Re: [linux-sunxi] [PATCH v2 1/2] media: v4l: Add definitions for the HEVC slice format and controls

[Tomasz Figa]

On Wed, Jan 30, 2019 at 3:28 PM Ayaka <ayaka@xxxxxxxxxxx> wrote:
>
>
>
> Sent from my iPad
>
> > On Jan 30, 2019, at 11:35 AM, Tomasz Figa <tfiga@xxxxxxxxxxxx> wrote:
> >
> > On Wed, Jan 30, 2019 at 11:29 AM Alexandre Courbot
> > <acourbot@xxxxxxxxxxxx> wrote:
> >>
> >>> On Wed, Jan 30, 2019 at 6:41 AM Nicolas Dufresne <nicolas@xxxxxxxxxxxx> wrote:
> >>>
> >>>> Le mardi 29 janvier 2019 Ã 16:44 +0900, Alexandre Courbot a Ãcrit :
> >>>> On Fri, Jan 25, 2019 at 10:04 PM Paul Kocialkowski
> >>>> <paul.kocialkowski@xxxxxxxxxxx> wrote:
> >>>>> Hi,
> >>>>>
> >>>>>> On Thu, 2019-01-24 at 20:23 +0800, Ayaka wrote:
> >>>>>> Sent from my iPad
> >>>>>>
> >>>>>>> On Jan 24, 2019, at 6:27 PM, Paul Kocialkowski <paul.kocialkowski@xxxxxxxxxxx> wrote:
> >>>>>>>
> >>>>>>> Hi,
> >>>>>>>
> >>>>>>>> On Thu, 2019-01-10 at 21:32 +0800, ayaka wrote:
> >>>>>>>> I forget a important thing, for the rkvdec and rk hevc decoder, it would
> >>>>>>>> requests cabac table, scaling list, picture parameter set and reference
> >>>>>>>> picture storing in one or various of DMA buffers. I am not talking about
> >>>>>>>> the data been parsed, the decoder would requests a raw data.
> >>>>>>>>
> >>>>>>>> For the pps and rps, it is possible to reuse the slice header, just let
> >>>>>>>> the decoder know the offset from the bitstream bufer, I would suggest to
> >>>>>>>> add three properties(with sps) for them. But I think we need a method to
> >>>>>>>> mark a OUTPUT side buffer for those aux data.
> >>>>>>>
> >>>>>>> I'm quite confused about the hardware implementation then. From what
> >>>>>>> you're saying, it seems that it takes the raw bitstream elements rather
> >>>>>>> than parsed elements. Is it really a stateless implementation?
> >>>>>>>
> >>>>>>> The stateless implementation was designed with the idea that only the
> >>>>>>> raw slice data should be passed in bitstream form to the decoder. For
> >>>>>>> H.264, it seems that some decoders also need the slice header in raw
> >>>>>>> bitstream form (because they take the full slice NAL unit), see the
> >>>>>>> discussions in this thread:
> >>>>>>> media: docs-rst: Document m2m stateless video decoder interface
> >>>>>>
> >>>>>> Stateless just mean it wonât track the previous result, but I donât
> >>>>>> think you can define what a date the hardware would need. Even you
> >>>>>> just build a dpb for the decoder, it is still stateless, but parsing
> >>>>>> less or more data from the bitstream doesnât stop a decoder become a
> >>>>>> stateless decoder.
> >>>>>
> >>>>> Yes fair enough, the format in which the hardware decoder takes the
> >>>>> bitstream parameters does not make it stateless or stateful per-se.
> >>>>> It's just that stateless decoders should have no particular reason for
> >>>>> parsing the bitstream on their own since the hardware can be designed
> >>>>> with registers for each relevant bitstream element to configure the
> >>>>> decoding pipeline. That's how GPU-based decoder implementations are
> >>>>> implemented (VAAPI/VDPAU/NVDEC, etc).
> >>>>>
> >>>>> So the format we have agreed on so far for the stateless interface is
> >>>>> to pass parsed elements via v4l2 control structures.
> >>>>>
> >>>>> If the hardware can only work by parsing the bitstream itself, I'm not
> >>>>> sure what the best solution would be. Reconstructing the bitstream in
> >>>>> the kernel is a pretty bad option, but so is parsing in the kernel or
> >>>>> having the data both in parsed and raw forms. Do you see another
> >>>>> possibility?
> >>>>
> >>>> Is reconstructing the bitstream so bad? The v4l2 controls provide a
> >>>> generic interface to an encoded format which the driver needs to
> >>>> convert into a sequence that the hardware can understand. Typically
> >>>> this is done by populating hardware-specific structures. Can't we
> >>>> consider that in this specific instance, the hardware-specific
> >>>> structure just happens to be identical to the original bitstream
> >>>> format?
> >>>
> >>> At maximum allowed bitrate for let's say HEVC (940MB/s iirc), yes, it
> >>> would be really really bad. In GStreamer project we have discussed for
> >>> a while (but have never done anything about) adding the ability through
> >>> a bitmask to select which part of the stream need to be parsed, as
> >>> parsing itself was causing some overhead. Maybe similar thing applies,
> >>> though as per our new design, it's the fourcc that dictate the driver
> >>> behaviour, we'd need yet another fourcc for drivers that wants the full
> >>> bitstream (which seems odd if you have already parsed everything, I
> >>> think this need some clarification).
> >>
> >> Note that I am not proposing to rebuild the *entire* bitstream
> >> in-kernel. What I am saying is that if the hardware interprets some
> >> structures (like SPS/PPS) in their raw format, this raw format could
> >> be reconstructed from the structures passed by userspace at negligible
> >> cost. Such manipulation would only happen on a small amount of data.
> >>
> >> Exposing finer-grained driver requirements through a bitmask may
> >> deserve more exploring. Maybe we could end with a spectrum of
> >> capabilities that would allow us to cover the range from fully
> >> stateless to fully stateful IPs more smoothly. Right now we have two
> >> specifications that only consider the extremes of that range.
> >
> > I gave it a bit more thought and if we combine what Nicolas suggested
> > about the bitmask control with the userspace providing the full
> > bitstream in the OUTPUT buffers, split into some logical units and
> > "tagged" with their type (e.g. SPS, PPS, slice, etc.), we could
> > potentially get an interface that would work for any kind of decoder I
> > can think of, actually eliminating the boundary between stateful and
> > stateless decoders.
> I agree with this idea, that is what I want calling memory region description while I am still struggling with userspace to post my driver demo.
> >
> > For example, a fully stateful decoder would have the bitmask control
> > set to 0 and accept data from all the OUTPUT buffers as they come. A
> > decoder that doesn't do any parsing on its own would have all the
> > valid bits in the bitmask set and ignore the data in OUTPUT buffers
> > tagged as any kind of metadata. And then, we could have any cases in
> > between, including stateful decoders which just can't parse the stream
> > on their own, but still manage anything else themselves, or stateless
> > ones which can parse parts of the stream, like the rk3399 vdec can
> > parse the H.264 slice headers on its own.
> >
> Actually not, the rkvdec and rkhevc can parse most but not all syntax sections.
> Besides the vp9 decoder of rkvdec wonât parse most of the syntax.
>
> I talked to some rockchip staff about the performance problem of reconstruction bitstream after yesterday arguing with tfiga at IRC yesterday. Although 1ms looks small to those decoder which can decode a picture of a UHD 4K HEVC videos in 9ms, it is enough for 60fps. But how about a higher frame rate like 120fps or 240fps and when it comes to 8K which is used in Japan broadcast.

