linux_dsm_epyc7002/Documentation/sound/designs/compress-offload.rst

=========================
ALSA Compress-Offload API
=========================

Pierre-Louis.Bossart <pierre-louis.bossart@linux.intel.com>

Vinod Koul <vinod.koul@linux.intel.com>


Overview
========
Since its early days, the ALSA API was defined with PCM support or
constant bitrates payloads such as IEC61937 in mind. Arguments and
returned values in frames are the norm, making it a challenge to
extend the existing API to compressed data streams.

In recent years, audio digital signal processors (DSP) were integrated
in system-on-chip designs, and DSPs are also integrated in audio
codecs. Processing compressed data on such DSPs results in a dramatic
reduction of power consumption compared to host-based
processing. Support for such hardware has not been very good in Linux,
mostly because of a lack of a generic API available in the mainline
kernel.

Rather than requiring a compatibility break with an API change of the
ALSA PCM interface, a new 'Compressed Data' API is introduced to
provide a control and data-streaming interface for audio DSPs.

The design of this API was inspired by the 2-year experience with the
Intel Moorestown SOC, with many corrections required to upstream the
API in the mainline kernel instead of the staging tree and make it
usable by others.


Requirements
============
The main requirements are:

- separation between byte counts and time. Compressed formats may have
  a header per file, per frame, or no header at all. The payload size
  may vary from frame-to-frame. As a result, it is not possible to
  estimate reliably the duration of audio buffers when handling
  compressed data. Dedicated mechanisms are required to allow for
  reliable audio-video synchronization, which requires precise
  reporting of the number of samples rendered at any given time.

- Handling of multiple formats. PCM data only requires a specification
  of the sampling rate, number of channels and bits per sample. In
  contrast, compressed data comes in a variety of formats. Audio DSPs
  may also provide support for a limited number of audio encoders and
  decoders embedded in firmware, or may support more choices through
  dynamic download of libraries.

- Focus on main formats. This API provides support for the most
  popular formats used for audio and video capture and playback. It is
  likely that as audio compression technology advances, new formats
  will be added.

- Handling of multiple configurations. Even for a given format like
  AAC, some implementations may support AAC multichannel but HE-AAC
  stereo. Likewise WMA10 level M3 may require too much memory and cpu
  cycles. The new API needs to provide a generic way of listing these
  formats.

- Rendering/Grabbing only. This API does not provide any means of
  hardware acceleration, where PCM samples are provided back to
  user-space for additional processing. This API focuses instead on
  streaming compressed data to a DSP, with the assumption that the
  decoded samples are routed to a physical output or logical back-end.

- Complexity hiding. Existing user-space multimedia frameworks all
  have existing enums/structures for each compressed format. This new
  API assumes the existence of a platform-specific compatibility layer
  to expose, translate and make use of the capabilities of the audio
  DSP, eg. Android HAL or PulseAudio sinks. By construction, regular
  applications are not supposed to make use of this API.


Design
======
The new API shares a number of concepts with the PCM API for flow
control. Start, pause, resume, drain and stop commands have the same
semantics no matter what the content is.

The concept of memory ring buffer divided in a set of fragments is
borrowed from the ALSA PCM API. However, only sizes in bytes can be
specified.

Seeks/trick modes are assumed to be handled by the host.

The notion of rewinds/forwards is not supported. Data committed to the
ring buffer cannot be invalidated, except when dropping all buffers.

The Compressed Data API does not make any assumptions on how the data
is transmitted to the audio DSP. DMA transfers from main memory to an
embedded audio cluster or to a SPI interface for external DSPs are
possible. As in the ALSA PCM case, a core set of routines is exposed;
each driver implementer will have to write support for a set of
mandatory routines and possibly make use of optional ones.

The main additions are

get_caps
  This routine returns the list of audio formats supported. Querying the
  codecs on a capture stream will return encoders, decoders will be
  listed for playback streams.

get_codec_caps
  For each codec, this routine returns a list of
  capabilities. The intent is to make sure all the capabilities
  correspond to valid settings, and to minimize the risks of
  configuration failures. For example, for a complex codec such as AAC,
  the number of channels supported may depend on a specific profile. If
  the capabilities were exposed with a single descriptor, it may happen
  that a specific combination of profiles/channels/formats may not be
  supported. Likewise, embedded DSPs have limited memory and cpu cycles,
  it is likely that some implementations make the list of capabilities
  dynamic and dependent on existing workloads. In addition to codec
  settings, this routine returns the minimum buffer size handled by the
  implementation. This information can be a function of the DMA buffer
  sizes, the number of bytes required to synchronize, etc, and can be
  used by userspace to define how much needs to be written in the ring
  buffer before playback can start.

