[PATCH weston] libweston: don't accumulate damage from transparent views

Fri May 4 13:26:13 UTC 2018

On Fri, 4 May 2018 12:09:55 +0000
"Ucan, Emre (ADITG/ESB)" <eucan at de.adit-jv.com> wrote:

> Hi Pekka,
> 
> Sorry for late response.

Hi Emre,

no worries!

> > -----Original Message-----
> > From: Pekka Paalanen [mailto:ppaalanen at gmail.com]
> > Sent: Donnerstag, 19. April 2018 12:36
> > To: Ucan, Emre (ADITG/ESB)
> > Cc: wayland-devel at lists.freedesktop.org
> > Subject: Re: [PATCH weston] libweston: don't accumulate damage from
> > transparent views
> > 
> > On Thu, 19 Apr 2018 09:26:59 +0000
> > "Ucan, Emre (ADITG/ESB)" <eucan at de.adit-jv.com> wrote:
> >   
> > > Hi Pekka,
> > >
> > > If we remove the view from scenegraph, application will be blocked.
> > > Because it is not getting any surface frame events. It is not OK to
> > > block unexpected applications. Especially if the application is
> > > sending output of a camera or digital TV, weston should always get
> > > latest buffer from the application.  
> > 
> > Hi Emre,
> > 
> > sounds like the application is broken if it gets literally blocked
> > by that.
> > 
> > Applications very much have to be able to deal with frame callbacks
> > not coming back in a timely fashion. This is what compositors use to
> > throttle down well-behaving clients. It's not blocking the app,
> > unless the app is badly written.
> > 
> > An invisible surface is a very good reason to try and throttle down
> > the client.
> >   
> > > You might say that camera application can use wl_display_sync
> > > instead wl_surface_frame.  
> > 
> > Not necessarily. It would certainly make the app greedy instead of
> > well-behaved in my opinion.
> > 
> > But, this would definitely make the latest buffer from the
> > application always available to the compositor. You shouldn't need
> > wl_display.sync even, the app can simply keep on committing new
> > buffers whenever it wants.  
> 
> You are right that this issue can be work around for an EGL
> application when we set swapInterval to 0, and send a new buffer
> every time internal state of the application requires to change its
> contents. But this does not fix the greediness problem. Because
> application state can change more often than display refresh rate.

The key with EGL is to never allow eglSwapBuffers() block. The app can
still ask for frame callbacks itself and throttle to those while
remaining otherwise responsive.

eglSwapBuffers() has been designed to be blocking in nature, i.e. it
has been designed to waste time. Any application that has more than one
window or needs to react to more than just Wayland input events or
needs to run a simulation (games!) would better set swapInterval to
zero. If they don't, and they also don't manually throttle to frame
callbacks, they are bound to waste time in eglSwapBuffers(), and this
is not only a Wayland issue. Wayland just allows us to make this bad
client design painfully obvious.

> > Or is the camera framerate an exact integer fraction of the display
> > refresh rate?  
> Ideally yes. Let's think that camera is sending buffer with 60 Hz and
> display has also the same refresh rate. But camera and display could
> be still out of sync, so that camera sends its buffer 8ms after
> Vsync. Therefore, it will miss the repaint window of weston.
> 
> It is also possible that camera buffers sometimes hit, sometimes miss
> the repaint window. This would cause visible video stutter. Because
> delay of displaying a buffer would swing between 0,5 Vsync to 1,5
> Vsync ( Repaint window is 0.5 Vsync). Therefore, it is not ok to send
> buffer every time when it is ready. We should send it directly after
> repaint, so that (hopefully) we won't miss the repaint window.

I don't think it changes much with respect to timings though. If the
client was posting frames as soon as it got it from the camera, the
compositor would be picking which ones make it to the screen. If the
client is throttled to frame callbacks, then yes, any frame the client
chooses will hit the screen, but now it is the client making the same
choice as the compositor in the former case. Except, the client needs
to make that decision earlier than the compositor.

Can the client here be smarter than the compositor to make a difference?
Given that we don't know about the deadline yet, discussed further
below.

> > I know there is a problem that when the surface becomes visible
> > again, it would take a frame cycle to have the client send an
> > updated buffer. This could be worked around on either side: the
> > compositor could be sending frame callbacks at a slow rate even if
> > the surface is not visible, or the client could be updating the
> > surface content at a slow rate even if it doesn't get frame
> > callbacks.  
> 
> But your proposal is also a hack.

Right, but what else could you do, when you do not know in advance when
the window will become visible and yet you want to minimize the number of
frames drawn and discarded?

The smartest option might be to postpone the actual show in the
compositor until the client has updated, but that's a trade-off between
latency from action to effect and a possible glitch (temporary outdated
content).

> Maybe better solution would be that
> we introduce a repaint_window event to wl_output interface, so that
> camera/video applications can synchronize themselves with this event
> instead of surface frame events. Compositor can send a timestamp and
> duration of repaint window to the clients.

(That proposal does not solve the above issue.)

Yeah, there has been talk about exposing the deadline timestamp for the
next vblank (if the compositor processes a commit before this time, it
will hit the refresh). It would help to minimize latency.

However, I would not like an event that gets sent on every output
refresh regardless of client updates. Instead, something like this
might be workable:
https://patchwork.freedesktop.org/patch/174555/

We still want to minimize client and server wake-ups when possible.

Mind, that you can also drive client updates by Presentation-time
feedback events. They are still lacking the deadline information
though, but they do carry the refresh period and phase.

The feedback events are intended for video players to accurately
predict the next time of presentation when they submit a frame as a
response to each frame callback event. When you predict the time the
next frame will become visible, you can choose and even interpolate a
frame for that exact time. After all, the app could be able to compute
intermediate frames, but the display hardware can only show discrete
frames at certain times.

> > One possibility would be to exclude the view from
> > weston_compositor::view_list which is used for both rendering and
> > input picking. This would happen in
> > weston_compositor_build_view_list(). It could also exclude
> > completely occluded views, but we can leave that for another time.
> > 
> > Excluding the view from the rendering list will avoid sending frame
> > callbacks. It will also affect Presentation feedback in the expected
> > way: the updates the compositor has decided to not show (e.g. by
> > setting view alpha to zero) will not result in a "presented" event.
> > 
> > Would you like to examine this path instead?  
> 
> You are right that it would be a better solution to not have
> invisible views in the compositor view list. Daniel also implemented
> to discard occluded views here: "
> https://patchwork.freedesktop.org/patch/202693/". I can reuse it. But
> my real intention for this patch was to mitigate the performance
> penalty of hiding a surface via using opacity instead of visibility.
> Technically, it is of course correct that invisible surfaces do not
> get any frame events. But we need some other solution to synchronize
> video applications with wayland compositors...

Yes. I think the best way to mitigate the performance penalty is to
completely exclude fully transparent surfaces from processing.

Is the synchronization of fully invisible video surfaces actually a
problem? There is the old content issue, but aside from that, it will
be in sync again after the next frame callback on becoming visible
again.

Thanks,
pq
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 833 bytes
Desc: OpenPGP digital signature
URL: <https://lists.freedesktop.org/archives/wayland-devel/attachments/20180504/19b42c98/attachment-0001.sig>