1 ms for a 500 MHz CPU (which is quite slow these days) is 500k
cycles. We don't have to reconstruct the whole bitstream, just the
parsed metadata and also we don't get a new PPS or SPS every frame.
Not sure where you have this 1 ms from. Most of the difference between
our structures and the bitstream is that the latter is packed and
could be variable length.

We actually have some sample bitstream assembly code for the rockchip encoder:

https://chromium.googlesource.com/chromiumos/third_party/libv4lplugins/+/5e6034258146af6be973fb6a5bb6b9d6e7489437/libv4l-rockchip_v2/libvepu/h264e/h264e.c#148
https://chromium.googlesource.com/chromiumos/third_party/libv4lplugins/+/5e6034258146af6be973fb6a5bb6b9d6e7489437/libv4l-rockchip_v2/libvepu/streams.c#36

Disassembling the stream_put_bits() gives 36 thumb2 instructions,
including 23 for the loop for each byte that is written.
stream_write_ue() is a bit more complicated, but in the worst case it
ends up with 4 calls to stream_put_bits(), each at most spanning 4
bytes for simplicity.

Let's count the operations for SPS then:
 (1) stream_put_bits() spanning 1 byte: 33 times
 (2) stream_put_bits() spanning up to 3 bytes: 4 times
 (3) stream_write_ue() up to 31 bits: 19 times

Adding it together:
(1) 33 * 36 +
(2) 4 * (36 + 2 * 23) +
(3) 19 * (4 * (36 + 3 * 23)) =

1188 + 328 + 7980 = 9496 ~= 10k instructions

The code above doesn't seem to contain any expensive instructions,
like division, so for a modern pipelined out of order core (e.g. A53),
it could be safe to assume 1 instruction per cycle. At 500 MHz that
gives you 20 usecs.

SPS is the most complex header and for H.264 we just do PPS and some
slice headers. Let's round it up a bit and we could have around 100
usecs for the complete frame metadata.

>
> I would bring more detail in the FOSDEM 2019, I may stay at graphics devroom at Saturday.
> > That could potentially let us completely eliminate the distinction
> > between the stateful and stateless interfaces and just have one that
> > covers both.
> >
> > Thoughts?

Any thoughts on my proposal to make the interface more flexible? Any
specific examples of issues that we could encounter that would prevent
it from working efficiently with Rockchip (or other) hardware?

Best regards,
Tomasz