Internet Engineering Task Force (IETF) S. Wenger
Request for Comments: 8082 J. Lennox
Updates: 5104 Vidyo, Inc.
Category: Standards Track B. Burman
ISSN: 2070-1721 M. Westerlund
March 2017 Using Codec Control Messages in the RTP Audio-Visual Profile with
Feedback with Layered Codecs
This document updates RFC 5104 by fixing a shortcoming in the
specification language of the Codec Control Message Full Intra
Request (FIR) description when using it with layered codecs. In
particular, a decoder refresh point needs to be sent by a media
sender when a FIR is received on any layer of the layered bitstream,
regardless of whether those layers are being sent in a single or in
multiple RTP flows. The other payload-specific feedback messages
defined in RFC 5104 and RFC 4585 (which was updated by RFC 5506) have
also been analyzed, and no corresponding shortcomings have been
Status of This Memo
This is an Internet Standards Track document.
This document is a product of the Internet Engineering Task Force
(IETF). It represents the consensus of the IETF community. It has
received public review and has been approved for publication by the
Internet Engineering Steering Group (IESG). Further information on
Internet Standards is available in Section 2 of RFC 7841.
Information about the current status of this document, any errata,
and how to provide feedback on it may be obtained at
1. Introduction and Problem Statement
The "Extended RTP Profile for Real-time Transport Control Protocol
(RTCP)-Based Feedback (RTP/AVPF)" [RFC4585] and "Codec Control
Messages in the RTP Audio-Visual Profile with Feedback (AVPF)"
[RFC5104] specify a number of payload-specific feedback messages that
a media receiver can use to inform a media sender of certain
conditions or to make certain requests. The feedback messages are
being sent as RTCP receiver reports, and RFC 4585 specifies timing
rules that make the use of those messages practical for time-
sensitive codec control.
Since the time those RFCs were developed, layered codecs have gained
in popularity and deployment. Layered codecs use multiple sub-
bitstreams called "layers" to represent the content in different
fidelities. Depending on the media codec and its RTP payload format
in use, a number of options exist on how to transport those layers in
RTP. Summarizing "A Taxonomy of Semantics and Mechanisms for Real-
Time Transport Protocol (RTP) Sources" [RFC7656]):
single layers or groups of layers may be sent in their own RTP
streams in Multiple RTP streams on a Single media Transport (MRST)
or Multiple RTP streams on Multiple media Transports (MRMT) mode;
using media-codec specific multiplexing mechanisms, multiple
layers may be sent in a single RTP stream in Single RTP stream on
a Single media Transport (SRST) mode.
The dependency relationship between layers in a truly layered,
pyramid-shaped bitstream forms a directed graph, with the base layer
at the root. Enhancement layers depend on the base layer and
potentially on other enhancement layers, and the target layer and all
layers it depends on have to be decoded jointly in order to recreate
the uncompressed media signal at the fidelity of the target layer.
Such a layering structure is assumed henceforth; for more exotic
layering structures, please see Section 5.
Implementation experience has shown that the Full Intra Request (FIR)
command as defined in [RFC5104] is underspecified when used with
layered codecs and when more than one RTP stream is used to transport
the layers of a layered bitstream at a given fidelity. In
particular, from the [RFC5104] specification language, it is not
clear whether a FIR received for only a single RTP stream of multiple
RTP streams covering the same layered bitstream necessarily triggers
the sending of a decoder refresh point (as defined in [RFC5104],
Section 2.2) for all layers, or only for the layer that is
transported in the RTP stream that the FIR request is associated
This document fixes this shortcoming by:
a. Updating the definition of the decoder refresh point (as defined
in [RFC5104], Section 2.2) to cover layered codecs, in line with
the corresponding definitions used in a popular layered codec
format, namely H.264/SVC (Scalable Video Coding) [H.264].
Specifically, a decoder refresh point, in conjunction with
layered codecs, resets the state of the whole decoder, which
implies that it includes hard or gradual single-layer decoder
refresh for all layers;
b. Requiring a media sender to send a decoder refresh point after
the media sender has received a FIR over an RTCP stream
associated with any of the RTP streams over which a part of the
layered bitstream is transported;
c. Requiring that a media receiver send the FIR on the RTCP stream
associated with the base layer. The option of receiving FIR on
the enhancement-layer-associated RTCP stream as specified in
point b) above is kept for backward compatibility; and
d. Providing guidance on how to detect that a layered bitstream is
in use for which the above rules apply.
