This document describes the audio streaming interface exposed by audio drivers in Fuchsia. It is meant to serve as a reference for both users and driver-authors, and to unambiguously define the interface contract that drivers must implement and users must follow.
Audio streams are device nodes published by driver services intended to be used by applications in order to capture or render audio on a Fuchsia device. Each stream in the system (input or output) represents a stream of digital audio information that may be either received or transmitted by device. Streams are dynamic and may created or destroyed by the system at any time. Which streams exist at any given point in time, and what controls their lifecycles are considered to be issues of audio policy and codec management and are not discussed in this document. Additionally, the information present in audio outputs streams is exclusive to the application owner of the stream. Mixing of audio is not a service provided by the audio stream interface.
|Sample||A representation of the sound rendered by a single speaker, or captured by a single microphone, at a single instant in time.|
|LPCM||Linear pulse code modulation. The specific representation of audio samples present in all Fuchsia uncompressed audio streams. LPCM audio samples are representations of the amplitude of the audio signal at an instant in time where the numeric values of the encoded audio are linearly distributed across the amplitude levels of the rendering or capture device. This is in contrast to A-law and μ-law encodings, which have non-linear mappings from numeric value to amplitude level.|
|Channel||Within an audio stream, the subset of information that will be rendered by a single speaker, or which was captured by a single microphone in a stream.|
|Frame||A set of audio samples for every channel of a audio stream captured/rendered at a single instant in time.|
|Frame Rate||a.k.a. "Sample Rate". The rate (in Hz) at which audio frames are produced or consumed. Common sample rates include 44.1 KHz, 48 KHz, 96 KHz, and so on.|
|Client or User or Application||These terms are used interchangeably in this document. They refer to modules that use these interfaces to communicate with an audio driver/device.|
Communication with an audio stream device is performed using messages sent over a channel. Applications open the device node for a stream and obtain a channel by issuing a FIDL request. After obtaining the channel, the device node may be closed. All subsequent communication with the stream occurs using channels.
The stream channel is used for most command and control tasks, including:
- Capability interrogation
- Format negotiation
- Hardware gain control
- Determining outboard latency
- Plug detection notification
- Access control capability detection and signalling
In order to actually send or receive audio information on the stream, the
specific format to be used must first be set. The response to a successful
CreateRingBuffer operation will contain a new "ring-buffer" channel. The ring-buffer
channel may be used to request a shared buffer from the stream (delivered in the
form of a VMO) that may be mapped into the address
space of the application and used to send or receive audio data as appropriate.
Generally, the operations conducted over the ring buffer channel include:
- Requesting a shared buffer
- Starting and Stopping stream playback and capture
- Receiving notifications of playback and capture progress
- Receiving clock recovery information in the case that the audio output clock is based on a different oscillator than the oscillator that backs the monotonic clock
Audio stream device nodes must be published by drivers using the protocol preprocessor symbol given in the table below. This will cause stream device nodes to be published in the locations given in the table. Applications can monitor these directories in order to discover new streams as they are published by the drivers.
Establishing the stream channel
After opening the device node, client applications may obtain a stream channel
for subsequent communication using the
fuchsia.hardware.audio.Device/GetChannel FIDL message.
Client side termination of the stream channel
Clients may terminate the connection to the stream at any time simply by calling zx_handle_close(...) on the stream channel. Drivers must close any active ring-buffer channels established using this stream channel and must make every attempt to gracefully quiesce any on-going streaming operations in the process.
Sending and receiving messages on the stream and ring-buffer channels
All of the messages and message payloads that may be sent or received over stream and ring buffer channels are defined in stream_config.fidl and ring_buffer.fidl. Messages may be sent to the driver using the zx_channel_write(...) syscall. If a response is expected, it may be read using the zx_channel_read(...) syscall. Best practice, however, is to queue packets for your channel(s) port using the zx_port_queue(...) syscall, and use the zx_port_wait(...) syscall to determine when your set of channels have messages (either expected responses or asynchronous notifications) to be read. There are bindings for different languages to facilitate sending and receiving FIDL messages, and in particular for C++ drivers there is also a library SimpleAudioStream that facilitates the creation of drivers in C++, this library uses the new C++ bindings to send and receive FIDL messages.
Format related protocol messages allow the driver to list its supported
formats to the client. The supported formats may include multiple rates, bit per sample,
etc. Each driver advertises what it can support and the client mandates what format
is to be used for each driver.
