qntz 0.1.8

Vector quantization primitives (RaBitQ, ternary, bit packing) for ANN systems.
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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259

use criterion::{black_box, criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};

use qntz::simd_ops::{hamming_distance, pack_binary_fast, unpack_binary_fast};

// ============================================================================
// Helpers
// ============================================================================

fn make_f32_vec(dim: usize, seed: u64) -> Vec<f32> {
    let mut state = seed;
    (0..dim)
        .map(|_| {
            state = state
                .wrapping_mul(6364136223846793005)
                .wrapping_add(1442695040888963407);
            let bits = (state >> 33) as u32;
            (bits as f32 / u32::MAX as f32) * 2.0 - 1.0
        })
        .collect()
}

fn make_binary_codes(dim: usize, seed: u64) -> Vec<u8> {
    let mut state = seed;
    (0..dim)
        .map(|_| {
            state = state
                .wrapping_mul(6364136223846793005)
                .wrapping_add(1442695040888963407);
            (state >> 63) as u8
        })
        .collect()
}

fn make_packed(dim: usize, seed: u64) -> Vec<u8> {
    let codes = make_binary_codes(dim, seed);
    let mut packed = vec![0u8; dim.div_ceil(8)];
    pack_binary_fast(&codes, &mut packed).unwrap();
    packed
}

// ============================================================================
// pack_binary_fast / unpack_binary_fast
// ============================================================================

fn bench_pack_binary(c: &mut Criterion) {
    let mut group = c.benchmark_group("pack_binary_fast");
    for &dim in &[128usize, 768] {
        let codes = make_binary_codes(dim, 42);
        let mut packed = vec![0u8; dim.div_ceil(8)];
        group.throughput(Throughput::Elements(dim as u64));
        group.bench_with_input(BenchmarkId::from_parameter(dim), &dim, |b, _| {
            b.iter(|| {
                pack_binary_fast(black_box(&codes), black_box(&mut packed)).unwrap();
            });
        });
    }
    group.finish();
}

fn bench_unpack_binary(c: &mut Criterion) {
    let mut group = c.benchmark_group("unpack_binary_fast");
    for &dim in &[128usize, 768] {
        let packed = make_packed(dim, 42);
        let mut codes = vec![0u8; dim];
        group.throughput(Throughput::Elements(dim as u64));
        group.bench_with_input(BenchmarkId::from_parameter(dim), &dim, |b, _| {
            b.iter(|| {
                unpack_binary_fast(black_box(&packed), black_box(&mut codes), dim).unwrap();
            });
        });
    }
    group.finish();
}

// ============================================================================
// hamming_distance
// ============================================================================

fn bench_hamming_distance(c: &mut Criterion) {
    let mut group = c.benchmark_group("hamming_distance");
    for &dim in &[128usize, 768] {
        let a = make_packed(dim, 11);
        let b = make_packed(dim, 22);
        group.throughput(Throughput::Elements(dim as u64));
        group.bench_with_input(BenchmarkId::from_parameter(dim), &dim, |b_bench, _| {
            b_bench.iter(|| hamming_distance(black_box(&a), black_box(&b)));
        });
    }
    group.finish();
}

// ============================================================================
// RaBitQ
// ============================================================================

#[cfg(feature = "rabitq")]
fn bench_rabitq(c: &mut Criterion) {
    use qntz::rabitq::{RaBitQConfig, RaBitQQuantizer};

    let mut group = c.benchmark_group("rabitq");
    for &dim in &[128usize, 768] {
        // --- 1-bit (binary) ---
        let q1 = RaBitQQuantizer::with_config(dim, 42, RaBitQConfig::binary()).unwrap();
        let doc = make_f32_vec(dim, 1);
        let query = make_f32_vec(dim, 2);
        let qv1 = q1.quantize(&doc).unwrap();
        let rotated_query1 = q1.rotate_query(&query).unwrap();

        group.throughput(Throughput::Elements(dim as u64));
        group.bench_with_input(BenchmarkId::new("quantize_1bit", dim), &dim, |b, _| {
            b.iter(|| q1.quantize(black_box(&doc)).unwrap());
        });
        group.bench_with_input(
            BenchmarkId::new("approximate_l2_sqr_1bit", dim),
            &dim,
            |b, _| {
                b.iter(|| {
                    q1.approximate_l2_sqr(black_box(&query), black_box(&qv1))
                        .unwrap()
                });
            },
        );
        // Pre-rotated distance -- the hot path when rotation is amortized.
        group.bench_with_input(
            BenchmarkId::new("l2_sqr_prerotated_1bit", dim),
            &dim,
            |b, _| {
                b.iter(|| {
                    RaBitQQuantizer::approximate_l2_sqr_prerotated(
                        black_box(&rotated_query1),
                        black_box(&qv1),
                    )
                });
            },
        );

