Contrary to all other hardware decoders, vaapi can't work with a
hwaccel_context anymore. It requires a hw_device_ctx or a hw_frames_ctx
to work.

This first implementation use a hw_device_ctx and let avcodec handle its
frames pool. It is also possible to handle hw_frames_ctx ourself to get
more controls.

Note: Most of the code could be reused by other va modules if we decide
to a hw_device_ctx.

Fixes #25707

The future vaapi va module will create AVFrame from hwframes_ctx with a
frame->buf[0] set internally by avcodec. If this is the case, don't
override it and use opaque_ref (an user field) to keep the reference on
the VLC picture_t.

frame->buf[0] must be valid but doesn't have to contain data since the
data will be accessed via frame->data[] directly.

Furthermore, all va modules are only setting frame->data[3], therefore,
frame->data[0] was always NULL.

Provide the AVCodecContext and the full AVFrame. It will be used by the
new vaapi va module.

It's already set as a read-only pointer in vd->cfg before calling Open().

This allows using COM helpers and shorter API calls.

It's already built with C++ from the sensor and DirectComposition handlers.

This is a C99 thing.

Fatih Uzunoğlu authored 3 years ago and Jean-Baptiste Kempf committed 3 years ago

qmlRegisterUncreatableType already registers the type
to the meta object system.

Fatih Uzunoğlu authored 3 years ago and Jean-Baptiste Kempf committed 3 years ago

this change allows selectedProfileControlListChanged
signal to be emitted after the initial reload happens.

Fatih Uzunoğlu authored 3 years ago and Jean-Baptiste Kempf committed 3 years ago

With Qt 5.0, a new syntax for signal/slot handling was introduced. This
patch updates the related function call that tests if the signal is
connected to its Qt 5 way of doing.

Pierre Lamot authored 3 years ago and Jean-Baptiste Kempf committed 3 years ago

   this allows to avoid code duplication between C++ and QML

Alexandre Janniaux authored 3 years ago and Jean-Baptiste Kempf committed 3 years ago

The event reporting must happen one event at a time and never trigger
behaviour not expected by the underlying application. It uses a custom
runloop mode to ensure only the VLC aware code can use the runloop (and
thus usually main thread) when reporting an event. It was ensured by
calling CFRunLoopRunInMode until the runloop is stopped.

But there were multiple undetected flaws in this:
 - CFRunLoopRunInMode will terminate (with timeout or finish) whenever
   one event have been processed, so it needs to be called in loop with
   the correct invariant.
 - CFRunLoopRunInMode with timeout=0 will exit even sooner if there are
   not events available.
 - CFRunLoopStop will interrupt CFRunLoopInMode only if it's actually
   running so there's an asynchronous design flaw where we try to stop
   the event loop from another thread without syncing with this event
   loop.

This patch fixes those issues, though it adds some callback nightmares,
by adding a new level of dispatch to the main CFRunLoop, which will be
the one to call CFRunLoopStop on itself. Thus, the CFRunLoop must be in
a running state when CFRunLoopStop is called.

It also ensure that we don't leave the loop for another reason than
calling CFRunLoopStop, eg. Timeout or Finish result code.

Alexandre Janniaux authored 3 years ago and Jean-Baptiste Kempf committed 3 years ago

Switch the event queue used to a dedicated event queue which will create
its own thread, instead of using QOS parallel queues, which prevents
mixing with non-VLC tasks and ensure tasks are ordered.

Steve Lhomme authored 3 years ago and Jean-Baptiste Kempf committed 3 years ago

And remove a useless cast from void*

Steve Lhomme authored 3 years ago and Jean-Baptiste Kempf committed 3 years ago

That's the preferred output format for most codec and also with better DXGI
display support.