On 7/25/23 20:15, Paul Kocialkowski wrote:
Subject:
Re: Stateless Encoding uAPI Discussion and Proposal
From:
Paul Kocialkowski <paul.kocialkowski@xxxxxxxxxxx>
Date:
7/25/23, 20:15
To:
Hsia-Jun Li <Randy.Li@xxxxxxxxxxxxx>
CC:
linux-kernel@xxxxxxxxxxxxxxx, Nicolas Dufresne
<nicolas.dufresne@xxxxxxxxxxxxx>, linux-media@xxxxxxxxxxxxxxx, Hans
Verkuil <hverkuil@xxxxxxxxx>, Sakari Ailus <sakari.ailus@xxxxxx>,
Andrzej Pietrasiewicz <andrzej.p@xxxxxxxxxxxxx>, Michael Tretter
<m.tretter@xxxxxxxxxxxxxx>, Jernej Škrabec <jernej.skrabec@xxxxxxxxx>,
Chen-Yu Tsai <wens@xxxxxxxx>, Samuel Holland <samuel@xxxxxxxxxxxx>,
Thomas Petazzoni <thomas.petazzoni@xxxxxxxxxxx>
Hey,
Long time, good to see you are still around and interested in these topics 😄
On Tue 25 Jul 23, 11:33, Hsia-Jun Li wrote:
On 7/12/23 22:07, Paul Kocialkowski wrote:
Hi Nicolas,
Thanks for the quick reply!
On Tue 11 Jul 23, 14:18, Nicolas Dufresne wrote:
Le mardi 11 juillet 2023 à 19:12 +0200, Paul Kocialkowski a écrit :
Hi everyone!
After various discussions following Andrzej's talk at EOSS, feedback from the
Media Summit (which I could not attend unfortunately) and various direct
discussions, I have compiled some thoughts and ideas about stateless encoders
support with various proposals. This is the result of a few years of interest
in the topic, after working on a PoC for the Hantro H1 using the hantro driver,
which turned out to have numerous design issues.
I am now working on a H.264 encoder driver for Allwinner platforms (currently
focusing on the V3/V3s), which already provides some usable bitstream and will
be published soon.
This is a very long email where I've tried to split things into distinct topics
and explain a few concepts to make sure everyone is on the same page.
# Bitstream Headers
Stateless encoders typically do not generate all the bitstream headers and
sometimes no header at all (e.g. Allwinner encoder does not even produce slice
headers). There's often some hardware block that makes bit-level writing to the
destination buffer easier (deals with alignment, etc).
The values of the bitstream headers must be in line with how the compressed
data bitstream is generated and generally follow the codec specification.
Some encoders might allow configuring all the fields found in the headers,
others may only allow configuring a few or have specific constraints regarding
which values are allowed.
As a result, we cannot expect that any given encoder is able to produce frames
for any set of headers. Reporting related constraints and limitations (beyond
profile/level) seems quite difficult and error-prone.
So it seems that keeping header generation in-kernel only (close to where the
hardware is actually configured) is the safest approach.
This seems to match with what happened with the Hantro VP8 proof of concept. The
encoder does not produce the frame header, but also, it produces 2 encoded
buffers which cannot be made contiguous at the hardware level. This notion of
plane in coded data wasn't something that blended well with the rest of the API
and we didn't want to copy in the kernel while the userspace would also be
forced to copy to align the headers. Our conclusion was that it was best to
generate the headers and copy both segment before delivering to userspace. I
suspect this type of situation will be quite common.
Makes sense! I guess the same will need to be done for Hantro H1 H.264 encoding
(in my PoC the software-generated headers were crafted in userspace and didn't
have to be part of the same buffer as the coded data).
We just need a method to indicate where the hardware could write its slice
data or compressed frame.
While we would decided which frame that the current frame should refer, the
(some) hardware may discard our decision, which reference picture set would
use less bits. Unless the codec supports a fill up method, it could lead to
a gap between header and frame data.
I think I would need a bit more context to understand this case, especially
what the hardware could decide to discard.
