#dag #rtl #hdl

starlight

experimental HDL and optimizer for DAGs of lookup tables

6 releases (breaking)

new 0.4.0 Feb 22, 2024
0.3.0 Jan 23, 2024
0.2.0 Dec 12, 2023
0.1.1 Dec 5, 2023
0.0.0 Jul 7, 2022

#245 in Algorithms

Download history 2/week @ 2023-11-03 2/week @ 2023-11-10 1/week @ 2023-11-17 3/week @ 2023-11-24 57/week @ 2023-12-01 48/week @ 2023-12-08 3/week @ 2023-12-15 9/week @ 2023-12-22 4/week @ 2023-12-29 2/week @ 2024-01-05 31/week @ 2024-01-19 7/week @ 2024-01-26 6/week @ 2024-02-02 9/week @ 2024-02-09 70/week @ 2024-02-16

92 downloads per month

MIT/Apache

575KB
12K SLoC

Starlight

This provides an HDL (Hardware Design Language), combinational and temporal logic simulator and optimizer, and general purpose router for FPGAs and more. The HDL is special in that it is written in ordinary Rust code with all the features that Rust provides.

Most of the MVP features of this crate are ready, except for the Router which is still a WIP and has a lot of todo!();

See the documentation of awint/awint_dag which is used as the backend for this. awint is the base library that operations are modeled off of. awint_dag allows for recording a DAG of arbitrary bitwidth integer operations. starlight lowers high level operations down into a DAG of simple lookup tables, and also adds on temporal structs like Loops. It can optimize, evaluate, and retroactively change values in the DAG for various purposes.

There are several features on this crate that enable awint features. The u32_ptrs feature reduces the memory consumption of the algorithms significantly, but limits the number of possible internal references to about 4 billion, which the largest circuits might not fit in.

use std::num::NonZeroUsize;
use starlight::{awi, dag, Epoch, EvalAwi, LazyAwi};

// in the scope where this is glob imported, all arbitrary width types, some primitives, and
// the mechanisms in the macros will use mimicking types and be lazily evaluated in general.
use dag::*;

// This is just some arbitrary example I coded up, note that you can use
// almost all of Rust's features that you can use on the normal types
struct StateMachine {
    data: inlawi_ty!(16),
    counter: Awi,
}

impl StateMachine {
    pub fn new(w: NonZeroUsize) -> Self {
        Self {
            data: inlawi!(0u16),
            counter: Awi::zero(w),
        }
    }

    pub fn update(&mut self, input: &Bits) -> Option<()> {
        self.counter.inc_(true);

        let mut s0 = inlawi!(0u4);
        let mut s1 = inlawi!(0u4);
        let mut s2 = inlawi!(0u4);
        let mut s3 = inlawi!(0u4);
        cc!(self.data; s3, s2, s1, s0)?;
        s2.xor_(&s0)?;
        s3.xor_(&s1)?;
        s1.xor_(&s2)?;
        s0.xor_(&s3)?;
        s3.rotl_(1)?;
        s2.mux_(&input, input.get(0)?)?;
        cc!(s3, s2, s1, s0; self.data)?;
        Some(())
    }
}

// First, create an epoch, this will live until this struct is dropped. The
// epoch needs to live until all mimicking operations are done and states are
// lowered. Manually drop it with the `drop` function to avoid mistakes.
let epoch = Epoch::new();

let mut m = StateMachine::new(bw(4));

// this is initially an opaque value that cannot be eagerly evaluated
let input = LazyAwi::opaque(bw(4));

// if we later retroactively assign this to an unequal value, the
// `assert_assertions_strict` call will error and show the location of the
// assertion that errored
mimick::assert_eq!(Awi::from(&input), awi!(0101));

// step the state machine forward
m.update(&input).unwrap();
m.update(&awi!(0110)).unwrap();
m.update(&awi!(0110)).unwrap();

// use `EvalAwi`s to evaluate the resulting values
let output_counter = EvalAwi::from(m.counter);
let output_data = EvalAwi::from(m.data);

{
    // switch back to normal structs
    use awi::*;

    // discard all unused mimicking states so the render is cleaner
    epoch.prune_unused_states().unwrap();

    // See the mimicking state DAG before it is lowered
    epoch
        .render_to_svgs_in_dir(std::path::PathBuf::from("./".to_owned()))
        .unwrap();

    // lower into purely static bit movements and lookup tables and optimize
    epoch.optimize().unwrap();

    // Now the combinational logic is described in a DAG of lookup tables that we
    // could use for various purposes
    epoch.ensemble(|ensemble| {
        for state in ensemble.stator.states.vals() {
            assert!(state.lowered_to_lnodes);
        }
    });

    // "retroactively" assign the input with a non-opaque value
    input.retro_(&awi!(0101)).unwrap();
    // check assertions (all `dag::assert*` functions and dynamic `unwrap`s done
    // during the current `Epoch`)
    epoch.assert_assertions(true).unwrap();
    // evaluate the outputs
    assert_eq!(output_counter.eval().unwrap(), awi!(0011));
    assert_eq!(output_data.eval().unwrap(), awi!(0xa505_u16));

    // reassign and reevaluate
    input.retro_(&awi!(1011)).unwrap();
    assert!(epoch.assert_assertions(true).is_err());
    assert_eq!(output_data.eval().unwrap(), awi!(0x7b0b_u16));
}
drop(epoch);
use starlight::{dag, awi, Epoch, EvalAwi};

use dag::*;

let epoch = Epoch::new();

let mut lhs = inlawi!(zero: ..8);
let rhs = inlawi!(umax: ..8);
let x = inlawi!(10101010);
let y = InlAwi::from_u64(4);

let mut output = inlawi!(0xffu8);

// error: expected `bool`, found struct `bool`
//if lhs.ult(&rhs).unwrap() {
//    output.xor_(&x).unwrap();
//} else {
//    output.lshr_(y.to_usize()).unwrap();
//};

// A little more cumbersome, but we get to use all the features of
// normal Rust in metaprogramming and don't have to support an entire DSL.
// In the future we will have more macros to help with this.

let lt = lhs.ult(&rhs).unwrap();

let mut tmp0 = output;
tmp0.xor_(&x).unwrap();
output.mux_(&tmp0, lt).unwrap();

let mut tmp1 = output;
tmp1.lshr_(y.to_usize()).unwrap();
output.mux_(&tmp1, !lt).unwrap();

let output_eval = EvalAwi::from(&output);

{
    use awi::*;
    assert_eq!(output_eval.eval().unwrap(), awi!(01010101));
}
drop(epoch);

Dependencies

~2–2.8MB
~57K SLoC