qdrant-edge 0.7.2

A lightweight, in-process vector search engine designed for embedded devices, autonomous systems, and mobile agents.
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

#[cfg(target_arch = "x86")]
use std::arch::x86::*;
#[cfg(target_arch = "x86_64")]
use std::arch::x86_64::*;

use crate::common::types::ScoreType;

use super::tools::is_length_zero_or_normalized;
use crate::segment::data_types::vectors::{DenseVector, VectorElementType};

#[target_feature(enable = "sse")]
#[allow(clippy::missing_safety_doc)]
pub unsafe fn hsum128_ps_sse(x: __m128) -> f32 {
    let x64: __m128 = _mm_add_ps(x, _mm_movehl_ps(x, x));
    let x32: __m128 = _mm_add_ss(x64, _mm_shuffle_ps(x64, x64, 0x55));
    _mm_cvtss_f32(x32)
}

#[target_feature(enable = "sse")]
pub(crate) unsafe fn euclid_similarity_sse(
    v1: &[VectorElementType],
    v2: &[VectorElementType],
) -> ScoreType {
    unsafe {
        let n = v1.len();
        let m = n - (n % 16);
        let mut ptr1: *const f32 = v1.as_ptr();
        let mut ptr2: *const f32 = v2.as_ptr();
        let mut sum128_1: __m128 = _mm_setzero_ps();
        let mut sum128_2: __m128 = _mm_setzero_ps();
        let mut sum128_3: __m128 = _mm_setzero_ps();
        let mut sum128_4: __m128 = _mm_setzero_ps();
        let mut i: usize = 0;
        while i < m {
            let sub128_1 = _mm_sub_ps(_mm_loadu_ps(ptr1), _mm_loadu_ps(ptr2));
            sum128_1 = _mm_add_ps(_mm_mul_ps(sub128_1, sub128_1), sum128_1);

            let sub128_2 = _mm_sub_ps(_mm_loadu_ps(ptr1.add(4)), _mm_loadu_ps(ptr2.add(4)));
            sum128_2 = _mm_add_ps(_mm_mul_ps(sub128_2, sub128_2), sum128_2);

            let sub128_3 = _mm_sub_ps(_mm_loadu_ps(ptr1.add(8)), _mm_loadu_ps(ptr2.add(8)));
            sum128_3 = _mm_add_ps(_mm_mul_ps(sub128_3, sub128_3), sum128_3);

            let sub128_4 = _mm_sub_ps(_mm_loadu_ps(ptr1.add(12)), _mm_loadu_ps(ptr2.add(12)));
            sum128_4 = _mm_add_ps(_mm_mul_ps(sub128_4, sub128_4), sum128_4);

            ptr1 = ptr1.add(16);
            ptr2 = ptr2.add(16);
            i += 16;
        }

        let mut result = hsum128_ps_sse(sum128_1)
            + hsum128_ps_sse(sum128_2)
            + hsum128_ps_sse(sum128_3)
            + hsum128_ps_sse(sum128_4);
        for i in 0..n - m {
            result += (*ptr1.add(i) - *ptr2.add(i)).powi(2);
        }
        -result
    }
}

#[target_feature(enable = "sse")]
pub(crate) unsafe fn manhattan_similarity_sse(
    v1: &[VectorElementType],
    v2: &[VectorElementType],
) -> ScoreType {
    unsafe {
        let mask: __m128 = _mm_set1_ps(-0.0f32); // 1 << 31 used to clear sign bit to mimic abs

        let n = v1.len();
        let m = n - (n % 16);
        let mut ptr1: *const f32 = v1.as_ptr();
        let mut ptr2: *const f32 = v2.as_ptr();
        let mut sum128_1: __m128 = _mm_setzero_ps();
        let mut sum128_2: __m128 = _mm_setzero_ps();
        let mut sum128_3: __m128 = _mm_setzero_ps();
        let mut sum128_4: __m128 = _mm_setzero_ps();
        let mut i: usize = 0;
        while i < m {
            let sub128_1 = _mm_sub_ps(_mm_loadu_ps(ptr1), _mm_loadu_ps(ptr2));
            sum128_1 = _mm_add_ps(_mm_andnot_ps(mask, sub128_1), sum128_1);

            let sub128_2 = _mm_sub_ps(_mm_loadu_ps(ptr1.add(4)), _mm_loadu_ps(ptr2.add(4)));
            sum128_2 = _mm_add_ps(_mm_andnot_ps(mask, sub128_2), sum128_2);

            let sub128_3 = _mm_sub_ps(_mm_loadu_ps(ptr1.add(8)), _mm_loadu_ps(ptr2.add(8)));
            sum128_3 = _mm_add_ps(_mm_andnot_ps(mask, sub128_3), sum128_3);

            let sub128_4 = _mm_sub_ps(_mm_loadu_ps(ptr1.add(12)), _mm_loadu_ps(ptr2.add(12)));
            sum128_4 = _mm_add_ps(_mm_andnot_ps(mask, sub128_4), sum128_4);

            ptr1 = ptr1.add(16);
            ptr2 = ptr2.add(16);
            i += 16;
        }

        let mut result = hsum128_ps_sse(sum128_1)
            + hsum128_ps_sse(sum128_2)
            + hsum128_ps_sse(sum128_3)
            + hsum128_ps_sse(sum128_4);
        for i in 0..n - m {
            result += (*ptr1.add(i) - *ptr2.add(i)).abs();
        }
        -result
    }
}

