[FFmpeg-devel] [PATCH 8/9] avfilter/scale_amf: Add AMF HW scaler & color converter

[Evgeny Pavlov]

On Wed, Feb 14, 2024 at 5:26 PM Dennis Mungai <dmngaie at gmail.com> wrote:

> On Wed, 14 Feb 2024, 18:28 Evgeny Pavlov, <lucenticus at gmail.com> wrote:
>
> > On Wed, Feb 14, 2024 at 4:08 PM Timo Rothenpieler <timo at rothenpieler.org
> >
> > wrote:
> >
> > > On 14/02/2024 02:55, Dmitrii Ovchinnikov wrote:
> > > > From: Evgeny Pavlov <lucenticus at gmail.com>
> > > >
> > > > This commit adds two AMF filters: scale_amf & scale_amf_hq.
> > > > Both filters are using AMF hardware acceleration.
> > > > scale_amf supports simple scaling algorithms & color conversion.
> > > > scale_amf_hq supports advanced scaling algorithms & might be used
> > > > for upscaling only.
> > >
> > > Haven't looked at the patch yet, but can't this be one filter, and it
> > > picks the best possible method depending on options/inputs/whatever?
> > > _______________________________________________
> > > ffmpeg-devel mailing list
> > > ffmpeg-devel at ffmpeg.org
> > > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> > >
> > > To unsubscribe, visit link above, or email
> > > ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
> > >
> >
> > AMF has 2 separate components for color conversion + simple scaling
> > (VideoConverter)  and for advanced scaling (HQScaler).
> > We've got a recommendation from the AMD AMF team to implement these
> > components as separate ffmpeg filters.
> >
>
>
> Still, this should be a single VIdeo Post-Processing (VPP) -style filter,
> exposing these scaling and video post processing options as tunables
> therein.
>
> A perfect example of such an implementation that excels in such an
> abstraction is intel's vpp_qsv filter, from which multiple compute, color
> space conversion methods, tonemapping, etc are made available through
> tunables.
>
> Another benefit of such an abstraction would be that re-using this filter
> on other GPU derivatives of the discrete silicon, eg in smaller IGPs
> sharing these offload blocks would be auto-detecting logic to ensure that
> even with defaults, the filter chains run.
>
> Taking another example from Intel, they have a full H/W path for low power
> encode and Post-Processing that can be automatically toggled on by specific
> filter options without user intervention, guaranteeing runtime safety for
> the same command(s) even on newer GPUs.
>
> Food for thought.
>
> >
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
>

AMF Video Converter (which is used in scale_amf) is the full analog of
Intel's VPP. And in the future, it may have an option for dedicated HW
block in GPU. It has scaling, color conversion, tonemapping, HDR <> SDR
conversion features.
But HQ Scaler brings advanced scaling algorithms that are similar to AMD
FSR (https://www.amd.com/en/technologies/fidelityfx-super-resolution ) and
don’t have color conversion and cannot be mapped to HW blocks.
Do you think that it would be better to rename scale_amf to vpp_amf or some
other name? Maybe we should rename scale_amf_hq as well for better
usability (e.g. sr_amf)?