[PATCH v1 0/9] AV1 stateless decoder for RK3588
Ezequiel Garcia
ezequiel at vanguardiasur.com.ar
Tue Dec 20 09:50:53 PST 2022
Bonjour Nicolas,
On Tue, Dec 20, 2022 at 2:15 PM Nicolas Dufresne <nicolas at ndufresne.ca> wrote:
>
> Le mardi 20 décembre 2022 à 14:40 +0100, Michael Grzeschik a écrit :
> > Hi Ezequiel,
> >
> > On Mon, Dec 19, 2022 at 10:52:02PM -0300, Ezequiel Garcia wrote:
> > > On Mon, Dec 19, 2022 at 6:54 PM Michael Grzeschik <mgr at pengutronix.de> wrote:
> > > > On Mon, Dec 19, 2022 at 06:07:38PM -0300, Ezequiel Garcia wrote:
> > > > > On Mon, Dec 19, 2022 at 12:56 PM Benjamin Gaignard
> > > > > <benjamin.gaignard at collabora.com> wrote:
> > > > > >
> > > > > > This series implement AV1 stateless decoder for RK3588 SoC.
> > > > > > The harware support 8 and 10 bits bitstreams up to 7680x4320.
> > > > > > AV1 feature like film grain or scaling are done by the postprocessor.
> > > > > > The driver can produce NV12_4L4 and NV12 pixel formats.
> > > > > > A native 10bits NV12_4L4 format is possible but need more investigation
> > > > > > to be completly documented and enabled.
> > > > > >
> > > > > > It is based on Daniel's "[RFC,v3] media: Add AV1 uAPI" [1] patches and
> > > > > > Sebastian's device-tree patches for RK3588.
> > > > > >
> > > > >
> > > > > I thought the AV1 decoder in RK3588 was really a separate hardware
> > > > > from the Hantro G1/G2.
> > > > >
> > > > > Shouldn't this need a new driver for this new hardware?
> > > >
> > > > Just jumping into this discussion as I am currently working on the rkvenc driver.
> > > >
> > >
> > > The more the merrier, there's always room for developers :-)
> > >
> > > > In my case I am extending the rkvdec driver to become more generic for
> > > > other rockchip specific enc/decoders.
> > > >
> > > > My first change looks like this:
> > > > ---
> > > > drivers/staging/media/rkvdec/Makefile | 4 +-
> > > > drivers/staging/media/rkvdec/rkvdec-h264.c | 100 ++++-----
> > > > drivers/staging/media/rkvdec/rkvdec-vp9.c | 142 ++++++-------
> > > > drivers/staging/media/rkvdec/{rkvdec.c => rkvpu.c} | 510 +++++++++++++++++++++++-----------------------
> > > > drivers/staging/media/rkvdec/{rkvdec.h => rkvpu.h} | 66 +++---
> > > > ---
> > > >
> > > > While working on other parts of the encoder I found many places in the
> > > > rkvdec driver (e.g. v4l2 and vb2 callbacks) that looked familiar to the hantro
> > > > functions but where limited to the decoder case.
> > > >
> > >
> > > Because stateless decoders devices are very similar in their general behavior,
> > > their drivers could be very similar.
> > >
> > > Hantro and Rkvdec could look similar because the same humans worked on them.
> > >
> > > Most boilerplate code, as well as V4L2 format negotiation, VB2 buffer handling
> > > could be shared among all stateless decoder drivers. I think even at one point
> > > we experimented with having a shared/common code base for all stateless codecs.
> > >
> > > In other words, it's entirely possible to support Hantro devices in
> > > the Cedrus driver
> > > and vice-versa, you would only have to write the hardware-specific bits.
> > >
> > > However, there is consensus to have a separate driver for each
> > > different hardware,
> > > even when the hardware is a bit similar. This may lead to some code duplication,
> > > but it's less fragile / more flexible. Maintaining drivers this way allows
> > > developers to evolve, testing on a small family of devices, without
> > > breaking support
> > > for other devices.
> > >
> > > This is important as sometimes it's hard to get the hardware,
> > > but we still don't want to break the support!
> > >
> > > > I think there are two options for the av1 codec.
> > > >
> > > > 1) If the vpu981 is a driver that has nothing to do with verisilicon but
> > > > works with this driver framework, then we should integrate vepu981 into it
> > > > but consider rename the verisilicon unrelated parts to something generic.