To find out what formats are supported by a given driver, the client uses the
GetSupportedFormats function. The driver replies with a vector of
SupportedFormats, where each
SupportedFormats includes a
- A vector of number of channels. This lists the number of channels supported
by the driver, for example
<2,4,6,8>. A driver that supports either two or four channels would report a vector with two elements
<2,4>. Must be in ascending order.
- A vector of sample formats, e.g.
- A vector of rates. Frame rates, for example 44100, 48000, and 96000. Must be in ascending order.
- A number of bytes per sample. Must be in ascending order.
- A vector of bits per sample. Sample widths, this could be smaller than the total available bytes, e.g. 24 bits in a 4-byte sample. Must be in ascending order.
When not all combinations supported by the driver can be described with one
PcmSupportedFormats, the driver returns more than one
the returned vector. For example, if one
PcmSupportedFormats allows for 16 or 32 bits samples at
48KHz, and 16 bits samples at 96KHz, but not 32 bits samples at 96KHz, then the driver
replies with 2
<<16bits>,<96KHz>>. For simplicity, this example ignores parameters other than
rate and bits per sample. In the case where the driver supports either 16 or 32
bits samples at either 48 or 96KHz, the driver would reply with 1
Additionally, it is assumed that bits per sample is always smaller or equal to
8 * bytes per sample. Hence, a driver can report
and this does not imply that it is reporting that 32 bits per sample on 2-byte
samples is valid, it specifies only the 3 valid combinations:
- 2-byte samples channels with 16 valid bits
- 4-byte samples with 32 valid bits
- 4-byte samples with 16 valid bits
The client specifies the format to use with the
CreateRingBuffer function based on
information that the driver provides in
GetSupportedFormats reply, what is supported
by the client, and any other requirements. This function takes a parameter that specifies:
- A number of channels. This is the number of channels available in the buffer.
- A bitmask of channels to use. These are the channels in the buffer to be used by
the driver. For instance for stereo this must be a bitmask with 2 bits enabled
0x3, i.e. both channels 0 and 1 are used.
- A sample format.
- A frame rate.
- A number of bytes per sample.
- A number of bits per sample.
- By default, multi-byte sample formats are assumed to use host-endianness.
PCM_FLOATencoding uses specifically the IEEE 754 floating point representation.
- By default, non-floating point PCM encodings are assumed expressed using
two's complement signed
integers. eg. the bit values for a 16 bit PCM sample format would range from
[0x8000, 0x7FFF] with 0x0000 representing zero speaker deflection. If the
PCM_UNSIGNEDsample format is used, the bit values would range from [0x0000, 0xFFFF] with 0x8000 representing zero deflection.
- When encoding a smaller sample size in a larger channel (e.g. 20 or 24-bit in 32), the most significant bits of the 32-bit container are used while the least significant bits will be ignored (left justified). e.g. a 20-bit sample would be mapped onto the range [12,31] (bits [0,11] would be ignored) of the 32-bit container.
Setting the desired stream format
In order to select a stream format, applications send a
CreateRingBuffer message over the
stream channel. In the message, the application specifies the format to be used.
The client specifies the new ring buffer channel over which streaming operations will be conducted. If a previous ring buffer channel had been established and was still active, the driver must close this channel and make every attempt to gracefully quiesce any on-going streaming operations in the process.
Determining external latency
The external latency of an audio stream is defined as the amount of time it takes outbound audio to travel from the system's interconnect to the speakers themselves, or inbound audio to travel from the microphone to the system's interconnect. As an example, consider an external codec connected to the system using a TDM interconnect: if this interconnect introduces a 4 frame delay between the reception of a TDM frame and the rendering of that frame at the speakers themselves, then the external delay of this audio path is the time duration equivalent to 4 audio frames.
External delay is reported in the
external_delay field of a
response to a
WatchDelayInfo. Drivers should make their best attempt to
accurately report the total of all of the sources of delay the driver knows about.
Information about this delay can frequently be found in codec data sheets,
dynamically reported as properties of codecs using protocols such as Intel HDA
or the USB Audio specifications, or reported by down stream devices using
mechanisms such as EDID when using HDMI or DisplayPort interconnects.
