DSLX Tutorial: Intro to procs

Up to this point, our tutorials have described stateless, non-communicating, combinational modules. To add state or communication with other actors, we need to venture into the exciting land of procs!

Procs, short for "communicating sequential processes", are the means by which DSLX models sequential and stateful modules. A proc contains:

A init function that initializes the proc state.
A config function that spawns any other dependent/child procs needed for execution.
A recurrent (i.e., infinitely looping) next function that contains the actual logic to be executed by the proc.

A critical component of procs is communication: every proc needs a means to share data with other procs, or else it'd just be spinning dead code. These means are channels: entities into which data can be sent and from which data can be received. Each channel has a send and a receive endpoint: data inserted into a channel by a send op can be pulled out by a recv op.

Example proc

Many concepts here are more easily explained via example, so here's a possible DSLX implementation of FMAC (fused multiply-accumulate), which computes C = A * B + C:

import float32;

type F32 = float32::F32;

proc Fmac {
  input_a_consumer: chan<F32> in;
  input_b_consumer: chan<F32> in;
  output_producer: chan<F32> out;

  init {
    float32::zero(false)
  }

  config(input_a_consumer: chan<F32> in, input_b_consumer: chan<F32> in,
         output_producer: chan<F32> out) {
    (input_a_consumer, input_b_consumer, output_producer)
  }

  next(tok: token, state: F32) {
    let (tok_a, input_a) = recv(tok, input_a_consumer);
    let (tok_b, input_b) = recv(tok, input_b_consumer);
    let result = float32::fma(input_a, input_b, state);
    let tok = join(tok_a, tok_b);
    let tok = send(tok, output_producer, result);
    result
  }
}

(In practice, a FMAC unit would want a reset signal, but we've leaving that out for simplicity.)

There's a lot to unpack here, so we'll walk through the example.

The first part of interest is the declaration of the proc member values: the three channels. These are similar to class data members in software. In DSLX, proc members are constant values, and are set by the output of the config function. Member values can be referred to inside the next function in the same way as locally-declared data.

Next up is the init function, defining the initial value for the proc state.

Following is the config function itself. When a proc is "spawned", its given a sets of values that get passed to the config function (following this example, we'll show how procs are spawned). Inside a config function, any constant values can be computed, any necessary procs can be spawned, and finally member values are set by the return value. Member values are assigned in declaration order: the first element of the return tuple corresponds to the first-declared member, and so on.

After that, we encounter next. This function serves as the real "body" of the Proc. The next function maintains and evolves the proc's recurrent state and is responsible for communicating with the outside world, as well. In our example, the first two lines are that communication, receiving the input values for the computation. The token elements are used to sequence events: since the receives can happen in parallel, they can share the same token, but since sending the output must happen after that, their result tokens are "joined" (think joining two threads of execution in software), and the result is used to sequence the send. Between the communication routines is the actual computation.

At the end of the proc, we terminate with the result value. This final value becomes the input state for the next iteration. This is how recurrent state is managed by procs: a state value is provided to the next function, and the result of that function is used as the next iteration's state input. Procs can have several state elements: in that case, there will be several input state args, e.g., next(tok: token, state_a: F32, state_b: F32), and the output will be a tuple of values, corresponding in order to the input state values. Regardless of that, a next function will always take a token as the first parameter.

Spawning procs

In any real design, procs will form a network¹, where one proc will spawn any number of child procs, which themselves might spawn other procs. (The "root" proc will be instantiated by some outside component in the outside RTL environment.) Procs may only be spawned in config functions, as they're part of statically configuring the hardware network to construct.

As an example, spawning a couple of our procs above would look as follows:

proc Spawner {
  fmac_1_a_producer: chan<F32> out;
  fmac_1_b_producer: chan<F32> out;
  fmac_1_output_consumer: chan<F32> in;
  fmac_2_a_producer: chan<F32> out;
  fmac_2_b_producer: chan<F32> out;
  fmac_2_output_consumer: chan<F32> in;

  // ...

  config() {
    let (fmac_1_a_p, fmac_1_a_c) = chan<F32>("fmac_1_a");
    let (fmac_1_b_p, fmac_1_b_c) = chan<F32>("fmac_1_b");
    let (fmac_1_output_p, fmac_1_output_c) = chan<F32>("fmac_1_output");
    spawn Fmac(fmac_1_a_c, fmac_1_b_c, fmac_1_output_p);

    let (fmac_2_a_p, fmac_2_a_c) = chan<F32>("fmac_2_a");
    let (fmac_2_b_p, fmac_2_b_c) = chan<F32>("fmac_2_b");
    let (fmac_2_output_p, fmac_2_output_c) = chan<F32>("fmac_2_output");
    spawn Fmac(fmac_2_a_c, fmac_2_b_c, fmac_2_output_p);

    (fmac_1_a_p, fmac_1_b_p, fmac_1_output_c,
     fmac_2_a_p, fmac_2_b_p, fmac_2_output_c)
  }

  // ...
}

For each child proc, we first declare the necessary channels (each channel declaration produces a producer and consumer channel, respectively), then we actually spawn it. The first set of arguments is passed to the child's config function. A spawn produces no value, hence no let on the left-hand side.