        // --- 4-bit ---
        let q4 = RaBitQQuantizer::with_config(dim, 42, RaBitQConfig::bits4()).unwrap();
        let qv4 = q4.quantize(&doc).unwrap();
        let rotated_query4 = q4.rotate_query(&query).unwrap();

        group.bench_with_input(BenchmarkId::new("quantize_4bit", dim), &dim, |b, _| {
            b.iter(|| q4.quantize(black_box(&doc)).unwrap());
        });
        group.bench_with_input(
            BenchmarkId::new("l2_sqr_prerotated_4bit", dim),
            &dim,
            |b, _| {
                b.iter(|| {
                    RaBitQQuantizer::approximate_l2_sqr_prerotated(
                        black_box(&rotated_query4),
                        black_box(&qv4),
                    )
                });
            },
        );
    }
    group.finish();
}

// ============================================================================
// TernaryQuantizer
// ============================================================================

#[cfg(feature = "ternary")]
fn bench_ternary(c: &mut Criterion) {
    use qntz::ternary::{TernaryConfig, TernaryQuantizer};

    let mut group = c.benchmark_group("ternary_quantize");
    for &dim in &[128usize, 768] {
        let config = TernaryConfig {
            threshold_high: 0.3,
            threshold_low: -0.3,
            normalize: true,
            target_sparsity: None,
        };
        let q = TernaryQuantizer::new(dim, config);
        let v = make_f32_vec(dim, 7);
        group.throughput(Throughput::Elements(dim as u64));
        group.bench_with_input(BenchmarkId::from_parameter(dim), &dim, |b, _| {
            b.iter(|| q.quantize(black_box(&v)).unwrap());
        });
    }
    group.finish();
}

// ============================================================================
// BinaryQuantizer (rotation-based)
// ============================================================================

#[cfg(feature = "binary")]
fn bench_binary_quantizer(c: &mut Criterion) {
    use qntz::binary::BinaryQuantizer;

    let dim = 128;
    let q = BinaryQuantizer::new(dim, dim, 42);
    let v = make_f32_vec(dim, 5);
    let code = q.quantize(&v).unwrap();

    let mut group = c.benchmark_group("binary_quantizer");
    group.throughput(Throughput::Elements(dim as u64));
    group.bench_function("quantize_dim128", |b| {
        b.iter(|| q.quantize(black_box(&v)).unwrap());
    });
    group.bench_function("asymmetric_distance_dim128", |b| {
        b.iter(|| {
            q.asymmetric_distance(black_box(&v), black_box(&code))
                .unwrap()
        });
    });
    group.finish();
}

// ============================================================================
// AdaptiveQuantizer
// ============================================================================

#[cfg(feature = "adaptive")]
fn bench_adaptive(c: &mut Criterion) {
    use qntz::adaptive::AdaptiveQuantizer;

    let dim = 128;
    let n = 64;
    let q = AdaptiveQuantizer::new(8).unwrap();
    let vectors: Vec<Vec<f32>> = (0..n).map(|i| make_f32_vec(dim, i as u64)).collect();
    let query = make_f32_vec(dim, 999);
    let packed = q.quantize_packed(&vectors).unwrap();

    let mut group = c.benchmark_group("adaptive_quantizer");
    group.throughput(Throughput::Elements((dim * n) as u64));
    group.bench_function("quantize_packed_64x128", |b| {
        b.iter(|| q.quantize_packed(black_box(&vectors)).unwrap());
    });
    group.bench_function("asymmetric_distances_64x128", |b| {
        b.iter(|| packed.asymmetric_distances(black_box(&query)).unwrap());
    });
    group.finish();
}

// ============================================================================
// Benchmark registration
// ============================================================================

fn base_benches(c: &mut Criterion) {
    bench_pack_binary(c);
    bench_unpack_binary(c);
    bench_hamming_distance(c);
    #[cfg(feature = "rabitq")]
    bench_rabitq(c);
    #[cfg(feature = "ternary")]
    bench_ternary(c);
    #[cfg(feature = "binary")]
    bench_binary_quantizer(c);
    #[cfg(feature = "adaptive")]
    bench_adaptive(c);
}

criterion_group!(benches, base_benches);
criterion_main!(benches);