[Mesa-dev] i965/radeonsi use STD430 packing of UBOs by default

[Marek Olšák]

On Mon, Aug 21, 2017 at 1:38 AM, Timothy Arceri <tarceri at itsqueeze.com> wrote:
>
>
> On 21/08/17 08:58, Marek Olšák wrote:
>>
>> On Mon, Aug 21, 2017 at 12:52 AM, Timothy Arceri <tarceri at itsqueeze.com>
>> wrote:
>>>
>>> On 21/08/17 03:25, Marek Olšák wrote:
>>>>
>>>>
>>>> On Thu, Aug 17, 2017 at 1:02 PM, Timothy Arceri <tarceri at itsqueeze.com>
>>>> wrote:
>>>>>
>>>>>
>>>>> Shared (the default) and packed layouts are decided by the
>>>>> implementation.
>>>>> Currently we just pack them using the std140 layout. This change makes
>>>>> it
>>>>> so
>>>>> we use the slightly more compact std430 layout on i965 and radeonsi.
>>>>>
>>>>> I doubt this will help many games, but it still seems worth
>>>>> implementing.
>>>>> I could only find shaders for a single game in my shader-db collection
>>>>> where STD140 layout wasn't explicitly defined for UBOs, and even there
>>>>> it was using vec4s so there would be no improvement.
>>>>
>>>>
>>>>
>>>> Why having a separate codepath that only 2 drivers will use If it
>>>> doesn't improve any app?
>>>
>>>
>>>
>>> I didn't say it doesn't improve any apps, presumably something uses the
>>> shared and packed layout provided by the spec. I just didn't seen it in
>>> my
>>> quick search of my shader-db collection.
>>>
>>> Ignoring how many apps use it, it's not much of a separate path it mostly
>>> just reuses the existing paths for SSBOs. Also if we can get SNB fixed up
>>> and add support to the remaining Gallium drivers that support ubos than
>>> we
>>> can drop the old paths entirely.
>>
>>
>> All Gallium drivers will never support it (unless you intent to add
>> support by yourself).
>>
>>>
>>> For radeonsi once I get LOAD working with my uniform packing series that
>>> would just leave immediates using the fetch_constant() code path.
>>
>>
>> What about Nine and VDPAU/VAAPI/OpenMAX also using CONST?
>>
>> Marek
>>
>
> I wrote the series based on Nicolai's feedback as a way to get more familiar
> with the LOAD implementation so I could use it with constant packing.
>
> Using LOAD for UBOs was suggested by Nicolai, the Intel guys are interested
> in enabling this feature. Can you please point out exactly what patches you
> have concern over and maybe we can focus on that?
>

I skimmed over the gallium patches and they looked OK to me. I just
wanted to understand the usefulness because it's a separate codepath
that we know for sure that most drivers won't use.

Marek