oxionnx 0.1.3

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

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

#[test]
fn test_op_placement_cpu_only() {
    use crate::execution_providers::{decide_placement, OpPlacement, ProviderKind};
    let placement = OpPlacement::CpuOnly;
    let ops = [
        OpKind::MatMul,
        OpKind::Conv,
        OpKind::Add,
        OpKind::Reshape,
        OpKind::Softmax,
        OpKind::Relu,
    ];
    for op in &ops {
        let result = decide_placement(op, 1_000_000, &placement);
        assert_eq!(
            result,
            ProviderKind::Cpu,
            "CpuOnly must always return Cpu for {:?}",
            op
        );
    }
}

#[test]
fn test_op_placement_auto_small_input() {
    use crate::execution_providers::{decide_placement, OpPlacement, ProviderKind};
    // Threshold 64KB; input is only 100 bytes → should stay on CPU
    let placement = OpPlacement::Auto {
        gpu_threshold_bytes: 65536,
    };
    let result = decide_placement(&OpKind::MatMul, 100, &placement);
    assert_eq!(result, ProviderKind::Cpu);
}

#[test]
fn test_op_placement_auto_threshold() {
    use crate::execution_providers::{decide_placement, OpPlacement, ProviderKind};
    let placement = OpPlacement::Auto {
        gpu_threshold_bytes: 1024,
    };

    // Below threshold → CPU
    let below = decide_placement(&OpKind::MatMul, 512, &placement);
    assert_eq!(below, ProviderKind::Cpu);

    // At threshold → GPU-capable op should request GPU (returns Cpu without feature)
    let at = decide_placement(&OpKind::MatMul, 1024, &placement);
    // Without the gpu feature, result is Cpu; with gpu feature, result is Gpu
    #[cfg(feature = "gpu")]
    assert_eq!(at, ProviderKind::Gpu);
    #[cfg(not(feature = "gpu"))]
    assert_eq!(at, ProviderKind::Cpu);

    // Non-GPU-capable op above threshold → still CPU
    let reshape = decide_placement(&OpKind::Reshape, 2048, &placement);
    assert_eq!(reshape, ProviderKind::Cpu);
}

#[test]
fn test_op_placement_manual() {
    use crate::execution_providers::{decide_placement, OpPlacement, ProviderKind};
    let mut map = HashMap::new();
    #[cfg(feature = "gpu")]
    {
        map.insert(OpKind::MatMul, ProviderKind::Gpu);
    }
    #[cfg(not(feature = "gpu"))]
    {
        // Without gpu feature, just map to Cpu to test lookup works
        map.insert(OpKind::MatMul, ProviderKind::Cpu);
    }
    let placement = OpPlacement::Manual(map);

    let matmul_result = decide_placement(&OpKind::MatMul, 0, &placement);
    #[cfg(feature = "gpu")]
    assert_eq!(matmul_result, ProviderKind::Gpu);
    #[cfg(not(feature = "gpu"))]
    assert_eq!(matmul_result, ProviderKind::Cpu);

    // Unmapped op defaults to Cpu
    let reshape_result = decide_placement(&OpKind::Reshape, 0, &placement);
    assert_eq!(reshape_result, ProviderKind::Cpu);
}

#[test]
fn test_decide_placement_default() {
    use crate::execution_providers::{decide_placement, OpPlacement, ProviderKind};
    let placement = OpPlacement::default();
    let result = decide_placement(&OpKind::Add, 999999, &placement);
    assert_eq!(result, ProviderKind::Cpu);
}

#[test]
fn test_is_gpu_capable_matmul() {
    use crate::execution_providers::is_gpu_capable;
    assert!(is_gpu_capable(&OpKind::MatMul));
    assert!(is_gpu_capable(&OpKind::Gemm));
    assert!(is_gpu_capable(&OpKind::Conv));
    assert!(is_gpu_capable(&OpKind::Softmax));
    assert!(is_gpu_capable(&OpKind::Relu));
    assert!(is_gpu_capable(&OpKind::ReduceMean));
}

#[test]
fn test_is_gpu_capable_reshape() {
    use crate::execution_providers::is_gpu_capable;
    assert!(!is_gpu_capable(&OpKind::Reshape));
    assert!(!is_gpu_capable(&OpKind::Squeeze));
    assert!(!is_gpu_capable(&OpKind::Flatten));
    assert!(!is_gpu_capable(&OpKind::Gather));
    assert!(!is_gpu_capable(&OpKind::Shape));
}

#[test]
fn test_builder_op_placement_api() {
    use crate::execution_providers::OpPlacement;
    let builder = SessionBuilder::new().with_op_placement(OpPlacement::Auto {
        gpu_threshold_bytes: 4096,
    });
    match &builder.op_placement {
        OpPlacement::Auto {
            gpu_threshold_bytes,
        } => {
            assert_eq!(*gpu_threshold_bytes, 4096);
        }
        other => panic!("Expected Auto, got {:?}", other),
    }

