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

//! Tests for MatMul-related fusion passes.

use crate::graph::OpKind;
use crate::optimizer::fusion::matmul::{
    fuse_add_matmul_to_gemm, fuse_layer_norm, fuse_matmul_add, fuse_matmul_transpose,
};
use crate::optimizer::test_utils::{make_layer_norm_pattern, make_node};
use crate::tensor::Tensor;
use std::collections::HashMap;

#[test]
fn test_fuse_matmul_add() {
    let nodes = vec![
        make_node(OpKind::MatMul, "mm", vec!["x", "w"], vec!["mm_out"]),
        make_node(OpKind::Add, "add", vec!["mm_out", "bias"], vec!["add_out"]),
    ];
    let mut weights = HashMap::new();
    weights.insert("w".to_string(), Tensor::new(vec![1.0; 4], vec![2, 2]));
    weights.insert("bias".to_string(), Tensor::new(vec![0.5, 0.5], vec![2]));

    let result = fuse_matmul_add(nodes, &weights);
    assert_eq!(result.len(), 1);
    assert!(matches!(result[0].op, OpKind::Gemm));
    assert_eq!(result[0].outputs[0], "add_out");
    assert_eq!(result[0].inputs.len(), 3);
    assert_eq!(result[0].inputs[0], "x");
    assert_eq!(result[0].inputs[1], "w");
    assert_eq!(result[0].inputs[2], "bias");
}

#[test]
fn test_fuse_matmul_add_single_node() {
    let nodes = vec![make_node(
        OpKind::MatMul,
        "mm",
        vec!["x", "w"],
        vec!["mm_out"],
    )];
    let weights = HashMap::new();
    let result = fuse_matmul_add(nodes, &weights);
    assert_eq!(result.len(), 1);
}

#[test]
fn test_fuse_matmul_add_bias_not_1d() {
    let nodes = vec![
        make_node(OpKind::MatMul, "mm", vec!["x", "w"], vec!["mm_out"]),
        make_node(OpKind::Add, "add", vec!["mm_out", "bias"], vec!["add_out"]),
    ];
    let mut weights = HashMap::new();
    weights.insert("w".to_string(), Tensor::new(vec![1.0; 4], vec![2, 2]));
    weights.insert("bias".to_string(), Tensor::new(vec![0.5; 4], vec![2, 2]));
    let result = fuse_matmul_add(nodes, &weights);
    assert_eq!(result.len(), 2);
}

#[test]
fn test_no_fusion_when_multiple_consumers() {
    let nodes = vec![
        make_node(OpKind::MatMul, "mm", vec!["x", "w"], vec!["mm_out"]),
        make_node(OpKind::Add, "add", vec!["mm_out", "bias"], vec!["add_out"]),
        make_node(OpKind::Relu, "relu", vec!["mm_out"], vec!["relu_out"]),
    ];
    let weights = {
        let mut w = HashMap::new();
        w.insert("w".to_string(), Tensor::new(vec![1.0; 4], vec![2, 2]));
        w.insert("bias".to_string(), Tensor::new(vec![0.5, 0.5], vec![2]));
        w
    };

    let result = fuse_matmul_add(nodes, &weights);
    assert_eq!(result.len(), 3);
}

#[test]
fn test_fuse_layer_norm_basic() {
    let (nodes, weights) = make_layer_norm_pattern(false);
    let result = fuse_layer_norm(nodes, &weights);

    assert_eq!(result.len(), 1);
    assert!(matches!(result[0].op, OpKind::LayerNorm));
    assert_eq!(result[0].inputs[0], "X");
    assert_eq!(result[0].outputs[0], "normalized");
    let eps = result[0].attrs.f("epsilon", 0.0);
    assert!((eps - 1e-5).abs() < 1e-8);
}

#[test]
fn test_fuse_layer_norm_with_scale_bias() {
    let (nodes, weights) = make_layer_norm_pattern(true);
    let result = fuse_layer_norm(nodes, &weights);

    assert_eq!(result.len(), 1);
    assert!(matches!(result[0].op, OpKind::LayerNorm));
    assert_eq!(result[0].inputs.len(), 3);
    assert_eq!(result[0].inputs[0], "X");
    assert_eq!(result[0].inputs[1], "scale");
    assert_eq!(result[0].inputs[2], "bias");
    assert_eq!(result[0].outputs[0], "output");
}

