doc/dev_guide/av1_encoder.dox

/*!\page encoder_guide AV1 ENCODER GUIDE

\tableofcontents

\section architecture_introduction Introduction

This document provides an architectural overview of the libaom AV1 encoder.

It is intended as a high level starting point for anyone wishing to contribute
to the project, that will help them to more quickly understand the structure
of the encoder and find their way around the codebase.

It stands above and will where necessary link to more detailed function
level documents.

\subsection  architecture_gencodecs Generic Block Transform Based Codecs

Most modern video encoders including VP8, H.264, VP9, HEVC and AV1
(in increasing order of complexity) share a common basic paradigm. This
comprises separating a stream of raw video frames into a series of discrete
blocks (of one or more sizes), then computing a prediction signal and a
quantized, transform coded, residual error signal. The prediction and residual
error signal, along with any side information needed by the decoder, are then
entropy coded and packed to form the encoded bitstream. See Figure 1: below,
where the blue blocks are, to all intents and purposes, the lossless parts of
the encoder and the red block is the lossy part.

This is of course a gross oversimplification, even in regard to the simplest
of the above codecs.  For example, all of them allow for block based
prediction at multiple different scales (i.e. different block sizes) and may
use previously coded pixels in the current frame for prediction or pixels from
one or more previously encoded frames. Further, they may support multiple
different transforms and transform sizes and quality optimization tools like
loop filtering.

\image html genericcodecflow.png "" width=70%

\subsection architecture_av1_structure AV1 Structure and Complexity

As previously stated, AV1 adopts the same underlying paradigm as other block
transform based codecs. However, it is much more complicated than previous
generation codecs and supports many more block partitioning, prediction and
transform options.

AV1 supports block partitions of various sizes from 128x128 pixels down to 4x4
pixels using a multi-layer recursive tree structure as illustrated in figure 2
below.

\image html av1partitions.png "" width=70%

AV1 also provides 71 basic intra prediction modes, 56 single frame inter prediction
modes (7 reference frames x 4 modes x 2 for OBMC (overlapped block motion
compensation)), 12768 compound inter prediction modes (that combine inter
predictors from two reference frames) and 36708 compound inter / intra
prediction modes. Furthermore, in addition to simple inter motion estimation,
AV1 also supports warped motion prediction using affine transforms.

In terms of transform coding, it has 16 separable 2-D transform kernels
\f$(DCT, ADST, fADST, IDTX)^2\f$ that can be applied at up to 19 different
scales from 64x64 down to 4x4 pixels.

When combined together, this means that for any one 8x8 pixel block in a
source frame, there are approximately 45,000,000 different ways that it can
be encoded.

Consequently, AV1 requires complex control processes. While not necessarily
a normative part of the bitstream, these are the algorithms that turn a set
of compression tools and a bitstream format specification, into a coherent
and useful codec implementation. These may include but are not limited to
things like :-

- Rate distortion optimization (The process of trying to choose the most
  efficient combination of block size, prediction mode, transform type
  etc.)
- Rate control (regulation of the output bitrate)
- Encoder speed vs quality trade offs.
- Features such as two pass encoding or optimization for low delay
  encoding.

For a more detailed overview of AV1's encoding tools and a discussion of some
of the design considerations and hardware constraints that had to be
accommodated, please refer to <a href="https://arxiv.org/abs/2008.06091">
A Technical Overview of AV1</a>.

Figure 3 provides a slightly expanded but still simplistic view of the
AV1 encoder architecture with blocks that relate to some of the subsequent
sections of this document. In this diagram, the raw uncompressed frame buffers
are shown in dark green and the reconstructed frame buffers used for
prediction in light green. Red indicates those parts of the codec that are
(or may be) lossy, where fidelity can be traded off against compression
efficiency, whilst light blue shows algorithms or coding tools that are
lossless. The yellow blocks represent non-bitstream normative configuration
and control algorithms.

\image html av1encoderflow.png "" width=70%

\section architecture_command_line The Libaom Command Line Interface

 Add details or links here: TODO ? elliotk@

\section architecture_enc_data_structures Main Encoder Data Structures

The following are the main high level data structures used by the libaom AV1
encoder and referenced elsewhere in this overview document:

- \ref AV1_PRIMARY
    - \ref AV1_PRIMARY.gf_group (\ref GF_GROUP)
    - \ref AV1_PRIMARY.lap_enabled
    - \ref AV1_PRIMARY.twopass (\ref TWO_PASS)
    - \ref AV1_PRIMARY.p_rc (\ref PRIMARY_RATE_CONTROL)
    - \ref AV1_PRIMARY.tf_info (\ref TEMPORAL_FILTER_INFO)

- \ref AV1_COMP
    - \ref AV1_COMP.oxcf (\ref AV1EncoderConfig)
    - \ref AV1_COMP.rc (\ref RATE_CONTROL)
    - \ref AV1_COMP.speed
    - \ref AV1_COMP.sf (\ref SPEED_FEATURES)

- \ref AV1EncoderConfig (Encoder configuration parameters)
    - \ref AV1EncoderConfig.pass
    - \ref AV1EncoderConfig.algo_cfg (\ref AlgoCfg)
    - \ref AV1EncoderConfig.kf_cfg (\ref KeyFrameCfg)
    - \ref AV1EncoderConfig.rc_cfg (\ref RateControlCfg)

- \ref AlgoCfg (Algorithm related configuration parameters)
    - \ref AlgoCfg.arnr_max_frames
    - \ref AlgoCfg.arnr_strength

- \ref KeyFrameCfg (Keyframe coding configuration parameters)
    - \ref KeyFrameCfg.enable_keyframe_filtering

- \ref RateControlCfg (Rate control configuration)
    - \ref RateControlCfg.mode
    - \ref RateControlCfg.target_bandwidth
    - \ref RateControlCfg.best_allowed_q
    - \ref RateControlCfg.worst_allowed_q
    - \ref RateControlCfg.cq_level
    - \ref RateControlCfg.under_shoot_pct
    - \ref RateControlCfg.over_shoot_pct
    - \ref RateControlCfg.maximum_buffer_size_ms
    - \ref RateControlCfg.starting_buffer_level_ms
    - \ref RateControlCfg.optimal_buffer_level_ms
    - \ref RateControlCfg.vbrbias
    - \ref RateControlCfg.vbrmin_section
    - \ref RateControlCfg.vbrmax_section

- \ref PRIMARY_RATE_CONTROL (Primary Rate control status)
    - \ref PRIMARY_RATE_CONTROL.gf_intervals[]
    - \ref PRIMARY_RATE_CONTROL.cur_gf_index

- \ref RATE_CONTROL (Rate control status)
    - \ref RATE_CONTROL.intervals_till_gf_calculate_due
    - \ref RATE_CONTROL.frames_till_gf_update_due
    - \ref RATE_CONTROL.frames_to_key

- \ref TWO_PASS (Two pass status and control data)

- \ref GF_GROUP (Data related to the current GF/ARF group)

- \ref FIRSTPASS_STATS (Defines entries in the first pass stats buffer)
    - \ref FIRSTPASS_STATS.coded_error

- \ref SPEED_FEATURES (Encode speed vs quality tradeoff parameters)
    - \ref SPEED_FEATURES.hl_sf (\ref HIGH_LEVEL_SPEED_FEATURES)

- \ref HIGH_LEVEL_SPEED_FEATURES
    - \ref HIGH_LEVEL_SPEED_FEATURES.recode_loop
    - \ref HIGH_LEVEL_SPEED_FEATURES.recode_tolerance

- \ref TplParams

\section architecture_enc_use_cases Encoder Use Cases

The libaom AV1 encoder is configurable to support a number of different use
cases and rate control strategies.

The principle use cases for which it is optimised are as follows:

 - <b>Video on Demand / Streaming</b>
 - <b>Low Delay or Live Streaming</b>
 - <b>Video Conferencing / Real Time Coding (RTC)</b>
 - <b>Fixed Quality / Testing</b>

Other examples of use cases for which the encoder could be configured but for
which there is less by way of specific optimizations include:

 - <b>Download and Play</b>
 - <b>Disk Playback</b>>
 - <b>Storage</b>
 - <b>Editing</b>
 - <b>Broadcast video</b>

Specific use cases may have particular requirements or constraints. For
example:

<b>Video Conferencing:</b>  In a video conference we need to encode the video
in real time and to avoid any coding tools that could increase latency, such
as frame look ahead.

<b>Live Streams:</b> In cases such as live streaming of games or events, it
may be possible to allow some limited buffering of the video and use of
lookahead coding tools to improve encoding quality. However,  whilst a lag of
a second or two may be fine given the one way nature of this type of video,
it is clearly not possible to use tools such as two pass coding.

<b>Broadcast:</b> Broadcast video (e.g. digital TV over satellite) may have
specific requirements such as frequent and regular key frames (e.g. once per
second or more) as these are important as entry points to users when switching
channels. There may also be  strict upper limits on bandwidth over a short
window of time.

