qdrant-edge 0.7.0

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

#if !defined PC_VER
#include <arm_neon.h>
#endif

#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
#include <arm_fp16.h>
float32_t dotProduct_half_4x4(const float16_t* pSrcA, const float16_t* pSrcB, uint32_t blockSize)
{
    float32_t dotProduct = 0.0f;
    float16x8_t sum1 = vdupq_n_f16(0.0f);
    float16x8_t sum2 = vdupq_n_f16(0.0f);
    float16x8_t sum3 = vdupq_n_f16(0.0f);
    float16x8_t sum4 = vdupq_n_f16(0.0f);

    for(uint32_t i=0; i < blockSize - (blockSize % 32); i+=32)
    {
        sum1 = vfmaq_f16(sum1, vld1q_f16(pSrcA), vld1q_f16(pSrcB));
        sum2 = vfmaq_f16(sum2, vld1q_f16(pSrcA+8), vld1q_f16(pSrcB+8));
        sum3 = vfmaq_f16(sum3, vld1q_f16(pSrcA+16), vld1q_f16(pSrcB+16));
        sum4 = vfmaq_f16(sum4, vld1q_f16(pSrcA+24), vld1q_f16(pSrcB+24));

        pSrcA += 32;
        pSrcB += 32;
    }

    float32x4_t sum = vcvt_f32_f16(vget_high_f16(sum1));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum1)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_high_f16(sum2)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum2)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_high_f16(sum3)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum3)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_high_f16(sum4)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum4)));

    dotProduct = vaddvq_f32(sum);
    for (uint32_t i=0; i < (blockSize % 32); i++) {
        dotProduct += (float32_t)((*pSrcA)*(*pSrcB));
        pSrcA += 1;
        pSrcB += 1;
    }

    return dotProduct;
}

float32_t euclideanDist_half_4x4(const float16_t* pSrcA, const float16_t* pSrcB, uint32_t blockSize)
{
    float32_t euclideanDistance = 0.0f;
    float16x8_t sum1 = vdupq_n_f16(0.0f);
    float16x8_t sub1 = vdupq_n_f16(0.0f);
    float16x8_t sum2 = vdupq_n_f16(0.0f);
    float16x8_t sub2 = vdupq_n_f16(0.0f);
    float16x8_t sum3 = vdupq_n_f16(0.0f);
    float16x8_t sub3 = vdupq_n_f16(0.0f);
    float16x8_t sum4 = vdupq_n_f16(0.0f);
    float16x8_t sub4 = vdupq_n_f16(0.0f);

    for(uint32_t i=0; i < blockSize - (blockSize % 32); i+=32)
    {
        sub1 = vsubq_f16(vld1q_f16(pSrcA), vld1q_f16(pSrcB));
        sum1 = vfmaq_f16(sum1, sub1, sub1);

        sub2 = vsubq_f16(vld1q_f16(pSrcA+8), vld1q_f16(pSrcB+8));
        sum2 = vfmaq_f16(sum2, sub2, sub2);

        sub3 = vsubq_f16(vld1q_f16(pSrcA+16), vld1q_f16(pSrcB+16));
        sum3 = vfmaq_f16(sum3, sub3, sub3);

        sub4 = vsubq_f16(vld1q_f16(pSrcA+24), vld1q_f16(pSrcB+24));
        sum4 = vfmaq_f16(sum4, sub4, sub4);

        pSrcA += 32;
        pSrcB += 32;
    }

    float32x4_t sum = vcvt_f32_f16(vget_high_f16(sum1));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum1)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_high_f16(sum2)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum2)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_high_f16(sum3)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum3)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_high_f16(sum4)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum4)));

    euclideanDistance = vaddvq_f32(sum);
    float16_t tmp = 0.0f;
    for (uint32_t i=0; i < (blockSize % 32); i++) {
        tmp = (*pSrcA - *pSrcB);
        euclideanDistance += (float32_t)(tmp * tmp);
        pSrcA += 1;
        pSrcB += 1;
    }

    return euclideanDistance;
}

float32_t manhattanDist_half_4x4(const float16_t* pSrcA, const float16_t* pSrcB, uint32_t blockSize)
{
    float32_t manhattanDistance = 0.0f;
    float16x8_t sum1 = vdupq_n_f16(0.0f);
    float16x8_t sum2 = vdupq_n_f16(0.0f);
    float16x8_t sum3 = vdupq_n_f16(0.0f);
    float16x8_t sum4 = vdupq_n_f16(0.0f);
    uint32_t i = 0;

    for(i=0; i < blockSize - (blockSize % 32); i+=32)
    {
        sum1 = vaddq_f16(sum1, vabdq_f16(vld1q_f16(pSrcA), vld1q_f16(pSrcB)));

        sum2 = vaddq_f16(sum2, vabdq_f16(vld1q_f16(pSrcA+8), vld1q_f16(pSrcB+8)));

        sum3 = vaddq_f16(sum3, vabdq_f16(vld1q_f16(pSrcA+16), vld1q_f16(pSrcB+16)));

        sum4 = vaddq_f16(sum4, vabdq_f16(vld1q_f16(pSrcA+24), vld1q_f16(pSrcB+24)));

        pSrcA += 32;
        pSrcB += 32;
    }

    float32x4_t sum = vcvt_f32_f16(vget_high_f16(sum1));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum1)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_high_f16(sum2)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum2)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_high_f16(sum3)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum3)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_high_f16(sum4)));
    sum = vaddq_f32(sum, vcvt_f32_f16(vget_low_f16(sum4)));

    manhattanDistance = vaddvq_f32(sum);
    for (i=0; i < (blockSize % 32); i++) {
        manhattanDistance += (float32_t)(vabsh_f16(*pSrcA - *pSrcB));
        pSrcA += 1;
        pSrcB += 1;
    }

    return manhattanDistance;
}
#endif