay with a fixed value of 10ms on construction. This wasn't a problem with the previous implementation (it would be now though) but seemed to me like the wrong place to be setting that value.
>
> * Made the render delay value in VideoRenderFrames, const.
>
> BUG=chromium:620232
> TBR=mflodman
>
> Committed: https://crrev.com/e03f8787377bbc03a4e00184bb14b7561b108cbb
> Cr-Commit-Position: refs/heads/master@{#13175}
TBR=mflodman@webrtc.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=chromium:620232
Review-Url: https://codereview.webrtc.org/2071093002
Cr-Commit-Position: refs/heads/master@{#13176}
Original issue's description:
Split IncomingVideoStream into two implementations, with smoothing and without.
This CL fixes an issue with the non-smoothing implementation where frames were delivered on the decoder thread. No-smoothing is now done in a separate class that uses a TaskQueue. The implementation may drop frames if the renderer doesn't keep up and it doesn't block the decoder thread.
Further work done:
* I added TODOs and documentation for VideoReceiveStream::OnFrame, where we today grab 5 locks.
* I removed the Start/Stop methods from the IncomingVideoStream implementations. Now, when an instance is created, it should be considered to be "running" and when it is deleted, it's "not running". This saves on resources and also reduces the amount of locking required and I could remove one critical section altogether.
* I changed the VideoStreamDecoder class to not depend on IncomingVideoStream but rather use the generic rtc::VideoSinkInterface<VideoFrame> interface. This means that any implementation of that interface can be used and the decoder can be made to just use the 'renderer' from the config. Once we do that, we can decouple the IncomingVideoStream implementations from the decoder and VideoReceiveStream implementations and leave it up to the application for how to do smoothing. The app can choose to use the Incoming* classes or roll its own (which may be preferable since applications often have their own scheduling mechanisms).
* The non-smoothing IncomingVideoStream implementation currently allows only 1 outstanding pending frame. If we exceed that, the current frame won't be delivered to the renderer and instead we deliver the next one (since when this happens, the renderer is falling behind).
* The lifetime of the VideoStreamDecoder instance is now bound to Start/Stop in VideoReceiveStream and not all of the lifetime of VideoReceiveStream.
* Fixed VideoStreamDecoder to unregister callbacks in the dtor that were registered in the ctor. (this was open to a use-after-free regression)
* Delay and callback pointers are now passed via the ctors to the IncomingVideoStream classes. The thread primitives in the IncomingVideoStream classes are also constructed/destructed at the same time as the owning object, which allowed me to remove one more lock.
* Removed code in the VideoStreamDecoder that could overwrite the VideoReceiveStream render delay with a fixed value of 10ms on construction. This wasn't a problem with the previous implementation (it would be now though) but seemed to me like the wrong place to be setting that value.
* Made the render delay value in VideoRenderFrames, const.
BUG=chromium:620232
TBR=mflodman
Review-Url: https://codereview.webrtc.org/2071473002
Cr-Commit-Position: refs/heads/master@{#13175}
This change reduces the number of times the Android hardware video
encoder is reconfigured when making an outgoing call. With this change,
the encoder should only be initialized once as opposed to the ~3 times
it happens currently.
Before the fix, the following sequence of events caused the extra
reconfigurations:
1. After the SetLocalDescription call, the WebRtcVideoSendStream is created.
All frames from the camera are dropped until the corresponding
VideoSendStream is created.
2. SetRemoteDescription() triggers the VideoSendStream creation. At
this point, the encoder is configured for the first time, with the
frame dimensions set to a low resolution default (176x144).
3. When the first video frame is received from the camera after the
VideoSendStreamIsCreated, the encoder is reconfigured to the correct
dimensions. If we are using the Android hardware encoder, the default
configuration is set to encode from a memory buffer (use_surface=false).
4. When the frame is passed down to the encoder in
androidmediaencoder_jni.cc EncodeOnCodecThread(), it may be stored in
a texture instead of a memory buffer. In this case, yet another
reconfiguration takes place to enable encoding from a texture.
5. Even if the resolution and texture flag were known at the start of
the call, there would be a reconfiguration involved if the camera is
rotated (such as when making a call from a phone in portrait orientation).
The reason for that is that at construction time, WebRtcVideoEngine2
sets the VideoSinkWants structure parameter to request frames rotated
by the source; the early frames will then arrive in portrait resolution.
