MIRR Documentation

    GitHub →

    HLS Optimization Guide

    The MIRR compiler includes a High-Level Synthesis (HLS) pass that optimizes hardware resource usage through scheduling, sharing, and binding.

    Overview

    The HLS pass operates on the guard/reflex DAG and provides:

    • ASAP scheduling — earliest cycle for each operation
    • ALAP scheduling — latest cycle given target latency
    • Resource sharing — time-multiplex compatible operations
    • Operation binding — assign to physical resources
    • FIFO streaming — bounded dataflow between operations

    Usage

    use mirrc::hls::{OpDag, ResourceKind, HlsConfig};

let mut dag = OpDag::new();
let a = dag.add_op(ResourceKind::Add, 8, vec![8, 8]).unwrap();
let b = dag.add_op(ResourceKind::Mul, 16, vec![8, 8]).unwrap();
dag.add_edge(a, b);

let config = HlsConfig { latency: 3, sharing: true, binding: true, fifo: true };
let result = mirrc::hls::run_hls_pass(&dag, &config).unwrap();

    Scheduling

    ASAP (As Soon As Possible)

    Each operation is scheduled to the earliest cycle where all dependencies are met.

    ALAP (As Late As Possible)

    Each operation is scheduled to the latest cycle that meets the target latency.

    Combined

    The scheduler computes both ASAP and ALAP, then uses the slack for resource sharing.

    Resource Sharing

    Compatible operations (same resource kind, non-overlapping time slots) share the same physical resource. This reduces area at the cost of multiplexing logic.

    Operation Binding

    Each operation is assigned to a physical resource ID. Operations sharing a resource get the same binding ID.

    See Also