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
//! # Transpose
//!
//! Transposes the input tensor by permuting its dimensions, similar to numpy.transpose.
//!
//! **ONNX Spec**: <https://onnx.ai/onnx/operators/onnx__Transpose.html>
//!
//! ## Type Constraints
//! - T: All tensor types (float16, float32, float64, int8, int16, int32, int64, uint8, uint16,
//! uint32, uint64, bool, complex64, complex128, bfloat16, string)
//!
//! ## Opset Versions
//! - **Opset 1**: Initial version supporting all tensor types and permutation.
//! - **Opset 13**: Added bfloat16 support.
//! - **Opset 21**: Added support for int4, uint4, and float8 types.
//!
//! ## Example
//! When `perm = [1, 0, 2]` and input shape is `(1, 2, 3)`, the output shape will be `(2, 1, 3)`.
use derive_new::new;
use onnx_ir_derive::NodeBuilder;
use crate::ir::{ArgType, Argument, Node, RawNode, TensorType};
use crate::processor::{
InputSpec, NodeProcessor, NodeSpec, OutputPreferences, OutputSpec, ProcessError,
};
/// Configuration for Transpose operations
#[derive(Debug, Clone, new)]
pub struct TransposeConfig {
/// Permutation of dimensions
pub perm: Vec<i64>,
}
/// Node representation for Transpose operation
#[derive(Debug, Clone, NodeBuilder)]
pub struct TransposeNode {
pub name: String,
pub inputs: Vec<Argument>,
pub outputs: Vec<Argument>,
pub config: TransposeConfig,
}
pub(crate) struct TransposeProcessor;
impl NodeProcessor for TransposeProcessor {
type Config = TransposeConfig;
fn spec(&self) -> NodeSpec {
NodeSpec {
min_opset: 1,
max_opset: None,
inputs: InputSpec::Exact(1),
outputs: OutputSpec::Exact(1),
}
}
fn infer_types(
&self,
node: &mut RawNode,
opset: usize,
_output_preferences: &OutputPreferences,
) -> Result<(), ProcessError> {
// Get reference to config for type inference
let config = self
.extract_config(node, opset)
.expect("Config extraction failed");
// TODO: Missing validation that perm is a valid permutation.
// Must verify: len(perm) == rank, all values in [0, rank-1], no duplicates.
// Invalid permutations could cause index out of bounds errors.
// TODO: Missing validation that perm contains each index exactly once.
// E.g., perm=[0, 0, 2] has duplicate 0, missing 1 - should be rejected.
// Validate perm length matches input rank
let input_rank = match &node.inputs[0].ty {
ArgType::Tensor(t) => t.rank,
_ => {
return Err(ProcessError::TypeMismatch {
expected: "Tensor".to_string(),
actual: format!("{:?}", node.inputs[0].ty),
});
}
};
if config.perm.len() != input_rank {
return Err(ProcessError::Custom(format!(
"Transpose: perm length {} doesn't match input rank {}",
config.perm.len(),
input_rank
)));
}
let input_tensor = match &node.inputs[0].ty {
ArgType::Tensor(t) => t,
_ => unreachable!(), // already validated above
};
// Permute static_shape according to perm
let permuted_static_shape = input_tensor
.static_shape
.as_ref()
.map(|shape| config.perm.iter().map(|&i| shape[i as usize]).collect());
node.outputs[0].ty = ArgType::Tensor(TensorType {
dtype: input_tensor.dtype,
rank: input_tensor.rank,
static_shape: permuted_static_shape,
});
Ok(())
}
fn is_noop(&self, node: &RawNode) -> bool {
// Transpose is a no-op when perm is the identity permutation [0, 1, 2, ...]
let rank = match &node.inputs[0].ty {
ArgType::Tensor(t) => t.rank,
_ => return false,
};
let perm: Vec<i64> = if let Some(axes) = node.attrs.get("perm") {
axes.clone().into_i64s()
} else {
// Default perm reverses dimensions, which is only identity for rank 0 or 1
return rank <= 1;
};
perm.iter().enumerate().all(|(i, &p)| p == i as i64)
}
fn extract_config(&self, node: &RawNode, _opset: usize) -> Result<Self::Config, ProcessError> {
// Extract the shape of the input tensor
let tensor = match &node.inputs.first().unwrap().ty {
ArgType::Tensor(tensor) => tensor.clone(),
_ => {
return Err(ProcessError::TypeMismatch {
expected: "Tensor".to_string(),
actual: format!("{:?}", node.inputs.first().unwrap().ty),
});
}
};
// Default: reverse the dimensions
let mut perm = (0..tensor.rank as i64).rev().collect::<Vec<i64>>();
if let Some(axes) = node.attrs.get("perm") {
perm = axes.clone().into_i64s();
// TODO: Validate perm values are in valid range [0, rank-1].
