The AVB enabled Ethernet MAC also includes a IEEE 1722 packet router that routes audio packets to the listener components in the system. It controls the routing by stream ID. This requires no configuration and is controlled implicitly via the AVB API described in Section AVB API.

Audio is transported in streams of data, where each stream may have multiple channels. Endpoints producing streams are called Talkers and those receiving them are called Listeners. Each stream on the network has a unique 64-bit stream ID.

A single endpoint can be a Talker, a Listener or both. In general each endpoint will have a number of sinks with the capacity to receive a number of incoming streams and a number of sources with the capacity to transmit a number of streams.

Routing is done using layer 2 Ethernet addresses. Each stream is sent from a particular source MAC address to a particular destination MAC address. The destination MAC address is a multicast address so that several Listeners may receive it. In addition, AVB switches can reserve an end-to-end path with guaranteed bandwidth for a stream. This is done by the Talker endpoint advertising the stream to the switches and the Listener(s) registering to receive it. If sufficient bandwidth is not available, this registration will fail.

Streams carry their own presentation time, the time that samples are due to be output, allowing multiple Listeners that receive the same stream to output in sync.

• Streams are encoded using the 1722 AVB transport protocol.
• All channels in a stream must be synchronized to the same sample clock.
• All the channels in a stream must come from the same Talker.
• Routing of audio streams uses Ethernet layer 2 routing based on a multicast destination MAC address
• Routing of channels is done at the stream level. All channels within a stream must be routed to the same place. However, a stream can be multicast to several Listeners, each of which picks out different channels.
• A single end point can be both a Talker and Listener.
• Information such as stream ID and destination MAC address of a Talker stream should be communicated to Listeners via 1722.1. (see Section Device Discovery, Connection Management and Control).

As described in the previous section, an IEEE 1722 audio stream may consist of many channels. These channels need to be routed to particular audio I/Os on the endpoint. To achieve maximum flexibility the XMOS design uses intermediate media FIFOs to route audio. Each FIFO contains a single channel of audio.

The above figure shows the breakdown of 1722 streams into local FIFOs. The figure shows four points where transitions to and from media FIFOs occur. For audio being received by an endpoint:

1. When a 1722 stream is received, its channels are mapped to output media FIFOs. This mapping can be configured dynamically so that it can be changed at runtime by the configuration component.
2. The digital hardware interface maps media FIFOs to audio outputs. This mapping is fixed and is configured statically in the software.

For audio being transmitted by an endpoint:

1. The digital hardware interface maps digital audio inputs to local media FIFOs. This mapping is fixed and cannot be changed at runtime.
2. Several input FIFOs can be combined into a 1722 stream. This mapping is dynamic.

The configuration of the mappings is handled through the API describe in AVB API.

Media FIFOs uses shared memory to move data between tasks, thus the filling and emptying of the FIFO must be on the same tile.

A talker unit consists of one logcial core which creates IEEE 1722 packets and passes the audio samples onto the MAC. Audio samples are passed to this component via input media FIFOs. Samples are pushed into this FIFO from a different task implementing the audio hardware interface. The packetizer task removes the samples and combines them into IEEE 1722 Ethernet packets to be transmitted via the MAC component.

When the packets are created the timestamps are converted to the time domain of the global clock provided by the PTP component, and a fixed offset is added to the timestamps to provide the presentation time of the samples (i.e the time at which the sample should be played by a Listener).

A system may have several Talker units. However, since samples are passed via a shared memory interface a talker can only combine input FIFOs that are created on the same tile as the talker. The instantiating of talker units is performed via the API described in Section Component functions. Once the talker unit starts, it registers with the main control task and is control via the main AVB API described in Section AVB API.

A Listener unit takes IEEE 1722 packets from the MAC and converts them into a sample stream to be fed into a media FIFOs. Each audio Listener component can listen to several IEEE 1722 streams.

A system may have several Listener units. The instantiating of Listener units is performed via the API described in Section Component functions. Once the Listener unit starts, it registers with the main control task and is controlled via the main AVB API described in Section AVB API.

Sometimes it is useful to convert the audio stream in a media FIFO into a sample stream over an XC channel. This may be needed to move samples off tile or if the audio interface task requires samples over a channel. Several functions are provided to do this and are described in Section Component functions.

The audio hardware interface components drive external audio hardware, pull audio out of media output FIFOs and push into media input FIFOs.