When the remote description is finally set, if the rotation RTP extension
is supported by the remote receiver, the source is asked to provide
non-rotated frames. The very next frame will then arrive in landscape
resolution with a non-zero rotation value to be applied by the receiver.
Since the encoder was configured with the last (portrait) frame size,
it's going to need to be reconfigured again.
The fix makes the following changes:
1. WebRtcVideoSendStream::OnFrame() now caches the last seen frame
dimensions, and whether the frame was stored in a texture.
2. When the encoder is configured the first time
(WebRtcVideoSendStream::SetCodec()) - the last seen frame dimensions
are used instead of the default dimensions.
3. A flag that indicates if encoding is to be done from a texture has
been added to the webrtc::VideoStream and webrtc::VideoCodec structs,
and it's been wired up to be passed down all the way to the JNI code in
androidmediaencoder_jni.cc.
4. MediaCodecVideoEncoder::InitEncode is now reading the is_surface
flag from the VideoCodec structure instead of guessing the default as
false. This way we end up with the correct encoder configuration the
first time around.
5. WebRtcVideoSendStream now takes an optimistic guess and requests non-
rotated frames when the supported RtpExtensions list is not available.
This makes the "early" frames arrive non-rotated, and the cached dimensions
will be correct for the common case when the rotation extension is supported.
If the other side is an older endpoint which does not support rotation,
the encoder will have to be reconfigured - but it's better to penalize the
uncommon case rather than the common one.
Review-Url: https://codereview.webrtc.org/2067103002
Cr-Commit-Position: refs/heads/master@{#13173}
Removes the need to use VoEVolume::SetInputMute()/GetInputMute().
BUG=webrtc:4690
NOTRY=true
Review-Url: https://codereview.webrtc.org/2066973002
Cr-Commit-Position: refs/heads/master@{#13172}
An rtc::Buffer is used to write output of RBSP parsing, usually one byte
at a time. It turns out that container will then expand its capacity one
byte at a time, for each byte reallocating the whole buffer and copying
the contents over, turning this into an O(n^2) operation.
Fix is for now only to preallocate the container storage. Longer term, I
think we should mull over if we really need custom containers...
R=pbos@webrtc.orgTBR=mflodman@webrtc.org
BUG=
Review URL: https://codereview.webrtc.org/2073763002 .
Cr-Commit-Position: refs/heads/master@{#13167}
The function GetExecutablePath is a hack with poor portability. Delete
it and its caller GetTestFilePath. The latter was used in
videoframe_unittest.h, where it is replaced by
webrtc::test::ResourcePath.
Delete unused functions declared in base/testutils.h: ReadFile,
GetSiblingDirectory, GetGoogle3Directory, GetTalkDirectory,
CmpHelperFileEq, EXPECT_FILEEQ, ASSERT_FILEEQ.
Delete unused functions declared in media/base/testutils.h:
GetTestFilePath (see above), LoadPlanarYuvTestImage,
DumpPlanarYuvTestImage, ComputePSNR, ComputeSumSquareError.
The functions LoadPlanarYuvTestImage, DumpPlanarYuvTestImage were used
in yuvscaler_unittests.cc and planarfunctions_unittests.cc, under
webrtc/pc. However, these tests are never compiled or run, and appear
not to have been for the last few years, and are therefore deleted
rather than updated. It might make sense to check if libyuv have
comparable tests, and if not, resurrect them as part of libyuv
unittests.
BUG=
R=perkj@webrtc.org
Review URL: https://codereview.webrtc.org/2058043002 .
Cr-Commit-Position: refs/heads/master@{#13163}
And implement SampleRateHz in a bunch of mocks.
BUG=webrtc:5801
NOTRY=true
Review-Url: https://codereview.webrtc.org/2029543002
Cr-Commit-Position: refs/heads/master@{#13161}
The test sent a media packet, then verified it was sent by checking the
"last packet sent"'s ID. But the last packet sent may have been
a STUN packet that came *after* the media packet.
BUG=webrtc:5978
Review-Url: https://codereview.webrtc.org/2071573002
Cr-Commit-Position: refs/heads/master@{#13156}
This is a somewhat involved refactoring of this class. Here's an overview of the changes:
* FileWrapper can now be used as a regular class and instances allocated on the stack.
* The type now has support for move semantics and copy isn't allowed.
* New public ctor with FILE* that can be used instead of OpenFromFileHandle.
* New static Open() method. The intent of this is to allow opening a file and getting back a FileWrapper instance. Using this method instead of Create(), will allow us in the future to make the FILE* member pointer, to be const and simplify threading (get rid of the lock).
