Re: [Dri-devel] pageflipping

[Jens O. <je...@tu...>]

Alan Hourihane wrote:
> On Tue, Sep 24, 2002 at 02:49:37 +0100, Alan Hourihane wrote:
>=20
>>On Tue, Sep 24, 2002 at 02:19:24 +0100, Keith Whitwell wrote:
>>
>>>Michel D=E4nzer wrote:
>>>
>>>>On Die, 2002-09-24 at 13:11, Keith Whitwell wrote:=20
>>>>
>>>>
>>>>>I've been thinking again about pageflipping & realized I can solve t=
he=20
>>>>>remaining few problems if I can tweak the behaviour of the 2d driver=
=20
>>>>>slightly.
>>>>>
>>>>>At the moment, the 2d driver always draws to buffer zero (the old fr=
ont=20
>>>>>buffer), and then at some point in the future, the dirty parts are=20
>>>>>blitted to buffer one (the old back buffer).
>>>>>
>>>>>However, this can be incorrect in a couple of circumstances, particu=
larly=20
>>>>>when the dirty regions (or even the drawing itself) overlaps with th=
e 3d=20
>>>>>window.
>>>>>
>>>>>I think all my problems can be solved if I do two things:
>>>>>	1) always have X draw to the *current* front buffer (buffer zero or=
=20
>>>>>	buffer one, depending)
>>>>>	2) subtract the 3d window from the dirty regions before blitting to=
=20
>>>>>	the current back buffer.
>>>>>
>>>>>I can probably figure out 2) without to much effort.
>>>>>
>>>>>For 1), I can adjust the accelerated functions fairly easily to get =
them=20
>>>>>to draw to buffer zero/one as appropriate.
>>>>>
>>>>
>>>>How do you deal with offscreen access?
>>>
>>>They should remain exactly as they are:  I want to change the offset o=
f the=20
>>>front buffer - I would hope that the two are decoupled, maybe that's a=
=20
>>>mistake on my behalf.
>>>
>>>
>>>
>>>
>>>>>What about software fallbacks?  What value do these use to get a poi=
nter
>>>>>to the framebuffer, and how can I go about changing it?
>>>>>
>>>>
>>>>It's passed to xxxScreenInit(), and even if there's a single place wh=
ere
>>>>it's stored afterwards, I'm not sure relying on that is a good idea. =

>>>>
>>>>Doesn't 2) alone work?
>>>
>>>No, because you should be able to render to the same window with both =
X and=20
>>>GL and see the results onscreen.  If X always renders to buffer zero, =

>>>sometimes that won't work.
>>
>>In theory,
>>
>>this should work, but I've not tested it.
>>
>>	{
>>		PixmapPtr pspix;
>>
>>    		pspix =3D (*pScreen->GetScreenPixmap) (pScreen);
>>
>>		(*pScreen->ModifyPixmapHeader)(pspix, pScreen->width,
>>		    pScreen->height, pScreen->rootDepth,
>>		    BitsPerPixel(pScreen->rootDepth),
>>		    PixmapBytePad(pScrn->displayWidth, pScreen->rootDepth),
>>		    NEWFBPOINTER);
>>	}
>>
>>Basically fbScreenInit, passes the framebuffer pointer down to=20
>>miScreenInit, which ends up in miCreateScreenResources later where
>>the screen pixmap is setup. It essentially does the above too, to setup=

>>this up.
>=20
>=20
> In fact, scratch that, looking furthur shows that the miCreateScreenRes=
ources
> stores the Pixmap pointer in the devPrivate.
>=20
> I'll do some digging.

It would be good to get Page Swapping working in the server.  Not only=20
is this very valuable for 3D performance, but the X double buffer=20
extension should utilize this as well.

Perhaps we'll need to move this discussion to the xpert list, but first, =

I'll ask here.  Does anybody know of any applications that use DBE?  If=20
not, we should remove it.  If so, we need to fix it when the DRI is=20
enabled.  The GLX spec clearly states:

"All GLX rendering contexts share the same notion of which are front=20
buffers and which are back buffers for a given drawable.  This notion is =

also shared with the X double buffer extension (DBE)."

Currently, DBE buffers are managed as X Pixmaps and DRI drivers have not =

notion of DBE buffers.  Any application that used OpenGL and DBE=20
together would be broken, but I haven't seen any bug reports along this=20
line.  I have to conclude that no applications are using OpenGL and DBE=20
together.

Note:  DBE and DRI cooperation has been broken for a long time.  Keith's =

work on page flipping is *not* the start of the problem, but it does=20
present an opportunity for us to fix DBE, or dump it.

--=20
                                /\
          Jens Owen            /  \/\ _
   je...@tu...  /    \ \ \   Steamboat Springs, Colorado