<b>Download and Play:</b> Download and play applications may have less strict
requirements in terms of local frame by frame rate control but there may be a
requirement to accurately hit a file size target for the video clip as a
whole. Similar considerations may apply to playback from mass storage devices
such as DVD or disk drives.

<b>Editing:</b> In certain special use cases such as offline editing, it may
be desirable to have very high quality and data rate but also very frequent
key frames or indeed to encode the video exclusively as key frames. Lossless
video encoding may also be required in this use case.

<b>VOD / Streaming:</b> One of the most important and common use cases for AV1
is video on demand or streaming, for services such as YouTube and Netflix. In
this use case it is possible to do two or even multi-pass encoding to improve
compression efficiency. Streaming services will often store many encoded
copies of a video at different resolutions and data rates to support users
with different types of playback device and bandwidth limitations.
Furthermore, these services support dynamic switching between multiple
streams, so that they can respond to changing network conditions.

Exact rate control when encoding for a specific format (e.g 360P or 1080P on
YouTube) may not be critical, provided that the video bandwidth remains within
allowed limits. Whilst a format may have a nominal target data rate, this can
be considered more as the desired average egress rate over the video corpus
rather than a strict requirement for any individual clip. Indeed, in order
to maintain optimal quality of experience for the end user, it may be
desirable to encode some easier videos or sections of video at a lower data
rate and harder videos or sections at a higher rate.

VOD / streaming does not usually require very frequent key frames (as in the
broadcast case) but key frames are important in trick play (scanning back and
forth to different points in a video) and for adaptive stream switching. As
such, in a use case like YouTube, there is normally an upper limit on the
maximum time between key frames of a few seconds, but within certain limits
the encoder can try to align key frames with real scene cuts.

Whilst encoder speed may not seem to be as critical in this use case, for
services such as YouTube, where millions of new videos have to be encoded
every day, encoder speed is still important, so libaom allows command line
control of the encode speed vs quality trade off.

<b>Fixed Quality / Testing Mode:</b> Libaom also has a fixed quality encoder
pathway designed for testing under highly constrained conditions.

\section architecture_enc_speed_quality Speed vs Quality Trade Off

In any modern video encoder there are trade offs that can be made in regard to
the amount of time spent encoding a video or video frame vs the quality of the
final encode.

These trade offs typically limit the scope of the search for an optimal
prediction / transform combination with faster encode modes doing fewer
partition, reference frame, prediction mode and transform searches at the cost
of some reduction in coding efficiency.

The pruning of the size of the search tree is typically based on assumptions
about the likelihood of different search modes being selected based on what
has gone before and features such as the dimensions of the video frames and
the Q value selected for encoding the frame. For example certain intra modes
are less likely to be chosen at high Q but may be more likely if similar
modes were used for the previously coded blocks above and to the left of the
current block.

The speed settings depend both on the use case (e.g. Real Time encoding) and
an explicit speed control passed in on the command line as <b>--cpu-used</b>
and stored in the \ref AV1_COMP.speed field of the main compressor instance
data structure (<b>cpi</b>).

The control flags for the speed trade off are stored the \ref AV1_COMP.sf
field of the compressor instancve and are set in the following functions:-

- \ref av1_set_speed_features_framesize_independent()
- \ref av1_set_speed_features_framesize_dependent()
- \ref av1_set_speed_features_qindex_dependent()

A second factor impacting the speed of encode is rate distortion optimisation
(<b>rd vs non-rd</b> encoding).

When rate distortion optimization is enabled each candidate combination of
a prediction mode and transform coding strategy is fully encoded and the
resulting error (or distortion) as compared to the original source and the
number of bits used, are passed to a rate distortion function. This function
converts the distortion and cost in bits to a single <b>RD</b> value (where
lower is better). This <b>RD</b> value is used to decide between different
encoding strategies for the current block where, for example, a one may
result in a lower distortion but a larger number of bits.

The calculation of this <b>RD</b> value is broadly speaking as follows:

\f[
  RD = (&lambda; * Rate) + Distortion
\f]

This assumes a linear relationship between the number of bits used and
distortion (represented by the rate multiplier value <b>&lambda;</b>) which is
not actually valid across a broad range of rate and distortion values.
Typically, where distortion is high, expending a small number of extra bits
will result in a large change in distortion. However, at lower values of
distortion the cost in bits of each incremental improvement is large.

To deal with this we scale the value of <b>&lambda;</b> based on the quantizer
value chosen for the frame. This is assumed to be a proxy for our approximate
position on the true rate distortion curve and it is further assumed that over
a limited range of distortion values, a linear relationship between distortion
and rate is a valid approximation.

Doing a rate distortion test on each candidate prediction / transform
combination is expensive in terms of cpu cycles. Hence, for cases where encode
speed is critical, libaom implements a non-rd pathway where the <b>RD</b>
value is estimated based on the prediction error and quantizer setting.

\section architecture_enc_src_proc Source Frame Processing

\subsection architecture_enc_frame_proc_data Main Data Structures

The following are the main data structures referenced in this section
(see also \ref architecture_enc_data_structures):

- \ref AV1_PRIMARY ppi (the primary compressor instance data structure)
    - \ref AV1_PRIMARY.tf_info (\ref TEMPORAL_FILTER_INFO)

- \ref AV1_COMP cpi (the main compressor instance data structure)
    - \ref AV1_COMP.oxcf (\ref AV1EncoderConfig)

- \ref AV1EncoderConfig (Encoder configuration parameters)
    - \ref AV1EncoderConfig.algo_cfg (\ref AlgoCfg)
    - \ref AV1EncoderConfig.kf_cfg (\ref KeyFrameCfg)

- \ref AlgoCfg (Algorithm related configuration parameters)
    - \ref AlgoCfg.arnr_max_frames
    - \ref AlgoCfg.arnr_strength

- \ref KeyFrameCfg (Keyframe coding configuration parameters)
    - \ref KeyFrameCfg.enable_keyframe_filtering

\subsection architecture_enc_frame_proc_ingest Frame Ingest / Coding Pipeline

 To encode a frame, first call \ref av1_receive_raw_frame() to obtain the raw
 frame data. Then call \ref av1_get_compressed_data() to encode raw frame data
 into compressed frame data. The main body of \ref av1_get_compressed_data()
 is \ref av1_encode_strategy(), which determines high-level encode strategy
 (frame type, frame placement, etc.) and then encodes the frame by calling
 \ref av1_encode(). In \ref av1_encode(), \ref av1_first_pass() will execute
 the first_pass of two-pass encoding, while \ref encode_frame_to_data_rate()
 will perform the final pass for either one-pass or two-pass encoding.

 The main body of \ref encode_frame_to_data_rate() is
 \ref encode_with_recode_loop_and_filter(), which handles encoding before
 in-loop filters (with recode loops \ref encode_with_recode_loop(), or
 without any recode loop \ref encode_without_recode()), followed by in-loop
 filters (deblocking filters \ref loopfilter_frame(), CDEF filters and
 restoration filters \ref cdef_restoration_frame()).

 Except for rate/quality control, both \ref encode_with_recode_loop() and
 \ref encode_without_recode() call \ref av1_encode_frame() to manage the
 reference frame buffers and \ref encode_frame_internal() to perform the
 rest of encoding that does not require access to external frames.
 \ref encode_frame_internal() is the starting point for the partition search
 (see \ref architecture_enc_partitions).

\subsection architecture_enc_frame_proc_tf Temporal Filtering

\subsubsection architecture_enc_frame_proc_tf_overview Overview

Video codecs exploit the spatial and temporal correlations in video signals to
achieve compression efficiency. The noise factor in the source signal
attenuates such correlation and impedes the codec performance. Denoising the
video signal is potentially a promising solution.

One strategy for denoising a source is motion compensated temporal filtering.
Unlike image denoising, where only the spatial information is available,
video denoising can leverage a combination of the spatial and temporal
information. Specifically, in the temporal domain, similar pixels can often be
tracked along the motion trajectory of moving objects. Motion estimation is
applied to neighboring frames to find similar patches or blocks of pixels that
can be combined to create a temporally filtered output.

AV1, in common with VP8 and VP9, uses an in-loop motion compensated temporal
filter to generate what are referred to as alternate reference frames (or ARF
frames). These can be encoded in the bitstream and stored as frame buffers for
use in the prediction of subsequent frames, but are not usually directly
displayed (hence they are sometimes referred to as non-display frames).

The following command line parameters set the strength of the filter, the
number of frames used and determine whether filtering is allowed for key
frames.