I known the Hantro can't do this. While such design is not unusal. HW
could tell we can't CU that we could do inter predict from one previous
reconstruction frame, then it is not necessary to have it in its RPS.
My understanding is that the VP8 encoder needs to write part of the header
separately from the coded data and uses distinct address registers for the two.
I don't think Hantro H1 would do that.
So the approach is to move the hw-generated headers and coded data before
returning to userspace.
# Codec Features
Codecs have many variable features that can be enabled or not and specific
configuration fields that can take various values. There is usually some
top-level indication of profile/level that restricts what can be used.
This is a very similar situation to stateful encoding, where codec-specific
controls are used to report and set profile/level and configure these aspects.
A particularly nice thing about it is that we can reuse these existing controls
and add new ones in the future for features that are not yet covered.
This approach feels more flexible than designing new structures with a selected
set of parameters (that could match the existing controls) for each codec.
Though, reading more into this emails, we still have a fair amount of controls
to design and add, probably some compound controls too ?
Yeah definitely. My point here is merely that we should reuse existing control
for general codec features, but I don't think we'll get around introducing new
ones for stateless-specific parts.
Things likes profile, level or tiers could be reused. It make no sense to
expose those vendor special feature.
Besides, profile, level or tiers are usually stored in the sequence header
or uncompressed header, hardware doesn't care about that.
I think we should go with the vendor registers buffer way that I always
said. There are many encoding tools that a codec offer, variants of hardware
may not support or use them all. The context switching between userspace and
kernel would drive you mad for so many controls.
I am strongly against this approach, instead I think we need to keep all
vendor-specific parts in the kernel driver and provide a clean unified userspace
API.
We are driving away vendor participation. Besides, the current design is
a performance bottleneck.
Also I think V4L2 has way to set multiple controls at once, so the
userspace/kernel context switching is rather minimal and within reasonable
expectations. Of course it will never be as efficient as userspace mapping the
hardware registers in virtual memory but there are so many problems with this
approach that it's really not worth it.
I am not talking about mapping the register to the userspace.
The userspace would generate a register set for the current frame, while
the kernel should fill that register set with buffer address and trigger
the hardware to apply the register.
Generating a register set from control or even fill partial slice header
cost many resources.
And what we try to define may not fit for real hardware design, you
could only cover the most of hardwares would require but not vendor
didn't have to follow that. Besides, codec spec could be updated even
after it has been released for a while.
# Reference and Reconstruction Management
With stateless encoding, we need to tell the hardware which frames need to be
used as references for encoding the current frame and make sure we have the
these references available as decoded frames in memory.
Regardless of references, stateless encoders typically need some memory space to
write the decoded (known as reconstructed) frame while it's being encoded.
One question here is how many slots for decoded pictures should be allocated
by the driver when starting to stream. There is usually a maximum number of
reference frames that can be used at a time, although perhaps there is a use
case to keeping more around and alternative between them for future references.
Another question is how the driver should keep track of which frame will be used
as a reference in the future and which one can be evicted from the pool of
decoded pictures if it's not going to be used anymore.
A restrictive approach would be to let the driver alone manage that, similarly
to how stateful encoders behave. However it might provide extra flexibility
(and memory gain) to allow userspace to configure the maximum number of possible
reference frames. In that case it becomes necessary to indicate if a given
frame will be used as a reference in the future (maybe using a buffer flag)
and to indicate which previous reference frames (probably to be identified with
the matching output buffer's timestamp) should be used for the current encode.
This could be done with a new dedicated control (as a variable-sized array of
timestamps). Note that userspace would have to update it for every frame or the
reference frames will remain the same for future encodes.
The driver will then make sure to keep the reconstructed buffer around, in one
of the slots. When there's no slot left, the driver will drop the oldest
reference it has (maybe with a bounce buffer to still allow it to be used as a
reference for the current encode).
With this behavior defined in the uAPI spec, userspace will also be able to
keep track of which previous frame is no longer allowed as a reference.
If we want, we could mirror the stateless decoders here. During the decoding, we
pass a "dpb" or a reference list, which represent all the active references.