Advanced features

Channel arrays and loop-based spawning

Note: This feature is currently WIP and is not yet available.

Many hardware layouts have regular arrays of components, such as systolic arrays, vector units, etc. Individually specifying these quickly grows cumbersome, so users can instead declare arrays of channels and spawn procs inside for loops. This looks as follows:

proc Spawner4x4 {
  input_producers: chan<F32>[4][4] out;
  output_consumers: chan<F32>[4][4] in;

  // ...

  config() {
    let (input_producers, input_consumers) = chan<F32>[4][4]("node_input");
    let (output_producers, output_consumers) = chan<F32>[4][4]("node_output");

    for (i, _) : (u32, ()) in range(u32:0, u32:4) {
      for (j, _) : (u32, ()) in range(u32:0, u32:4) {
        spawn Node(input_consumers[i][j],
                   output_producers[i][j]);
      }(());
    }(());

    (input_producers, output_consumers)
  }

  // ...
}

Parametric procs

Just as with other DSLX constructs, Procs can be parameterized. Parametrics must be specified at the proc level, and not at the component function level (i.e., not on config or next). Building off of the previous example, this looks as follows:

proc Parametric<N: u32, M: u32> {
  input_producers: chan<F32>[N][M] out;
  output_consumers: chan<F32>[N][M] in;

  // ...

  config() {
    let (input_producers, input_consumers) = chan<F32>[N][M]("node_input");
    let (output_producers, output_consumers) = chan<F32>[N][M]("node_output");

    for (i, _) : (u32, ()) in range(u32:0, N) {
      for (j, _) : (u32, ()) in range(u32:0, M) {
        spawn Node(input_consumers[i][j],
                   output_producers[i][j]);
      }(());
    }(());

    (input_producers, output_consumers)
  }

  // ...
}

These two features are very powerful together: users can specify a broad variety of designs simply by adjusting a few parametric values.

Proc testing

The DSLX interpreter supports testing procs via the test_proc construct. A test proc is very similar to a normal proc with the following changes:

A test proc is preceded by the #[test_proc] directive. This directive, as one might expect, notifies the interpreter that the following proc is a test proc.
A test proc's config function must accept a single argument: a boolean input channel for terminating interpretation. When the test is complete, the proc should send the test's status (true on success, false on failure) on that channel (commonly called the "terminator" channel).

A skeletal example:

#[test_proc]
proc Tester {
  terminator: chan<bool> out;

  // ...

  config(terminator: chan<bool> out) {
    spawn proc_under_test(...);
    (terminator,)
  }

  next(tok: token, state: ()) {
    // send and recv message to the proc under test.
    // ...
    // terminate the test interpretation.
    send(tok, terminator, true);
  }
}

The FP32 fmac module has a more complete proc test that may be used for reference.

Scheduling constraints

If you want to interface with something in the outside world that is latency sensitive (for example, an SRAM -- though we have separate infrastructure for making SRAMs work that builds on top of this feature), you can create external channels representing the interface you want to use, e.g.:

proc main {
  req: chan<u32> out;
  resp: chan<u32> in;

  init { u32: 0 }

  config(req: chan<u32> out, resp: chan<u32> in) {
    (req, resp)
  }

  next(tok: token, state: u32) {
    let request = state * state;
    let tok = send(tok, req, request);
    let (tok, response) = recv(tok, resp);
    state + u32:1
  }
}

where the fact that req and resp are parameters of config, and main is the top proc during IR conversion, is what makes them "external".

Then when you codegen this, you can pass in --io_constraints=foo__req:send:foo__resp:recv:2:2 where foo__req is the mangled name of the channel, which you can see by examining the generated IR prior to codegen. That constraint means "a send on any channel named req must occur exactly two cycles before a receive on any channel named resp"; the 2 is specified twice because it is possible to give a range of allowed cycle differences.

For more details on --io_constraints, check out the docs. For a complete example, see //xls/examples:constraint_sv and associated build targets; the target you'd build to get the mangled channel names is :constraint_ir.

The proc network itself will form a tree, but channels may make point-to-point connections between any two procs. ↩