oxionnx 0.1.2

Pure Rust ONNX inference engine — zero C/C++ dependencies
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

//! Native f16 element-wise operation implementations.
//!
//! Converts inputs to `half::f16`, performs the operation, and returns
//! results as standard f32 `Tensor` values with f16-rounded precision.

use crate::tensor::Tensor;
use crate::OnnxError;

use super::broadcast::broadcast_binary_f16;

/// Execute an element-wise operation natively in f16 precision.
///
/// Converts inputs to `half::f16`, performs the operation, and returns
/// the result as a standard f32 `Tensor` with f16-rounded values.
///
/// Returns `None` if the op is not supported for native f16 execution,
/// in which case the caller should fall back to the normal f32 path
/// (optionally with f16 rounding of outputs).
pub fn execute_elementwise_f16(
    op_type: &str,
    inputs: &[&Tensor],
) -> Option<Result<Vec<Tensor>, OnnxError>> {
    match op_type {
        "Relu" => Some(execute_relu_f16(inputs)),
        "Add" => Some(execute_add_f16(inputs)),
        "Mul" => Some(execute_mul_f16(inputs)),
        "Sub" => Some(execute_sub_f16(inputs)),
        "Sigmoid" => Some(execute_sigmoid_f16(inputs)),
        "Tanh" => Some(execute_tanh_f16(inputs)),
        "Neg" => Some(execute_neg_f16(inputs)),
        "Abs" => Some(execute_abs_f16(inputs)),
        _ => None,
    }
}

fn execute_relu_f16(inputs: &[&Tensor]) -> Result<Vec<Tensor>, OnnxError> {
    let input = inputs.first().ok_or_else(|| {
        OnnxError::ShapeMismatch("Relu f16: expected at least 1 input".to_string())
    })?;
    let zero = half::f16::ZERO;
    let data: Vec<f32> = input
        .data
        .iter()
        .map(|&v| {
            let h = half::f16::from_f32(v);
            if h < zero { zero } else { h }.to_f32()
        })
        .collect();
    Ok(vec![Tensor::new(data, input.shape.clone())])
}

fn execute_add_f16(inputs: &[&Tensor]) -> Result<Vec<Tensor>, OnnxError> {
    if inputs.len() < 2 {
        return Err(OnnxError::ShapeMismatch(
            "Add f16: expected 2 inputs".to_string(),
        ));
    }
    let a = inputs[0];
    let b = inputs[1];

    let out_shape = Tensor::broadcast_shape(&a.shape, &b.shape)
        .map_err(|e| OnnxError::ShapeMismatch(format!("Add f16 broadcast: {e}")))?;
    let out_size: usize = out_shape.iter().product();

    let data = if a.shape == b.shape {
        // Fast path: same shape, no broadcasting needed
        a.data
            .iter()
            .zip(b.data.iter())
            .map(|(&va, &vb)| {
                let ha = half::f16::from_f32(va);
                let hb = half::f16::from_f32(vb);
                (ha + hb).to_f32()
            })
            .collect()
    } else {
        broadcast_binary_f16(
            &a.data,
            &a.shape,
            &b.data,
            &b.shape,
            &out_shape,
            out_size,
            |ha, hb| ha + hb,
        )
    };

    Ok(vec![Tensor::new(data, out_shape)])
}

fn execute_mul_f16(inputs: &[&Tensor]) -> Result<Vec<Tensor>, OnnxError> {
    if inputs.len() < 2 {
        return Err(OnnxError::ShapeMismatch(
            "Mul f16: expected 2 inputs".to_string(),
        ));
    }
    let a = inputs[0];
    let b = inputs[1];

    let out_shape = Tensor::broadcast_shape(&a.shape, &b.shape)
        .map_err(|e| OnnxError::ShapeMismatch(format!("Mul f16 broadcast: {e}")))?;
    let out_size: usize = out_shape.iter().product();

    let data = if a.shape == b.shape {
        a.data
            .iter()
            .zip(b.data.iter())
            .map(|(&va, &vb)| {
                let ha = half::f16::from_f32(va);
                let hb = half::f16::from_f32(vb);
                (ha * hb).to_f32()
            })
            .collect()
    } else {
        broadcast_binary_f16(
            &a.data,
            &a.shape,
            &b.data,
            &b.shape,
            &out_shape,
            out_size,
            |ha, hb| ha * hb,
        )
    };

