Alexandre Janniaux authored 3 years ago and Felix Paul Kühne committed 2 years ago

When the encoder is opened, it cannot change format anymore, and we
cannot change it behind its back.

Alexandre Janniaux authored 3 years ago and Felix Paul Kühne committed 2 years ago

This test scenario for the transcode checks that the error state is
correctly reported back to the stream output pipeline even if it
happened asynchronously.

Alexandre Janniaux authored 4 years ago and Felix Paul Kühne committed 2 years ago

Some decoders are pushing format update synchronously during their
activation but others have an asynchronous behaviour and may push it
before their activation function has finished, or even after the first
decode call when the decoder is asynchronously configured and ready.

Thus, previously, the decoder format from the decoder was not
necessarily known when opening the filters and the encoder. This
patchset shifts the initialization of previously mentioned objects to
the decoder update callback.

In addition, encoders will now need to get the final video context to
configure and it cannot open correctly without it, so the encoder test
typically makes no sense anymore for video. In this patch, since we
create the encoder only when the final format is finally supplied by the
decoder, we can directly create the encoder that will be used for
transcoding.

In addition, since format update would still happen at activation for
the synchronous decoders, the format and encoders will already be
available, effectively replacing the encoder test afterwards.

Alexandre Janniaux authored 3 years ago and Felix Paul Kühne committed 2 years ago

In future commits, we'll create the filter chain and encoder as soon as
the decoder outputs the format it will generate through
decoder_UpdateVideoOutput. It means that if the decoder doesn't report
any format, we won't have an encoder anymore.

Alexandre Janniaux authored 3 years ago and Felix Paul Kühne committed 2 years ago

If transcode failed, we want to forward the error to the stream output
pipeline, so we need to store it for later and to abort the transcoding
process.

Alexandre Janniaux authored 4 years ago and Felix Paul Kühne committed 2 years ago

They will be reused differently when refactoring the initialization
later at the format update site.

Alexandre Janniaux authored 4 years ago and Felix Paul Kühne committed 2 years ago

The object is needed to forward the elementary streams and data to the
sout stream components following the transcode.

Note that having the stream object here means that we must ensure we
don't call it after the transcode is supposed to be closed, so we must
ensure every asynchronous components are drained and destroyed before
leaving the close functions.

The constraint was already there though, given p_obj actually pointed to
the p_stream object too.

Alexandre Janniaux authored 4 years ago and Felix Paul Kühne committed 2 years ago

A unit-test for the transcoder code currently is hard to achieve, as it
involves decoders, encoders, filters, sout_stream_t object, and general
transcode code.

This commit introduces an integration test which mocks the extern
components (filters, decoders, encoders) used by the transcode pipeline
and starts this pipeline using the usual input item properties.

The main goals are to:
 - check whether no part of the stream is dropped.
 - check whether push model with video context is supported everywhere.
 - check the different kind of decoder behaviour, in particular those
   for which decode is asynchronous with the decoder_QueueVideo call,
   those for which Open() is asynchronous with the decoder_UpdateVideo*
   call and their synchronous variants.
 - check memory leaks and use-after-free in the pipeline (through asan).
 - check threading within the pipeline (through tsan).
 - check decoder, encoder and filter loading failure.
 - check format adaptation through filters before the encoder.
 - check format conversion (called final_conv_static) before the
   encoder, needed for encoders requesting a conversion.

Alexandre Janniaux authored 3 years ago and Felix Paul Kühne committed 2 years ago

Missing part from 5140b25e.

Rémi Denis-Courmont authored 2 years ago and Felix Paul Kühne committed 2 years ago

This code is useless as the window cannot be NULL, and it was already
optimised out.

Window can no longer be NULL here.

User pointer allows to read the data into a memory allocation made by
the application. In the context of VLC, this serves no purpose since
memory mappings from the V4L2 device node work just as well as anonymous
memory mappings.

Furthermore most V4L2 drivers do not even support user pointer mode, so
this was a useless and untested code path.

Now that buffers are tracked by pointers rather than structures, the
pointer itself can be used as a flag, with NULL denoting that the
buffer is in flight.

Memory synchronisation becomes implicit, with (re)queuing as the
release side and dequeuing as the acquire side.

To support dynamic buffer allocation later on, buffers need to be
allocated separately. Otherwise adding a buffer would potentially
change the address of all existing buffers, including in-flight ones.

Keep the pool lock while requeuing a buffer, so streaming cannot be
stopped in the interval.

As of the previous changeset, we no longer need a detailled bit-mask
with one distinct bit for each buffer.

For each buffer, track its in-flight flag independently of the other
buffers. There are no needs to fight for a single atomic bit-mask.

This flag is purely temporary and solely for clean bisection. It is
removed later on.

In the corner case that the very first buffer allocation fails, the
(empty) pool was never freed, since the buffer release callback was not
invoked even once.

This adds a proper reference count with one reference for the capture
process and one reference for each allocate buffer.

The array may be reallocated later on, so it ought to be allocated
separately from its container.

Alexandre Janniaux authored 3 years ago and Jean-Baptiste Kempf committed 2 years ago

vout_display_opengl code has separate prepare() and display(), which
might or might not match vout_display's prepare()/display() calls, now
that the opengl swap() is not done by the vout_display_opengl code.

Alexandre Janniaux authored 3 years ago and Jean-Baptiste Kempf committed 2 years ago

Now that we splitted the vlc_gl_Swap() call from the rendering, there's
no excuse for starting rendering sooner so that the result starts to be
available when executing display() and will almost be ready for
presentation.

is_dirty is there to indicate whether vlc_gl_MakeCurrent failed and thus
the frame could not be rendered, to avoid swapping undefined content.

Alexandre Janniaux authored 3 years ago and Jean-Baptiste Kempf committed 2 years ago

The display code is better suited to execute Swap() than the
vout_display_opengl object, since:
 - It actually knows whether it's an on-screen or off-screen
   implementation.
 - It could add code to render more things between swapping.
 - In particular, it could render, but discard the rendering instead of
   swapping depending on the time spent or GPU state/chrono.

Alexandre Janniaux authored 3 years ago and Jean-Baptiste Kempf committed 2 years ago

Since the vout display work is splitted between prepare() and display(),
the flush is useful to ensure everything has been submitted by the end
of prepare() so that the result can be used in display().

Alexandre Janniaux authored 3 years ago and Jean-Baptiste Kempf committed 2 years ago

The vlc_gl_t API doesn't require the context to be current to call
vlc_gl_Swap(), but EGL needs it. Allowing vlc_gl_Swap from being called
outside of vlc_gl_MakeCurrent/ReleaseCurrent will allow to remove the
MakeCurrent/ReleaseCurrent mechanism and have the OpenGL provider pull
for a frame rendering instead of the code waiting on draw commands for
OpenGL to be ready to render, which is functionnally the same, but will
allow implementations like macosx.m or caopengllayer.m to be a real
OpenGL implementation instead of dummy provider around a display module.

Thomas Guillem authored 2 years ago and Rémi Denis-Courmont committed 2 years ago

0 and INT64_MAX make the graph difficult to scale.

Thomas Guillem authored 2 years ago and Rémi Denis-Courmont committed 2 years ago

Use the same order than the clock API (system, then pts).

Thomas Guillem authored 2 years ago and Hugo Beauzée-Luyssen committed 2 years ago

Specially late/early events.

Thomas Guillem authored 2 years ago and Hugo Beauzée-Luyssen committed 2 years ago

Specially late and toolate.