The destruction order is important, but tricky, because the screen is
open/close by the decoder, but destroyed by scrcpy.c on the main thread.
Add assertions to guarantee that the screen is not destroyed before
being closed.
The video buffer is now an internal detail of the screen component.
Since the screen is plugged to the decoder via the frame sink trait, the
decoder does not access to the video buffer anymore.
The fact that the recorder uses a separate thread is an internal detail,
so the functions _start(), _stop() and _join() should not be exposed.
Instead, start the thread on _open() and _stop()+_join() on close().
This paves the way to expose the recorder as a packet sink trait.
The video buffer took ownership of the producer frame (so that it could
swap frames quickly).
In order to support multiple sinks plugged to the decoder, the decoded
frame must not be consumed by the display video buffer.
Therefore, move the producer and consumer frames out of the video
buffer, and use FFmpeg AVFrame refcounting to share ownership while
avoiding copies.
This flag forced the decoder to wait for the previous frame to be
consumed by the display.
It was initially implemented as a compilation flag for testing, not
intended to be exposed at runtime. But to remove ifdefs and to allow
users to test this flag easily, it had finally been exposed by commit
ebccb9f6cc.
In practice, it turned out to be useless: it had no practical impact,
and it did not solve or mitigate any performance issues causing frame
skipping.
But that added some complexity to the codebase: it required an
additional condition variable, and made video buffer calls possibly
blocking, which in turn required code to interrupt it on exit.
To prepare support for multiple sinks plugged to the decoder (display
and v4l2 for example), the blocking call used for pacing the decoder
output becomes unacceptable, so just remove this useless "feature".
The screen receives callbacks from the decoder, fed by the stream.
The decoder is run from the stream thread, so waiting for the end of
stream is sufficient to avoid possible use-after-destroy.
When --no-display was passed, screen_destroy() was called while
screen_init() was never called.
In practice, it did not crash because it just freed NULL pointers, but
it was still incorrect.
A skipped frame is detected when the producer offers a frame while the
current pending frame has not been consumed.
However, the producer (in practice the decoder) is not interested in the
fact that a frame has been skipped, only the consumer (the renderer) is.
Therefore, notify frame skip via a consumer callback. This allows to
manage the skipped and rendered frames count at the same place, and
remove fps_counter from decoder.
As soon as the stream is started, the video buffer could notify a new
frame available.
In order to pass this event to the screen without race condition, the
screen must be initialized before the screen is started.
Video buffer is a tool between a frame producer and a frame consumer.
For now, it is used between a decoder and a renderer, but in the future
another instance might be used to swscale decoded frames.
It makes sense to extract default values for bitrate and port range
(which are arbitrary and might be changed in the future).
However, the default values for "max size" and "lock video orientation"
are naturally unlimited/unlocked, and will never be changed. Extracting
these options just added complexity for no benefit, so hardcode them.
After the struct screen is initialized, the window, the renderer and the
texture are necessarily valid, so there is no need to check in
screen_destroy().