#[test]
fn test_fuse_layer_norm_no_match_wrong_pow() {
    let (nodes, mut weights) = make_layer_norm_pattern(false);
    weights.insert("pow_exp".to_string(), Tensor::new(vec![3.0], vec![1]));

    let original_len = nodes.len();
    let result = fuse_layer_norm(nodes, &weights);

    assert_eq!(result.len(), original_len);
}

// --- fuse_matmul_transpose tests ---

#[test]
fn test_fuse_matmul_transpose_2d() {
    let matmul = make_node(OpKind::MatMul, "mm", vec!["a", "b"], vec!["mm_out"]);
    let mut transpose = make_node(OpKind::Transpose, "t", vec!["mm_out"], vec!["t_out"]);
    transpose
        .attrs
        .int_lists
        .insert("perm".to_string(), vec![1, 0]);

    let nodes = vec![matmul, transpose];
    let result = fuse_matmul_transpose(nodes);

    assert_eq!(result.len(), 1);
    assert!(matches!(result[0].op, OpKind::Gemm));
    // Should be Gemm(B, A) with transA=1, transB=1
    assert_eq!(result[0].inputs[0], "b");
    assert_eq!(result[0].inputs[1], "a");
    assert_eq!(result[0].attrs.i("transA", 0), 1);
    assert_eq!(result[0].attrs.i("transB", 0), 1);
    assert_eq!(result[0].outputs[0], "t_out");
}

#[test]
fn test_fuse_matmul_transpose_3d_last_two() {
    let matmul = make_node(OpKind::MatMul, "mm", vec!["a", "b"], vec!["mm_out"]);
    let mut transpose = make_node(OpKind::Transpose, "t", vec!["mm_out"], vec!["t_out"]);
    transpose
        .attrs
        .int_lists
        .insert("perm".to_string(), vec![0, 2, 1]);

    let nodes = vec![matmul, transpose];
    let result = fuse_matmul_transpose(nodes);

    assert_eq!(result.len(), 1);
    assert!(matches!(result[0].op, OpKind::Gemm));
}

#[test]
fn test_fuse_matmul_transpose_no_fusion_wrong_perm() {
    let matmul = make_node(OpKind::MatMul, "mm", vec!["a", "b"], vec!["mm_out"]);
    let mut transpose = make_node(OpKind::Transpose, "t", vec!["mm_out"], vec!["t_out"]);
    // Perm [2, 0, 1] doesn't just swap last two dims
    transpose
        .attrs
        .int_lists
        .insert("perm".to_string(), vec![2, 0, 1]);

    let nodes = vec![matmul, transpose];
    let result = fuse_matmul_transpose(nodes);

    assert_eq!(result.len(), 2);
}

#[test]
fn test_fuse_matmul_transpose_no_fusion_multiple_consumers() {
    let matmul = make_node(OpKind::MatMul, "mm", vec!["a", "b"], vec!["mm_out"]);
    let mut transpose = make_node(OpKind::Transpose, "t", vec!["mm_out"], vec!["t_out"]);
    transpose
        .attrs
        .int_lists
        .insert("perm".to_string(), vec![1, 0]);
    let relu = make_node(OpKind::Relu, "relu", vec!["mm_out"], vec!["relu_out"]);

    let nodes = vec![matmul, transpose, relu];
    let result = fuse_matmul_transpose(nodes);

    assert_eq!(result.len(), 3);
}

// --- fuse_add_matmul_to_gemm tests ---

#[test]
fn test_fuse_add_matmul_to_gemm() {
    let add = make_node(OpKind::Add, "add", vec!["x", "bias"], vec!["add_out"]);
    let matmul = make_node(OpKind::MatMul, "mm", vec!["add_out", "w"], vec!["mm_out"]);

    let nodes = vec![add, matmul];
    let mut weights = HashMap::new();
    // bias: [2], W: [2, 3]
    weights.insert("bias".to_string(), Tensor::new(vec![1.0, 2.0], vec![2]));
    weights.insert(
        "w".to_string(),
        Tensor::new(vec![1.0, 0.0, 0.0, 0.0, 1.0, 0.0], vec![2, 3]),
    );

    let result = fuse_add_matmul_to_gemm(nodes, &mut weights);

    assert_eq!(result.len(), 1);
    assert!(matches!(result[0].op, OpKind::Gemm));
    assert_eq!(result[0].inputs[0], "x");
    assert_eq!(result[0].inputs[1], "w");
    assert_eq!(result[0].inputs.len(), 3);
    assert_eq!(result[0].outputs[0], "mm_out");
    assert_eq!(result[0].attrs.f("alpha", 0.0), 1.0);
    assert_eq!(result[0].attrs.f("beta", 0.0), 1.0);

