Re: [PATCH 0/5] media: iris: HFI gen2 features and fixes for VA-API
From: bod
Date: Tue Nov 17 2026 - 08:04:12 EST
On 2026-11-04 10:40 +0800, Xilin Wu wrote:
> On 11/4/2026 6:29 AM, Nicolas Dufresne wrote:
> > Le mardi 03 novembre 2026 à 23:00 +0100, Bryan O'Donoghue a écrit :
> >> On 29/10/2026 04:58, Xilin Wu wrote:
> >>> This series adds features and fixes to Iris HFI gen2 to support
> >>> qcom-va-driver [1], a userspace VA-API backend for the V4L2 stateful codec
> >>> interface.
> >>>
> >>> It enables decode-order output and larger CAPTURE buffer pools for VA-API
> >>> clients, and exposes encoder deblocking controls. It also fixes
> >>> error-buffer timestamps and handling of non-display AV1 and VP9 pictures,
> >>> which affect both VA-API and direct V4L2 clients.
> >>>
> >>> [1]https://github.com/radxa-pkg/qcom-va-driver
> >>
> >> Is this code based on https://github.com/intel/libva and is there a
> >> route to getting it into the upstream of that repository ?
> >
> > libva is the runtime, qcom-va-driver is a driver, so that won't happen. Also,
> > qcom-va-drivers does all sort of guessing heuristic, bitstream crafting and the
> > readme show that it cost as much as 15% in throuput performance. Let's be
> > honest, this is a bit of hack, not really a solution. A proper stateless
> > firmware and driver would do a way better job and reduce complexity instead of
> > increasing it. The provided patches though seems to all have merit, matching
> > feature with venus, fixing bugs. A bit of spec work would not hurt though.
> >
> > Nicolas
>
> Hi Nicolas,
>
> The userspace software ecosystem around stateful V4L2 has long been a
> challenge. With qcom-va-driver, we're trying to provide users with an
> optional, as-out-of-the-box-as-possible solution that works with
> existing VA-API applications.
>
> I'm not sure whether the Qualcomm Iris hardware is capable of supporting
> a stateless interface.
>
> As far as I know, some folks at Qualcomm are also working on getting
> qcom-va-driver into Debian:
> https://tracker.debian.org/pkg/qcom-va-driver
>
> Thanks for reviewing the patches!
>
> --
> Best regards,
> Xilin Wu <sophon@xxxxxxxxx>
Iris can do some kind of stateless but that very much needs to be scoped
out and developed further. Windows DXVA is public, some scoping work has
been done on that too.
Putative future work on Iris to enable stateless need not block taking in
these patches in the meantime.
As the proverb goes "A bird in the hand is worth two in the bush".
Generally speaking I think these patches are close to mergable so if you
can address the outstanding comments I see no barrier in principle to
taking them.
The gap in browser support in particular is a pain point I'd like to see
removed for the community.
---
bod
---
bod