These do not have to be used by the current frame, but the driver is allowed to
use this list to cleanup and free unused memory (or reuse in case it has a fixed
slot model, like mtk vcodec).
On top of this, we add a list of references to be used for producing the current
frame. Usually, the picture references are indices into the dpb/reference list
of timestamp. This makes validation easier. We'll have to define how many
reference can be used I think since unlike decoders, encoders don't have to
fully implement levels and profiles.
So that would be a very explicit description instead of expecting drivers to
do the maintainance and userspace to figure out which frame was evicted from
the list. So yeah this feels more robust!
Regarding the number of reference frames, I think we need to specify both
how many references can be used at a time (number of hardware slots) and how
many total references can be in the reference list (number of rec buffers to
keep around).
We could also decide that making the current frame part of the global reference
list is a way to indicate that its reconstruction buffer must be kept around,
or we could have a separate way to indicate that. I lean towards the former
since it would put all reference-related things in one place and avoid coming
up with a new buffer flag or such.
Also we would probably still need to do some validation driver-side to make
sure that userspace doesn't put references in the list that were not marked
as such when encoded (and for which the reconstruction buffer may have been
recycled already).
DPB is the only thing we need to decide any API here under the vendor
registers buffer way. We need the driver to translate the buffer reference
to the address that hardware could use and in its right registers.
The major problem is how to export the reconstruction buffer which was
hidden for many years.
This could be disccused in the other thread like V4L2 ext buffer api.
Following my previous point, I am also strongly against exposing the
reconstruction buffer to userspace.
Android hates peoploe allocating a huge of memory without
userspace's(Android's core system) awareness.
Whether a reconstruction frame would be used for long term reference(or
golden frame) is complete up to userspace decision. For example, when we
encoding a part of SVC layer 1, we may not reference to a frame in layer
0, should we let the hardware discard that? Later, we may decide to
refernce it again.
Besides, I don't like the timetstamp way to refer a buffer here, one
input graphics buffer could produce multiple reconstruction buffer(with
different coding options), which is common for SVC case.
# Frame Types
Stateless encoder drivers will typically instruct the hardware to encode either
an intra-coded or an inter-coded frame. While a stream composed only of a single
intra-coded frame followed by only inter-coded frames is possible, it's
generally not desirable as it is not very robust against data loss and makes
seeking difficult.
Let's avoid this generalization in our document and design. In RTP streaming,
like WebRTP or SIP, it is desirable to use open GOP (with nothing else then P
frames all the time, except the very first one). The FORCE_KEY_FRAME is meant to
allow handling RTP PLI (and other similar feedback). Its quite rare an
application would mix close GOP and FORCE_KEY_FRAME, but its allowed though.
What I've seen the most, is the FORCE_KEY_FRAME would just start a new GOP,
following size and period from this new point.
Okay fair enough, thanks for the details!
As a result, the frame type is usually decided based on a given GOP size
(the frequency at which a new intra-coded frame is produced) while intra-coded
frames can be explicitly requested upon request. Stateful encoders implement
these through dedicated controls:
- V4L2_CID_MPEG_VIDEO_FORCE_KEY_FRAME
- V4L2_CID_MPEG_VIDEO_GOP_SIZE
- V4L2_CID_MPEG_VIDEO_H264_I_PERIOD
It seems that reusing them would be possible, which would let the driver decide
of the particular frame type.
However it makes the reference frame management a bit trickier since reference
frames might be requested from userspace for a frame that ends up being
intra-coded. We can either allow this and silently ignore the info or expect
that userspace keeps track of the GOP index and not send references on the first
frame.
In some codecs, there's also a notion of barrier key-frames (IDR frames in
H.264) that strictly forbid using any past reference beyond the frame.
There seems to be an assumption that the GOP start uses this kind of frame
(and not any intra-coded frame), while the force key frame control does not
particularly specify it.
