native_neural_network 0.3.1

Lib no_std Rust for native neural network (.rnn)
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

use super::qtypes::QuantError;

pub fn quantize_i8_symmetric_f32(input: &[f32], output: &mut [i8]) -> Result<f32, QuantError> {
    if input.is_empty() {
        return Err(QuantError::Empty);
    }
    if output.len() < input.len() {
        return Err(QuantError::ShapeMismatch);
    }

    let mut scale_from_contract = 0.0f32;
    if crate::engine::try_invoke_gpu_quantize_i8_f32(input, output, &mut scale_from_contract) {
        if !scale_from_contract.is_finite() || scale_from_contract <= 0.0 {
            return Err(QuantError::InvalidScale);
        }
        return Ok(scale_from_contract);
    }

    let mut max_abs = 0.0f32;
    for &v in input {
        let a = v.abs();
        if a > max_abs {
            max_abs = a;
        }
    }

    let scale = if max_abs <= 0.0 { 1.0 } else { max_abs / 127.0 };
    if !scale.is_finite() || scale <= 0.0 {
        return Err(QuantError::InvalidScale);
    }

    for i in 0..input.len() {
        let v = crate::math::roundf(input[i] / scale);
        output[i] = v.clamp(-127.0, 127.0) as i8;
    }

    Ok(scale)
}

pub fn quantize_i8_symmetric(input: &[f32], output: &mut [i8]) -> Result<f32, QuantError> {
    quantize_i8_symmetric_f32(input, output)
}

pub fn dequantize_i8_symmetric_f32(
    input: &[i8],
    output: &mut [f32],
    scale: f32,
) -> Result<(), QuantError> {
    if input.is_empty() {
        return Err(QuantError::Empty);
    }
    if output.len() < input.len() {
        return Err(QuantError::ShapeMismatch);
    }
    if !scale.is_finite() || scale <= 0.0 {
        return Err(QuantError::InvalidScale);
    }

    if crate::engine::try_invoke_gpu_dequantize_i8_f32(input, output, scale) {
        return Ok(());
    }

    for i in 0..input.len() {
        output[i] = input[i] as f32 * scale;
    }
    Ok(())
}

pub fn dequantize_i8_symmetric(
    input: &[i8],
    output: &mut [f32],
    scale: f32,
) -> Result<(), QuantError> {
    dequantize_i8_symmetric_f32(input, output, scale)
}

pub fn dequantize_i8_symmetric_f64(
    input: &[i8],
    output: &mut [f64],
    scale: f64,
) -> Result<(), QuantError> {
    if input.is_empty() {
        return Err(QuantError::Empty);
    }
    if output.len() < input.len() {
        return Err(QuantError::ShapeMismatch);
    }
    if !scale.is_finite() || scale <= 0.0 {
        return Err(QuantError::InvalidScale);
    }

    if crate::engine::try_invoke_gpu_dequantize_i8_f64(input, output, scale) {
        return Ok(());
    }

    for i in 0..input.len() {
        output[i] = input[i] as f64 * scale;
    }
    Ok(())
}

pub fn quantize_i8_symmetric_f64(input: &[f64], output: &mut [i8]) -> Result<f64, QuantError> {
    if input.is_empty() {
        return Err(QuantError::Empty);
    }
    if output.len() < input.len() {
        return Err(QuantError::ShapeMismatch);
    }

    let mut scale_from_contract = 0.0f64;
    if crate::engine::try_invoke_gpu_quantize_i8_f64(input, output, &mut scale_from_contract) {
        if !scale_from_contract.is_finite() || scale_from_contract <= 0.0 {
            return Err(QuantError::InvalidScale);
        }
        return Ok(scale_from_contract);
    }

    let mut max_abs = 0.0f64;
    for &v in input {
        let a = v.abs();
        if a > max_abs {
            max_abs = a;
        }
    }

    let scale = if max_abs <= 0.0 { 1.0 } else { max_abs / 127.0 };
    if !scale.is_finite() || scale <= 0.0 {
        return Err(QuantError::InvalidScale);
    }

    for i in 0..input.len() {
        let v = crate::math::roundd(input[i] / scale);
        output[i] = v.clamp(-127.0, 127.0) as i8;
    }

    Ok(scale)
}