set_params
  This routine sets the configuration chosen for a specific codec. The
  most important field in the parameters is the codec type; in most
  cases decoders will ignore other fields, while encoders will strictly
  comply to the settings

get_params
  This routines returns the actual settings used by the DSP. Changes to
  the settings should remain the exception.

get_timestamp
  The timestamp becomes a multiple field structure. It lists the number
  of bytes transferred, the number of samples processed and the number
  of samples rendered/grabbed. All these values can be used to determine
  the average bitrate, figure out if the ring buffer needs to be
  refilled or the delay due to decoding/encoding/io on the DSP.

Note that the list of codecs/profiles/modes was derived from the
OpenMAX AL specification instead of reinventing the wheel.
Modifications include:
- Addition of FLAC and IEC formats
- Merge of encoder/decoder capabilities
- Profiles/modes listed as bitmasks to make descriptors more compact
- Addition of set_params for decoders (missing in OpenMAX AL)
- Addition of AMR/AMR-WB encoding modes (missing in OpenMAX AL)
- Addition of format information for WMA
- Addition of encoding options when required (derived from OpenMAX IL)
- Addition of rateControlSupported (missing in OpenMAX AL)


Gapless Playback
================
When playing thru an album, the decoders have the ability to skip the encoder
delay and padding and directly move from one track content to another. The end
user can perceive this as gapless playback as we don't have silence while
switching from one track to another

Also, there might be low-intensity noises due to encoding. Perfect gapless is
difficult to reach with all types of compressed data, but works fine with most
music content. The decoder needs to know the encoder delay and encoder padding.
So we need to pass this to DSP. This metadata is extracted from ID3/MP4 headers
and are not present by default in the bitstream, hence the need for a new
interface to pass this information to the DSP. Also DSP and userspace needs to
switch from one track to another and start using data for second track.

The main additions are:

set_metadata
  This routine sets the encoder delay and encoder padding. This can be used by
  decoder to strip the silence. This needs to be set before the data in the track
  is written.

set_next_track
  This routine tells DSP that metadata and write operation sent after this would
  correspond to subsequent track

partial drain
  This is called when end of file is reached. The userspace can inform DSP that
  EOF is reached and now DSP can start skipping padding delay. Also next write
  data would belong to next track

Sequence flow for gapless would be:
- Open
- Get caps / codec caps
- Set params
- Set metadata of the first track
- Fill data of the first track
- Trigger start
- User-space finished sending all,
- Indicate next track data by sending set_next_track
- Set metadata of the next track
- then call partial_drain to flush most of buffer in DSP
- Fill data of the next track
- DSP switches to second track

(note: order for partial_drain and write for next track can be reversed as well)


Not supported
=============
- Support for VoIP/circuit-switched calls is not the target of this
  API. Support for dynamic bit-rate changes would require a tight
  coupling between the DSP and the host stack, limiting power savings.

- Packet-loss concealment is not supported. This would require an
  additional interface to let the decoder synthesize data when frames
  are lost during transmission. This may be added in the future.

- Volume control/routing is not handled by this API. Devices exposing a
  compressed data interface will be considered as regular ALSA devices;
  volume changes and routing information will be provided with regular
  ALSA kcontrols.

- Embedded audio effects. Such effects should be enabled in the same
  manner, no matter if the input was PCM or compressed.

- multichannel IEC encoding. Unclear if this is required.

- Encoding/decoding acceleration is not supported as mentioned
  above. It is possible to route the output of a decoder to a capture
  stream, or even implement transcoding capabilities. This routing
  would be enabled with ALSA kcontrols.

- Audio policy/resource management. This API does not provide any
  hooks to query the utilization of the audio DSP, nor any preemption
  mechanisms.

- No notion of underrun/overrun. Since the bytes written are compressed
  in nature and data written/read doesn't translate directly to
  rendered output in time, this does not deal with underrun/overrun and
  maybe dealt in user-library


Credits
=======
- Mark Brown and Liam Girdwood for discussions on the need for this API
- Harsha Priya for her work on intel_sst compressed API
- Rakesh Ughreja for valuable feedback
- Sing Nallasellan, Sikkandar Madar and Prasanna Samaga for
  demonstrating and quantifying the benefits of audio offload on a
  real platform.