embedded-nn 0.2.0

A pure Rust, #![no_std] neural network inference library inspired by CMSIS-NN for microcontrollers and embedded targets.
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

use embedded_nn::{
    pack_q15x2_32x1, pack_s8x4_32x1, requantize_s64,
    simd::{vec_dot_s16, vec_dot_s8},
    support::{clamp, divide_by_power_of_two, doubling_high_mult_no_sat, requantize},
    Activation, Context, Dims, Error, PerChannelQuantParams, PerTensorQuantParams, QuantParams,
    Tile,
};

#[test]
fn test_clamp_boundaries_and_ordering() {
    assert_eq!(clamp(100, -10, 10), 10);
    assert_eq!(clamp(-100, -10, 10), -10);
    assert_eq!(clamp(0, -10, 10), 0);
    assert_eq!(clamp(-10, -10, 10), -10);
    assert_eq!(clamp(10, -10, 10), 10);
}

#[test]
fn test_doubling_high_mult_no_sat_cases() {
    // Q31 multiplication
    let res1 = doubling_high_mult_no_sat(1073741824, 1073741824); // 0.5 * 0.5 = 0.25 -> 536870912
    assert!((res1 - 536870912).abs() <= 1);

    let res2 = doubling_high_mult_no_sat(2147483647, 1073741824); // ~1.0 * 0.5 = ~0.5 -> 1073741823
    assert!((res2 - 1073741823).abs() <= 2);

    let res_neg = doubling_high_mult_no_sat(-1073741824, 1073741824); // -0.5 * 0.5 = -0.25
    assert!((res_neg - (-536870912)).abs() <= 1);
}

#[test]
fn test_divide_by_power_of_two_rounding_and_edge_cases() {
    assert_eq!(divide_by_power_of_two(100, 0), 100);
    assert_eq!(divide_by_power_of_two(100, -1), 100);
    assert_eq!(divide_by_power_of_two(100, 31), 0);
    assert_eq!(divide_by_power_of_two(100, 35), 0);

    // Positive dividends
    assert_eq!(divide_by_power_of_two(16, 2), 4);
    assert_eq!(divide_by_power_of_two(18, 2), 5); // 4.5 -> rounds to 5

    // Negative dividends rounding
    assert_eq!(divide_by_power_of_two(-16, 2), -4);
    assert_eq!(divide_by_power_of_two(-18, 2), -5); // -4.5 -> rounds away from zero (-5)
}

#[test]
fn test_requantize_positive_and_negative_shifts() {
    let val = 1000;
    let mult = 1073741824; // 0.5

    // Right shift (shift < 0)
    let res_right = requantize(val, mult, -1); // 1000 * 0.5 / 2 = 250
    assert_eq!(res_right, 250);

    // Left shift (shift >= 0)
    let res_left = requantize(val, mult, 1); // (1000 << 1) * 0.5 = 1000
    assert_eq!(res_left, 1000);
}

#[test]
fn test_requantize_s64_math() {
    let val: i64 = 1000000;
    let reduced_mult = 16384;
    let shift = 0;
    let res = requantize_s64(val, reduced_mult, shift);
    assert!(res > 0);
}

#[test]
fn test_packing_helpers() {
    let packed_s8 = pack_s8x4_32x1(-1i8, 0i8, 127i8, -128i8);
    assert_eq!(packed_s8 as u32 & 0xFF, 0xFF);
    assert_eq!((packed_s8 as u32 >> 8) & 0xFF, 0x00);
    assert_eq!((packed_s8 as u32 >> 16) & 0xFF, 0x7F);
    assert_eq!((packed_s8 as u32 >> 24) & 0xFF, 0x80);

    let packed_q15 = pack_q15x2_32x1(-1000i16, 2000i16);
    assert_eq!((packed_q15 & 0xFFFF) as i16, -1000);
    assert_eq!((packed_q15 >> 16) as i16, 2000);
}

#[test]
fn test_vec_dot_s8_length_variations() {
    // Test exact chunk boundaries: lengths 0, 1, 3, 7, 8, 9, 16, 17
    let lhs = [1i8, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17];
    let rhs = [1i8; 17];

    assert_eq!(vec_dot_s8(&lhs[..0], &rhs[..0], 0), 0);
    assert_eq!(vec_dot_s8(&lhs[..1], &rhs[..1], 0), 1);
    assert_eq!(vec_dot_s8(&lhs[..3], &rhs[..3], 0), 6);
    assert_eq!(vec_dot_s8(&lhs[..7], &rhs[..7], 0), 28);
    assert_eq!(vec_dot_s8(&lhs[..8], &rhs[..8], 0), 36);
    assert_eq!(vec_dot_s8(&lhs[..9], &rhs[..9], 0), 45);
    assert_eq!(vec_dot_s8(&lhs[..16], &rhs[..16], 0), 136);
    assert_eq!(vec_dot_s8(&lhs[..17], &rhs[..17], 0), 153);

    // Test with lhs_offset = 2
    // Sum for length 8 with offset 2: 36 + 8*2 = 52
    assert_eq!(vec_dot_s8(&lhs[..8], &rhs[..8], 2), 52);
}

#[test]
fn test_vec_dot_s16_length_variations() {
    let lhs = [10i16, 20, 30, 40, 50, 60, 70];
    let rhs = [1i16; 7];

    assert_eq!(vec_dot_s16(&lhs[..0], &rhs[..0]), 0);
    assert_eq!(vec_dot_s16(&lhs[..1], &rhs[..1]), 10);
    assert_eq!(vec_dot_s16(&lhs[..4], &rhs[..4]), 100);
    assert_eq!(vec_dot_s16(&lhs[..7], &rhs[..7]), 280);
}

#[test]
fn test_types_and_error_display() {
    let err1 = Error::ArgumentError;
    let err2 = Error::NoImplementation;
    let err3 = Error::Failure;

    assert_eq!(format!("{}", err1), "Invalid or incompatible arguments");
    assert_eq!(format!("{}", err2), "No implementation available");
    assert_eq!(format!("{}", err3), "Operation failure");

    let dims = Dims::new(2, 3, 4, 5);
    assert_eq!(dims.total_size(), 120);

    let tile = Tile::new(3, 3);
    assert_eq!(tile.w, 3);
    assert_eq!(tile.h, 3);

    let pt = PerTensorQuantParams::new(100, 2);
    let mults = [100, 200];
    let shifts = [1, 2];
    let pc = PerChannelQuantParams::new(&mults, &shifts);

    let q1 = QuantParams::PerTensor(pt);
    let q2 = QuantParams::PerChannel(pc);

    match q1 {
        QuantParams::PerTensor(p) => assert_eq!(p.multiplier, 100),
        _ => panic!(),
    }
    match q2 {
        QuantParams::PerChannel(p) => assert_eq!(p.multiplier.len(), 2),
        _ => panic!(),
    }

    let mut buf = [0u8; 128];
    let ctx_empty = Context::empty();
    assert!(ctx_empty.buf.is_none());

    let ctx_new = Context::new(&mut buf);
    assert!(ctx_new.buf.is_some());

    let act8 = Activation::int8_unconstrained();
    assert_eq!(act8.min, -128);
    assert_eq!(act8.max, 127);

    let act16 = Activation::int16_unconstrained();
    assert_eq!(act16.min, -32768);
    assert_eq!(act16.max, 32767);
}