- <b>--arnr-strength</b> (\ref AlgoCfg.arnr_strength)
- <b>--arnr-maxframes</b> (\ref AlgoCfg.arnr_max_frames)
- <b>--enable-keyframe-filtering</b>
  (\ref KeyFrameCfg.enable_keyframe_filtering)

Note that in AV1, the temporal filtering scheme is designed around the
hierarchical ARF based pyramid coding structure. We typically apply denoising
only on key frame and ARF frames at the highest (and sometimes the second
highest) layer in the hierarchical coding structure.

\subsubsection architecture_enc_frame_proc_tf_algo Temporal Filtering Algorithm

Our method divides the current frame into "MxM" blocks. For each block, a
motion search is applied on frames before and after the current frame. Only
the best matching patch with the smallest mean square error (MSE) is kept as a
candidate patch for a neighbour frame. The current block is also a candidate
patch. A total of N candidate patches are combined to generate the filtered
output.

Let f(i) represent the filtered sample value and \f$p_{j}(i)\f$ the sample
value of the j-th patch. The filtering process is:

\f[
  f(i) = \frac{p_{0}(i) + \sum_{j=1}^{N} &omega;_{j}(i).p_{j}(i)}
              {1 + \sum_{j=1}^{N} &omega;_{j}(i)}
\f]

where \f$ &omega;_{j}(i) \f$ is the weight of the j-th patch from a total of
N patches. The weight is determined by the patch difference as:

\f[
  &omega;_{j}(i) = exp(-\frac{D_{j}(i)}{h^2})
\f]

where \f$ D_{j}(i) \f$ is the sum of squared difference between the current
block and the j-th candidate patch:

\f[
  D_{j}(i) = \sum_{k\in&Omega;_{i}}||p_{0}(k) - p_{j}(k)||_{2}
\f]

where:
- \f$p_{0}\f$ refers to the current frame.
- \f$&Omega;_{i}\f$ is the patch window, an "LxL" pixel square.
- h is a critical parameter that controls the decay of the weights measured by
  the Euclidean distance. It is derived from an estimate of noise amplitude in
  the source. This allows the filter coefficients to adapt for videos with
  different noise characteristics.
- Usually, M = 32, N = 7, and L = 5, but they can be adjusted.

It is recommended that the reader refers to the code for more details.

\subsubsection architecture_enc_frame_proc_tf_funcs Temporal Filter Functions

The main entry point for temporal filtering is \ref av1_temporal_filter().
This function returns 1 if temporal filtering is successful, otherwise 0.
When temporal filtering is applied, the filtered frame will be held in
the output_frame, which is the frame to be
encoded in the following encoding process.

Almost all temporal filter related code is in av1/encoder/temporal_filter.c
and av1/encoder/temporal_filter.h.

Inside \ref av1_temporal_filter(), the reader's attention is directed to
\ref tf_setup_filtering_buffer() and \ref tf_do_filtering().

- \ref tf_setup_filtering_buffer(): sets up the frame buffer for
  temporal filtering, determines the number of frames to be used, and
  calculates the noise level of each frame.

- \ref tf_do_filtering(): the main function for the temporal
  filtering algorithm. It breaks each frame into "MxM" blocks. For each
  block a motion search \ref tf_motion_search() is applied to find
  the motion vector from one neighboring frame. tf_build_predictor() is then
  called to build the matching patch and \ref av1_apply_temporal_filter_c() (see
  also optimised SIMD versions) to apply temporal filtering. The weighted
  average over each pixel is accumulated and finally normalized in
  \ref tf_normalize_filtered_frame() to generate the final filtered frame.

- \ref av1_apply_temporal_filter_c(): the core function of our temporal
  filtering algorithm (see also optimised SIMD versions).

\subsection architecture_enc_frame_proc_film Film Grain Modelling

 Add details here.

\section architecture_enc_rate_ctrl Rate Control

\subsection architecture_enc_rate_ctrl_data Main Data Structures

The following are the main data structures referenced in this section
(see also \ref architecture_enc_data_structures):

 - \ref AV1_PRIMARY ppi (the primary compressor instance data structure)
    - \ref AV1_PRIMARY.twopass (\ref TWO_PASS)

 - \ref AV1_COMP cpi (the main compressor instance data structure)
    - \ref AV1_COMP.oxcf (\ref AV1EncoderConfig)
    - \ref AV1_COMP.rc (\ref RATE_CONTROL)
    - \ref AV1_COMP.sf (\ref SPEED_FEATURES)

 - \ref AV1EncoderConfig (Encoder configuration parameters)
    - \ref AV1EncoderConfig.rc_cfg (\ref RateControlCfg)

 - \ref FIRSTPASS_STATS *frame_stats_buf (used to store per frame first
   pass stats)

 - \ref SPEED_FEATURES (Encode speed vs quality tradeoff parameters)
    - \ref SPEED_FEATURES.hl_sf (\ref HIGH_LEVEL_SPEED_FEATURES)

\subsection architecture_enc_rate_ctrl_options Supported Rate Control Options

Different use cases (\ref architecture_enc_use_cases) may have different
requirements in terms of data rate control.

The broad rate control strategy is selected using the <b>--end-usage</b>
parameter on the command line, which maps onto the field
\ref aom_codec_enc_cfg_t.rc_end_usage in \ref aom_encoder.h.

The four supported options are:-

- <b>VBR</b> (Variable Bitrate)
- <b>CBR</b> (Constant Bitrate)
- <b>CQ</b> (Constrained Quality mode ; A constrained variant of VBR)
- <b>Fixed Q</b> (Constant quality of Q mode)

The value of \ref aom_codec_enc_cfg_t.rc_end_usage is in turn copied over
into the encoder rate control configuration data structure as
\ref RateControlCfg.mode.

In regards to the most important use cases above, Video on demand uses either
VBR or CQ mode. CBR is the preferred rate control model for RTC and Live
streaming and Fixed Q is only used in testing.

The behaviour of each of these modes is regulated by a series of secondary
command line rate control options but also depends somewhat on the selected
use case, whether 2-pass coding is enabled and the selected encode speed vs
quality trade offs (\ref AV1_COMP.speed and \ref AV1_COMP.sf).

The list below gives the names of the main rate control command line
options together with the names of the corresponding fields in the rate
control configuration data structures.

- <b>--target-bitrate</b> (\ref RateControlCfg.target_bandwidth)
- <b>--min-q</b> (\ref RateControlCfg.best_allowed_q)
- <b>--max-q</b> (\ref RateControlCfg.worst_allowed_q)
- <b>--cq-level</b> (\ref RateControlCfg.cq_level)
- <b>--undershoot-pct</b> (\ref RateControlCfg.under_shoot_pct)
- <b>--overshoot-pct</b> (\ref RateControlCfg.over_shoot_pct)

The following control aspects of vbr encoding

- <b>--bias-pct</b> (\ref RateControlCfg.vbrbias)
- <b>--minsection-pct</b> ((\ref RateControlCfg.vbrmin_section)
- <b>--maxsection-pct</b> ((\ref RateControlCfg.vbrmax_section)

The following relate to buffer and delay management in one pass low delay and
real time coding

- <b>--buf-sz</b> (\ref RateControlCfg.maximum_buffer_size_ms)
- <b>--buf-initial-sz</b> (\ref RateControlCfg.starting_buffer_level_ms)
- <b>--buf-optimal-sz</b> (\ref RateControlCfg.optimal_buffer_level_ms)

\subsection architecture_enc_vbr Variable Bitrate (VBR) Encoding

For streamed VOD content the most common rate control strategy is Variable
Bitrate (VBR) encoding. The CQ mode mentioned above is a variant of this
where additional quantizer and quality constraints are applied.  VBR
encoding may in theory be used in conjunction with either 1-pass or 2-pass
encoding.

VBR encoding varies the number of bits given to each frame or group of frames
according to the difficulty of that frame or group of frames, such that easier
frames are allocated fewer bits and harder frames are allocated more bits. The
intent here is to even out the quality between frames. This contrasts with
Constant Bitrate (CBR) encoding where each frame is allocated the same number
of bits.

Whilst for any given frame or group of frames the data rate may vary, the VBR
algorithm attempts to deliver a given average bitrate over a wider time
interval. In standard VBR encoding, the time interval over which the data rate
is averaged is usually the duration of the video clip.  An alternative
approach is to target an average VBR bitrate over the entire video corpus for
a particular video format (corpus VBR).

\subsubsection architecture_enc_1pass_vbr 1 Pass VBR Encoding

The command line for libaom does allow 1 Pass VBR, but this has not been
properly optimised and behaves much like 1 pass CBR in most regards, with bits
allocated to frames by the following functions:

- \ref av1_calc_iframe_target_size_one_pass_vbr(
           const struct AV1_COMP *const cpi)
       "av1_calc_iframe_target_size_one_pass_vbr()"
- \ref av1_calc_pframe_target_size_one_pass_vbr(
           const struct AV1_COMP *const cpi,
           FRAME_UPDATE_TYPE frame_update_type)
       "av1_calc_pframe_target_size_one_pass_vbr()"

\subsubsection architecture_enc_2pass_vbr 2 Pass VBR Encoding

The main focus here will be on 2-pass VBR encoding (and the related CQ mode)
as these are the modes most commonly used for VOD content.

