The Chromium Projects

Search this site

Except as otherwise noted, the content of this page is licensed under a Creative Commons Attribution 2.5 license, and examples are licensed under the BSD License.
Chromium OS‎ > ‎Design Documents‎ > ‎

CRAS: ChromeOS Audio Server

Why


ChromeOS needs a way to route audio to different devices.  Currently ChromeOS only plays audio to the first ALSA device in the system.  Devices attached after boot are ignored, and no priority is given to the built in sound device.
The goal of the new server is to allow for sound to be routed dynamically to newly attached audio-capable monitors(DisplayPort and HDMI), USB webcam, USB speakers, bluetooth headsets, etc.  And to do this in a way that requires as little CPU as possible and that adds little or no latency.

How

Put an audio server in between Chrome and ALSA that can change how audio flows on the fly.  Chrome will connect to this audio server, send audio data to it, and the audio server will render this to the selected ALSA device through alsa-lib.  For input the process will be reversed, the server will get audio samples from alsa and forward them to Chrome.

Details

The design decisions were driven by three main requirements: Low Latency, Low CPU usage, and dynamic audio routing.  This will handle the streaming of PCM only; audio decode will all be done in Chrome.

The basic design combines the timer-driven wakeup scheme of pulseaudio with the synchronous audio callbacks of JACK.  This synchronous callback eliminates local buffering of data in the server, allowing the server to wake up less for a given latency.  The server wakes up when a timer fires, sometime before there is an ALSA over/underrun and calls back to Chrome to get/put audio data.  This minimizes latency as the only additional latency needed is that to exchange IPC messages between the server and client.  Because ALSA wakeups aren’t used, on some devices the minimal latency will be lower than without the server; there is no need to wake up at an ALSA period boundary.  The timer can be armed to wake up when there are an arbitrary number of samples left in the hardware buffer.  ALSA will always be configured with the largest possible buffer size, yet only some fraction of that will be used at any given time.

Audio data will be exchanged through shared memory so that no copying is needed (except into the ALSA buffer).  Communication will happen through UNIX domain sockets.  There will be separate communication connections for control (high latency, lower priority) and audio data (low latency and higher priority).

At each wakeup a timestamp will be updated that, for playback streams, will indicate the exact time at which the next written sample will be rendered to the DAC.  For capture, the timestamp will reflect the actual time the sample was captured at the ADC.

The key to keeping system resource usage low is to do as little as possible in the server when it wakes up and similar for the callback in the client.  Less than two percent of the CPU of a CR48 is used while playing back audio with 20ms latency and waking up every 10ms.  The minimum the server can do is read in the streams, mix them, and write the result to ALSA. Other processing can take place in the server; DSP blocks can be configured to process the buffers as they pass through, this will add to the processing time and increase the minimum latency.

The lowest latency stream will be used to drive the timer wakeup interval (and will have the most accurate latency).  Only clients that need to send more data will be woken.

Device Enumeration

The server keeps a separate list of inputs and outputs.  The server will listen to udev events that notify it of newly added or removed devices. Whenever a device appears or disappears, the priority list of devices will be re-evaluated and the highest priority device will be used.

Devices will be given a priority based on the device type and the time they are inserted. The priority is as follows (for both input and output):
1) Headset jack and USB audio devices
2) DisplayPort and HDMI outputs (when EDID indicates the sink is audio capable).
3) Built-in speakers.

If two devices of the same priority are attached, the device that was plugged most recently wins.

On device removal, all streams are removed from the device and the device is closed.  The streams will be moved to the next device in the list when this happens.

Mixing

The server is required to mix for the case where multiple streams are being played at once.   The mixing will be done immediately before the samples are sent to the playback hardware. If a DSP block is enabled, it will be applied post-mix.

Per-Board Configuration

Without a config, the server will automatically discover devices and mute/volume controls. This provides all that is need for x86 most of the time. On x86 the main use of config files is to set volume level mappings between ui and hardware, and to configure codec-specific features.

The volume curve file is per-codec, per-device and the format is described in the main README file.

ALSA UCM is used to do low level codec config. On arm systems, this setup is necessary to get audio paths routed correctly when an headset is attached and to initialize codec setup after start up. In addition to this, the UCM file is used to map jacks to audio devices.



Comments