oxfuse-runtime 0.4.0

Tensors, execution plans and bit-exact CPU backends (reference and fast) for oxfuse graphs.
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
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180

//! Compare the reference backend, the fast backend and ndarray's `dot` (safe
//! code over `matrixmultiply`) on the same work. Backends are compiled once;
//! only `run` is timed, so the numbers include the executor's buffer copies,
//! which is what a user pays.

use criterion::{Criterion, Throughput, criterion_group, criterion_main};
use ndarray::Array2;
use oxfuse_core::passes::default_pipeline;
use oxfuse_core::{Graph, GraphBuilder, Limits, Op};
use oxfuse_runtime::{Backend, CpuBackend, ReferenceBackend, ReferenceEvaluator, Tensor};
use std::hint::black_box;
use std::sync::Arc;

const CPU: &str = if cfg!(feature = "parallel") {
    "cpu-threads"
} else {
    "cpu"
};

/// Deterministic values in [-1, 1).
fn data(len: usize, seed: u32) -> Vec<f32> {
    let mut state = seed.wrapping_mul(2_654_435_761).wrapping_add(1);
    (0..len)
        .map(|_| {
            state = state.wrapping_mul(1_664_525).wrapping_add(1_013_904_223);
            (state >> 8) as f32 / (1u32 << 23) as f32 - 1.0
        })
        .collect()
}

fn tensor(shape: &[usize], seed: u32) -> Tensor {
    Tensor::new(shape.to_vec(), data(shape.iter().product(), seed)).unwrap()
}

fn binary_graph(op: Op, shape: &[usize]) -> Graph {
    let mut b = GraphBuilder::new(Limits::default());
    let x = b.input("x", shape.to_vec()).unwrap();
    let y = b.input("y", shape.to_vec()).unwrap();
    let z = b.node("z", op, &[x, y]).unwrap();
    b.finish(&[z]).unwrap()
}

fn unary_graph(op: Op, shape: &[usize]) -> Graph {
    let mut b = GraphBuilder::new(Limits::default());
    let x = b.input("x", shape.to_vec()).unwrap();
    let z = b.node("z", op, &[x]).unwrap();
    b.finish(&[z]).unwrap()
}

fn time_backends(c: &mut Criterion, name: &str, size: usize, graph: &Graph, inputs: &[Tensor]) {
    let mut group = c.benchmark_group(name);
    let reference = ReferenceBackend.compile(graph).unwrap();
    group.bench_function(format!("reference/{size}"), |b| {
        b.iter(|| black_box(reference.run(black_box(inputs)).unwrap()))
    });
    let cpu = CpuBackend.compile(graph).unwrap();
    group.bench_function(format!("{CPU}/{size}"), |b| {
        b.iter(|| black_box(cpu.run(black_box(inputs)).unwrap()))
    });
    group.finish();
}

fn matmul(c: &mut Criterion) {
    for n in [64usize, 128, 256, 512] {
        let graph = binary_graph(Op::MatMul, &[n, n]);
        let inputs = [tensor(&[n, n], 1), tensor(&[n, n], 2)];
        let mut group = c.benchmark_group("matmul");
        group.throughput(Throughput::Elements(2 * (n * n * n) as u64));
        let a = Array2::from_shape_vec((n, n), data(n * n, 1)).unwrap();
        let b = Array2::from_shape_vec((n, n), data(n * n, 2)).unwrap();
        group.bench_function(format!("ndarray/{n}"), |bench| {
            bench.iter(|| black_box(black_box(&a).dot(black_box(&b))))
        });
        let reference = ReferenceBackend.compile(&graph).unwrap();
        group.bench_function(format!("reference/{n}"), |bench| {
            bench.iter(|| black_box(reference.run(black_box(&inputs)).unwrap()))
        });
        let cpu = CpuBackend.compile(&graph).unwrap();
        group.bench_function(format!("{CPU}/{n}"), |bench| {
            bench.iter(|| black_box(cpu.run(black_box(&inputs)).unwrap()))
        });
        group.finish();
    }
}