2-pass encoding is selected on the command line by setting --passes=2
(or -p 2).

Generally speaking, in 2-pass encoding, an encoder will first encode a video
using a default set of parameters and assumptions. Depending on the outcome
of that first encode, the baseline assumptions and parameters will be adjusted
to optimize the output during the second pass.  In essence the first pass is a
fact finding mission to establish the complexity and variability of the video,
in order to allow a better allocation of bits in the second pass.

The libaom 2-pass algorithm is unusual in that the first pass is not a full
encode of the video. Rather it uses a limited set of prediction and transform
options and a fixed quantizer,  to generate statistics about each frame. No
output bitstream is created and the per frame first pass statistics are stored
entirely in volatile memory. This has some disadvantages when compared to a
full first pass encode, but avoids the need for file I/O and improves speed.

For two pass encoding, the function \ref av1_encode() will first be called
for each frame in the video with the value \ref AV1EncoderConfig.pass = 1.
This will result in calls to \ref av1_first_pass().

Statistics for each frame are stored in \ref FIRSTPASS_STATS frame_stats_buf.

After completion of the first pass, \ref av1_encode() will be called again for
each frame with \ref AV1EncoderConfig.pass = 2.  The frames are then encoded in
accordance with the statistics gathered during the first pass by calls to
\ref encode_frame_to_data_rate() which in turn calls
 \ref av1_get_second_pass_params().

In summary the second pass code :-

- Searches for scene cuts (if auto key frame detection is enabled).
- Defines the length of and hierarchical structure to be used in each
  ARF/GF group.
- Allocates bits based on the relative complexity of each frame, the quality
  of frame to frame prediction and the type of frame (e.g. key frame, ARF
  frame, golden frame or normal leaf frame).
- Suggests a maximum Q (quantizer value) for each ARF/GF group, based on
  estimated complexity and recent rate control compliance
  (\ref RATE_CONTROL.active_worst_quality)
- Tracks adherence to the overall rate control objectives and adjusts
  heuristics.

The main two pass functions in regard to the above include:-

- \ref find_next_key_frame()
- \ref define_gf_group()
- \ref calculate_total_gf_group_bits()
- \ref get_twopass_worst_quality()
- \ref av1_gop_setup_structure()
- \ref av1_gop_bit_allocation()
- \ref av1_twopass_postencode_update()

For each frame, the two pass algorithm defines a target number of bits
\ref RATE_CONTROL.base_frame_target,  which is then adjusted if necessary to
reflect any undershoot or overshoot on previous frames to give
\ref RATE_CONTROL.this_frame_target.

As well as \ref RATE_CONTROL.active_worst_quality, the two pass code also
maintains a record of the actual Q value used to encode previous frames
at each level in the current pyramid hierarchy
(\ref PRIMARY_RATE_CONTROL.active_best_quality). The function
\ref rc_pick_q_and_bounds(), uses these values to set a permitted Q range
for each frame.

\subsubsection architecture_enc_1pass_lagged 1 Pass Lagged VBR Encoding

1 pass lagged encode falls between simple 1 pass encoding and full two pass
encoding and is used for cases where it is not possible to do a full first
pass through the entire video clip, but where some delay is permissible. For
example near live streaming where there is a delay of up to a few seconds. In
this case the first pass and second pass are in effect combined such that the
first pass starts encoding the clip and the second pass lags behind it by a
few frames.  When using this method, full sequence level statistics are not
available, but it is possible to collect and use frame or group of frame level
data to help in the allocation of bits and in defining ARF/GF coding
hierarchies.  The reader is referred to the \ref AV1_PRIMARY.lap_enabled field
in the main compressor instance (where <b>lap</b> stands for
<b>look ahead processing</b>). This encoding mode for the most part uses the
same rate control pathways as two pass VBR encoding.

\subsection architecture_enc_rc_loop The Main Rate Control Loop

Having established a target rate for a given frame and an allowed range of Q
values, the encoder then tries to encode the frame at a rate that is as close
as possible to the target value, given the Q range constraints.

There are two main mechanisms by which this is achieved.

The first selects a frame level Q, using an adaptive estimate of the number of
bits that will be generated when the frame is encoded at any given Q.
Fundamentally this mechanism is common to VBR, CBR and to use cases such as
RTC with small adjustments.

As the Q value mainly adjusts the precision of the residual signal, it is not
actually a reliable basis for accurately predicting the number of bits that
will be generated across all clips. A well predicted clip, for example, may
have a much smaller error residual after prediction.  The algorithm copes with
this by adapting its predictions on the fly using a feedback loop based on how
well it did the previous time around.

The main functions responsible for the prediction of Q and the adaptation over
time, for the two pass encoding pipeline are:

- \ref rc_pick_q_and_bounds()
    - \ref get_q()
        - \ref av1_rc_regulate_q(
                   const struct AV1_COMP *cpi, int target_bits_per_frame,
                   int active_best_quality, int active_worst_quality,
                   int width, int height) "av1_rc_regulate_q()"
        - \ref get_rate_correction_factor()
        - \ref set_rate_correction_factor()
        - \ref find_closest_qindex_by_rate()
- \ref av1_twopass_postencode_update()
    - \ref av1_rc_update_rate_correction_factors()

A second mechanism for control comes into play if there is a large rate miss
for the current frame (much too big or too small). This is a recode mechanism
which allows the current frame to be re-encoded one or more times with a
revised Q value. This obviously has significant implications for encode speed
and in the case of RTC latency (hence it is not used for the RTC pathway).

Whether or not a recode is allowed for a given frame depends on the selected
encode speed vs quality trade off. This is set on the command line using the
--cpu-used parameter which maps onto the \ref AV1_COMP.speed field in the main
compressor instance data structure.

The value of \ref AV1_COMP.speed, combined with the use case, is used to
populate the speed features data structure AV1_COMP.sf. In particular
\ref HIGH_LEVEL_SPEED_FEATURES.recode_loop determines the types of frames that
may be recoded and \ref HIGH_LEVEL_SPEED_FEATURES.recode_tolerance is a rate
error trigger threshold.

For more information the reader is directed to the following functions:

- \ref encode_with_recode_loop()
- \ref encode_without_recode()
- \ref recode_loop_update_q()
- \ref recode_loop_test()
- \ref av1_set_speed_features_framesize_independent()
- \ref av1_set_speed_features_framesize_dependent()

\subsection architecture_enc_fixed_q Fixed Q Mode

There are two main fixed Q cases:
-# Fixed Q with adaptive qp offsets: same qp offset for each pyramid level
   in a given video, but these offsets are adaptive based on video content.
-# Fixed Q with fixed qp offsets: content-independent fixed qp offsets for
   each pyramid level.

The reader is also refered to the following functions:
- \ref av1_rc_pick_q_and_bounds()
- \ref rc_pick_q_and_bounds_no_stats_cbr()
- \ref rc_pick_q_and_bounds_no_stats()
- \ref rc_pick_q_and_bounds()

\section architecture_enc_frame_groups GF/ ARF Frame Groups & Hierarchical Coding

\subsection architecture_enc_frame_groups_data Main Data Structures

The following are the main data structures referenced in this section
(see also \ref architecture_enc_data_structures):

- \ref AV1_COMP cpi (the main compressor instance data structure)
    - \ref AV1_COMP.rc (\ref RATE_CONTROL)

- \ref FIRSTPASS_STATS *frame_stats_buf (used to store per frame first pass
stats)

\subsection architecture_enc_frame_groups_groups Frame Groups

To process a sequence/stream of video frames, the encoder divides the frames
into groups and encodes them sequentially (possibly dependent on previous
groups). In AV1 such a group is usually referred to as a golden frame group
(GF group) or sometimes an Alt-Ref (ARF) group or a group of pictures (GOP).
A GF group determines and stores the coding structure of the frames (for
example, frame type, usage of the hierarchical structure, usage of overlay
frames, etc.) and can be considered as the base unit to process the frames,
therefore playing an important role in the encoder.

The length of a specific GF group is arguably the most important aspect when
determining a GF group. This is because most GF group level decisions are
based on the frame characteristics, if not on the length itself directly.
Note that the GF group is always a group of consecutive frames, which means
the start and end of the group (so again, the length of it) determines which
frames are included in it and hence determines the characteristics of the GF
group. Therefore, in this document we will first discuss the GF group length
decision in Libaom, followed by frame structure decisions when defining a GF
group with a certain length.

\subsection architecture_enc_gf_length GF / ARF Group Length Determination

The basic intuition of determining the GF group length is that it is usually
desirable to group together frames that are similar. Hence, we may choose
longer groups when consecutive frames are very alike and shorter ones when
they are very different.