    Ok(vec![Tensor::new(data, out_shape)])
}

fn execute_sub_f16(inputs: &[&Tensor]) -> Result<Vec<Tensor>, OnnxError> {
    if inputs.len() < 2 {
        return Err(OnnxError::ShapeMismatch(
            "Sub f16: expected 2 inputs".to_string(),
        ));
    }
    let a = inputs[0];
    let b = inputs[1];

    let out_shape = Tensor::broadcast_shape(&a.shape, &b.shape)
        .map_err(|e| OnnxError::ShapeMismatch(format!("Sub f16 broadcast: {e}")))?;
    let out_size: usize = out_shape.iter().product();

    let data = if a.shape == b.shape {
        a.data
            .iter()
            .zip(b.data.iter())
            .map(|(&va, &vb)| {
                let ha = half::f16::from_f32(va);
                let hb = half::f16::from_f32(vb);
                (ha - hb).to_f32()
            })
            .collect()
    } else {
        broadcast_binary_f16(
            &a.data,
            &a.shape,
            &b.data,
            &b.shape,
            &out_shape,
            out_size,
            |ha, hb| ha - hb,
        )
    };

    Ok(vec![Tensor::new(data, out_shape)])
}

fn execute_sigmoid_f16(inputs: &[&Tensor]) -> Result<Vec<Tensor>, OnnxError> {
    let input = inputs.first().ok_or_else(|| {
        OnnxError::ShapeMismatch("Sigmoid f16: expected at least 1 input".to_string())
    })?;
    let data: Vec<f32> = input
        .data
        .iter()
        .map(|&v| {
            // Compute sigmoid in f16: 1 / (1 + exp(-x))
            let h = half::f16::from_f32(v);
            let neg_h = -h;
            let exp_neg = half::f16::from_f32(neg_h.to_f32().exp());
            let one = half::f16::ONE;
            let denom = one + exp_neg;
            half::f16::from_f32(one.to_f32() / denom.to_f32()).to_f32()
        })
        .collect();
    Ok(vec![Tensor::new(data, input.shape.clone())])
}

fn execute_tanh_f16(inputs: &[&Tensor]) -> Result<Vec<Tensor>, OnnxError> {
    let input = inputs.first().ok_or_else(|| {
        OnnxError::ShapeMismatch("Tanh f16: expected at least 1 input".to_string())
    })?;
    let data: Vec<f32> = input
        .data
        .iter()
        .map(|&v| {
            let h = half::f16::from_f32(v);
            half::f16::from_f32(h.to_f32().tanh()).to_f32()
        })
        .collect();
    Ok(vec![Tensor::new(data, input.shape.clone())])
}

fn execute_neg_f16(inputs: &[&Tensor]) -> Result<Vec<Tensor>, OnnxError> {
    let input = inputs.first().ok_or_else(|| {
        OnnxError::ShapeMismatch("Neg f16: expected at least 1 input".to_string())
    })?;
    let data: Vec<f32> = input
        .data
        .iter()
        .map(|&v| (-half::f16::from_f32(v)).to_f32())
        .collect();
    Ok(vec![Tensor::new(data, input.shape.clone())])
}

fn execute_abs_f16(inputs: &[&Tensor]) -> Result<Vec<Tensor>, OnnxError> {
    let input = inputs.first().ok_or_else(|| {
        OnnxError::ShapeMismatch("Abs f16: expected at least 1 input".to_string())
    })?;
    let data: Vec<f32> = input
        .data
        .iter()
        .map(|&v| {
            let h = half::f16::from_f32(v);
            half::f16::from_f32(h.to_f32().abs()).to_f32()
        })
        .collect();
    Ok(vec![Tensor::new(data, input.shape.clone())])
}