AGL Advanced Audio Architecture (4a)


4a is an audio framework made to handle automotive use cases like stream priority and interruptions. It does not deal with the audio by itself but rather with the policies and signaling around it. This means that 4a must be used in conjunction with an other API that is in charge of the audio stuff, like alsa or pulseaudio.


Global architecture

applications architecture diagram

This detailed diagram shows the interactions between all components.

applications architecture diagram

This diagram shows the global architecture that any application playing audio should follow. As a developer, when you want to play audio, you rely on APIs such as alsa, pulseaudio or higher level API(s). To play audio you usually have some kind of device URI that you have to open in order to write audio data to it.

4a does not change anything about that. You still have to do all of this. 4a provides permissions, signals and policies. This means that all compliant applications have to ask 4a for a ‘role’ when they want to play audio. If 4a grants them the permission, it returns a device URI that applications have to open.

Nothing prevent an application to directly open a device, but in this case no policies can be applied and access permissions are handled by system policies.

4a Architecture

4a's architecture diagram

Even if applications only deal with the high level API, 4a is made of multiple components that rely on each other.

The high level API uses the hal-manager to list enabled HALs then it uses directly those HALs.

HALs rely directly on drivers and/or softmixer.


The 4a softmixer is a sound server. It is a more secured, more CPU-friendly server than PulseAudio.

It is mandatory to have a sound server to share a sound card between several applications, that may write to the sound card at different sound rates and formats.

It is the responsibility of the 4a-softmixer to instantiate the dmix alsa plugin, based on the 4A config files.


snd-aloop, which has recently been replaced by a more flexible implementation (called avirt), proxifies the access from playback application to sound cards.

The 4a-softmixer reads from snd-aloop (or avirt) captures, and will perform the rate conversion, soft volume control for hardware that does not implement it, routing to configured audio channels, and mixing at the end.

A security label (SMACK) can be put on the snd-aloop PCM playback devices, and since each these devices matches an audio role, that applications have to claim to the 4A high level API, this makes a gatekeeper for audio apps.

snd-aloop introduces one more buffer copy; but the measured performances and CPU load on the tested boards do not have revealed any unacceptable levels, up to now.

The CPU load (which could be even better with some unimplemented optimizations today) is typically less than 5% on the Renesas RCAR board, and similar numbers have been seen on Intel Atom boards.


Introduced recently on the master branch of AGL, it provides to 4A high level a common interface for both softmixing, and DSP-accelerated mixing. I also sets user-readable names to the dynamically-configured playback devices, that match the 4A audio roles (aka streams).

The avirt-loop backend is a loop implementation that provides the same functionality as snd-aloop, with enhanced security (each playback has a dedicated PCM device, on which a SMACK label can be put), and with the flexibility offered by the top avirt layer.

It is not limited to 8 channels like snd-aloop.