* Rename the Open() method to is_open() and make it inline.
* The FileWrapper interface is no longer a pure virtual interface. There's only one implementation so there's no need to go through a vtable for everything.
* Functionality offered by the class, is now reduced. No support for looping (not clear if that was actually useful to users of that flag), no need to implement the 'read_only_' functionality in the class, since file APIs implement that already, no support for *not* managing the file handle (this wasn't used). OpenFromFileHandle always "manages" the file.
* Delete the unused WriteText() method and don't support opening files in text mode. Text mode is only different on Windows and on Windows it translates \n to \r\n, which means that files such as log files, could have a slightly different format on Windows than other platforms. Besides, tools on Windows can handle UNIX line endings.
* Remove FileName(), change Trace code to manage its own path.
* Rename id_ member variable to file_.
* Removed the open_ member variable since the same functionality can be gotten from just checking the file pointer.
* Don't call CloseFile inside of Write. Write shouldn't be changing the state of the class beyond just attempting to write.
* Remove concept of looping from FileWrapper and never close inside of Read()
* Changed stream base classes to inherit from a common base class instead of both defining the Rewind method. Ultimately, Id' like to remove these interfaces and just have FileWrapper.
* Remove read_only param from OpenFromFileHandle
* Renamed size_in_bytes_ to position_, since it gets set to 0 when Rewind() is called (and the size actually does not change).
* Switch out rw lock for CriticalSection. The r/w lock was only used for reading when checking the open_ flag.
BUG=
Review-Url: https://codereview.webrtc.org/2054373002
Cr-Commit-Position: refs/heads/master@{#13155}
1. It moves calculation of the needed padding to VideoSendStream instead of ViEEncoder and only does it once per send Stream instead of every time the network estimate changes.
2. The maximum amount of padding sent was prior to this cl calculated and updated based on network estimate changes. However, it can only change based on encoder configuration changes and if send streams are added or removed. This cl change the VideoSendStream/VieEncoder to notify the BitrateAllocator of changes to the needed padding bitrate and for BitrateAllocator to notify Call of these changes.
3. Fixed an issue in the SendPacer where it could send a padding packet before sending a real packet. This caused the test EndToEndTest.RestartingSendStreamPreservesRtpStatesWithRtx to fail with these refactorings since the pacer suddenly could send a padding packet before the encoder had produced its first frame.
BUG=webrtc:5687
Review-Url: https://codereview.webrtc.org/1993113003
Cr-Commit-Position: refs/heads/master@{#13149}
Reason for revert:
Reverting after experiment. What seems to happen is that the thread can't be stopped when PostThreadMessage fails with this error code (or lack thereof).
Original issue's description:
> Attempt to figure out what the issue is on the Win10 FYI build bot in content_browsertests.
> Check for the case when PostThreadMessage returns false but GetLastError returns ERROR_SUCCESS.
>
> TBR=olka
> NOTRY=true
>
> Committed: https://crrev.com/e11041159d6e66fb25c2e045ad1558fc99d2d6cd
> Cr-Commit-Position: refs/heads/master@{#13143}
TBR=olka@webrtc.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
Review-Url: https://codereview.webrtc.org/2067013003
Cr-Commit-Position: refs/heads/master@{#13147}
Reason for revert:
Reverting while we track down the issue on the Win10 bot.
Original issue's description:
> Split IncomingVideoStream into two implementations, with smoothing and without.
>
> This CL fixes an issue with the non-smoothing implementation where frames were delivered on the decoder thread. No-smoothing is now done in a separate class that uses a TaskQueue. The implementation may drop frames if the renderer doesn't keep up and it doesn't block the decoder thread.
>
> Further work done:
>
> * I added TODOs and documentation for VideoReceiveStream::OnFrame, where we today grab 5 locks.
>
> * I removed the Start/Stop methods from the IncomingVideoStream implementations. Now, when an instance is created, it should be considered to be "running" and when it is deleted, it's "not running". This saves on resources and also reduces the amount of locking required and I could remove one critical section altogether.
>
> * I changed the VideoStreamDecoder class to not depend on IncomingVideoStream but rather use the generic rtc::VideoSinkInterface<VideoFrame> interface. This means that any implementation of that interface can be used and the decoder can be made to just use the 'renderer' from the config. Once we do that, we can decouple the IncomingVideoStream implementations from the decoder and VideoReceiveStream implementations and leave it up to the application for how to do smoothing. The app can choose to use the Incoming* classes or roll its own (which may be preferable since applications often have their own scheduling mechanisms).
