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
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307

use super::super::types::OptLevel;
use super::super::Session;
use crate::graph::{Attributes, Graph, Node, OpKind};
use crate::tensor::Tensor;
use std::collections::HashMap;

/// Basic mixed precision: Relu → Add → MatMul → Relu graph.
/// Verify output matches f32-only within tolerance.
#[test]
fn test_mixed_precision_relu_add_matmul_relu() {
    // Graph: input [2,3] → Relu → relu_out → Add(relu_out, bias) → add_out
    //        → MatMul(add_out, weight [3,2]) → mm_out → Relu → output [2,2]
    let relu1 = Node {
        op: OpKind::Relu,
        name: "relu1".to_string(),
        inputs: vec!["input".to_string()],
        outputs: vec!["relu_out".to_string()],
        attrs: Attributes::default(),
    };
    let add = Node {
        op: OpKind::Add,
        name: "add1".to_string(),
        inputs: vec!["relu_out".to_string(), "bias".to_string()],
        outputs: vec!["add_out".to_string()],
        attrs: Attributes::default(),
    };
    let matmul = Node {
        op: OpKind::MatMul,
        name: "matmul1".to_string(),
        inputs: vec!["add_out".to_string(), "weight".to_string()],
        outputs: vec!["mm_out".to_string()],
        attrs: Attributes::default(),
    };
    let relu2 = Node {
        op: OpKind::Relu,
        name: "relu2".to_string(),
        inputs: vec!["mm_out".to_string()],
        outputs: vec!["output".to_string()],
        attrs: Attributes::default(),
    };

    let graph = Graph {
        nodes: vec![relu1, add, matmul, relu2],
        input_names: vec!["input".to_string()],
        output_names: vec!["output".to_string()],
        ..Default::default()
    };

    let mut weights = HashMap::new();
    weights.insert(
        "bias".to_string(),
        Tensor::new(vec![0.5, 0.5, 0.5], vec![3]),
    );
    weights.insert(
        "weight".to_string(),
        Tensor::new(vec![1.0, 0.0, 0.0, 1.0, 1.0, 1.0], vec![3, 2]),
    );

    // Run with mixed precision
    let session_mp = Session::builder()
        .with_optimization_level(OptLevel::None)
        .with_mixed_precision(true)
        .build_from_graph(graph.clone(), weights.clone())
        .expect("build mixed precision session");

    // Run without mixed precision (reference)
    let session_f32 = Session::builder()
        .with_optimization_level(OptLevel::None)
        .build_from_graph(graph, weights)
        .expect("build f32 session");

    let input = Tensor::new(vec![-1.0, 2.0, 0.5, 3.0, -0.5, 1.0], vec![2, 3]);
    let out_mp = session_mp.run_one("input", input.clone()).expect("run mp");
    let out_f32 = session_f32.run_one("input", input).expect("run f32");

    let mp_data = &out_mp.get("output").expect("mp output").data;
    let f32_data = &out_f32.get("output").expect("f32 output").data;

    // Mixed precision should match f32 within tolerance (f16 has ~0.1% relative error)
    assert_eq!(mp_data.len(), f32_data.len());
    for (i, (&mp_val, &f32_val)) in mp_data.iter().zip(f32_data.iter()).enumerate() {
        let abs_err = (mp_val - f32_val).abs();
        let rel_tol = f32_val.abs() * 0.01 + 0.01; // 1% relative + 0.01 absolute
        assert!(
            abs_err < rel_tol,
            "Output[{i}]: mp={mp_val}, f32={f32_val}, err={abs_err} > tol={rel_tol}"
        );
    }
}

/// Verify f16-safe ops actually execute in f16 by checking profiling data.
#[test]
fn test_mixed_precision_profiling_shows_f16() {
    let relu = Node {
        op: OpKind::Relu,
        name: "relu1".to_string(),
        inputs: vec!["input".to_string()],
        outputs: vec!["relu_out".to_string()],
        attrs: Attributes::default(),
    };
    let add = Node {
        op: OpKind::Add,
        name: "add1".to_string(),
        inputs: vec!["relu_out".to_string(), "bias".to_string()],
        outputs: vec!["output".to_string()],
        attrs: Attributes::default(),
    };

    let graph = Graph {
        nodes: vec![relu, add],
        input_names: vec!["input".to_string()],
        output_names: vec!["output".to_string()],
        ..Default::default()
    };

    let mut weights = HashMap::new();
    weights.insert(
        "bias".to_string(),
        Tensor::new(vec![1.0, 2.0, 3.0], vec![3]),
    );

    let session = Session::builder()
        .with_optimization_level(OptLevel::None)
        .with_profiling()
        .with_mixed_precision(true)
        .build_from_graph(graph, weights)
        .expect("build");

    let input = Tensor::new(vec![-1.0, 2.0, 3.0], vec![1, 3]);
    let _outputs = session.run_one("input", input).expect("run");

    let profiles = session.profiling_results().expect("profiling enabled");
    assert_eq!(profiles.len(), 2);
    // Relu has native f16 path — profiled as "Relu(f16)"
    assert_eq!(profiles[0].op_type, "Relu(f16)");
    // Add has native f16 path — profiled as "Add(f16)"
    assert_eq!(profiles[1].op_type, "Add(f16)");
}

