[Mesa-dev] [RFC] egl: Add DRI3 support to the EGL backend.

[Joonas Lahtinen]

Hi,

Comments below.

On ti, 2014-10-28 at 12:44 +0100, Axel Davy wrote:
> On 28/10/2014 10:19, Joonas Lahtinen wrote :
> > Hi,
> >
> > On pe, 2014-10-24 at 17:37 +0000, Emil Velikov wrote:
> >> Hi Joonas,
> >>
> >> On 22/10/14 18:17, Joonas Lahtinen wrote:
> >>> Hi,
> >>>
> >>> This patch introduced DRI3 support to the EGL backend.
> >>>
> >>> Patch is on top of current master. With the patch you can observe
> >>> reduced CPU stress when many glViewport calls are made.
> >>>
> >>> Notice that the DRI3 extension is only exposed if the DRI3 interface is
> >>> found working too instead of just existing. So you need to enable DRI3
> >>> from all three; Mesa, X driver and X server, to get the benefit.
> >>>
> >> It seems that you introduce yet another dri extension
> >> (DRIdri3Extension) which afaict should not be needed ?
> >>
> >> Am I missing something here or did you get carried away by looking at
> >> __DRIcoreExtension, __DRIdri2Extension and __DRIswrastExtension :P
> >>
> > The DRI3 (just like the USE_INVALIDATE extension) extension is needed
> > for the drivers to know that they do not need to invalidate buffers on
> > each glViewport call.
> >
> > One could use USE_INVALIDATE extension for same functionality, but I see
> > it would not be very good idea as the naming vs. context would be
> > confusing.
> >
> > DRI3 is exposed dynamically depending on if DRI3 is usable or not. The
> > USE_INVALIDATE is exposed as long as GLX is used. So that also makes
> > them bit different (in logic too, not just naming).
> >
> > Regards, Joonas
> >
> >> Thanks
> >> Emil
> >>
> >>
> Hi,
> 
> Looking at the patch it seems you are not really using DRI3,
> and you are just using the Present extension to listen when the window 
> is resized.
> 

That seems quite true. I'll have to refer to the original author and ask
what's with the DRI3 requirement.

Best I know it's been done following the the GLX DRI3 backend.

> A complete implementation, that would use several back buffers and use 
> the Present
> extension for swapbuffers instead of copying on the front buffer,
> would be preferable. (vsync/tear-free support + the advantages Present 
> will bring to
> compositors when they can handle Present events and avoid useless copies)
> 
> Since you use Present over the DRI2 backend, that explains why you have 
> to use hacks to use
> invalidate differently. With complete DRI3/Present implementation, you 
> wouldn't need these hacks.
> 
> Like for the egl Wayland backend you would use several back buffers and 
> invalidate only for a resize
> and after swapbuffers.
> 

Even with the current patch, invalidate only happens on resizes. So it
should be a substantial advantage. I don't see it as a hack, as it only
gets rid of the USE_INVALIDATE hack (intel_viewport call).

Removing the DRI3 test parts, and changing the naming to refer to
present extension, might be next steps. After I get to know why the DRI3
part was there originally.

Regards, Joonas

> Axel Davy