>
> * The non-smoothing IncomingVideoStream implementation currently allows only 1 outstanding pending frame. If we exceed that, the current frame won't be delivered to the renderer and instead we deliver the next one (since when this happens, the renderer is falling behind).
>
> * The lifetime of the VideoStreamDecoder instance is now bound to Start/Stop in VideoReceiveStream and not all of the lifetime of VideoReceiveStream.
>
> * Fixed VideoStreamDecoder to unregister callbacks in the dtor that were registered in the ctor. (this was open to a use-after-free regression)
>
> * Delay and callback pointers are now passed via the ctors to the IncomingVideoStream classes. The thread primitives in the IncomingVideoStream classes are also constructed/destructed at the same time as the owning object, which allowed me to remove one more lock.
>
> * Removed code in the VideoStreamDecoder that could overwrite the VideoReceiveStream render delay with a fixed value of 10ms on construction. This wasn't a problem with the previous implementation (it would be now though) but seemed to me like the wrong place to be setting that value.
>
> * Made the render delay value in VideoRenderFrames, const.
>
> BUG=
>
> Committed: https://crrev.com/1c7eef652b0aa22d8ebb0bfe2b547094a794be22
> Cr-Commit-Position: refs/heads/master@{#13129}
TBR=mflodman@webrtc.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=
Review-Url: https://codereview.webrtc.org/2061363002
Cr-Commit-Position: refs/heads/master@{#13146}
EncodedFrameCallbackAdapter was used VideoSendStream and
VideoReceiveStream, but there is no reason to have it as these classes
can call EncodedFrameObserver directly.
Review-Url: https://codereview.webrtc.org/2068463004
Cr-Commit-Position: refs/heads/master@{#13145}
Check for the case when PostThreadMessage returns false but GetLastError returns ERROR_SUCCESS.
TBR=olka
NOTRY=true
Review-Url: https://codereview.webrtc.org/2063313003
Cr-Commit-Position: refs/heads/master@{#13143}
Current number of threads selection code does not work well
for Android builds - middle and low end devices are having hard time
encoding VGA and QVGA with just one thread.
Increase the amount of vp8 encoder threads for 180p and above resolution.
Also limit maximum number of thread to 3, since for 8 core devices
most of time 4 cores are idle when thermal throttling kicks in.
BUG=b/27946721
R=marpan@webrtc.org
Review URL: https://codereview.webrtc.org/2058753003 .
Cr-Commit-Position: refs/heads/master@{#13142}
The RtpReceiverObserverInterface is created.
The SignalFirstPacketReceived will be forwarded from BaseChannel to WebRtcSession.
WebRtcSession will forward SignalFirstAudioPacketReceived and SignalFirstVideoPacketReceived to the RtpReceiverInterface.
The application can listen to the Signal by implementing and registering a RtpReceiverObserver.
Review-Url: https://codereview.webrtc.org/1999853002
Cr-Commit-Position: refs/heads/master@{#13139}
It was already disabled for browsers by design, and for everyone else
because of a bug.
BUG=webrtc:5922
Review-Url: https://codereview.webrtc.org/2055493003
Cr-Commit-Position: refs/heads/master@{#13138}
It appears there the encode and send operation can happen over multiple
threads. Also, padding data itself may be sent on a different thread.
Remove thread checker and protect all data with crit_sect.
BUG=webrtc:5998
TBR=stefan@webrtc.org
Review-Url: https://codereview.webrtc.org/2066863002
Cr-Commit-Position: refs/heads/master@{#13137}
Added the plumbing necessary to get two different lists of codecs from
WebRtcVoiceEngine up to MediaSessionDescriptionFactory.
This should be the last step in this set of CLs. Once
https://codereview.webrtc.org/1991233004/ has landed, it's possible to
implement the ReceiveCodecs getter with the info from the
AudioDecoderFactory. The factory needs to be updated to actually
produce the correct list, as well.
BUG=webrtc:5805
Review-Url: https://codereview.webrtc.org/2013053002
Cr-Commit-Position: refs/heads/master@{#13131}
Camera1 tests are now separated from general CameraVideoCapturer tests.
Main motivation behind these changes is that Camera2 implementation can
be tested using the same tests.
CL also reduces code duplication on tests using textures.
BUG=webrtc:5519
Review-Url: https://codereview.webrtc.org/2024843002
Cr-Commit-Position: refs/heads/master@{#13130}
