[PATCH v6 11/17] media: uapi: Add V4L2_CID_STATELESS_HEVC_ENTRY_POINT_OFFSETS control

Nicolas Dufresne nicolas.dufresne at collabora.com
Wed Jun 1 10:07:54 PDT 2022


Le mercredi 01 juin 2022 à 18:35 +0200, Jernej Škrabec a écrit :
> > I believe its defined following "Table A.8 – General tier and level limits".
> > With the assumption there will never be a level 7 (which I think is fair).
> > If anyone saw other reasons for this limit, let me know.
> > 
> > This is a worse case scenario, this is quite unlikely in practice, so while
> > performance might be a disaster if your craft a stream for that case, I
> > don't think it will ever happen in real life.
> 
> But do we really need to cover worst case scenario? In theory, one driver can 
> set limit to (for example) max 100 slices and if there is a frame with 600 
> slices, userspace app would submit 6 decode requests. Basically the same way 
> it's done today. While not as performant, it would be good compromise between 
> resources and speed.

The limit here is to prevent userland from tricking the kernel into doing very
big allocation. But with dynamic array, you'll allocate just the right amount.

Nicolas




More information about the linux-arm-kernel mailing list