/// A batch of `m` rows against a fixed weight: the shape of inference with a
/// small batch. `matmul-skinny` has a 1024 x 1024 weight (a hidden layer),
/// `matmul-head` a 784 x 10 one (a classifier head, 10 is not a multiple of
/// the vector width). The size in the id is the number of rows.
fn skinny(c: &mut Criterion) {
    for (name, k, n) in [
        ("matmul-skinny", 1024usize, 1024usize),
        ("matmul-head", 784, 10),
    ] {
        for m in [1usize, 2, 3, 5, 8, 32] {
            let graph = binary_graph_shapes(Op::MatMul, &[m, k], &[k, n]);
            let inputs = [tensor(&[m, k], 1), tensor(&[k, n], 2)];
            let mut group = c.benchmark_group(name);
            group.throughput(Throughput::Elements(2 * (m * k * n) as u64));
            let a = Array2::from_shape_vec((m, k), data(m * k, 1)).unwrap();
            let b = Array2::from_shape_vec((k, n), data(k * n, 2)).unwrap();
            group.bench_function(format!("ndarray/{m}"), |bench| {
                bench.iter(|| black_box(black_box(&a).dot(black_box(&b))))
            });
            let cpu = CpuBackend.compile(&graph).unwrap();
            group.bench_function(format!("{CPU}/{m}"), |bench| {
                bench.iter(|| black_box(cpu.run(black_box(&inputs)).unwrap()))
            });
            group.finish();
        }
    }
}

fn binary_graph_shapes(op: Op, a: &[usize], b: &[usize]) -> Graph {
    let mut g = GraphBuilder::new(Limits::default());
    let x = g.input("x", a.to_vec()).unwrap();
    let y = g.input("y", b.to_vec()).unwrap();
    let z = g.node("z", op, &[x, y]).unwrap();
    g.finish(&[z]).unwrap()
}

fn elementwise(c: &mut Criterion) {
    let n = 1 << 20;
    let inputs = [tensor(&[n], 3), tensor(&[n], 4)];
    time_backends(c, "add", n, &binary_graph(Op::Add, &[n]), &inputs);
    time_backends(c, "relu", n, &unary_graph(Op::Relu, &[n]), &inputs[..1]);
    time_backends(
        c,
        "sigmoid",
        n,
        &unary_graph(Op::Sigmoid, &[n]),
        &inputs[..1],
    );
}

fn transpose(c: &mut Criterion) {
    let n = 1024;
    let graph = unary_graph(Op::Transpose { perm: vec![1, 0] }, &[n, n]);
    time_backends(c, "transpose", n, &graph, &[tensor(&[n, n], 5)]);
}

/// A graph shaped like an exporter's output: the weight is stored transposed
/// and scaled by a constant, and a dead debug branch hangs off the input.
fn exporter_style_graph(n: usize) -> Graph {
    let mut b = GraphBuilder::new(Limits::default());
    let x = b.input("x", vec![64, n]).unwrap();
    let w = b.constant("w", vec![n, n], data(n * n, 6)).unwrap();
    let scale = b.constant("scale", vec![], vec![0.5]).unwrap();
    let wt = b
        .node("wt", Op::Transpose { perm: vec![1, 0] }, &[w])
        .unwrap();
    let ws = b.node("ws", Op::Mul, &[wt, scale]).unwrap();
    let y = b.node("y", Op::MatMul, &[x, ws]).unwrap();
    let out = b.node("out", Op::Relu, &[y]).unwrap();
    b.node("debug", Op::Sigmoid, &[x]).unwrap();
    b.finish(&[out]).unwrap()
}

fn pipeline(c: &mut Criterion) {
    let n = 256;
    let graph = exporter_style_graph(n);
    let optimized = default_pipeline(Limits::default(), Arc::new(ReferenceEvaluator))
        .run(&graph)
        .unwrap()
        .graph;
    let inputs = [tensor(&[64, n], 7)];
    let mut group = c.benchmark_group("pipeline");
    let plain = CpuBackend.compile(&graph).unwrap();
    group.bench_function(format!("{CPU}-unoptimized/{n}"), |b| {
        b.iter(|| black_box(plain.run(black_box(&inputs)).unwrap()))
    });
    let fast = CpuBackend.compile(&optimized).unwrap();
    group.bench_function(format!("{CPU}-optimized/{n}"), |b| {
        b.iter(|| black_box(fast.run(black_box(&inputs)).unwrap()))
    });
    group.finish();
}

criterion_group!(benches, matmul, skinny, elementwise, transpose, pipeline);
criterion_main!(benches);