mnemonist-quant 0.4.3

TurboQuant vector quantization for mnemonist — near-optimal MSE and inner-product quantizers
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

use criterion::{BenchmarkId, Criterion, criterion_group, criterion_main};
use mnemonist_quant::mse::TurboQuantMse;
use mnemonist_quant::pack;
use mnemonist_quant::prod::TurboQuantProd;

fn random_unit_vector(dim: usize, seed: u64) -> Vec<f32> {
    // Simple deterministic pseudo-random vector using xorshift
    let mut state = seed;
    let mut v: Vec<f32> = (0..dim)
        .map(|_| {
            state ^= state << 13;
            state ^= state >> 7;
            state ^= state << 17;
            (state as f32 / u64::MAX as f32) * 2.0 - 1.0
        })
        .collect();
    let norm: f32 = v.iter().map(|x| x * x).sum::<f32>().sqrt();
    if norm > 0.0 {
        for x in &mut v {
            *x /= norm;
        }
    }
    v
}

fn bench_mse_quantize(c: &mut Criterion) {
    let mut group = c.benchmark_group("turbo_quant_mse/quantize");

    for dim in [64, 128, 384] {
        let quant = TurboQuantMse::new(dim, 2, 42).unwrap();
        let x = random_unit_vector(dim, 7);

        group.bench_with_input(BenchmarkId::new("dim", dim), &dim, |bench, _| {
            bench.iter(|| quant.quantize(&x).unwrap());
        });
    }

    group.finish();
}

fn bench_mse_dequantize(c: &mut Criterion) {
    let mut group = c.benchmark_group("turbo_quant_mse/dequantize");

    for dim in [64, 128, 384] {
        let quant = TurboQuantMse::new(dim, 2, 42).unwrap();
        let x = random_unit_vector(dim, 7);
        let q = quant.quantize(&x).unwrap();

        group.bench_with_input(BenchmarkId::new("dim", dim), &dim, |bench, _| {
            bench.iter(|| quant.dequantize(&q).unwrap());
        });
    }

    group.finish();
}

fn bench_mse_bits(c: &mut Criterion) {
    let mut group = c.benchmark_group("turbo_quant_mse/bits");
    let dim = 128;
    let x = random_unit_vector(dim, 7);

    for bits in 1..=4u8 {
        let quant = TurboQuantMse::new(dim, bits, 42).unwrap();

        group.bench_with_input(BenchmarkId::new("quantize", bits), &bits, |bench, _| {
            bench.iter(|| quant.quantize(&x).unwrap());
        });

        let q = quant.quantize(&x).unwrap();
        group.bench_with_input(BenchmarkId::new("dequantize", bits), &bits, |bench, _| {
            bench.iter(|| quant.dequantize(&q).unwrap());
        });
    }

    group.finish();
}

fn bench_prod(c: &mut Criterion) {
    let mut group = c.benchmark_group("turbo_quant_prod");
    let dim = 128;
    let x = random_unit_vector(dim, 7);
    let query = random_unit_vector(dim, 99);

    for bits in 2..=4u8 {
        let quant = TurboQuantProd::new(dim, bits, 42, 99).unwrap();

        group.bench_with_input(BenchmarkId::new("quantize", bits), &bits, |bench, _| {
            bench.iter(|| quant.quantize(&x).unwrap());
        });

        let q = quant.quantize(&x).unwrap();

        group.bench_with_input(BenchmarkId::new("dequantize", bits), &bits, |bench, _| {
            bench.iter(|| quant.dequantize(&q).unwrap());
        });

        group.bench_with_input(
            BenchmarkId::new("inner_product_estimate", bits),
            &bits,
            |bench, _| {
                bench.iter(|| quant.inner_product_estimate(&query, &q).unwrap());
            },
        );
    }

    group.finish();
}

fn bench_pack(c: &mut Criterion) {
    let mut group = c.benchmark_group("pack");

    for (dim, bits) in [(128, 2u8), (384, 2), (384, 4)] {
        let indices: Vec<u8> = (0..dim).map(|i| (i % (1 << bits)) as u8).collect();

        group.bench_with_input(
            BenchmarkId::new("pack", format!("{dim}x{bits}b")),
            &dim,
            |bench, _| {
                bench.iter(|| pack::pack_indices(&indices, bits).unwrap());
            },
        );

        let packed = pack::pack_indices(&indices, bits).unwrap();
        group.bench_with_input(
            BenchmarkId::new("unpack", format!("{dim}x{bits}b")),
            &dim,
            |bench, _| {
                bench.iter(|| pack::unpack_indices(&packed, bits, dim).unwrap());
            },
        );
    }

    group.finish();
}

criterion_group!(
    benches,
    bench_mse_quantize,
    bench_mse_dequantize,
    bench_mse_bits,
    bench_prod,
    bench_pack
);
criterion_main!(benches);