We're going to do this in Mesa instead: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22771
Marek
On Fri, Apr 28, 2023 at 6:36 PM Marek Olšák <maraeo@xxxxxxxxx> wrote:
On Fri, Apr 28, 2023, 16:14 Joshua Ashton <joshua@xxxxxxxxx> wrote:I mean I would also like power and perf numbers for Vangogh given you referenced 10.3.
Generic "power consumption is better" isn't enough to convince me that this is the right call.Raphael and Mendocino have worse power consumption with retiled displayable DCC and modifiers, and that can also be due to how retiling is implemented for modifiers.Marek
- Joshie 🐸✨
On Friday, 28 April 2023, Marek Olšák <maraeo@xxxxxxxxx> wrote:
> I thought the same thing initially, but then realized that's not how modifiers were designed to work.
> Mesa should expose all modifiers it wants to allow for 3D and it doesn't care which ones are displayable.
> The kernel should expose all modifiers it wants to allow for display.
> With that, Mesa can still use theoretically displayable DCC, but it will only be used for anything that's not the display.
> We can, of course, disable it in Mesa instead to get the same effect.
> We would need perf numbers for dGPUs to be able to tell whether it's beneficial with the cost of DCC retiling.
> Marek
>
> On Fri, Apr 28, 2023, 12:11 Joshua Ashton <joshua@xxxxxxxxx> wrote:
>>
>> I really don't think the kernel isn't the right place to do this.
>> Is there any reason to not just disable it from the Mesa side?
>>
>> We can already disable displayable DCC there, so I don't see why you are even touching the kernel.
>> It makes it infinitely harder for anyone to evaluate perf and power tradeoffs if you disable it at this level.
>>
>> The whole power vs perf trade is also not a big deal on dGPUs compared to APUs. Probably needs a better heuristic either way to avoid regressing perf.
>>
>> - Joshie 🐸✨
>>
>> On 28 April 2023 10:47:17 BST, "Marek Olšák" <maraeo@xxxxxxxxx> wrote:
>>>
>>> Hi,
>>> It's attached for review.
>>>
>>> Thanks,
>>> Marek
>