While, clearly, the reaction to FIR for layered codecs in [RFC5104]
and the companion documents is underspecified, it appears that this
is not the case for any of the other payload-specific codec control
messages defined in [RFC4585] and [RFC5104]. A brief summary of the
analysis that led to this conclusion is also included in this
2. Requirements Language
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this
document are to be interpreted as described in RFC 2119 [RFC2119].
3. Updated Definition of Decoder Refresh Point
The remainder of this section replaces the definition of decoder
refresh point in Section 2.2 of [RFC5104] in its entirety.
Decoder Refresh Point: A bit string, packetized in one or more RTP
packets, that completely resets the decoder to a known state.
Examples for "hard" single-layer decoder refresh points are Intra
pictures in H.261 [H.261], H.263 [H.263], MPEG-1 [MPEG-1], MPEG-2
[MPEG-2], and MPEG-4 [MPEG-4]; Instantaneous Decoder Refresh (IDR)
pictures in H.264 [H.264] and H.265 [H.265]; and keyframes in VP8
[RFC6386] and VP9 [VP9-BITSTREAM]. "Gradual" decoder refresh points
may also be used; see, for example, H.264 [H.264]. While both "hard"
and "gradual" decoder refresh points are acceptable in the scope of
this specification, in most cases the user experience will benefit
from using a "hard" decoder refresh point.
A decoder refresh point also contains all header information above
the syntactical level of the picture layer that is conveyed in-band.
In [H.264], for example, a decoder refresh point contains those
parameter set Network Adaptation Layer (NAL) units that generate
parameter sets necessary for the decoding of the following slice/data
partition NAL units. (That is, assuming the parameter sets have not
been conveyed out of band.)
When a layered codec is in use, the above definition -- in
particular, the requirement to completely reset the decoder to a
known state -- implies that the decoder refresh point includes hard
or gradual single-layer decoder refresh points for all layers.
4. Full Intra Request for Layered Codecs
A media receiver or middlebox may decide to send a FIR command based
on the guidance provided in Section 4.3.1 of [RFC5104]. When sending
the FIR command, it MUST target the RTP stream that carries the base
layer of the layered bitstream, and this is done by setting the
Feedback Control Information (FCI) (and, in particular, the
synchronization source (SSRC) field therein) to refer to the SSRC of
the forward RTP stream that carries the base layer.
When a Full Intra Request command is received by the designated media
sender in the RTCP stream associated with any of the RTP streams in
which any layer of a layered bitstream are sent, the designated media
sender MUST send a decoder refresh point (Section 3) as defined above
at its earliest opportunity. The requirements related to congestion
control on the forward RTP streams as specified in Sections 3.5.1 and
5 of [RFC5104] apply for the RTP streams both in isolation and
Note: the requirement to react to FIR commands associated with
enhancement layers is included for robustness and backward-
5. Identifying the Use of Layered Bitstreams (Informative)
The above modifications to RFC 5104 unambiguously define how to deal
with FIR commands when layered bitstreams are in use. However, it is
surprisingly difficult to identify the use of a layered bitstream.
In general, it is expected that implementers know when layered
bitstreams (in its commonly understood sense: with inter-layer
prediction between pyramid-arranged layers) are in use and when not
and can therefore implement the above updates to RFC 5104 correctly.
However, there are scenarios in which layered codecs are employed
creating non-pyramid-shaped bitstreams. Those scenarios may be
viewed as somewhat exotic today but clearly are supported by certain
video coding syntaxes, such as H.264/SVC. When blindly applying the
above rules to those non-pyramid-arranged layering structures,
suboptimal system behavior would result. Nothing would break, and
there would not be an interoperability failure, but the user
experience may suffer through the sending or receiving of decoder
refresh points at times or on parts of the bitstream that are
unnecessary from a user experience viewpoint. Therefore, this
informative section is included that provides the current
understanding of when a layered bitstream is in use and when not.
The key observation made here is that the RTP payload format
negotiated for the RTP streams, in isolation, is not necessarily an
indicator for the use of a layered bitstream. Some layered codecs
(including H.264/SVC) can form decodable bitstreams including only
(one or more) enhancement layers, without the base layer, effectively
creating simulcastable sub-bitstreams within a single scalable
bitstream (as defined in the video coding standard), but without
inter-layer prediction. In such a scenario, it is potentially,
though not necessarily, counterproductive to send a decoder refresh
point on all layers for that payload format and media source. It is
beyond the scope of this document to discuss optimized reactions to
FIRs received on RTP streams carrying such exotic bitstreams.