/// Verify MatMul uses f32 accumulation even with mixed precision enabled.
#[test]
fn test_mixed_precision_matmul_stays_f32() {
    let matmul = Node {
        op: OpKind::MatMul,
        name: "mm".to_string(),
        inputs: vec!["input".to_string(), "weight".to_string()],
        outputs: vec!["output".to_string()],
        attrs: Attributes::default(),
    };

    let graph = Graph {
        nodes: vec![matmul],
        input_names: vec!["input".to_string()],
        output_names: vec!["output".to_string()],
        ..Default::default()
    };

    let mut weights = HashMap::new();
    weights.insert(
        "weight".to_string(),
        Tensor::new(vec![1.0, 0.0, 0.0, 1.0], vec![2, 2]),
    );

    let session = Session::builder()
        .with_optimization_level(OptLevel::None)
        .with_profiling()
        .with_mixed_precision(true)
        .build_from_graph(graph, weights)
        .expect("build");

    let input = Tensor::new(vec![3.0, 7.0, 5.0, 11.0], vec![2, 2]);
    let outputs = session.run_one("input", input.clone()).expect("run");

    let out = outputs.get("output").expect("output");
    // Identity matrix multiplication: result == input (exact f32)
    assert_eq!(out.data, vec![3.0, 7.0, 5.0, 11.0]);

    // Profiling should show "MatMul" (not "MatMul(f16)")
    let profiles = session.profiling_results().expect("profiling enabled");
    assert_eq!(profiles.len(), 1);
    assert_eq!(profiles[0].op_type, "MatMul");
}

/// Verify mixed precision session builds and runs without error.
#[test]
fn test_mixed_precision_builder() {
    let session = Session::builder()
        .with_mixed_precision(true)
        .load_from_bytes(&[]);
    assert!(session.is_ok());
    let session = session.expect("should build");
    assert!(session.mixed_precision);
}

/// Verify f16 rounding for ops without native f16 path (e.g., Softmax).
#[test]
fn test_mixed_precision_f16_rounding_fallback() {
    // Softmax is f16-safe but has no native f16 path in execute_elementwise_f16.
    // With mixed precision, its output should be rounded to f16 precision.
    let softmax = Node {
        op: OpKind::Softmax,
        name: "sm".to_string(),
        inputs: vec!["input".to_string()],
        outputs: vec!["output".to_string()],
        attrs: Attributes::default(),
    };

    let graph = Graph {
        nodes: vec![softmax],
        input_names: vec!["input".to_string()],
        output_names: vec!["output".to_string()],
        ..Default::default()
    };

    let session_mp = Session::builder()
        .with_optimization_level(OptLevel::None)
        .with_mixed_precision(true)
        .build_from_graph(graph.clone(), HashMap::new())
        .expect("build mp");

    let session_f32 = Session::builder()
        .with_optimization_level(OptLevel::None)
        .build_from_graph(graph, HashMap::new())
        .expect("build f32");

    let input = Tensor::new(vec![1.0, 2.0, 3.0], vec![1, 3]);
    let out_mp = session_mp.run_one("input", input.clone()).expect("run mp");
    let out_f32 = session_f32.run_one("input", input).expect("run f32");

    let mp_data = &out_mp.get("output").expect("mp output").data;
    let f32_data = &out_f32.get("output").expect("f32 output").data;

    // Outputs should be close but mp data should be f16-rounded
    for (&mp_val, &f32_val) in mp_data.iter().zip(f32_data.iter()) {
        let abs_err = (mp_val - f32_val).abs();
        assert!(abs_err < 0.01, "mp={mp_val}, f32={f32_val}, err={abs_err}");
        // Verify mp_val is exactly representable in f16
        let roundtrip = half::f16::from_f32(mp_val).to_f32();
        assert_eq!(
            mp_val, roundtrip,
            "mp output should be exactly f16-representable"
        );
    }
}

/// Chain of consecutive f16-safe ops: Relu → Add → Sigmoid.
/// Verifies multi-op f16 execution works end-to-end.
#[test]
fn test_mixed_precision_consecutive_f16_ops() {
    let relu = Node {
        op: OpKind::Relu,
        name: "relu".to_string(),
        inputs: vec!["input".to_string()],
        outputs: vec!["relu_out".to_string()],
        attrs: Attributes::default(),
    };
    let add = Node {
        op: OpKind::Add,
        name: "add".to_string(),
        inputs: vec!["relu_out".to_string(), "bias".to_string()],
        outputs: vec!["add_out".to_string()],
        attrs: Attributes::default(),
    };
    let sigmoid = Node {
        op: OpKind::Sigmoid,
        name: "sig".to_string(),
        inputs: vec!["add_out".to_string()],
        outputs: vec!["output".to_string()],
        attrs: Attributes::default(),
    };

    let graph = Graph {
        nodes: vec![relu, add, sigmoid],
        input_names: vec!["input".to_string()],
        output_names: vec!["output".to_string()],
        ..Default::default()
    };

    let mut weights = HashMap::new();
    weights.insert(
        "bias".to_string(),
        Tensor::new(vec![-0.5, 0.0, 0.5], vec![3]),
    );

    let session = Session::builder()
        .with_optimization_level(OptLevel::None)
        .with_profiling()
        .with_mixed_precision(true)
        .build_from_graph(graph, weights)
        .expect("build");

    let input = Tensor::new(vec![-2.0, 1.0, 3.0], vec![1, 3]);
    let outputs = session.run_one("input", input).expect("run");
    let out = outputs.get("output").expect("output");

    // All ops are f16-safe and have native f16 paths
    let profiles = session.profiling_results().expect("profiling");
    assert_eq!(profiles.len(), 3);
    assert_eq!(profiles[0].op_type, "Relu(f16)");
    assert_eq!(profiles[1].op_type, "Add(f16)");
    assert_eq!(profiles[2].op_type, "Sigmoid(f16)");

    // Output should be sigmoid values in [0, 1]
    for &v in &out.data {
        assert!((0.0..=1.0).contains(&v), "sigmoid output {v} out of range");
    }
}