Re: new hyperz patch

[Roland S. <rsc...@hi...>]

Keith Whitwell wrote:
> Roland Scheidegger wrote:
> 
>> This is a new version of the hyperz patch. It should have a better
>>  chance of running on all cards.
> 
> 
> I get the feeling that hyperz was designed with a private z-buffer in
>  mind. It would be interesting to see if the X server can dynamically
>  un-hyperz the backbuffer when it detects a second context being
> created - have a look at the TransitionTo/From functions in
> radeon_dri.c - currently they are mainly concerned with page
> flipping.
In fact, that was already discussed briefly at irc. For now it just 
seemed more important to get it working on more cards and fix the 
rendering problems than to worry about "minor" issues like multiple 
rendering apps :). I did get clearing only the needed tiles working 
(minus some off-by-one issues currently) though at least on my rv250 
(with a 32x8 z pixel granularity however - for clearing the hardware 
seems to work on 8x2 4x4 "macro tiles", it looks like it might be even 
possible to get granularity down to 4x4 for clearing, at least on my card).

> Typically this is resolved by clearing by drawing two big triangles
> to cover the area, enabling only depth or stencil drawing as
> required.
That's what I was thinking too, but I'd certainly avoid it if possible - 
that's just basically switching hyperz off (well at least fast z-clear, 
z-buffer compression ought to be still possible). I just don't know if 
the hardware (could even be different for rv cards?) can do it or not. 
And even with the latest fglrx driver I was unsuccesful in measuring 
clear performance with clearspd, I've even converted it to SDL instead 
of GLUT (due to fullscreen problems) and fglrx still obviously never 
used fast z-clear - dri is now 80 times faster than fglrx, isn't that 
nice ;).
Though some optimization could be done in the driver (i.e. if only 
stencil buffer is cleared, but between the last z-buffer clear and this 
clear the z-buffer was never enabled for writing, then still hyperz 
clear could be used, though care would have to be taken so the right z 
clear value is still used and similar issues). It probably means that 
z/stencil clear can only be done with hyperz clear if the stencil write 
mask contains all stencil bits too. That's lots of possibilities for 
clear fallbacks...

Roland