Determining turn on delay
turn_on_delay of an audio stream is defined as the amount of time it may
take the audio samples on the ring buffer to actually start playing on the
speakers after a
Start command has been issued, or audio from the microphone
to start getting recorded into the ring buffer also after the
command. For instance, if we have an external codec connected to the system and
we have turned it off to save power, then after a
Start command is issued, the
driver providing the ring buffer will reply with a
start_time indicating when
the position on the ring buffer started moving and audio samples started to be
sent to the external codec. However, independent of the
Determining external latency above), the external codec may still be powering
up from a lower power mode. The
turn_on_delay, specified in the
RingBufferProperties table, is the driver's best upper bound estimation of the
amount of time it takes to get the external codec to actually output the audio
samples. The delay may also be due to other reasons, for instance, the codec's
built in ramp up delay to avoid glitches, or if the driver abstracts Bluetooth
communications the delay getting the remote device to start playing the audio
turn_on_delay is an upper limit estimation, audio may start
playing or being-captured before
turn_on_delay has passed, this is why this
delay must be taken into account if not getting the initial audio samples played
or captured is not acceptable.
external_delay represents an amount of time the audio samples get delayed
getting to the speakers or from microphones.
turn_on_delay does not delay the
audio samples and it does not indicate any buffering, but rather indicates an
amount of time during which audio samples may not actually be played/recorded.
turn_on_delay does not affect the calculation of the presentation time, but it
does affect if presentation is happening at all. For playback, we can visualize
these delays for getting audio samples into a speaker as:
|<--- external delay --->| |S|--|T|----------------------|P|----|O| |<------- turn on delay ------->|
S indicates the time when
Start was issued,
T indicates the time
when the position in the ring buffer started to move i.e.
returned from the
P indicates the presentation time in the
O indicates the time the amplifier completed turning on and audio
can be heard. Since
O is past the
turn_on_delay here affected actual
presentation on the speaker.
As time passes, we can visualize the current position on the ring buffer
below advancing and the samples now being presented in the speaker at time
past the time
O when the amplifier turned on.
|<--- external delay --->| |S|--|T|---------|C|-----------------|O|--|P| |<------- turn on delay ------->| |<--- external delay --->| |S|--|T|-----------------------------|O|-|C|---------------------|P| |<------- turn on delay ------->|
Hardware gain control
Hardware gain control capability reporting
In order to determine a stream's gain control capabilities, if it has not done
so yet, an application sends a
GetProperties message over the stream channel.
No parameters need to be supplied with this message. The driver replies with a
StreamProperties including gain capabilities among others. All stream drivers
must respond to this message, regardless of whether or not the stream
hardware is capable of any gain control. All gain values are expressed using 32
bit floating point numbers expressed in dB.
Drivers respond to this message with values that indicate the current stream's
gain control capabilities. Current gain settings are expressed using a bool
indicating whether the stream can be muted, a bool that indicates whether the
stream can AGC, the minimum and maximum gain settings, and a
gain_step_db indicates the smallest increment with which the gain can be
controlled counting from the minimum gain value.
For example, an amplifier that has 5 gain steps of 7.5 dB each and a maximum 0 dB gain would indicate a range of (-30.0, 0.0) and a step size of 7.5. Amplifiers capable of functionally continuous gain control may encode their gain step size as 0.0.
Regardless of mute capabilities, drivers for fixed gain streams must report
their min and max gain as (0.0, 0.0).
gain_step_db is meaningless in this
situation, but drivers should report it as 0.0.
Setting hardware gain control levels
In order to change a stream's current gain settings, applications send a
SetGain message over the stream channel. This message include a parameter
GainState indicating gain parameters to be configured including the dB gain that
should be applied to the stream, muted and AGC enablement.
Presuming that the request is valid, drivers should round the request to the nearest supported gain step size. For example, if a stream can control its gain on the range from -60.0 to 0.0 dB, using a gain step size of 0.5 dB, then a request to set the gain to -33.3 dB should result in a gain of -33.5 being applied. A request to that same stream for a gain of -33.2 dB should result in a gain of -33.0 being applied.
Gain state notifications
Clients may request that streams send them asynchronous notifications of
gain state changes by using the
WatchGainState command. The driver will reply to the
first |WatchGainState| sent by the client and will not respond to subsequent
client |WatchGainState| calls until the gain state changes from what was most recently
In addition to streams being published/unpublished in response to being connected or disconnected to/from their bus, streams may have the ability to be plugged or unplugged at any given point in time. For example, a set of USB headphones may publish a new output stream when connected to USB, but choose to be "hardwired" from a plug detection standpoint. A different USB audio adapter with a standard 3.5mm phono jack might publish an output stream when connected with USB, but choose to change its plugged/unplugged state as the user plugs and unplugs an analog device with the 3.5mm jack.