In that case we should flush the list of references and userspace should no
longer provide references to them for future frames. This puts a requirement on
userspace to keep track of GOP start in order to know when to flush its
reference list. It could also check if V4L2_BUF_FLAG_KEYFRAME is set, but this
could also indicate a general intra-coded frame that is not a barrier.
So another possibility would be for userspace to explicitly indicate which
frame type to use (in a codec-specific way) and act accordingly, leaving any
notion of GOP up to userspace. I feel like this might be the easiest approach
while giving an extra degree of control to userspace.
I also lean toward this approach ...
# Rate Control
Another important feature of encoders is the ability to control the amount of
data produced following different rate control strategies. Stateful encoders
typically do this in-firmware and expose controls for selecting the strategy
and associated targets.
It seems desirable to support both automatic and manual rate-control to
userspace.
Automatic control would be implemented kernel-side (with algos possibly shared
across drivers) and reuse existing stateful controls. The advantage is
simplicity (userspace does not need to carry its own rate-control
implementation) and to ensure that there is a built-in mechanism for common
strategies available for every driver (no mandatory dependency on a proprietary
userspace stack). There may also be extra statistics or controls available to
the driver that allow finer-grain control.
Though not controlling the GOP (or no gop) might require a bit more work on
driver side. Today, we do have queues of request, queues of buffer etc. But it
is still quite difficult to do lookahead these queues. That is only useful if
rate control algorithm can use future frame type (like keyframe) to make
decisions. That could be me pushing to far here though.
Yes I agree the interaction between userspace GOP control and kernel-side
rate-contrly might be quite tricky without any indication of what the next frame
types will be.
Maybe we could only allow explicit frame type configuration when using manual
rate-control and have kernel-side GOP management when in-kernel rc is used
(and we can allow it with manual rate-control too). I like having this option
because it allows for simple userspace implementations.
Note that this could perhaps also be added as an optional feature
for stateful encoders since some of them seem to be able to instruct the
firmware what frame type to use (in addition to directly controlling QP).
There's also a good chance that this feature is not available when using
a firmware-backed rc algorithm.
Manual control allows userspace to get creative and requires the ability to set
the quantization parameter (QP) directly for each frame (controls are already
as many stateful encoders also support it).
# Regions of Interest
Regions of interest (ROIs) allow specifying sub-regions of the frame that should
be prioritized for quality. Stateless encoders typically support a limited
number and allow setting specific QP values for these regions.
While the QP value should be used directly in manual rate-control, we probably
want to have some "level of importance" setting for kernel-side rate-control,
along with the dimensions/position of each ROI. This could be expressed with
a new structure containing all these elements and presented as a variable-sized
array control with as many elements as the hardware can support.
Do you see any difference in ROI for stateful and stateless ? This looks like a
feature we could combined. Also, ROI exist for cameras too, I'd probably try and
keep them separate though.
I feel like the stateful/stateless behavior should be the same, so that could be
a shared control too. Also we could use a QP delta which would apply to both
manual and in-kernel rate-control, but maybe that's too low-level in the latter
case (not very obvious when a relevant delta could be when userspace has no idea
of the current frame-wide QP value).
This is a very good overview of the hard work ahead of us. Looking forward on
this journey and your Allwinner driver.
Thanks a lot for your input!
Honestly I was expecting that it would be more difficult than decoding, but it
turns out it might not be the case.
Such rate control or quailty report would be complete vendor special.
We just need a method that let driver report those encoding statistic to the
userspace.
Returning the encoded bitstream size is perfectly generic and available to
every encoder. Maybe we could also return some average QP value since that
seems quite common. Other than that the rest should be kept in-kernel so we
can have a generic API.
You just throw the tools that a hardware could offer away.
Also it seems that the Hantro H1 specific mechanism (checkpoint-based) is not
necessarily a lot better than regular QP-wide settings.
Macroblock level QP control in Hantro H1 is very useful. For FOSS, those
vendor special statistic or controlling may not be necessary, while the
real product is not simple.
Cheers,
Paul
-- Paul Kocialkowski, Bootlin Embedded Linux and kernel engineering
https://bootlin.com
--
Hsia-Jun(Randy) Li
