Re: [RFC PATCH 0/1] Add LTR controls

From: dikshita
Date: Thu Aug 06 2020 - 08:18:21 EST

On 2020-07-22 19:56, Hans Verkuil wrote:
On 20/07/2020 16:33, dikshita@xxxxxxxxxxxxxx wrote:
On 2020-07-16 14:25, Hans Verkuil wrote:
On 16/06/2020 20:41, dikshita@xxxxxxxxxxxxxx wrote:
Hi Hans, Nicolas,

Thanks for your comments.

On 2020-06-12 14:41, Hans Verkuil wrote:
Hi Dikshita, Nicolas,

On 11/06/2020 16:22, Nicolas Dufresne wrote:
Le jeudi 11 juin 2020 à 15:55 +0530, Dikshita Agarwal a écrit :
LTR (Long Term Reference) frames are the frames that are encoded
sometime in the past
and stored in the DPB buffer list to be used as reference to encode
future frames.
One usage of LTR encoding is to reduce error propagation for video
in packet lossy networks.  For example, encoder may want to specify
some key frames as
LTR pictures and use them as reference frames for encoding. With
extra protection
selectively on these LTR frames or synchronization with the receiver
of reception of
the LTR frames during transmission, decoder can receive reference
frames more reliably
than other non-reference frames. As a result, transmission error can
be effectively
restricted within certain frames rather than propagated to future

We are introducing below V4l2 Controls for this feature
    a. This is used to query or configure the number of LTR frames.
       This is a static control and is controlled by the client.
    b. The LTR index varies from 0 to the max LTR-1.
    c. If LTR Count is more than max supported LTR count (max LTR) by
driver, it will be rejected.
    d. Auto Marking : If LTR count is non zero,
        1) first LTR count frames would be mark as LTR automatically
             every IDR frame (inclusive).
        2) For multilayer encoding: first LTR count base layer
reference frames starting after
           every IDR frame (inclusive) in encoding order would be
marked as LTR frames by the encoder.
        3) Auto marking of LTR due to IDR should consider following
            1. The frame is not already set to be marked as LTR.
            2. The frame is part of the base layer in the
hierarchical layer case.
            3. The number of frames currently marked as LTR is less
than the maximum LTR frame index plus 1.
    e. Encoder needs to handle explicit Mark/Use command when encoder
is still doing "auto" marking

I don't follow this, quite possibly due to lack of experience with

I kind of would expect to see two modes: either automatic where
encoders can
mark up to LTR_COUNT frames as long term reference, and userspace just
LTR_COUNT and doesn't have to do anything else.

Or it is manual mode where userspace explicitly marks long term

From the proposal above it looks like you can mix auto and manual

BTW, how do you 'unmark' long term reference frames?

This feature is for stateful encoders, right?

Perhaps we are missing a LONG_TERM_REFERENCE_MODE ? I bet some encoder
can select by themself long term references and even some encoders may
not let the user decide.

(not huge han of LTR acronyme, but that could be fine too, assuming
add more _).

Userspace sets LTR count which signifies the number of LTR frames
encoder needs to generate or keep.
The encoder has to build-up its internal buffer reference list (aka DBP
list or recon buffer list).
In order to achieve that encoder will fill It's LTR (long term
references) list and STR (short term references) list
by auto marking n frames as LTR frames(n is equal to LTR count) based on
auto-marking dictated by the encoder spec.
The client then can replace those automatically marked frames with new
frames using V4L2_CID_MPEG_VIDEO_MARKLTRFRAME and can ask
encoder to refer the newly marked frame for encoding the next frame

    a. This signals to mark the current frame as LTR frame. It is a
dynamic control and also provide the LTR index to be used.
    b. the LTR index provided by this control should never exceed the
max LTR-1. Else it will be rejected.

The "current" frame seems a bit loose. Perhaps you wanted to use
flags ? A bit like what we have to signal TOP/BOTTOM fields in
alternate interlacing.

I was thinking the same thing. Using a control for this doesn't seem

the client sets this to replace automatically marked frames by the
encoder with a particular frame.
this provides an index that ranges from 0 to LTR count-1 and then the
particular frame will be marked with that index.
this can be achieved through request by associating this control with a
specific buffer to make it synchronized.

    a. This specifies the LTR frame(s) to be used for encoding the
current frame. This is a dynamic control.
    b. LTR Use Bitmap : this consists of bits [0, 15]. A total of N
LSB bits of this field are valid,
       where N is the maximum number of LTRs supported. All the other
bits are invalid and should be rejected.
       The LSB corresponds to the LTR index 0. Bit N-1 from the LSB
corresponds to the LTR index max LTR-1.