> > > >
> > > > 2) Move the vepu981 av1 driver into the rkvdec instead.
> > > >
> > > > If 1) is the way to go, we can even think of moving the staging code parts from
> > > > rkvdec to the verisilicon code. Likewise to the vepu981-av1.
> > > >
> > >
> > > The Hantro driver should only support G1, G2, and VC8000D;
> > > which can be said to belong to the same family.
> > >
> > > The RKVDEC driver supports Rockchip vdpu34x core. I have to admit
> > > I'm not exactly sure if we support anything else than vdpu34x.
> >
> > Currently the rkvdec is only supporting vdpu34x. My work would integrate
> > vepu54x into the rkvdec boilerplate and so it would support encode as decode.
>
> Which CODEC do you currently work on ? We are about to send a first RFC for a
> VP8 stateless encoder API (with a rk3399 driver to test), but haven't written
> the Stateless Encoder API spec yet, so still some work there. And was planning
> to make an H.264 Sateless Encoder soon. Would be nice to avoid duplicating the
> effort.
>
> >
> > > I'm not familiar with the AV1 support provided by this patch,
> > > but looking at the mpp code:
> > >
> > > ...
> > > "rk3588",
> > > ROCKCHIP_SOC_RK3588,
> > > HAVE_VDPU2 | HAVE_VDPU2_PP | HAVE_VEPU2 | HAVE_RKVDEC | HAVE_RKVENC |
> > > HAVE_JPEG_DEC | HAVE_AV1DEC | HAVE_AVSDEC | HAVE_VEPU2_JPEG,
> > > { &vdpu38x, &rkjpegd, &vdpu2, &vdpu2_jpeg_pp, &av1d, &avspd},
> > > { &vepu58x, &vepu2, &vepu2_jpeg, NULL, },
> > >
> > > Seems RK3588 supports a Hantro core (VDPU2), a vdpu38x core and this AV1 core,
> > > which according to this patchset is vdpu981 (?)
> > >
> > > If the vdpu38x device interface, configuration, buffer handling and
> > > registers are
> > > similar enough with vdpu34x, adding vdpu38x to the Rkvdec driver
> > > should be straightforward.
> > > If the vdpu38x core differs, it may be reason enough to consider a new driver.
> > >
> > > As for vdpu981 (AV1), I'm inclined to think it deserves its own driver.
>
> Well, it has its own driver, Hantro (which is not rkvdec). But maybe you could
> extend on why you think VC9000D decoder have no place in the hantro/verisilicon
> family ?
>
Oh good, then the AV1 core in RK3588 is actually a Verisilicon VC9000D core,
I didn't know that.
Maybe the naming in the driver should be vc9000? Or maybe it should be noted
in comments and documentation. If that was already documented somewhere
in this patchset, then I definitely missed it.
> > >
> > > Again, I'm far less worried for a little code duplication in the
> > > boilerplate (which can be solved
> > > with helpers, etc.) and more worried about making sure we can evolve
> > > drivers easily,
> > > while minimizing regressions.
> >
> > Thanks for the explanation.
> >
> > As I agree that not breaking current drivers is a strong argument. Also
> > rkvdec is still in staging, which makes it less harmful for the
> > integration of the encoder path.
>
> We are working on unstaging patches.
>
Rkvdec can be unstaged.
Sebastian: I remember you were working on Rkvdec.
If you want to unstage this driver now, and work on features later,
I believe that would make a lot of sense!
Thanks!
Ezequiel
> >
> > Since we can not ensure that the rkvenc/rkvdec is not another unknown
> > verisilicon core, going the way of working on a common rkvpu driver is
> > probably the best for now.
>
> We can collectively share our knowledge (to the limit of our legal rights to
> share) make the right call. In the case of this VC9000D decoder, there is a
> massive amount of registers that aren't AV1 specific, and existed in VC8000
> cores as it, same offset, same size. Hantro designs have this very specific
> style, which is to share register, giving it a meaning for multiple CODECs.
>
> I've commented about that in my review, but until we have more codecs support on
> VC9000 cores, generalizing the register definition is premature.
>
> Though, an typical example of things that are Hantro specific and common to G1,
> VC8000 and VC9000, is the handling of references for H.264 decoding. This
> differs massively from how it works with rkvdec here.
