[Mesa-dev] [PATCH] [rfc] radv: add initial prime support.

Dave Airlie airlied at gmail.com
Wed Nov 23 19:52:57 UTC 2016


On 23 November 2016 at 20:46, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
> On 23.11.2016 06:28, Dave Airlie wrote:
>>
>> From: Dave Airlie <airlied at redhat.com>
>>
>> This is kind of a gross hacks, but vulkan doesn't specify anything
>> but it would be nice to let people with prime systems at least
>> see some stuff rendering for now.
>>
>> This creates a linear shadow image in GART that gets blitted to at the
>> image transition.
>>
>> Now ideally:
>> this would use SDMA - but we want to use SDMA for transfer queues
>> maybe we don't expose a transfer queue on prime cards who knows.
>
>
> Why is there a problem here? What stops you from just having multiple SDMA
> queues?
>
>
>> we wouldn't have to add two pointers to every image, but my other
>> attempts at this were ugly.
>>
>> Is the image transition the proper place to hack this in? not
>> really sure anywhere else is appropriate.
>
>
> So I'm not too deep into Vulkan, but isn't conceptually the blit to shadow
> image a part of the present operation, not of the image transition? I'd
> somehow expect the present part to have its own queue (which could be SDMA)
> on which the blit happens.

Well in theory this has no place in Vulkan at all, since apps should
be doing it,
but we can all guess how that ends up.

Well doing it in an image transition allows you to put it in the proper place in
a command stream, but I might look into doing it in the present area instead.

Maybe I should implement transfer queues first.

Dave.


More information about the mesa-dev mailing list