#[target_feature(enable = "sse")]
pub(crate) unsafe fn cosine_preprocess_sse(vector: DenseVector) -> DenseVector {
    unsafe {
        let n = vector.len();
        let m = n - (n % 16);
        let mut ptr: *const f32 = vector.as_ptr();
        let mut sum128_1: __m128 = _mm_setzero_ps();
        let mut sum128_2: __m128 = _mm_setzero_ps();
        let mut sum128_3: __m128 = _mm_setzero_ps();
        let mut sum128_4: __m128 = _mm_setzero_ps();

        let mut i: usize = 0;
        while i < m {
            let m128_1 = _mm_loadu_ps(ptr);
            sum128_1 = _mm_add_ps(_mm_mul_ps(m128_1, m128_1), sum128_1);

            let m128_2 = _mm_loadu_ps(ptr.add(4));
            sum128_2 = _mm_add_ps(_mm_mul_ps(m128_2, m128_2), sum128_2);

            let m128_3 = _mm_loadu_ps(ptr.add(8));
            sum128_3 = _mm_add_ps(_mm_mul_ps(m128_3, m128_3), sum128_3);

            let m128_4 = _mm_loadu_ps(ptr.add(12));
            sum128_4 = _mm_add_ps(_mm_mul_ps(m128_4, m128_4), sum128_4);

            ptr = ptr.add(16);
            i += 16;
        }

        let mut length = hsum128_ps_sse(sum128_1)
            + hsum128_ps_sse(sum128_2)
            + hsum128_ps_sse(sum128_3)
            + hsum128_ps_sse(sum128_4);
        for i in 0..n - m {
            length += (*ptr.add(i)).powi(2);
        }
        if is_length_zero_or_normalized(length) {
            return vector;
        }
        length = length.sqrt();
        vector.into_iter().map(|x| x / length).collect()
    }
}

#[target_feature(enable = "sse")]
pub(crate) unsafe fn dot_similarity_sse(
    v1: &[VectorElementType],
    v2: &[VectorElementType],
) -> ScoreType {
    unsafe {
        let n = v1.len();
        let m = n - (n % 16);
        let mut ptr1: *const f32 = v1.as_ptr();
        let mut ptr2: *const f32 = v2.as_ptr();
        let mut sum128_1: __m128 = _mm_setzero_ps();
        let mut sum128_2: __m128 = _mm_setzero_ps();
        let mut sum128_3: __m128 = _mm_setzero_ps();
        let mut sum128_4: __m128 = _mm_setzero_ps();

        let mut i: usize = 0;
        while i < m {
            sum128_1 = _mm_add_ps(_mm_mul_ps(_mm_loadu_ps(ptr1), _mm_loadu_ps(ptr2)), sum128_1);

            sum128_2 = _mm_add_ps(
                _mm_mul_ps(_mm_loadu_ps(ptr1.add(4)), _mm_loadu_ps(ptr2.add(4))),
                sum128_2,
            );

            sum128_3 = _mm_add_ps(
                _mm_mul_ps(_mm_loadu_ps(ptr1.add(8)), _mm_loadu_ps(ptr2.add(8))),
                sum128_3,
            );

            sum128_4 = _mm_add_ps(
                _mm_mul_ps(_mm_loadu_ps(ptr1.add(12)), _mm_loadu_ps(ptr2.add(12))),
                sum128_4,
            );

            ptr1 = ptr1.add(16);
            ptr2 = ptr2.add(16);
            i += 16;
        }

        let mut result = hsum128_ps_sse(sum128_1)
            + hsum128_ps_sse(sum128_2)
            + hsum128_ps_sse(sum128_3)
            + hsum128_ps_sse(sum128_4);
        for i in 0..n - m {
            result += (*ptr1.add(i)) * (*ptr2.add(i));
        }
        result
    }
}

#[cfg(test)]
mod tests {
    #[test]
    fn test_spaces_sse() {
        use super::*;
        use crate::segment::spaces::simple::*;

        if is_x86_feature_detected!("sse") {
            let v1: Vec<f32> = vec![
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                26., 27., 28., 29., 30., 31.,
            ];
            let v2: Vec<f32> = vec![
                40., 41., 42., 43., 44., 45., 46., 47., 48., 49., 50., 51., 52., 53., 54., 55.,
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                10., 11., 12., 13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
                56., 57., 58., 59., 60., 61.,
            ];

            let euclid_simd = unsafe { euclid_similarity_sse(&v1, &v2) };
            let euclid = euclid_similarity(&v1, &v2);
            assert_eq!(euclid_simd, euclid);

            let manhattan_simd = unsafe { manhattan_similarity_sse(&v1, &v2) };
            let manhattan = manhattan_similarity(&v1, &v2);
            assert_eq!(manhattan_simd, manhattan);

            let dot_simd = unsafe { dot_similarity_sse(&v1, &v2) };
            let dot = dot_similarity(&v1, &v2);
            assert_eq!(dot_simd, dot);

            let cosine_simd = unsafe { cosine_preprocess_sse(v1.clone()) };
            let cosine = cosine_preprocess(v1);
            assert_eq!(cosine_simd, cosine);
        } else {
            println!("sse test skipped");
        }
    }
}