How would userspace know this? Esp. with auto marking since userspace
would have
to predict how auto marking works (if I understand this correctly).

Client sets LTR count which tells about the number of LTR frames
automatically marked by the encoder.
so client can use LTR index (0 to LTR count -1) to ask encoder to refer
any particular
frame (marked automatically by driver or marked by client with
V4L2_CID_MPEG_VIDEO_MARKLTRFRAME) as a reference to encode the next

For which HW encoder is this meant?

This is primarily meant for H264 and HEVC.

The venus encoder?

Some more questions:

1) How many LTR frames do h.264 and hevc allow?
According to spec, MAX LTR allowed by H.264 is 15 and HEVC allows max 32 LTR frames.
2) Given N LTR frames, is there a ordering of those frames? E.g.
   the LTR frame with index 0 is processed/used differently from
   LTR frame with index 1? Or are they all equal in that it is just a pool
   of LTR frames that the encoder can use as it wishes?
they are different frames marked to be used as LTR and stored in available indices.
Userspace notifies encoder which LTR frame index to use via USE LTR control.

One more question: if a frame is marked as a LTR frame, does that mean
that userspace
can't reuse the buffer containing that frame? I assume that's the case.

One thing that I don't like about this API is that it introduces the
LTR indices.
Would it be possible to instead use buffer indices? (i.e. the
v4l2_buffer index field).
Then you can use a buffer flag to indicate that a buffer should be an
LTR buffer,
or (for automarking) the driver can mark a buffer as an LTR buffer by
setting this flag
before returning the buffer to userspace.

You would still need a V4L2_CID_MPEG_VIDEO_USELTRFRAME, but that can
be a bitmask
using the buffer index.

This would avoid introducing a second method of referring to buffers
(LTR index),
and I think that V4L2_CID_MPEG_VIDEO_MARKLTRFRAME can be dropped.



Hi Hans,
LTR indices are part of the spec, we are not introducing anything new.
Also, there is no relation between v4l2 buffer index and LTR index. So we can't use V4l2 index as LTR index.

below are some details about LTR from spec:

Regarding LTR count:
max_num_ref_frames specifies the maximum number of short-term and long-term reference frames, complementary reference field pairs, and non-paired reference fields that may be used by the decoding process for inter prediction of any picture in the coded video sequence. The value of max_num_ref_frames shall be in the range of 0 to MaxDpbFrames (as specified in clause A.3.1 or A.3.2), inclusive.
max_long_term_frame_idx_plus1 minus 1 specifies the maximum value of long-term frame index allowed for long-term reference pictures (until receipt of another value of max_long_term_frame_idx_plus1). The value of max_long_term_frame_idx_plus1 shall be in the range of 0 to max_num_ref_frames, inclusive.

Regarding LTR index which will be specified by V4L2_CID_MPEG_VIDEO_MARKLTRFRAM :
long_term_pic_num specifies the long-term picture number of the picture being moved to the current index in the list. When decoding a coded frame, long_term_pic_num shall be equal to a LongTermPicNum assigned to one of the reference frames or complementary reference field pairs marked as "used for long-term reference".
long_term_frame_idx is used (with memory_management_control_operation equal to 3 or 6) to assign a long-term frame index to a picture. When the associated memory_management_control_operation is processed by the decoding process, the value of long_term_frame_idx shall be in the range of 0 to MaxLongTermFrameIdx, inclusive.

Also As there can be multiple LTR present. We can instruct encoder to use one and/or multiple of them using V4L2_CID_MPEG_VIDEO_USELTRFRAME with the bitmask.






Note, I haven't captured very well the userspace control flow, perhaps
this could be enhanced through writing some documentation.

As per all other generic encoder controls, we need to make sure it
be usable and flexible enough for multiple HW blocks, as it can be
tedious to extend later otherwise. It is important that along with
RFC you provide some comparisons with with other HW / SW APIs in order
to help justify the design decisions. I also think there should be
link made V4L2_CID_MPEG_VIDEO_GOP_* , number of B-Frames etc.

I agree with Nicolas.




Dikshita Agarwal (1):
  media: v4l2-ctrls:  add control for ltr

 drivers/media/v4l2-core/v4l2-ctrls.c | 6 ++++++
 include/uapi/linux/v4l2-controls.h   | 4 ++++
 2 files changed, 10 insertions(+)