The determination of the GF group length is done in function \ref
calculate_gf_length(). The following encoder use cases are supported:

<ul>
  <li><b>Single pass with look-ahead disabled(\ref has_no_stats_stage()):
  </b> in this case there is no information available on the following stream
  of frames, therefore the function will set the GF group length for the
  current and the following GF groups (a total number of MAX_NUM_GF_INTERVALS
  groups) to be the maximum value allowed.</li>

  <li><b>Single pass with look-ahead enabled (\ref AV1_PRIMARY.lap_enabled):</b>
  look-ahead processing is enabled for single pass, therefore there is a
  limited amount of information available regarding future frames. In this
  case the function will determine the length based on \ref FIRSTPASS_STATS
  (which is generated when processing the look-ahead buffer) for only the
  current GF group.</li>

  <li><b>Two pass:</b> the first pass in two-pass encoding collects the stats
  and will not call the function. In the second pass, the function tries to
  determine the GF group length of the current and the following GF groups (a
  total number of MAX_NUM_GF_INTERVALS groups) based on the first-pass
  statistics. Note that as we will be discussing later, such decisions may not
  be accurate and can be changed later.</li>
</ul>

Except for the first trivial case where there is no prior knowledge of the
following frames, the function \ref calculate_gf_length() tries to determine the
GF group length based on the first pass statistics. The determination is divided
into two parts:

<ol>
   <li>Baseline decision based on accumulated statistics: this part of the function
   iterates through the firstpass statistics of the following frames and
   accumulates the statistics with function accumulate_next_frame_stats.
   The accumulated statistics are then used to determine whether the
   correlation in the GF group has dropped too much in function detect_gf_cut.
   If detect_gf_cut returns non-zero, or if we've reached the end of
   first-pass statistics, the baseline decision is set at the current point.</li>

   <li>If we are not at the end of the first-pass statistics, the next part will
   try to refine the baseline decision. This algorithm is based on the analysis
   of firstpass stats. It tries to cut the groups in stable regions or
   relatively stable points. Also it tries to avoid cutting in a blending
   region.</li>
</ol>

As mentioned, for two-pass encoding, the function \ref
calculate_gf_length() tries to determine the length of as many as
MAX_NUM_GF_INTERVALS groups. The decisions are stored in
\ref PRIMARY_RATE_CONTROL.gf_intervals[]. The variables
\ref RATE_CONTROL.intervals_till_gf_calculate_due and
\ref PRIMARY_RATE_CONTROL.gf_intervals[] help with managing and updating the stored
decisions. In the function \ref define_gf_group(), the corresponding
stored length decision will be used to define the current GF group.

When the maximum GF group length is larger or equal to 32, the encoder will
enforce an extra layer to determine whether to use maximum GF length of 32
or 16 for every GF group. In such a case, \ref calculate_gf_length() is
first called with the original maximum length (>=32). Afterwards,
\ref av1_tpl_setup_stats() is called to analyze the determined GF group
and compare the reference to the last frame and the middle frame. If it is
decided that we should use a maximum GF length of 16, the function
\ref calculate_gf_length() is called again with the updated maximum
length, and it only sets the length for a single GF group
(\ref RATE_CONTROL.intervals_till_gf_calculate_due is set to 1). This process
is shown below.

\image html tplgfgroupdiagram.png "" width=40%

Before encoding each frame, the encoder checks
\ref RATE_CONTROL.frames_till_gf_update_due. If it is zero, indicating
processing of the current GF group is done, the encoder will check whether
\ref RATE_CONTROL.intervals_till_gf_calculate_due is zero. If it is, as
discussed above, \ref calculate_gf_length() is called with original
maximum length. If it is not zero, then the GF group length value stored
in \ref PRIMARY_RATE_CONTROL.gf_intervals[\ref PRIMARY_RATE_CONTROL.cur_gf_index] is used
(subject to change as discussed above).

\subsection architecture_enc_gf_structure Defining a GF Group's Structure

The function \ref define_gf_group() defines the frame structure as well
as other GF group level parameters (e.g. bit allocation) once the length of
the current GF group is determined.

The function first iterates through the first pass statistics in the GF group to
accumulate various stats, using accumulate_this_frame_stats() and
accumulate_next_frame_stats(). The accumulated statistics are then used to
determine the use of the use of ALTREF frame along with other properties of the
GF group. The values of \ref PRIMARY_RATE_CONTROL.cur_gf_index, \ref
RATE_CONTROL.intervals_till_gf_calculate_due and \ref
RATE_CONTROL.frames_till_gf_update_due are also updated accordingly.

The function \ref av1_gop_setup_structure() is called at the end to determine
the frame layers and reference maps in the GF group, where the
construct_multi_layer_gf_structure() function sets the frame update types for
each frame and the group structure.

- If ALTREF frames are allowed for the GF group: the first frame is set to
  KF_UPDATE, GF_UPDATE or ARF_UPDATE. The last frames of the GF group is set to
  OVERLAY_UPDATE.  Then in set_multi_layer_params(), frame update
  types are determined recursively in a binary tree fashion, and assigned to
  give the final IBBB structure for the group.  - If the current branch has more
  than 2 frames and we have not reached maximum layer depth, then the middle
  frame is set as INTNL_ARF_UPDATE, and the left and right branches are
  processed recursively.  - If the current branch has less than 3 frames, or we
  have reached maximum layer depth, then every frame in the branch is set to
  LF_UPDATE.

- If ALTREF frame is not allowed for the GF group: the frames are set
  as LF_UPDATE. This basically forms an IPPP GF group structure.

As mentioned, the encoder may use Temporal dependancy modelling (TPL - see \ref
architecture_enc_tpl) to determine whether we should use a maximum length of 32
or 16 for the current GF group. This requires calls to \ref define_gf_group()
but should not change other settings (since it is in essence a trial). This
special case is indicated by the setting parameter <b>is_final_pass</b> for to
zero.

For single pass encodes where look-ahead processing is disabled
(\ref AV1_PRIMARY.lap_enabled = 0), \ref define_gf_group_pass0() is used
instead of \ref define_gf_group().

\subsection architecture_enc_kf_groups Key Frame Groups

A special constraint for GF group length is the location of the next keyframe
(KF). The frames between two KFs are referred to as a KF group. Each KF group
can be encoded and decoded independently. Because of this, a GF group cannot
span beyond a KF and the location of the next KF is set as a hard boundary
for GF group length.

<ul>
   <li>For two-pass encoding \ref RATE_CONTROL.frames_to_key controls when to
   encode a key frame. When it is zero, the current frame is a keyframe and
   the function \ref find_next_key_frame() is called. This in turn calls
   \ref define_kf_interval() to work out where the next key frame should
   be placed.</li>

   <li>For single-pass with look-ahead enabled, \ref define_kf_interval()
   is called whenever a GF group update is needed (when
   \ref RATE_CONTROL.frames_till_gf_update_due is zero). This is because
   generally KFs are more widely spaced and the look-ahead buffer is usually
   not long enough.</li>

   <li>For single-pass with look-ahead disabled, the KFs are placed according
   to the command line parameter <b>--kf-max-dist</b> (The above two cases are
   also subject to this constraint).</li>
</ul>

The function \ref define_kf_interval() tries to detect a scenecut.
If a scenecut within kf-max-dist is detected, then it is set as the next
keyframe. Otherwise the given maximum value is used.

\section architecture_enc_tpl Temporal Dependency Modelling

The temporal dependency model runs at the beginning of each GOP. It builds the
motion trajectory within the GOP in units of 16x16 blocks. The temporal
dependency of a 16x16 block is evaluated as the predictive coding gains it
contributes to its trailing motion trajectory. This temporal dependency model
reflects how important a coding block is for the coding efficiency of the
overall GOP. It is hence used to scale the Lagrangian multiplier used in the
rate-distortion optimization framework.

\subsection architecture_enc_tpl_config Configurations

The temporal dependency model and its applications are by default turned on in
libaom encoder for the VoD use case. To disable it, use --tpl-model=0 in the
aomenc configuration.

\subsection architecture_enc_tpl_algoritms Algorithms

The scheme works in the reverse frame processing order over the source frames,
propagating information from future frames back to the current frame. For each
frame, a propagation step is run for each MB. it operates as follows:

<ul>
   <li> Estimate the intra prediction cost in terms of sum of absolute Hadamard
   transform difference (SATD) noted as intra_cost. It also loads the motion
   information available from the first-pass encode and estimates the inter
   prediction cost as inter_cost. Due to the use of hybrid inter/intra
   prediction mode, the inter_cost value is further upper bounded by
   intra_cost. A propagation cost variable is used to collect all the
   information flowed back from future processing frames. It is initialized as
   0 for all the blocks in the last processing frame in a group of pictures
   (GOP).</li>

   <li> The fraction of information from a current block to be propagated towards
   its reference block is estimated as:
\f[
   propagation\_fraction = (1 - inter\_cost/intra\_cost)
\f]
   It reflects how much the motion compensated reference would reduce the
   prediction error in percentage.</li>

   <li> The total amount of information the current block contributes to the GOP
   is estimated as intra_cost + propagation_cost. The information that it
   propagates towards its reference block is captured by:

\f[
   propagation\_amount =
   (intra\_cost + propagation\_cost) * propagation\_fraction
\f]</li>

   <li> Note that the reference block may not necessarily sit on the grid of
   16x16 blocks. The propagation amount is hence dispensed to all the blocks
   that overlap with the reference block. The corresponding block in the
   reference frame accumulates its own propagation cost as it receives back
   propagation.

