[Intel-gfx] [PATCH i-g-t 1/3] lib/igt_fb: Allow specifiying object tiling when creating frame buffers

Tvrtko Ursulin tvrtko.ursulin at linux.intel.com
Fri Oct 16 06:06:00 PDT 2015

On 16/10/15 14:01, Ville Syrjälä wrote:
> On Fri, Oct 16, 2015 at 01:54:51PM +0100, Tvrtko Ursulin wrote:
>> On 16/10/15 13:29, Ville Syrjälä wrote:
>>> On Fri, Oct 16, 2015 at 01:19:35PM +0100, Tvrtko Ursulin wrote:
>>>> Hi,
>>>> On 16/10/15 13:03, Ville Syrjälä wrote:
>>>>> On Fri, Oct 16, 2015 at 11:59:47AM +0100, Tvrtko Ursulin wrote:
>>>>>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>>> Currently object tiling is inferred from the frame buffer modifier
>>>>>> and only for legacy X scanout.
>>>>>> It is useful to support overriding this selection for certain tests
>>>>>> so add the capability.
>>>>> So you want to set up the object tiling differently from the fb
>>>>> tiling? Why is that? And don't we reject it in the kernel? If we don't
>>>>> need a fence for scanout (ie. FBC or gen2/3) we could allow it I
>>>>> suppose, but not sure it it really helps with anything.
>>>> Hm, yes and no. Only different in a sense that currently igt_fb leaves
>>>> object tiling at linear regardless of the fb modifier tiling. (Apart for
>>>> the legacy X where it requires that they match.)
>>>> I needed a way of having Y tiled fb modifier and Y tiled object to hit a
>>>> warning in i915_gem_object_get_fence when only the rotated view exists.
>>> Is there a problem of just doing what X tiled did also for Y tiled?
>> What do you mean? In the kernel or igt_fb? If latter then it needs to be
>> able to have object tiling as linear by default since that is how fb
>> modifiers were intended to be used - decoupled from obj tiling. Is that
>> what you meant?
> Just setting obj to Y tiled when fb is Y tiled from igt_fb.

Right, so I explained that - we want to have Y tiled fb modifier + 
linear object in the majority of cases, _apart_ from this special test 
case which tries to hit a WARN_ON on !obj->map_and_fenceable. Which 
happens when obj tiling is Y, and fb tiling is Y, _and_ no normal VMA 
exists, just the rotate one.



More information about the Intel-gfx mailing list