tenferro-runtime 0.9.0

Concrete tensor helpers, traced tensors, graph execution, and extension runtime support for tenferro.
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

use criterion::{black_box, criterion_group, criterion_main, BenchmarkId, Criterion};
use tenferro_cpu::CpuEngine;
use tenferro_runtime::Runtime;
use tenferro_runtime::{
    CompiledGraph, DType, ExecutionEngine, GraphCompiler, Tensor, TensorRead, TracedTensor,
};

fn cpu_engine_and_runtime() -> (CpuEngine, Runtime) {
    let engine = CpuEngine::serial();
    let mut builder = Runtime::builder();
    builder
        .register_engine(tenferro_cpu::runtime_engine_registration(&engine).unwrap())
        .unwrap();
    (engine, builder.build().unwrap())
}

fn input_tensor(n: usize, scale: f64) -> Tensor {
    let data = (0..n)
        .map(|i| ((i % 97) as f64 + 1.0) * scale)
        .collect::<Vec<_>>();
    Tensor::from_vec_col_major(vec![n], data).expect("benchmark tensor")
}

fn compile_add_mul(n: usize) -> CompiledGraph {
    let a = TracedTensor::input_concrete_shape(DType::F64, &[n]).expect("a placeholder");
    let b = TracedTensor::input_concrete_shape(DType::F64, &[n]).expect("b placeholder");
    let sum = (&a + &b).expect("sum graph");
    let out = (&sum * &a).expect("multiply graph");
    let mut compiler = GraphCompiler::new();
    compiler
        .compile_with_input_specs(&out, &[(&a, DType::F64, &[n]), (&b, DType::F64, &[n])])
        .expect("compiled add-mul graph")
}

fn compile_broadcast_mul(rows: usize, cols: usize) -> CompiledGraph {
    let a = TracedTensor::input_concrete_shape(DType::F64, &[rows]).expect("a placeholder");
    let b = TracedTensor::input_concrete_shape(DType::F64, &[cols]).expect("b placeholder");
    let a_bc = a
        .broadcast_in_dim(&[rows, cols], &[0])
        .expect("lhs broadcast graph");
    let b_bc = b
        .broadcast_in_dim(&[rows, cols], &[1])
        .expect("rhs broadcast graph");
    let out = (&a_bc * &b_bc).expect("multiply graph");
    let mut compiler = GraphCompiler::new();
    compiler
        .compile_with_input_specs(
            &out,
            &[(&a, DType::F64, &[rows]), (&b, DType::F64, &[cols])],
        )
        .expect("compiled broadcast-multiply graph")
}

fn compile_broadcast_mul_add(rows: usize, cols: usize) -> CompiledGraph {
    let a = TracedTensor::input_concrete_shape(DType::F64, &[rows]).expect("a placeholder");
    let b = TracedTensor::input_concrete_shape(DType::F64, &[cols]).expect("b placeholder");
    let a_bc = a
        .broadcast_in_dim(&[rows, cols], &[0])
        .expect("lhs broadcast graph");
    let b_bc = b
        .broadcast_in_dim(&[rows, cols], &[1])
        .expect("rhs broadcast graph");
    let product = (&a_bc * &b_bc).expect("multiply graph");
    let out = (&product + &a_bc).expect("add graph");
    let mut compiler = GraphCompiler::new();
    compiler
        .compile_with_input_specs(
            &out,
            &[(&a, DType::F64, &[rows]), (&b, DType::F64, &[cols])],
        )
        .expect("compiled broadcast-multiply-add graph")
}

/// Benchmark one graph through the prepared path.
///
/// Every production run now goes through the scheduled executor that
/// `prepared_graph` measures, so there is one timing per case: the former
/// `unprepared_graph` label measured the same executor.
fn bench_both_paths(
    group: &mut criterion::BenchmarkGroup<'_, criterion::measurement::WallTime>,
    program: &CompiledGraph,
    inputs: &[&Tensor],
    case: String,
) {
    group.bench_function(BenchmarkId::new("prepared_graph", case), |bench| {
        let (mut engine, runtime) = cpu_engine_and_runtime();
        let reads: Vec<TensorRead<'_>> = inputs
            .iter()
            .map(|input| TensorRead::from_tensor(input))
            .collect();
        let prepared = runtime
            .prepare(program, &reads, &[engine.target()])
            .expect("graph should prepare");
        bench.iter(|| {
            let out = runtime
                .run(&mut [&mut engine], black_box(&prepared), &reads)
                .expect("graph run");
            black_box(out);
        });
    });
}

fn bench_runtime_add_mul(c: &mut Criterion) {
    let mut group = c.benchmark_group("runtime_elementwise_chain/f64/add_mul");
    for &n in &[4_096_usize, 65_536, 1_048_576] {
        let program = compile_add_mul(n);
        let a = input_tensor(n, 0.5);
        let b = input_tensor(n, 1.25);
        group.throughput(criterion::Throughput::Elements(n as u64));
        bench_both_paths(&mut group, &program, &[&a, &b], n.to_string());
    }
    group.finish();
}

fn bench_runtime_broadcast_mul(c: &mut Criterion) {
    let mut group = c.benchmark_group("runtime_elementwise_chain/f64/broadcast_mul");
    for &(rows, cols) in &[(256_usize, 256_usize), (1024, 1024)] {
        let program = compile_broadcast_mul(rows, cols);
        let a = input_tensor(rows, 0.5);
        let b = input_tensor(cols, 1.25);
        let elements = rows * cols;
        group.throughput(criterion::Throughput::Elements(elements as u64));
        bench_both_paths(&mut group, &program, &[&a, &b], format!("{rows}x{cols}"));
    }
    group.finish();
}

fn bench_runtime_broadcast_mul_add(c: &mut Criterion) {
    let mut group = c.benchmark_group("runtime_elementwise_chain/f64/broadcast_mul_add");
    for &(rows, cols) in &[(256_usize, 256_usize), (1024, 1024)] {
        let program = compile_broadcast_mul_add(rows, cols);
        let a = input_tensor(rows, 0.5);
        let b = input_tensor(cols, 1.25);
        let elements = rows * cols;
        group.throughput(criterion::Throughput::Elements(elements as u64));
        bench_both_paths(&mut group, &program, &[&a, &b], format!("{rows}x{cols}"));
    }
    group.finish();
}

criterion_group!(
    benches,
    bench_runtime_add_mul,
    bench_runtime_broadcast_mul,
    bench_runtime_broadcast_mul_add
);
criterion_main!(benches);