Many hardware acceleration problems can be expressed in the dataflow paradigm. It models a program as a directed graph of the data flowing between functions. The nodes of the graph are functional units called actors, and the edges represent the connections (transporting data) between them.
Actors communicate by exchanging data units called tokens. A token contains arbitrary (user-defined) data, which is a record containing one or many fields, a field being a bit vector or another record. Token exchanges are atomic (i.e. all fields are transferred at once from the transmitting actor to the receiving actor).
Actors in Migen are implemented in FHDL. This low-level approach maximizes the practical flexibility: for example, an actor can manipulate the bus signals to implement a DMA master in order to read data from system memory.
Token exchange ports of actors are called endpoints. Endpoints are unidirectional and can be sources (which transmit tokens out of the actor) or sinks (which receive tokens into the actor).
..figure:: actors_endpoints.png
:scale:50 %
Actors and endpoints.
The flow of tokens is controlled using two handshake signals (strobe and acknowledgement) which are implemented by every endpoint. The strobe signal is driven by sources, and the acknowledgement signal by sinks.
It is permitted to generate an ``ack`` signal combinatorially from one or several ``stb`` signals. However, there should not be any combinatorial path from an ``ack`` to a ``stb`` signal.
Actors are derived from the the ``migen.flow.actor.Actor`` base class. The constructor of this base class takes a variable number of parameters, each describing one endpoint of the actor.
An endpoint description is a triple consisting of:
* The endpoint's name.
* A reference to the ``migen.flow.actor.Sink`` or the ``migen.flow.actor.Source`` class, defining the token direction of the endpoint.
* The layout of the data record that the endpoint is dealing with.
Record layouts are a list of fields. Each field is described by a pair consisting of:
* The field's name.
* Either a BV object (see :ref:`bv`) if the field is a bit vector, or another record layout if the field is a lower-level record.
* One sink named ``operands`` accepting data structured as a 16-bit field ``a`` and a 16-bit field ``b``. Note that this is functionally different from having two endpoints ``a`` and ``b``, each accepting a single 16-bit field. With a single endpoint, the data is strobed when *both*``a`` and ``b`` are valid, and ``a`` and ``b`` are *both* acknowledged *atomically*. With two endpoints, the actor has to deal with accepting ``a`` and ``b`` independently. Plumbing actors (see :ref:`plumbing`) and abstract networks (see :ref:`actornetworks`) provide a systematic way of converting between these two behaviours, so user actors should implement the behaviour that results in the simplest or highest performance design.
* One source named ``result`` transmitting a single 17-bit field named ``r``.
Implementing the functionality of the actor can be done in two ways:
* Overloading the ``get_fragment`` method.
* Overloading both the ``get_control_fragment`` and ``get_process_fragment`` methods. The ``get_control_fragment`` method should return a fragment that manipulates the control signals (strobes, acknowledgements and the actor's busy signal) while ``get_process_fragment`` should return a fragment that manipulates the token payload. Overloading ``get_control_fragment`` alone allows you to define abstract actor classes implementing a given scheduling model. Migen comes with a library of such abstract classes for the most common schedules (see :ref:`schedmod`).
Accessing the endpoints is done via the ``endpoints`` dictionary, which is keyed by endpoint names and contains instances of the ``migen.flow.actor.Endpoint`` class. The latter holds:
* A signal object ``stb``.
* A signal object ``ack``.
* The data payload ``token``. The individual fields are the items (in the Python sense) of this object.
The actor datapath is made entirely of combinatorial logic. The handshake signals pass through. A small integer adder would use this model.
N-sequential
------------
The actor consumes one token at its input, and it produces one output token after N cycles. It cannot accept new input tokens until it has produced its output. A multicycle integer divider would use this model.
N-pipelined
-----------
This is similar to the sequential model, but the actor can always accept new input tokens. It produces an output token N cycles of latency after accepting an input token. A pipelined multiplier would use this model.
It is conceivable that a CAL [cal]_ to FHDL compiler be implemented at some point, to support higher level descriptions of some actors and reuse of third-party RVC-CAL applications. [orcc]_[orcapps]_[opendf]_