\f[
   propagation\_cost = propagation\_cost +
                       (\frac{overlap\_area}{(16*16)} * propagation\_amount)
\f]</li>

   <li> In the final encoding stage, the distortion propagation factor of a block
   is evaluated as \f$(1 + \frac{propagation\_cost}{intra\_cost})\f$, where the second term
   captures its impact on later frames in a GOP.</li>

   <li> The Lagrangian multiplier is adapted at the 64x64 block level. For every
   64x64 block in a frame, we have a distortion propagation factor:

\f[
  dist\_prop[i] = 1 + \frac{propagation\_cost[i]}{intra\_cost[i]}
\f]

   where i denotes the block index in the frame. We also have the frame level
   distortion propagation factor:

\f[
  dist\_prop = 1 +
  \frac{\sum_{i}propagation\_cost[i]}{\sum_{i}intra\_cost[i]}
\f]

   which is used to normalize the propagation factor at the 64x64 block level. The
   Lagrangian multiplier is hence adapted as:

\f[
  &lambda;[i] = &lambda;[0] * \frac{dist\_prop}{dist\_prop[i]}
\f]

   where &lambda;0 is the multiplier associated with the frame level QP. The
   64x64 block level QP is scaled according to the Lagrangian multiplier.
</ul>

\subsection architecture_enc_tpl_keyfun Key Functions and data structures

The reader is also refered to the following functions and data structures:

- \ref TplParams
- \ref av1_tpl_setup_stats() builds the TPL model.
- \ref setup_delta_q() Assign different quantization parameters to each super
  block based on its TPL weight.

\section architecture_enc_partitions Block Partition Search

 A frame is first split into tiles in \ref encode_tiles(), with each tile
 compressed by av1_encode_tile(). Then a tile is processed in superblock rows
 via \ref av1_encode_sb_row() and then \ref encode_sb_row().

 The partition search processes superblocks sequentially in \ref
 encode_sb_row(). Two search modes are supported, depending upon the encoding
 configuration, \ref encode_nonrd_sb() is for 1-pass and real-time modes,
 while \ref encode_rd_sb() performs more exhaustive rate distortion based
 searches.

 Partition search over the recursive quad-tree space is implemented by
 recursive calls to \ref av1_nonrd_use_partition(),
 \ref av1_rd_use_partition(), or av1_rd_pick_partition() and returning best
 options for sub-trees to their parent partitions.

 In libaom, the partition search lays on top of the mode search (predictor,
 transform, etc.), instead of being a separate module. The interface of mode
 search is \ref pick_sb_modes(), which connects the partition_search with
 \ref architecture_enc_inter_modes and \ref architecture_enc_intra_modes. To
 make good decisions, reconstruction is also required in order to build
 references and contexts. This is implemented by \ref encode_sb() at the
 sub-tree level and \ref encode_b() at coding block level.

 See also \ref partition_search

\section architecture_enc_intra_modes Intra Mode Search

AV1 also provides 71 different intra prediction modes, i.e. modes that predict
only based upon information in the current frame with no dependency on
previous or future frames. For key frames, where this independence from any
other frame is a defining requirement and for other cases where intra only
frames are required, the encoder need only considers these modes in the rate
distortion loop.

Even so, in most use cases, searching all possible intra prediction modes for
every block and partition size is not practical and some pruning of the search
tree is necessary.

For the Rate distortion optimized case, the main top level function
responsible for selecting the intra prediction mode for a given block is
\ref av1_rd_pick_intra_mode_sb(). The readers attention is also drawn to the
functions \ref hybrid_intra_mode_search() and \ref av1_nonrd_pick_intra_mode()
which may be used where encode speed is critical. The choice between the
rd path and the non rd or hybrid paths depends on the encoder use case and the
\ref AV1_COMP.speed parameter. Further fine control of the speed vs quality
trade off is provided by means of fields in \ref AV1_COMP.sf (which has type
\ref SPEED_FEATURES).

Note that some intra modes are only considered for specific use cases or
types of video. For example the palette based prediction modes are often
valueable for graphics or screen share content but not for natural video.
(See \ref av1_search_palette_mode())

See also \ref intra_mode_search for more details.

\section architecture_enc_inter_modes Inter Prediction Mode Search

For inter frames, where we also allow prediction using one or more previously
coded frames (which may chronologically speaking be past or future frames or
non-display reference buffers such as ARF frames), the size of the search tree
that needs to be traversed, to select a prediction mode, is considerably more
massive.

In addition to the 71 possible intra modes we also need to consider 56 single
frame inter prediction modes (7 reference frames x 4 modes x 2 for OBMC
(overlapped block motion compensation)), 12768 compound inter prediction modes
(these are modes that combine inter predictors from two reference frames) and
36708 compound inter / intra prediction modes.

As with the intra mode search, libaom supports an RD based pathway and a non
rd pathway for speed critical use cases.  The entry points for these two cases
are \ref av1_rd_pick_inter_mode() and \ref av1_nonrd_pick_inter_mode_sb()
respectively.

Various heuristics and predictive strategies are used to prune the search tree
with fine control provided through the speed features parameter in the main
compressor instance data structure \ref AV1_COMP.sf.

It is worth noting, that some prediction modes incurr a much larger rate cost
than others (ignoring for now the cost of coding the error residual). For
example, a compound mode that requires the encoder to specify two reference
frames and two new motion vectors will almost inevitable have a higher rate
cost than a simple inter prediction mode that uses a predicted or 0,0 motion
vector. As such, if we have already found a mode for the current block that
has a low RD cost, we can skip a large number of the possible modes on the
basis that even if the error residual is 0 the inherent rate cost of the
mode itself will garauntee that it is not chosen.

See also \ref inter_mode_search for more details.

\section architecture_enc_tx_search Transform Search

AV1 implements the transform stage using 4 seperable 1-d transforms (DCT,
ADST, FLIPADST and IDTX, where FLIPADST is the reversed version of ADST
and IDTX is the identity transform) which can be combined to give 16 2-d
combinations.

These combinations can be applied at 19 different scales from 64x64 pixels
down to 4x4 pixels.

This gives rise to a large number of possible candidate transform options
for coding the residual error after prediction. An exhaustive rate-distortion
based evaluation of all candidates would not be practical from a speed
perspective in a production encoder implementation. Hence libaom addopts a
number of strategies to prune the selection of both the transform size and
transform type.

There are a number of strategies that have been tested and implememnted in
libaom including:

- A statistics based approach that looks at the frequency with which certain
  combinations are used in a given context and prunes out very unlikely
  candidates. It is worth noting here that some size candidates can be pruned
  out immediately based on the size of the prediction partition. For example it
  does not make sense to use a transform size that is larger than the
  prediction partition size but also a very large prediction partition size is
  unlikely to be optimally pared with small transforms.

- A Machine learning based model

- A method that initially tests candidates using a fast algorithm that skips
  entropy encoding and uses an estimated cost model to choose a reduced subset
  for full RD analysis. This subject is covered more fully in a paper authored
  by Bohan Li, Jingning Han, and Yaowu Xu titled: <b>Fast Transform Type
  Selection Using Conditional Laplace Distribution Based Rate Estimation</b>

<b>TODO Add link to paper when available</b>

See also \ref transform_search for more details.

\section architecture_post_enc_filt Post Encode Loop Filtering

AV1 supports three types of post encode <b>in loop</b> filtering to improve
the quality of the reconstructed video.

- <b>Deblocking Filter</b> The first of these is a farily traditional boundary
  deblocking filter that attempts to smooth discontinuities that may occur at
  the boundaries between blocks. See also \ref in_loop_filter.

- <b>CDEF Filter</b> The constrained directional enhancement filter (CDEF)
  allows the codec to apply a non-linear deringing filter along certain
  (potentially oblique) directions. A primary filter is applied along the
  selected direction, whilst a secondary filter is applied at 45 degrees to
  the primary direction. (See also \ref in_loop_cdef and
  <a href="https://arxiv.org/abs/2008.06091"> A Technical Overview of AV1</a>.

- <b>Loop Restoration Filter</b> The loop restoration filter is applied after
  any prior post filtering stages. It acts on units of either 64 x 64,
  128 x 128, or 256 x 256 pixel blocks, refered to as loop restoration units.
  Each unit can independently select either to bypass filtering, use a Wiener
  filter, or use a self-guided filter. (See also \ref in_loop_restoration and
  <a href="https://arxiv.org/abs/2008.06091"> A Technical Overview of AV1</a>.

