[PATCH 1/2] present: Ignore TILE_SPLIT tiling flags in page flip check.

[Mario Kleiner]

On 06/05/2015 03:10 PM, Alex Deucher wrote:
> On Fri, Jun 5, 2015 at 8:33 AM, Mario Kleiner
> <mario.kleiner.de at gmail.com> wrote:
>> Mismatched tile split bits between source pixmap
>> and target screen pixmap caused present_check_flip
>> to reject page flips on at least RV730 (Radeon HD 4670).
>> (Source pixmap had 0x1, Target screen pixmap had 0x6000001)
>>
>> Mask those tile split bits out to make the check pass and
>> page flipping under DRI3/Present work.
>>
>> Maybe we should mask out more bits?
>
> I'm not sure if it's safe to do that.  Tile split is part of the
> tiling configuration in the display registers on evergreen and newer.
> It's probably better to try and figure out why the surfaces end up
> with different tiling parameters and fix that.
>
> Alex

Alex, you're right. Just sent out a new patch which hopefully fixes it 
properly for pre-evergreen, at least it seems to do so in my testing.

While looking at it, i found one minor thing which i don't know if it 
matters, probably not, so just fyi:


eg_tile_split() in evergreen_accel.c and eg_tile_split_opp() in 
radeon_bo_helper.c of the ddx assign the "default" case in the the 
switch(tile_split) statements inconsistently, iow. eg_tile_split_opp() 
isn't the "inverse" of eg_tile_split():

eg_tile_split(): default: and 4096 bytes map to flag 6.
eg_tile_split_opp(): default and flag 4 map to 1024 bytes.

Mesa otoh does it consistently:

src/gallium/drivers/r600/evergreen_state.c:
eg_tile_split(): default and 1024 bytes map to flag 4.

src/gallium/winsys/radeon/drm/radeon_drm_bo.c:

eg_tile_split(): default and flag 4 map to 1024 bytes.
eg_tile_split_rev(): default and 1024 bytes map to flag 4.

So it seems the eg_tile_split() "default" case in the ddx 
evergreen_accel.c is inconsistent with other definitions.

-mario