Different interfaces interact in different ways, some directly push and pull from the media FIFOs, whereas some for performance reasons require samples to be provided of an XC channel.

The following diagram shows one potential layout of the I2S component which pushes its input directly to media input FIFOs but takes output FIFOs from an XC channel. The diagram shows the supporting task that takes samples out of the media output FIFOs and serializes them over an XC channel:

A high quality, low jitter master clock is often required to drive an audio CODEC and must be synchronized with an AVB media clock. The XS1 chip cannot provide this clock directly but can provide a lower frequency source for a frequency synthesizer chip or external PLL chip. The frequency synthesizer chip must be able to generate a high frequency clock based on a lower frequency signal, such as the Cirrus Logic CS2100-CP. The recommended configuration is as in the block diagram below:

The XS1 device provides control to the frequency synthesizer and the frequency synthesizer provides the audio master clock to the CODEC and XS1 device. The sample bit and word clocks are then provided to the CODEC by the XS1 device.

In addition to components described in previous sections, an AVB endpoint application requires a task to control and configure the system. This control task varies across applications but the protocol to provide device discovery, connection management and control services has been standardised by the IEEE in 1722.1.

The 1722.1 standard defines four independent steps that can be used to connect end stations that use 1722 streams to transport media across a LAN. The steps are:

1. Discovery
2. Enumeration
3. Connection Management
4. Control

These steps can be used together to form a system of end stations that interoperate with each other in a standards compliant way. The application that will use these individual steps is called a Controller and is the third member in the Talker, Listener and Controller device relationship.

A Controller may exist within a Talker, a Listener, or exist remotely within the network in a separate endpoint or general purpose computer.

The Controller can use the individual steps to find, connect and control entities on the network but it may choose to not use all of the steps if the Controller already knows some of the information (e.g. hard coded values assigned by user/hardware switch or values from previous session establishment) that can be gained in using the steps. The only required step is connection management because this is the step that establishes the bandwidth usage and reservations across the AVB cloud.

The four steps are broken down as follows:

• Discovery is the process of finding AVB endpoints on the LAN that have services that are useful to the other AVB endpoints on the network. The discovery process also covers the termination of the publication of those services on the network.
• Enumeration is the process of the collection of information from the AVB endpoint that could help an 1722.1 Controller to use the capabilities of the AVB endpoint. This information can be used for connection management.
• Connection management is the process of connecting or disconnecting one or more streams between two or more AVB endpoint.
• Control is the process of adjusting a parameter on the endpoint from another Entity. There are a number of standard types of controls used in media devices like volume control, mute control and so on. A framework of basic commands allows the control process to be extended by the endpoint.

The XMOS endpoint provides full support for Talker and Listener 1722.1 services. Basic 1722.1 Controller functionality is available to allow ‘plug and play’ connection between two XMOS endpoints, however, it is expected that GUI Controller software will be available on the network for setting up larger topologies.

To assist in this task a unified control API is presented in Section AVB API.

Each XMOS device has a set of resources detailed in the following table. The resources are split amongst different tiles on the device which may affect how resources can be used:

Note that some ports overlap on the device so, for example, using a 16 bit port may make some 1 bit ports unavailable. See the device datasheets for details.

The following sections detail the resource required for each component. Please note that the memory requirements for code size should be taken as a rough guide since exact memory usage depends on the integration of components (which components are on which tile etc.) in the final build of the application.

Each endpoint requires an Ethernet MAC layer.

Every AVB endpoint must include a PTP component.

Every AVB endpoint must include a media clock server.

If the endpoint drives an external PLL, a PLL driver component is required.

PTP, Media Clock Server and PLL driver components may be combined into a single logical core running at 100 MIPS if the number of channels is constrained.

Each endpoint may have several listener and talker components. Each listener/talker component is capable of handling four IEEE 1722 streams and up to 12 channels of audio.

The Talker and Listener components may be combined into a single logical core running at 100 MIPS if the number of streams is 1 and the number of channels is <= 4 per stream.

The amount of resource required for audio processing depends on the interface and the number of audio channels required. The overheads for the interface are:

The following table shows that number of channels an interface can handle per logical core:

Note that several instances of the audio interface component can be made e.g. you could use 2 logical cores to handle 16 channels of I2S. The following table shows how much buffering memory is required depending on the number of audio channels.

In addition to the other components there are application dependant tasks that control other I/O. For general configuration and slow I/O a minimum of 1 logical core (50 MIPS) should be reserved.