[Mesa-stable] [Mesa-dev] [PATCH] st/mesa: use first image's dimensions when finalizing texture

[Nicolai Hähnle]

On 06.06.2016 23:58, Ilia Mirkin wrote:
> On Mon, Jun 6, 2016 at 5:37 PM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
>> On 05.06.2016 08:24, Ilia Mirkin wrote:
>>>
>>> In the case where we can't guess the base level size, just use the first
>>> image's dims. The width0/height0/depth0 on stObj may not have been set
>>> at this point. Observed in a trace that set up levels 2..9 of a 2d
>>> texture,
>>> and set the base level to 2, with height 1. This made the guess logic
>>> always bail.
>>
>>
>> So by coincidence, I've been digging around this code today as well, though
>> I think the issues we were looking at are unrelated.
>>
>> I'm a bit confused. As far as I understand the code, the allocated pipe
>> texture always has levels corresponding to the GL texture (ignoring texture
>> views, at least...). So in the described situation, BaseLevel == 2,
>> stObj->lastLevel == 9, and we allocate a 10-level texture object.
>>
>> The pipe texture dimensions should be the dimensions of a hypothetical level
>> 0, but with your change they will end up being the dimension of the GL level
>> 2, leading to a texture that is too small? How does this work?
>>
>> I agree though that the current code can't be correct either...
>
> Well, since it's *hypothetical*, I was kinda hoping everything would
> work out starting from firstImage and ignore the hypothetical higher
> levels. However perhaps that hope isn't rooted in reality, and is the
> reason why the test fails.

I mean, aren't the level sizes wrong?

E.g. the program sets GL texture level 2 to 128x1, level 3 to 64x1, etc. 
Now the pipe texture width0 should be 512 and the height0 should be 1, 
but will end up being 128 and 1, respectively.

>>
>> Perhaps guess_base_level_size shouldn't bail on dim == 1 in the first place?
>
> Well, it'll bail for *some* cases, so we might as well make it work. I
> think this was to avoid the situation where you end up over-guessing
> for a narrow image, such as this one, which only ever has height 1,
> but width up to 512. If we had started guessing at the 1x1 level, we'd
> come up with 512x512.

Yeah, I guess there is an argument in favor of bailing in this case when 
called from guess_and_alloc_texture.

But when called from st_finalize_texture, I think it should just do

width = width > 1 ? width << level : 1;

etc. Does that fix the test?

Nicolai