systile 1.0.0

Matmul-native data structures & algorithms: 17 structures (holographic memory, semiring graphs, automata, attention, DFT, Viterbi, k-NN, Bloom, sort, scan, …) whose core operation reduces to a dense matmul, built on a TPU-style tiled tensor with a systolic reference engine.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

//! Prefix sums as a triangular matmul.
//!
//! Run with `cargo run --release --example scan_prefix`.

use systile::prelude::*;

fn main() {
    let scan = TensorScan::new();
    let x = [3.0f32, 1.0, 4.0, 1.0, 5.0, 9.0, 2.0, 6.0];

    println!("input:      {x:?}");
    println!("inclusive:  {:?}", scan.inclusive(&x));
    println!("exclusive:  {:?}", scan.exclusive(&x));
    println!("suffix:     {:?}", scan.suffix(&x));
    println!("total:      {}", scan.total(&x));

    // Cross-check the inclusive scan against a sequential cumsum.
    let mut acc = 0.0;
    let expected: Vec<f32> = x
        .iter()
        .map(|&v| {
            acc += v;
            acc
        })
        .collect();
    assert_eq!(scan.inclusive(&x), expected);
    println!("\n✓ cumulative sum computed as L·x — O(1) depth, no carried loop.");
}