wl_surface_commit from different thread

Fri Feb 7 17:11:22 PST 2014

On Fri, Feb 7, 2014 at 12:57 PM, Prabhu S <prabhusundar at gmail.com> wrote:

> Please find my comments inline.
>
>
>
>
> On Fri, Feb 7, 2014 at 8:18 AM, Jason Ekstrand <jason at jlekstrand.net>wrote:
>
>> Hi Prabhu,
>> Could you be a little more specific as to what you are doing.  It sounds
>> like you are either writing a client or trying to write the client-side
>> wayland bits for a driver stack.  However, it's kind of hard from you
>> description to tell exactly what you're working on.  Nevertheless, I will
>> try to answer your question as best I can.
>>
> => Adding support to EGL for wayland
>
>>
>> While Weston itself is single-threaded, the libwayland client library can
>> handle multiple threads rather well.  Look into wl_event_queue which allows
>> you to manage what events get called on what thread.  Also, any request can
>> be called from any thread.  The only issue is in the synchronization that
>> your app may need to do (know when wl_surface.commit occurs relative to
>> wl_shell_surface.set_maximized for instance).  If you want to get a frame
>> completion event in another thread, simply call wl_surface.frame,
>> wl_surface.commit and then add the callback you got from wl_surface.frame
>> to the wl_event_queue for that thread.
>>
>
>> Also, I'm confused by the relationship between your eglSwapBuffers calls
>> and your wl_surface.commit calls.  eglSwapBuffers should be calling
>> wl_surface.commit before it returns.  Sometimes this means the underlying
>> graphics drivers pass sync fences or do other things to keep from doing a
>> full glFinish and stalling the GPU. Therefore, in most cases, you shouldn't
>> need to be calling wl_surface.commit manually.  Perhaps I'm just
>> misunderstanding something?
>>
>
> => I'm taking reference of dri2_swap_buffers_with_damage from
> http://cgit.freedesktop.org/mesa/mesa/tree/src/egl/drivers/dri2/platform_wayland.c.
> Assume 3D frame is complex and really takes more time for the 3D GPU to
> complete the frame, GPU will be still processing the frame after return
> from eglSwapBuffers. I will get notification of the frame completion in
> different thread. In my assumption calls following wl_surface.attach,
> wl_surface.damage, wl_surface.commit, wl_display.flush need to be called
> from in different thread. Can this possible by creating callback and
> setevent from different thread(frame completion thread). I assumg
> wl_surface.frame is the compositor frame completion event??
>

Ok, I think I am beginning to see what you mean.  First, there are a few
guarantees that your EGL implementation needs to provide to Wayland
clients/servers:

1) eglSwapBuffers (or eglSwapBuffersWithDamageEXT) must call
wl_display.attach, wl_display.damage, and wl_display.commit before it
returns.  This is so that the client can synchronize various bits of
surface state with the wl_display.commit request.  It doesn't matter
whether or not it happens in another thread, just that it happens by the
time eglSwapBuffers reutnrs.

2) As soon as a compositor gets a wl_surface.attach, it is free to pass
that buffer resource into eglCreateImageKHR with type
EGL_WAYLAND_BUFFER_WL, convert it to a texture, and start drawing with it
immediately.

Exactly how these tow constraints are met is up to your EGL
implementation.  On some particular graphics stacks, glFinish is called
inside of eglSwapBuffers to ensure that the graphics card has finished with
the buffer before passing it to the compositor.  As you have already
mentioned, this causes the GPU to stall and decreases performance.  The way
to solve this is to have some sort of a fence associated with each buffer.
Instead of eglSwapBuffers waiting for the buffer to be finished before
calling wl_surface.attach/damage/commit, it sets up a sync fence, calls
wl_surface.attach/damage/commit, and returns immediately.  Then it uses
that fence to serialize rendering commands somewhere further down the
stream.  This can mean that eglCreateImageKHR blocks waiting for the buffer
to be finished or that you pass the fence further down the line and block
in glBindTexture or even just use it to serialize commands to the GPU.  How
that fence is created, passed around, and waited upon is an internal detail
of your EGL implementation and is outside the scope of the core wayland
protocol.

A final note concerning the wl_surface.frame event.  This event has nothing
to do with your buffer or with something happening on the GPU as such.  It
exists for the sole purpose of telling the client that the compositor has
started to render using the currently attached buffer and that it can go
ahead and start rendering for it's next frame.  The only use your EGL
implementation should have for wl_surface.frame is to implement
eglSwapInterval(1) or similar.

I hope this helps,
--Jason Ekstrand

>
>>
>> I hope that helps,
>> --Jason Ekstrand
>>
>>
>> On Fri, Feb 7, 2014 at 8:03 AM, Prabhu S <prabhusundar at gmail.com> wrote:
>>
>>> Hello,
>>> eglSwapBuffers is a non blocking call. Normally, 3D GPU will be used
>>> both for content creation and composition, in that case everything will end
>>> up the GPU unit. So the end frame will be will be synchronized. (assuming 1
>>> GPU unit).
>>>
>>> I have a scenario, where 3D GPU being used for creating content and blit
>>> engine used for Weston composition. I'm getting frame completion from 3D
>>> engine in different thread (not in the same thread eglSwapBuffers calling
>>> thread). Currently eglSwapBuffers calling thread need to wait for the frame
>>> completion and then call wl_surface_commit to notify Wayland compositor.
>>> This takes out most of the performance since GPUs are not active all the
>>> time.
>>>
>>> wl_surface_commit cannot be called from different thread AFAIK. Some
>>> advise, suggestions would be appreciated.
>>>
>>> Thanks
>>> Prabhu
>>>
>>> _______________________________________________
>>> wayland-devel mailing list
>>> wayland-devel at lists.freedesktop.org
>>> http://lists.freedesktop.org/mailman/listinfo/wayland-devel
>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/wayland-devel/attachments/20140207/ce74f130/attachment.html>