The ability to query the currently plugged or unplugged state of a stream, and to register for asynchonous notifications of plug state changes (if supported) is handled through plug detection messages.
Plug detect capabilities
In order to determine a stream's plug detection capabilities, if it has not done
so yet, an application sends a
GetProperties command over the stream channel.
The driver replies with a
StreamProperties including plug detect capabilities
plug_detect_capabilities among others fields.
Valid plug-detect capabilities flags currently defined are:
HARDWIREDis set when the stream hardware is considered to be "hardwired". In other words, the stream is considered to be connected as long as the device is published. Examples include a set of built-in speakers, a pair of USB headphones, or a pluggable audio device with no plug detection functionality.
CAN_ASYNC_NOTIFYis set when the stream hardware is capable of both asynchronously detecting that a device's plug state has changed, and sending a notification message if the client has requested these notifications.
Plug state notifications
Clients may request that streams send them asynchronous notifications of
plug state changes by using the
WatchPlugState command if the
flag was sent by the driver in
StreamProperties. I.e. drivers for streams that
do not set the
CAN_ASYNC_NOTIFY flag are free to ignore the
by applications. Driver with
CAN_ASYNC_NOTIFY set will reply to the first
|WatchPlugState| sent by the client and will not respond to subsequent client
|WatchPlugState| calls until the plug state changes from what was most recently reported.
Stream purpose and association
Once an application has successfully set the format of a stream, it receives in the response a new channel representing its connection to the stream's ring-buffer. Clients use the ring-buffer channel to establish a shared memory buffer and start and stop playback and capture of audio stream data.
The ring buffer contents are produced by the client side (for playback) and the driver side (for recording). Hence, a client is a producer for playback and a consumer for recording and a driver is a producer for recording and a consumer for playback. The ring buffer contents may be directly consumed or produced by the audio hardware, or it may go through software processing of each sample done by the driver.
Ring buffer data production proceeds at the nominal rate from the point in time
given in a successful response to the
Start command. Note though that the ring-buffer
will almost certainly have some form of FIFO buffer
between the memory bus and the audio hardware, which causes it to either
read-ahead in the stream (in the case of playback), or potentially hold onto
data (in the case of capturing). It is important for clients to query the size
of this buffer before beginning
operation so they know how far ahead/behind the stream's nominal inferred
read/write position they need to stay in order to prevent audio glitching.
Also note that because of the shared buffer nature of the system, and the fact
that drivers are likely to be DMA-ing directly from this buffer to hardware, it
is important for clients running on architectures that are not automatically
cache coherent to be sure that they have properly written-back their cache after
writing playback data to the buffer, or invalidated their cache before reading
captured data. See
for a description of ring buffer data transfers.
Obtaining a shared buffer
To send or receive audio, the application must first establish a shared memory
buffer. This is done by sending an
CreateRingBuffer request over the
ring-buffer channel. This may only be done while the ring-buffer is stopped.
If the channel created with
CreateRingBuffer is closed by the driver for instance
because a buffer has already been established and the ring-buffer has already
been started, it must not either stop the ring-buffer, or discard the
existing shared memory. If the application requests a new buffer after having
already established a buffer while the ring buffer is stopped, it must
consider the existing buffer ii has to be invalid, the old buffer is now gone.
Applications must specify two parameters when requesting a ring buffer:
The minimum number of frames of audio the client needs allocated for the ring buffer. Drivers may make this buffer larger to meet hardware requirements. Clients must use the returned VMOs size (in bytes) to determine the actual size of the ring buffer. Clients must not assume that the size of the buffer (as determined by the driver) is exactly the size they requested. Drivers must ensure that the size of the ring buffer is an integral number of audio frames.
Optional number of position update notifications the client would like the driver to
send per cycle through the ring buffer, these notifications are meant to be used for clock
recovery. Drivers must only send these as a reply to a
Drivers should attempt to space notifications uniformly throughout the ring; however clients
must not rely on perfectly uniform spacing of the update notifications.