One good indication of the likely use of pyramid-shaped layering with
inter-layer prediction is when the various RTP streams are "bound"
together on the signaling level. In an SDP environment, this would
be the case if they are marked as being dependent on each other using
"The Session Description Protocol (SDP) Grouping Framework" [RFC5888]
and layer dependency [RFC5583].
6. Layered Codecs and Non-FIR Codec Control Messages (Informative)
Between them, AVPF [RFC4585] and Codec Control Messages [RFC5104]
define a total of seven payload-specific feedback messages. For the
FIR command message, guidance has been provided above. In this
section, some information is provided with respect to the remaining
six codec control messages.
6.1. Picture Loss Indication (PLI)
PLI is defined in Section 6.3.1 of [RFC4585]. The prudent response
to a PLI message received for an enhancement layer is to "repair"
that enhancement layer and all dependent enhancement layers through
appropriate source-coding-specific means. However, the reference
layer or layers used by the enhancement layer for which the PLI was
received do not require repair. The encoder can figure out by itself
what constitutes a dependent enhancement layer and does not need help
from the system stack in doing so. Thus, there is nothing that needs
to be specified herein.
6.2. Slice Loss Indication (SLI)
SLI is defined in Section 6.3.2 of [RFC4585]. The current
understanding is that the prudent response to an SLI message received
for an enhancement layer is to "repair" the affected spatial area of
that enhancement layer and all dependent enhancement layers through
appropriate source-coding-specific means. As in PLI, the reference
layers used by the enhancement layer for which the SLI was received
do not need to be repaired. Again, as in PLI, the encoder can
determine by itself what constitutes a dependent enhancement layer
and does not need help from the system stack in doing so. Thus,
there is nothing that needs to be specified herein. SLI has seen
very little implementation and, as far as it is known, none in
conjunction with layered systems.
6.3. Reference Picture Selection Indication (RPSI)
RPSI is defined in Section 6.3.3 of [RFC4585]. While a technical
equivalent of RPSI has been in use with non-layered systems for many
years, no implementations are known in conjunction of layered codecs.
The current understanding is that the reception of an RPSI message on
any layer indicating a missing reference picture forces the encoder
to appropriately handle that missing reference picture in the layer
indicated, and in all dependent layers. Thus, RPSI should work
without further need for specification language.
6.4. Temporal-Spatial Trade-Off Request and Notification (TSTR/TSTN)
TSTR/TSTN are defined in Sections 4.3.2 and 4.3.3 of [RFC5104],
respectively. The TSTR request communicates guidance of the
preferred trade-off between spatial quality and frame rate. A
technical equivalent of TSTR/TSTN has seen deployment for many years
in non-scalable systems.
TSTR and TSTN messages include an SSRC target, which, similarly to
FIR, may refer to an RTP stream carrying a base layer, an enhancement
layer, or multiple layers. Therefore, the current understanding is
that the semantics of the message applies to the layers present in
the targeted RTP stream.
It is noted that per-layer TSTR/TSTN is a mechanism that is, in some
ways, counterproductive in a system using layered codecs. Given a
sufficiently complex layered bitstream layout, a sending system has
flexibility in adjusting the spatio/temporal quality balance by
adding and removing temporal, spatial, or quality enhancement layers.
At present, it is unclear whether an allowed (or even recommended)
option to the reception of a TSTR is to adjust the bit allocation
within the layer(s) present in the addressed RTP stream or to adjust
the layering structure accordingly -- which can involve more than
just the addressed RTP stream.
Until there is a sufficient critical mass of implementation practice,
it is probably prudent for an implementer not to assume either of the
two options or any middle ground that may exist between the two.
Instead, it is suggested that an implementation be liberal in
accepting TSTR messages and upon receipt, responding in TSTN
indicating "no change". Further, it is suggested that new
implementations do not send TSTR messages except when operating in
SRST mode as defined in [RFC7656]. Finally, implementers are
encouraged to contribute to the IETF documentation of any
implementation requirements that make per-layer TSTR/TSTN useful.
6.5. H.271 Video Back Channel Message (VBCM)
VBCM is defined in Section 4.3.4 of [RFC5104]. What was said above
for RPSI (Section 6.3) applies here as well.
7. IANA Considerations
This memo includes no request to IANA.
The authors want to thank Mo Zanaty for useful discussions.
SE - 164 80 Kista
SE - 164 80 Kista