[Piglit] [RFC] Trying to fix dmabuf/yuv tests

[Rob Clark]

On Tue, Sep 6, 2016 at 4:53 PM, Kristian Høgsberg <hoegsberg at gmail.com> wrote:
> On Tue, Sep 6, 2016 at 1:32 PM, Rob Clark <robdclark at gmail.com> wrote:
>> On Tue, Sep 6, 2016 at 4:23 PM, Rob Clark <robdclark at gmail.com> wrote:
>>>>> @@ -212,28 +247,65 @@ piglit_gbm_buf_create(unsigned w, unsigned h, unsigned cpp,
>>>>>         void *dst_data;
>>>>>         void *map_data = NULL;
>>>>>         enum gbm_bo_format format;
>>>>> +       unsigned cpp = 0, src_stride;
>>>>> +       unsigned buf_w = w;
>>>>> +       unsigned buf_h = h;
>>>>>
>>>>> -       if (!gbm || h % 2)
>>>>> +       if (!gbm || h % 2 || w % 2)
>>>>>                 return false;
>>>>>
>>>>> -       /* It would be nice if we took in a fourcc instead of a cpp */
>>>>> -       switch (cpp) {
>>>>
>>>> Can we just switch over to always taking a fourcc? This cpp_or_fourcc is messy.
>>>
>>> I think it made a bit more sense when I started (and thought that this
>>> would be a smaller patch) to avoid extra churn..
>>>
>>> Maybe in the end since you have the U/V interleaved by row, we might
>>> not end up needing my patch.  But otherwise, at this point it wouldn't
>>> make the patch that much larger to fix up the handful of rgb tests
>>
>>
>> oh, no, spoke too soon, we still have an issue for UV plane when hw
>> has alignment restriction in pixels instead of bytes..
>
> We can just rearrange the input data for the yuv test to not be laid
> out like that. There was no specific reason that I did it that way.

So, fwiw, I was trying a less intrusive patch to keep the layout the
same, but use cpp=2 and w/=2 for allocating in the NV12 case.

But the problem with the row-interleaved YV12/YU12 cases is that the
2nd plane start position isn't aligned to the required pitch.

So I think I am back to the original idea of passing fourcc all the
way down to the dmabuf allocator (and re-arranging things so U/V
planes are not row-interleaved).  I'll go ahead and push the mesa
patch that adds the additional GBM formats that the original patch
depended on, and then clean things up to use fourcc everywhere
(instead of cpp_or_fourcc) and then resend.

BR,
-R