\section architecture_entropy Entropy Coding

\subsection architecture_entropy_aritmetic Arithmetic Coder

VP9, used a binary arithmetic coder to encode symbols, where the propability
of a 1 or 0 at each descision node was based on a context model that took
into account recently coded values (for example previously coded coefficients
in the current block). A mechanism existed to update the context model each
frame, either explicitly in the bitstream, or implicitly at both the encoder
and decoder based on the observed frequency of different outcomes in the
previous frame. VP9 also supported seperate context models for different types
of frame (e.g. inter coded frames and key frames).

In contrast, AV1 uses an M-ary symbol arithmetic coder to compress the syntax
elements, where integer \f$M\in[2, 14]\f$. This approach is based upon the entropy
coding strategy used in the Daala video codec and allows for some bit-level
parallelism in its implementation. AV1 also has an extended context model and
allows for updates to the probabilities on a per symbol basis as opposed to
the per frame strategy in VP9.

To improve the performance / throughput of the arithmetic encoder, especially
in hardware implementations, the probability model is updated and maintained
at 15-bit precision, but the arithmetic encoder only uses the most significant
9 bits when encoding a symbol. A more detailed discussion of the algorithm
and design constraints can be found in
<a href="https://arxiv.org/abs/2008.06091"> A Technical Overview of AV1</a>.

TODO add references to key functions / files.

As with VP9, a mechanism exists in AV1 to encode some elements into the
bitstream as uncrompresed bits or literal values, without using the arithmetic
coder. For example, some frame and sequence header values, where it is
beneficial to be able to read the values directly.

TODO add references to key functions / files.

\subsection architecture_entropy_coef Transform Coefficient Coding and Optimization
\image html coeff_coding.png "" width=70%

\subsubsection architecture_entropy_coef_what Transform coefficient coding
Transform coefficient coding is where the encoder compresses a quantized version
of prediction residue into the bitstream.

\paragraph architecture_entropy_coef_prepare Preparation - transform and quantize
Before the entropy coding stage, the encoder decouple the pixel-to-pixel
correlation of the prediction residue by transforming the residue from the
spatial domain to the frequency domain. Then the encoder quantizes the transform
coefficients to make the coefficients ready for entropy coding.

\paragraph architecture_entropy_coef_coding The coding process
The encoder uses \ref av1_write_coeffs_txb() to write the coefficients of
a transform block into the bitstream.
The coding process has three stages.
1. The encoder will code transform block skip flag (txb_skip). If the skip flag is
off, then the encoder will code the end of block position (eob) which is the scan
index of the last non-zero coefficient plus one.
2. Second, the encoder will code lower magnitude levels of each coefficient in
reverse scan order.
3. Finally, the encoder will code the sign and higher magnitude levels for each
coefficient if they are available.

Related functions:
- \ref av1_write_coeffs_txb()
- write_inter_txb_coeff()
- \ref av1_write_intra_coeffs_mb()

\paragraph architecture_entropy_coef_context Context information
To improve the compression efficiency, the encoder uses several context models
tailored for transform coefficients to capture the correlations between coding
symbols. Most of the context models are built to capture the correlations
between the coefficients within the same transform block. However, transform
block skip flag (txb_skip) and the sign of dc coefficient (dc_sign) require
context info from neighboring transform blocks.

Here is how context info spread between transform blocks. Before coding a
transform block, the encoder will use get_txb_ctx() to collect the context
information from neighboring transform blocks. Then the context information
will be used for coding transform block skip flag (txb_skip) and the sign of
dc coefficient (dc_sign). After the transform block is coded, the encoder will
extract the context info from the current block using
\ref av1_get_txb_entropy_context(). Then encoder will store the context info
into a byte (uint8_t) using av1_set_entropy_contexts(). The encoder will use
the context info to code other transform blocks.

Related functions:
- \ref av1_get_txb_entropy_context()
- av1_set_entropy_contexts()
- get_txb_ctx()
- \ref av1_update_intra_mb_txb_context()

\subsubsection architecture_entropy_coef_rd RD optimization
Beside the actual entropy coding, the encoder uses several utility functions
to make optimal RD decisions.

\paragraph architecture_entropy_coef_cost Entropy cost
The encoder uses \ref av1_cost_coeffs_txb() or \ref av1_cost_coeffs_txb_laplacian()
to estimate the entropy cost of a transform block. Note that
\ref av1_cost_coeffs_txb() is slower but accurate whereas
\ref av1_cost_coeffs_txb_laplacian() is faster but less accurate.

Related functions:
- \ref av1_cost_coeffs_txb()
- \ref av1_cost_coeffs_txb_laplacian()
- av1_cost_coeffs_txb_estimate() (see av1/encoder/txb_rdopt.c)

\paragraph architecture_entropy_coef_opt Quantized level optimization
Beside computing entropy cost, the encoder also uses \ref av1_optimize_txb()
to adjust the coefficient’s quantized levels to achieve optimal RD trade-off.
In \ref av1_optimize_txb(), the encoder goes through each quantized
coefficient and lowers the quantized coefficient level by one if the action
yields a better RD score.

Related functions:
- \ref av1_optimize_txb()

All the related functions are listed in \ref coefficient_coding.

\section architecture_simd SIMD usage

In order to efficiently encode video on modern platforms, it is necessary to
implement optimized versions of many core encoding and decoding functions using
architecture-specific SIMD instructions.

Functions which have optimized implementations will have multiple variants
in the code, each suffixed with the name of the appropriate instruction set.
There will additionally be an `_c` version, which acts as a reference
implementation which the SIMD variants can be tested against.

As different machines with the same nominal architecture may support different
subsets of SIMD instructions, we have dynamic CPU detection logic which chooses
the appropriate functions to use at run time. This process is handled by
`cmake/rtcd.pl`, with function definitions in the files
`*_rtcd_defs.pl` elsewhere in the codebase.

Currently SIMD is supported on the following platforms:

- x86: Requires SSE4.1 or above

- Arm: Requires Neon (Armv7-A and above)

We aim to provide implementations of all performance-critical functions which
are compatible with the instruction sets listed above. Additional SIMD
extensions (e.g. AVX on x86, SVE on Arm) are also used to provide even
greater performance where available.

*/

/*!\defgroup encoder_algo Encoder Algorithm
 *
 * The encoder algorithm describes how a sequence is encoded, including high
 * level decision as well as algorithm used at every encoding stage.
 */

/*!\defgroup high_level_algo High-level Algorithm
 * \ingroup encoder_algo
 * This module describes sequence level/frame level algorithm in AV1.
 * More details will be added.
 * @{
 */

/*!\defgroup speed_features Speed vs Quality Trade Off
 * \ingroup high_level_algo
 * This module describes the encode speed vs quality tradeoff
 * @{
 */
/*! @} - end defgroup speed_features */

/*!\defgroup src_frame_proc Source Frame Processing
 * \ingroup high_level_algo
 * This module describes algorithms in AV1 assosciated with the
 * pre-processing of source frames. See also \ref architecture_enc_src_proc
 *
 * @{
 */
/*! @} - end defgroup src_frame_proc */

/*!\defgroup rate_control Rate Control
 * \ingroup high_level_algo
 * This module describes rate control algorithm in AV1.
 *  See also \ref architecture_enc_rate_ctrl
 * @{
 */
/*! @} - end defgroup rate_control */

/*!\defgroup tpl_modelling Temporal Dependency Modelling
 * \ingroup high_level_algo
 * This module includes algorithms to implement temporal dependency modelling.
 *  See also \ref architecture_enc_tpl
 * @{
 */
/*! @} - end defgroup tpl_modelling */

/*!\defgroup two_pass_algo Two Pass Mode
   \ingroup high_level_algo

 In two pass mode, the input file is passed into the encoder for a quick
 first pass, where statistics are gathered. These statistics and the input
 file are then passed back into the encoder for a second pass. The statistics
 help the encoder reach the desired bitrate without as much overshooting or
 undershooting.

 During the first pass, the codec will return "stats" packets that contain
 information useful for the second pass. The caller should concatenate these
 packets as they are received. In the second pass, the concatenated packets
 are passed in, along with the frames to encode. During the second pass,
 "frame" packets are returned that represent the compressed video.

 A complete example can be found in `examples/twopass_encoder.c`. Pseudocode
 is provided below to illustrate the core parts.

