[Intel-gfx] [PATCH 1/2] drm: Introduce plane SIZE_HINTS property

[Ville Syrjälä]

On Tue, Feb 14, 2023 at 11:42:27AM +0200, Pekka Paalanen wrote:
> On Thu, 9 Feb 2023 13:51:05 +0200
> Pekka Paalanen <pekka.paalanen at collabora.com> wrote:
> 
> > Maybe we could refine this so that userspace uses the stride and height
> > implied by the caps for allocation, and then use the exact cursor image
> > size for AddFB2? And have drivers pick any size between those two they
> > can use. The kernel would need the userspace to promise that the
> > padding is always zero-initialized, so the driver can simply scan out
> > any area of the buffer it needs.
> > 
> > Then we don't need SIZE_HINTS.
> 
> Would there be any problem with this?
> 
> If this works, it would seem the superior solution to me, because
> userspace does not need to guess or test for the exact right size.
> Simply allocate at the CAP size, pad the cursor image with transparent
> pixels, and let the kernel scan out the optimal area.

No, the hardware cannot scan out a smaller area because the
stride will be wrong.

> 
> And if the kernel needs to do a pixel format conversion, it only needs
> to do the optimal minimum amount of work.

Involving the CPU (or GPU I suppose but that could involve big extra
latencies) in the kernel to massage the pixels around every time
seems extremely sub-optimal. Seems like we might as well use a
software cursor at that point.

-- 
Ville Syrjälä
Intel