If the request succeeds, the driver must return a handle to a VMO with permissions that allow applications to map the VMO into their address space using zx_vmar_map, and to read/write data in the buffer in the case of playback, or simply to read the data in the buffer in the case of capture.
If the request succeeds, the driver will also return the actual number of frames of audio
it will use in the buffer. The size of the VMO returned (as reported
by zx_vmo_get_size()) must not be larger than
this number of frames (when converted to bytes). This number may be larger
min_frames request from the client but must not be smaller than this number.
Starting and Stopping the ring-buffer
Clients may request that a ring-buffer start or stop using the
commands. Attempting to start a stream
which is already started must be considered a failure. Attempting to stop a
stream that is already stopped should be considered a success. Ring-buffers
cannot be either stopped or started until after a shared buffer has been
established using the
Upon successfully starting a stream, drivers must provide their best estimate of
the time at which their hardware began to transmit or capture the stream in the
start_time field of the response. This time stamp must be taken from the clock
exposed with the
syscall. Along with the FIFO depth property of the ring buffer, this timestamp
allows applications to send or receive stream data without the need for periodic
position updates from the driver. Along with the outboard latency estimate
provided by the stream channel, this timestamp allows applications to
synchronize presentation of audio information across multiple streams, or even
multiple devices (provided that an external time synchronization protocol is
used to synchronize the
monotonic timelines across
the cohort of synchronized devices).
Upon successfully starting a stream, drivers must guarantee that no position notifications will be sent before the start response has been enqueued into the ring-buffer channel.
Upon successfully stopping a stream, drivers must guarantee that no position notifications will be enqueued into the ring-buffer channel after the stop response has been enqueued.
If requested by the client through a non-zero
clock_recovery_notifications_per_ring in the
CreateRingBuffer operation, the driver will
periodically send updates to the client informing it of its current production
or consumption position in the buffer. This position is expressed in bytes in
position field of the
RingBufferPositionInfo struct sent on
a reply to the
WatchClockRecoveryPositionInfo message. The
message also includes a
timestamp field that contains the time (as
zx::time) that this byte position was valid. A
request must only be sent after
clock_recovery_notifications_per_ring has been
specified in the
GetVmo function and the
GetVmo function has returned. Note,
notifications indicate where in the buffer the driver has consumed or produced
data, not the nominal playback or capture position (sometimes called the
"write cursor" or "read cursor" respectively). The timing of their arrival is
not guaranteed to be perfectly uniform and should not be used to effect clock
recovery. However, the correspondence pair (
values themselves ARE intended to be used to recover the clock for the audio
stream. If a client discovers that a driver has consumed past the point in the
ring buffer where that client has written playback data, audio presentation is
undefined. Clients should increase their clock lead time and be certain to stay
ahead of this point in the stream in the future. Likewise, clients that capture
audio should not attempt to read beyond the point in the ring buffer
indicated by the most recent position notification sent by the driver.
Driver playback and capture position must always begin at ring buffer byte 0,
immediately following a successful
Start command. When the ring
buffer position reaches the end of the VMO (as indicated by
zx_vmo_get_size(...)), the ring buffer position
wraps back to zero. Drivers are not required to consume or produce data in
integral numbers of audio frames. Clients whose notion of stream position
depends on position notifications should take care to request that a sufficient
number of notifications per ring be sent (minimum 2) and to process them quickly
enough that aliasing does not occur.
Clock recovery and synchronization
AUDIO_STREAM_CMD_GET_CLOCK_DOMAIN message, the driver must
respond with the identifier of the clock domain containing that device. If the
audio device is locked to the local system monotonic clock and does not expose a
mechanism by which its rate would be fine-tuned, then it should return the value
0 to represent the local CLOCK_MONOTONIC domain. A client may use this
information (in addition to
AUDIO_RB_POSITION_NOTIFY messages) to simplify the
process of recovering the audio device's clock.
Unexpected client termination
If the client side of a ring buffer control channel is closed for any reason, drivers must immediately close the control channel and shut down the ring buffer, such that no further audio is emitted nor captured. While drivers are encouraged to do so in a way that produces a graceful transition to silence, they must ensure that the audio stream goes silent instead of looping. Once the transition to silence is complete, resources associated with playback or capture may be released and reused by the driver.
This way, if a playback client teminates unexpectedly, the system will close the client channels, causing audio playback to stop instead of continuing to loop.