    // fused_bias = [1, 2] @ [[1,0,0],[0,1,0]] = [1, 2, 0]
    let fused_bias_name = &result[0].inputs[2];
    let fused_bias = weights
        .get(fused_bias_name)
        .expect("fused bias should exist");
    assert_eq!(fused_bias.shape, vec![3]);
    assert!((fused_bias.data[0] - 1.0).abs() < 1e-6);
    assert!((fused_bias.data[1] - 2.0).abs() < 1e-6);
    assert!((fused_bias.data[2] - 0.0).abs() < 1e-6);
}

#[test]
fn test_fuse_add_matmul_to_gemm_bias_first_input() {
    // Add with bias as first input
    let add = make_node(OpKind::Add, "add", vec!["bias", "x"], vec!["add_out"]);
    let matmul = make_node(OpKind::MatMul, "mm", vec!["add_out", "w"], vec!["mm_out"]);

    let nodes = vec![add, matmul];
    let mut weights = HashMap::new();
    weights.insert("bias".to_string(), Tensor::new(vec![3.0, 4.0], vec![2]));
    weights.insert(
        "w".to_string(),
        Tensor::new(vec![1.0, 0.0, 0.0, 1.0], vec![2, 2]),
    );

    let result = fuse_add_matmul_to_gemm(nodes, &mut weights);

    assert_eq!(result.len(), 1);
    assert!(matches!(result[0].op, OpKind::Gemm));
    assert_eq!(result[0].inputs[0], "x");
}

#[test]
fn test_fuse_add_matmul_no_fusion_bias_not_1d() {
    let add = make_node(OpKind::Add, "add", vec!["x", "bias"], vec!["add_out"]);
    let matmul = make_node(OpKind::MatMul, "mm", vec!["add_out", "w"], vec!["mm_out"]);

    let nodes = vec![add, matmul];
    let mut weights = HashMap::new();
    weights.insert("bias".to_string(), Tensor::new(vec![1.0; 4], vec![2, 2]));
    weights.insert("w".to_string(), Tensor::new(vec![1.0; 4], vec![2, 2]));

    let result = fuse_add_matmul_to_gemm(nodes, &mut weights);
    assert_eq!(result.len(), 2);
}

#[test]
fn test_fuse_add_matmul_no_fusion_w_not_in_weights() {
    let add = make_node(OpKind::Add, "add", vec!["x", "bias"], vec!["add_out"]);
    let matmul = make_node(OpKind::MatMul, "mm", vec!["add_out", "w"], vec!["mm_out"]);

    let nodes = vec![add, matmul];
    let mut weights = HashMap::new();
    weights.insert("bias".to_string(), Tensor::new(vec![1.0, 2.0], vec![2]));
    // w not in weights

    let result = fuse_add_matmul_to_gemm(nodes, &mut weights);
    assert_eq!(result.len(), 2);
}

#[test]
fn test_fuse_add_matmul_no_fusion_shape_mismatch() {
    let add = make_node(OpKind::Add, "add", vec!["x", "bias"], vec!["add_out"]);
    let matmul = make_node(OpKind::MatMul, "mm", vec!["add_out", "w"], vec!["mm_out"]);

    let nodes = vec![add, matmul];
    let mut weights = HashMap::new();
    // bias [3] vs W [2, 2] — K mismatch
    weights.insert(
        "bias".to_string(),
        Tensor::new(vec![1.0, 2.0, 3.0], vec![3]),
    );
    weights.insert("w".to_string(), Tensor::new(vec![1.0; 4], vec![2, 2]));

    let result = fuse_add_matmul_to_gemm(nodes, &mut weights);
    assert_eq!(result.len(), 2);
}