>
> >
> > Also, since I have already done some work into that direction, it sounds
> > good for me. :)
>
> Great. For you interest, the modified Hantro H1 encoder is an information that
> Rockchip disclosed to us directly. And that whys vepu121 (if my memory is right)
> is implemented in Hantro driver. The register layout have been altered by RK but
> that's all there is, it does share semantic (and a lot of code) with the "real"
> H1 found on RK3288, IMX8M Mini and others.
>
> >
> >
> > > > I could also keep on integrating the rkvenc on that base instead.
> > > >
> > > > Regards,
> > > > Michael
> > > >
> > > > > > The full branch can be found here:
> > > > > > https://gitlab.collabora.com/linux/for-upstream/-/commits/rk3588_av1_decoder_v1
> > > > > >
> > > > > > Fluster score is: 151/239 while testing AV1-TEST-VECTORS with GStreamer-AV1-V4L2SL-Gst1.0.
> > > > > > The failing tests are:
> > > > > > - 10bits bitstream because 10bits output formats aren't yet implemented.
> > > > > > - the 2 tests with 2 spatial layers: few errors in luma/chroma values
> > > > > > - tests with resolution < hardware limit (64x64)
> > > > > >
> > > > > > Benjamin
> > > > > >
> > > > > > Benjamin Gaignard (9):
> > > > > > dt-bindings: media: rockchip-vpu: Add rk3588 vpu compatible
> > > > > > media: verisilicon: Add AV1 decoder mode and controls
> > > > > > media: verisilicon: Save bit depth for AV1 decoder
> > > > > > media: verisilicon: Check AV1 bitstreams bit depth
> > > > > > media: verisilicon: Compute motion vectors size for AV1 frames
> > > > > > media: verisilicon: Add AV1 entropy helpers
> > > > > > media: verisilicon: Add Rockchip AV1 decoder
> > > > > > media: verisilicon: Add film grain feature to AV1 driver
> > > > > > media: verisilicon: Enable AV1 decoder on rk3588
> > > > > >
> > > > > > .../bindings/media/rockchip-vpu.yaml | 1 +
> > > > > > drivers/media/platform/verisilicon/Makefile | 3 +
> > > > > > drivers/media/platform/verisilicon/hantro.h | 5 +
> > > > > > .../media/platform/verisilicon/hantro_drv.c | 54 +
> > > > > > .../media/platform/verisilicon/hantro_hw.h | 102 +
> > > > > > .../platform/verisilicon/hantro_postproc.c | 3 +
> > > > > > .../media/platform/verisilicon/hantro_v4l2.c | 5 +
> > > > > > .../verisilicon/rockchip_av1_entropymode.c | 4536 +++++++++++++++++
> > > > > > .../verisilicon/rockchip_av1_entropymode.h | 272 +
> > > > > > .../verisilicon/rockchip_av1_filmgrain.c | 401 ++
> > > > > > .../verisilicon/rockchip_av1_filmgrain.h | 36 +
> > > > > > .../verisilicon/rockchip_vpu981_hw_av1_dec.c | 2280 +++++++++
> > > > > > .../verisilicon/rockchip_vpu981_regs.h | 477 ++
> > > > > > .../platform/verisilicon/rockchip_vpu_hw.c | 116 +
> > > > > > 14 files changed, 8291 insertions(+)
> > > > > > create mode 100644 drivers/media/platform/verisilicon/rockchip_av1_entropymode.c
> > > > > > create mode 100644 drivers/media/platform/verisilicon/rockchip_av1_entropymode.h
> > > > > > create mode 100644 drivers/media/platform/verisilicon/rockchip_av1_filmgrain.c
> > > > > > create mode 100644 drivers/media/platform/verisilicon/rockchip_av1_filmgrain.h
> > > > > > create mode 100644 drivers/media/platform/verisilicon/rockchip_vpu981_hw_av1_dec.c
> > > > > > create mode 100644 drivers/media/platform/verisilicon/rockchip_vpu981_regs.h
> > > > > >
> > > > > > --
> > > > > > 2.34.1
> > > > > >
> > > > >
> > > > > _______________________________________________
> > > > > linux-arm-kernel mailing list
> > > > > linux-arm-kernel at lists.infradead.org
> > > > > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
> > > > >
> > > >
> > > > --
> > > > Pengutronix e.K. | |
> > > > Steuerwalder Str. 21 | http://www.pengutronix.de/ |
> > > > 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
> > > > Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
> > >
> >
>
More information about the Linux-rockchip
mailing list