[virglrenderer-devel] Towards gles 3.1

[Dave Airlie]

On 11 June 2018 at 10:20, Dave Airlie <airlied at gmail.com> wrote:
> On 11 June 2018 at 09:28, Dave Airlie <airlied at gmail.com> wrote:
>> On 10 June 2018 at 20:30, Gert Wollny <gert.wollny at collabora.com> wrote:
>>> Am Samstag, den 09.06.2018, 06:23 +1000 schrieb Dave Airlie:
>>>> On 9 June 2018 at 01:11, Gert Wollny <gert.wollny at collabora.com>
>>>> wrote:
>>>> > Any comments are welcome, I'll probably start with this on Monday
>>>>
>>>> First up I expect for this we want to always be rebasing, like keep
>>>> a branch, but be ready for it to have pieces merged into master and
>>>> rebased the whole time. I expect we won't just merge the completed
>>>> branch, but will have to go back and make proper commits out of it,
>>>> like I've done with the tessellation work.
>>> +1
>>>
>>>>
>>>> I've been playing around with this stuff on the sidelines:
>>>> https://gitlab.freedesktop.org/airlied/virglrenderer/tree/gl-4.5
>>>> https://cgit.freedesktop.org/~airlied/mesa/log/?h=virgl-gl-4.5
>>>>
>>>> Is my rebase of my parts onto my current desktop GL 4.1 work on
>>>> master.
>>>>
>>>> I need to rebase Gurchetan's ssbo fixes on top as well.
>>> So you will do this? In this case I would look into some other work to
>>> be done (probably fixing segfailts when the deqp-gles31 is run).
>>>
>>
>> I'll try and drop a branch today with Gurchetan's fixes merged into mine
>> and maybe a bit of cleaner history.
>
> Okay I've pushed a gl-4.5 branch to my tree[1].

This branch is updated again, I've merged all the tess shader fixes I got from
CTS into that,

but more importantly I realised my ssbo design was bogus (after
reading Gurchetan's
patches and thinking about it a bit more). Unfortunately I'm not sure
there's a brillantly
optimal mapping from TGSI to GLSL here.

I ended up with a unsized uint array in std430 mode, and if tgsi
requests multi channel
load stores, then I write multiple reads/writes to the uint array.

I'm sure this will break in a bunch of other ways, but it at least
passes a bunch of
dEQP ssbo.layout* tests for me now.

Dave.