    // Build a simple session with placement to verify end-to-end wiring
    let graph = Graph {
        nodes: vec![Node {
            name: "relu0".to_string(),
            op: OpKind::Relu,
            inputs: vec!["input".to_string()],
            outputs: vec!["output".to_string()],
            attrs: Attributes::default(),
        }],
        input_names: vec!["input".to_string()],
        output_names: vec!["output".to_string()],
        ..Default::default()
    };
    let session = SessionBuilder::new()
        .with_optimization_level(OptLevel::None)
        .with_op_placement(OpPlacement::Auto {
            gpu_threshold_bytes: 1024,
        })
        .build_from_graph(graph, HashMap::new())
        .expect("build with op placement");

    // Session should run correctly with placement configured
    let input = Tensor::new(vec![-1.0, 2.0, -3.0], vec![1, 3]);
    let out = session.run_one("input", input).expect("run");
    let y = out.get("output").expect("output");
    assert_eq!(y.data, vec![0.0, 2.0, 0.0]);
}

/// Verify that `with_provider_kinds([ProviderKind::Cpu])` stores one provider.
///
/// This test does NOT require GPU hardware — it only verifies the API stores
/// the provider list correctly and that a CPU-only session executes correctly.
#[test]
fn test_with_provider_kinds_cpu_stores_and_runs() {
    use crate::execution_providers::ProviderKind;
    let builder = SessionBuilder::new().with_provider_kinds([ProviderKind::Cpu]);
    assert_eq!(builder.providers.len(), 1, "providers must have 1 element");
    assert_eq!(
        builder.providers[0],
        ProviderKind::Cpu,
        "first provider must be Cpu"
    );

    // Build and run a simple session to verify CPU fallback works correctly
    // when the provider list is [Cpu].
    let graph = Graph {
        nodes: vec![Node {
            name: "relu_ep".to_string(),
            op: OpKind::Relu,
            inputs: vec!["x".to_string()],
            outputs: vec!["y".to_string()],
            attrs: Attributes::default(),
        }],
        input_names: vec!["x".to_string()],
        output_names: vec!["y".to_string()],
        ..Default::default()
    };
    let session = SessionBuilder::new()
        .with_optimization_level(OptLevel::None)
        .with_provider_kinds([ProviderKind::Cpu])
        .build_from_graph(graph, HashMap::new())
        .expect("build with CPU provider kind");

    let input = Tensor::new(vec![-2.0f32, 1.0, -3.0, 4.0], vec![4]);
    let out = session.run_one("x", input).expect("run with provider-list");
    let y = out.get("y").expect("output y");
    assert_eq!(y.data, vec![0.0, 1.0, 0.0, 4.0]);
    assert_eq!(y.shape, vec![4]);
}

/// Verify that an empty provider list (default) preserves legacy behavior.
///
/// Calling `build_from_graph` without `with_provider_kinds` must still work
/// correctly — the providers list is empty and the legacy dispatch path is used.
#[test]
fn test_empty_provider_list_uses_legacy_dispatch() {
    let graph = Graph {
        nodes: vec![Node {
            name: "relu_legacy".to_string(),
            op: OpKind::Relu,
            inputs: vec!["x".to_string()],
            outputs: vec!["y".to_string()],
            attrs: Attributes::default(),
        }],
        input_names: vec!["x".to_string()],
        output_names: vec!["y".to_string()],
        ..Default::default()
    };
    let session = SessionBuilder::new()
        .with_optimization_level(OptLevel::None)
        // No with_provider_kinds call — providers is empty Vec
        .build_from_graph(graph, HashMap::new())
        .expect("build with empty provider list (legacy)");

    // providers must be empty
    assert!(
        session.providers.is_empty(),
        "default build must have empty providers list"
    );

    let input = Tensor::new(vec![-5.0f32, 0.0, 3.0], vec![3]);
    let out = session
        .run_one("x", input)
        .expect("run with empty providers");
    let y = out.get("y").expect("output y");
    assert_eq!(y.data, vec![0.0, 0.0, 3.0]);
}

/// Verify `with_provider_kinds` with multiple providers stores them in order.
#[test]
fn test_with_provider_kinds_multiple_providers_order() {
    use crate::execution_providers::ProviderKind;
    // Only Cpu is guaranteed to be available without feature flags, but we
    // can verify the storage order by using Cpu twice (or once and checking).
    let builder = SessionBuilder::new().with_provider_kinds([ProviderKind::Cpu, ProviderKind::Cpu]);
    assert_eq!(builder.providers.len(), 2);
    assert_eq!(builder.providers[0], ProviderKind::Cpu);
    assert_eq!(builder.providers[1], ProviderKind::Cpu);
}