 During the first pass, the uncompressed frames are passed in and stats
 information is appended to a byte array.

~~~~~~~~~~~~~~~{.c}
// For simplicity, assume that there is enough memory in the stats buffer.
// Actual code will want to use a resizable array. stats_len represents
// the length of data already present in the buffer.
void get_stats_data(aom_codec_ctx_t *encoder, char *stats,
                    size_t *stats_len, bool *got_data) {
  const aom_codec_cx_pkt_t *pkt;
  aom_codec_iter_t iter = NULL;
  while ((pkt = aom_codec_get_cx_data(encoder, &iter))) {
    *got_data = true;
    if (pkt->kind != AOM_CODEC_STATS_PKT) continue;
    memcpy(stats + *stats_len, pkt->data.twopass_stats.buf,
           pkt->data.twopass_stats.sz);
    *stats_len += pkt->data.twopass_stats.sz;
  }
}

void first_pass(char *stats, size_t *stats_len) {
  struct aom_codec_enc_cfg first_pass_cfg;
  ... // Initialize the config as needed.
  first_pass_cfg.g_pass = AOM_RC_FIRST_PASS;
  aom_codec_ctx_t first_pass_encoder;
  ... // Initialize the encoder.

  while (frame_available) {
    // Read in the uncompressed frame, update frame_available
    aom_image_t *frame_to_encode = ...;
    aom_codec_encode(&first_pass_encoder, img, pts, duration, flags);
    get_stats_data(&first_pass_encoder, stats, stats_len);
  }
  // After all frames have been processed, call aom_codec_encode with
  // a NULL ptr repeatedly, until no more data is returned. The NULL
  // ptr tells the encoder that no more frames are available.
  bool got_data;
  do {
    got_data = false;
    aom_codec_encode(&first_pass_encoder, NULL, pts, duration, flags);
    get_stats_data(&first_pass_encoder, stats, stats_len, &got_data);
  } while (got_data);

  aom_codec_destroy(&first_pass_encoder);
}
~~~~~~~~~~~~~~~

 During the second pass, the uncompressed frames and the stats are
 passed into the encoder.

~~~~~~~~~~~~~~~{.c}
// Write out each encoded frame to the file.
void get_cx_data(aom_codec_ctx_t *encoder, FILE *file,
                 bool *got_data) {
  const aom_codec_cx_pkt_t *pkt;
  aom_codec_iter_t iter = NULL;
  while ((pkt = aom_codec_get_cx_data(encoder, &iter))) {
   *got_data = true;
   if (pkt->kind != AOM_CODEC_CX_FRAME_PKT) continue;
   fwrite(pkt->data.frame.buf, 1, pkt->data.frame.sz, file);
  }
}

void second_pass(char *stats, size_t stats_len) {
  struct aom_codec_enc_cfg second_pass_cfg;
  ... // Initialize the config file as needed.
  second_pass_cfg.g_pass = AOM_RC_LAST_PASS;
  cfg.rc_twopass_stats_in.buf = stats;
  cfg.rc_twopass_stats_in.sz = stats_len;
  aom_codec_ctx_t second_pass_encoder;
  ... // Initialize the encoder from the config.

  FILE *output = fopen("output.obu", "wb");
  while (frame_available) {
    // Read in the uncompressed frame, update frame_available
    aom_image_t *frame_to_encode = ...;
    aom_codec_encode(&second_pass_encoder, img, pts, duration, flags);
    get_cx_data(&second_pass_encoder, output);
  }
  // Pass in NULL to flush the encoder.
  bool got_data;
  do {
    got_data = false;
    aom_codec_encode(&second_pass_encoder, NULL, pts, duration, flags);
    get_cx_data(&second_pass_encoder, output, &got_data);
  } while (got_data);

  aom_codec_destroy(&second_pass_encoder);
}
~~~~~~~~~~~~~~~
 */

 /*!\defgroup look_ahead_buffer The Look-Ahead Buffer
    \ingroup high_level_algo

 A program should call \ref aom_codec_encode() for each frame that needs
 processing. These frames are internally copied and stored in a fixed-size
 circular buffer, known as the look-ahead buffer. Other parts of the code
 will use future frame information to inform current frame decisions;
 examples include the first-pass algorithm, TPL model, and temporal filter.
 Note that this buffer also keeps a reference to the last source frame.

 The look-ahead buffer is defined in \ref av1/encoder/lookahead.h. It acts as an
 opaque structure, with an interface to create and free memory associated with
 it. It supports pushing and popping frames onto the structure in a FIFO
 fashion. It also allows look-ahead when using the \ref av1_lookahead_peek()
 function with a non-negative number, and look-behind when -1 is passed in (for
 the last source frame; e.g., firstpass will use this for motion estimation).
 The \ref av1_lookahead_depth() function returns the current number of frames
 stored in it. Note that \ref av1_lookahead_pop() is a bit of a misnomer - it
 only pops if either the "flush" variable is set, or the buffer is at maximum
 capacity.

 The buffer is stored in the \ref AV1_PRIMARY::lookahead field.
 It is initialized in the first call to \ref aom_codec_encode(), in the
 \ref av1_receive_raw_frame() sub-routine. The buffer size is defined by
 the g_lag_in_frames parameter set in the
 \ref aom_codec_enc_cfg_t::g_lag_in_frames struct.
 This can be modified manually but should only be set once. On the command
 line, the flag "--lag-in-frames" controls it. The default size is 19 for
 non-realtime usage and 1 for realtime. Note that a maximum value of 35 is
 enforced.

 A frame will stay in the buffer as long as possible. As mentioned above,
 the \ref av1_lookahead_pop() only removes a frame when either flush is set,
 or the buffer is full. Note that each call to \ref aom_codec_encode() inserts
 another frame into the buffer, and pop is called by the sub-function
 \ref av1_encode_strategy(). The buffer is told to flush when
 \ref aom_codec_encode() is passed a NULL image pointer. Note that the caller
 must repeatedly call \ref aom_codec_encode() with a NULL image pointer, until
 no more packets are available, in order to fully flush the buffer.

 */

/*! @} - end defgroup high_level_algo */

/*!\defgroup partition_search Partition Search
 * \ingroup encoder_algo
 * For and overview of the partition search see \ref architecture_enc_partitions
 * @{
 */

/*! @} - end defgroup partition_search */

/*!\defgroup intra_mode_search Intra Mode Search
 * \ingroup encoder_algo
 * This module describes intra mode search algorithm in AV1.
 * More details will be added.
 * @{
 */
/*! @} - end defgroup intra_mode_search */

/*!\defgroup inter_mode_search Inter Mode Search
 * \ingroup encoder_algo
 * This module describes inter mode search algorithm in AV1.
 * More details will be added.
 * @{
 */
/*! @} - end defgroup inter_mode_search */

/*!\defgroup palette_mode_search Palette Mode Search
 * \ingroup intra_mode_search
 * This module describes palette mode search algorithm in AV1.
 * More details will be added.
 * @{
 */
/*! @} - end defgroup palette_mode_search */

/*!\defgroup transform_search Transform Search
 * \ingroup encoder_algo
 * This module describes transform search algorithm in AV1.
 * @{
 */
/*! @} - end defgroup transform_search */

/*!\defgroup coefficient_coding Transform Coefficient Coding and Optimization
 * \ingroup encoder_algo
 * This module describes the algorithms of transform coefficient coding and optimization in AV1.
 * More details will be added.
 * @{
 */
/*! @} - end defgroup coefficient_coding */

/*!\defgroup in_loop_filter In-loop Filter
 * \ingroup encoder_algo
 * This module describes in-loop filter algorithm in AV1.
 * More details will be added.
 * @{
 */
/*! @} - end defgroup in_loop_filter */

/*!\defgroup in_loop_cdef CDEF
 * \ingroup encoder_algo
 * This module describes the CDEF parameter search algorithm
 * in AV1. More details will be added.
 * @{
 */
/*! @} - end defgroup in_loop_restoration */

/*!\defgroup in_loop_restoration Loop Restoration
 * \ingroup encoder_algo
 * This module describes the loop restoration search
 * and estimation algorithm in AV1.
 * More details will be added.
 * @{
 */
/*! @} - end defgroup in_loop_restoration */

/*!\defgroup cyclic_refresh Cyclic Refresh
 * \ingroup encoder_algo
 * This module describes the cyclic refresh (aq-mode=3) in AV1.
 * More details will be added.
 * @{
 */
/*! @} - end defgroup cyclic_refresh */

/*!\defgroup SVC Scalable Video Coding
 * \ingroup encoder_algo
 * This module describes scalable video coding algorithm in AV1.
 * More details will be added.
 * @{
 */
/*! @} - end defgroup SVC */
/*!\defgroup variance_partition Variance Partition
 * \ingroup encoder_algo
 * This module describes variance partition algorithm in AV1.
 * More details will be added.
 * @{
 */
/*! @} - end defgroup variance_partition */
/*!\defgroup nonrd_mode_search NonRD Optimized Mode Search
 * \ingroup encoder_algo
 * This module describes NonRD Optimized Mode Search used in Real-Time mode.
 * More details will be added.
 * @{
 */
/*! @} - end defgroup nonrd_mode_search */