[PATCH v7 11/11] media: hantro: Support NV12 on the G2 core

[Jernej Škrabec]

Dne sreda, 20. oktober 2021 ob 17:25:40 CEST je Ezequiel Garcia napisal(a):
> On Wed, 20 Oct 2021 at 12:04, Jernej Škrabec <jernej.skrabec at gmail.com> 
wrote:
> >
> > Dne sreda, 20. oktober 2021 ob 13:06:59 CEST je Ezequiel Garcia 
napisal(a):
> > > Hi Jernej,
> > >
> > > On Tue, 19 Oct 2021 at 13:38, Jernej Škrabec <jernej.skrabec at gmail.com>
> > wrote:
> > > >
> > > > Hi Andrzej!
> > > >
> > > > Dne petek, 15. oktober 2021 ob 19:19:47 CEST je Andrzej Pietrasiewicz
> > > > napisal(a):
> > > > > Hi Jernej,
> > > > >
> > > > > W dniu 14.10.2021 o 19:42, Jernej Škrabec pisze:
> > > > > > Hi Andrzej!
> > > > > >
> > > > > > Dne sreda, 29. september 2021 ob 18:04:39 CEST je Andrzej
> > Pietrasiewicz
> > > > > > napisal(a):
> > > > > >> The G2 decoder block produces NV12 4x4 tiled format (NV12_4L4).
> > > > > >> Enable the G2 post-processor block, in order to produce regular 
NV12.
> > > > > >>
> > > > > >> The logic in hantro_postproc.c is leveraged to take care of
> > allocating
> > > > > >> the extra buffers and configure the post-processor, which is
> > > > > >> significantly simpler than the one on the G1.
> > > > > >
> > > > > > Quick summary of discussion on LibreELEC Slack:
> > > > > > When using NV12 format on Allwinner H6 variant of G2 (needs some
> > driver
> > > > > > changes), I get frames out of order. If I use native NV12 tiled
> > format,
> > > > frames
> > > > > > are ordered correctly.
> > > > > >
> > > > > > Currently I'm not sure if this is issue with my changes or is this
> > general
> > > > > > issue.
> > > > > >
> > > > > > I would be grateful if anyone can test frame order with and 
without
> > > > > > postprocessing enabled on imx8. Take some dynamic video with a lot 
of
> > > > short
> > > > > > scenes. It's pretty obvious when frames are out of order.
> > > > > >
> > > > >
> > > > > I checked on imx8 and cannot observe any such artifacts.
> > > >
> > > > I finally found the issue. As you mentioned on Slack, register write 
order
> > once
> > > > already affected decoding. Well, it's the case again. I made hacky test 
and
> > > > moved postproc enable call after output buffers are set and it worked. 
So,
> > this
> > > > is actually core quirk which is obviously fixed in newer variants.
> > > >
> > >
> > > Ugh, good catch.
> > >
> > > What happens if you move all the calls to HANTRO_PP_REG_WRITE_S
> > > (HANTRO_PP_REG_WRITE does a relaxed write)?
> > >
> > > Or what happens if the HANTRO_PP_REG_WRITE(vpu, out_luma_base, dst_dma)
> > > is moved to be done after all the other registers?
> >
> > Those two macros aren't used on G2. Andrzej introduced new postproc 
helpers
> > for G2.
> >
> 
> Ah, so the issue is specific on the G2 post-processor.

To be more precise, issue is specific only to old G2 post-processor, found in 
Allwinner H6. Andrzej tested code with newer G2 core and both locations worked 
fine.

> 
> > This commit solves issue for H6:
> > https://github.com/jernejsk/linux-1/commit/
> > a783a977c0843bb4b555dc9d0b5d64915cd219e7
> >
> 
> Right, but see this comment:
> 
>     /* Turn on pipeline mode. Must be done first. */
>     HANTRO_PP_REG_WRITE_S(vpu, pipeline_en, 0x1);
> 
> I have vague recollection of why we have that comment,
> but I'm reluctant to move post-proc enable to the end.
> (or at least not do it on G1?).

I missed that. Any idea what would be the cleanest way to move code for G2 
only? I can only think of quirk flag in platform specific structure.

Best regards,
Jernej

> 
> > >
> > > > This makes this series with minor adaptations completely working on 
H6. I
> > see
> > > > no reason not to merge whole series.
> > > >
> > >
> > > Do you have plans to submit your H6 work on top of this?
> >
> > Of course, why would I work on this otherwise? :) But before I do that, I 
have
> > to clean up and split one commit, which adapts VP9 G2 code for H6 variant.
> >
> 
> OK, sounds good.
> 
> > If you're interested in changes, take a look here:
> > https://github.com/jernejsk/linux-1/commits/vp9
> >
> 
> Will take a look.
> 
> Thanks,
> Ezequiel
>