new radeon tiling patch

[Roland Scheidegger]

Michel Dänzer wrote:
> On Tue, 2005-01-18 at 20:43 +0100, Roland Scheidegger wrote:
> 
>>>> The DRM could update the register in the vblank interrupt
>>>> handler?
> 
> 
> [...]
> 
> 
>> How would you do that in-kernel? There is vblank interrupt related
>> stuff (radeon_driver_vblank_wait for instance), but that only is
>> called when a user has requested a wait for vblank, as far as I can
>> see (which is all the time with dri clients).
> 
> 
> You'd have to do it in radeon_driver_irq_handler() or a bottom half
> (but I don't know if these will meet the timing requirements either,
> in particular the latter...).
What about the 2nd head? I guess it would be necessary to update its
offset at a different time, is there something like a 
RADEON_CRTC2_VBLANK_STAT bit?
Also, this looks like it would get really ugly. Use an ioctl to tell drm 
it needs to update the offsets, then it will update them in the irq 
handler? That would mean we need something different for dispatch_flip 
(since more flips can happen while waiting for vblanks, in that time the 
old offsets have to be used), and furthermore there needs to be some 
code to deal with disabled irqs.
Is this such a big issue? I'm happy to leave it broken.
> 
>> Also, if doing that in the drm, we'd need to mess with OFFSET_CNTL
>>  there too (i.e. messy calculation or another field in the sarea).
> 
> 
> You mean CRTC_OFFSET? I'm not sure the calculation is that big an 
> issue... it'll never happen more than a couple thousand times per
> second anyway. ;)
Well, both CRTC_OFFSET and CRTC_OFFSET_CNTL. It's not really the time 
which the calculation uses, but I just don't like additional ugly code.

> 
> 
>> I also figured maybe RADEON_WAIT_CRTC_PFLIP would stop the cp so
>> you could just put the offset_cntl and offset values in the ring
>> buffer after that wait command, so I tried hacking in some out_ring
>> commands in the DoAdjustFrame code, but that didn't work neither.
> 
> 
> Curious, what happened? I think AdjustFrame can be called 
> asynchronously?
I even tried to hack it into the drm in dispatch_flip (so new offsets 
now _only_ were written when pageflip was enabled and a 3d client was 
running...). Well, the most important issue with that is that 
RADEON_WAIT_CRTC_PFLIP doesn't actually wait for a pageflip! It made 
absolutely no difference. Maybe with FLIP_CNTL set, it now considers it 
as a pageflip whenever it hits a new line, dunno.

Roland