// Out-of-bounds values in perm should be rejected early.
}
let config = TransposeConfig { perm };
Ok(config)
}
fn build_node(&self, builder: RawNode, opset: usize) -> Node {
let config = self
.extract_config(&builder, opset)
.expect("Config extraction failed");
Node::Transpose(TransposeNode {
name: builder.name,
inputs: builder.inputs,
outputs: builder.outputs,
config,
})
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::ir::NodeType;
use crate::node::test_utils::TestNodeBuilder;
fn create_test_node(perm: Option<Vec<i64>>, rank: usize) -> RawNode {
let mut builder = TestNodeBuilder::new(NodeType::Transpose, "test_transpose")
.input_tensor_f32("data", rank, None)
.output_tensor_f32("transposed", rank, None);
if let Some(perm_val) = perm {
builder = builder.attr_ints("perm", perm_val);
}
builder.build()
}
#[test]
fn test_transpose_config_default() {
let node = create_test_node(None, 3);
let mut node = node;
let processor = TransposeProcessor;
let prefs = OutputPreferences::new();
let config = processor.extract_config(&node, 16).unwrap();
processor.infer_types(&mut node, 16, &prefs).unwrap();
assert_eq!(config.perm, vec![2, 1, 0]); // Default is to reverse the dimensions
}
#[test]
fn test_transpose_config_with_perm() {
let node = create_test_node(Some(vec![0, 2, 1]), 3);
let mut node = node;
let processor = TransposeProcessor;
let prefs = OutputPreferences::new();
let config = processor.extract_config(&node, 16).unwrap();
processor.infer_types(&mut node, 16, &prefs).unwrap();
assert_eq!(config.perm, vec![0, 2, 1]);
}
#[test]
fn test_transpose_config_multiple_inputs() {
let mut node = create_test_node(None, 3);
// Add an extra input to cause the expected error
node.inputs.push(Argument {
name: "extra".to_string(),
ty: crate::ir::ArgType::Tensor(crate::ir::TensorType {
dtype: crate::ir::DType::F32,
rank: 3,
static_shape: None,
}),
value_source: crate::ir::ValueSource::Dynamic,
value_store: None,
});
let processor = TransposeProcessor;
let spec = processor.spec();
let result = crate::processor::validate_node_spec(&node, 16, &spec);
assert!(matches!(
result,
Err(ProcessError::InvalidInputCount {
expected: 1,
actual: 2
})
));
}
#[test]
fn test_transpose_identity_perm_is_noop() {
let node = create_test_node(Some(vec![0, 1, 2]), 3);
let processor = TransposeProcessor;
assert!(processor.is_noop(&node));
}
#[test]
fn test_transpose_non_identity_perm_is_not_noop() {
let node = create_test_node(Some(vec![0, 2, 1]), 3);
let processor = TransposeProcessor;
assert!(!processor.is_noop(&node));
}
#[test]
fn test_transpose_default_perm_rank1_is_noop() {
// Default perm for rank 1 is [0] which is identity
let node = create_test_node(None, 1);
let processor = TransposeProcessor;
assert!(processor.is_noop(&node));
}
#[test]
fn test_transpose_default_perm_rank3_is_not_noop() {
// Default perm for rank 3 is [2, 1, 0] which is NOT identity
let node = create_test_node(None, 3);
let processor = TransposeProcessor;
assert!(!processor.is_noop(&node));
}
// TODO: Missing test for invalid permutations - duplicate indices.
// E.g., perm=[0, 0, 2] should be rejected.
// TODO: Missing test for out-of-bounds indices in perm.
// E.g., perm=[0, 5, 2] for rank-3 tensor should be rejected.
// TODO: Missing test for perm length mismatch.
// E.g., perm=[0, 1] for rank-3 tensor should be rejected.
// TODO: Missing test for 1D tensor transpose - perm=[0] should be identity.
// TODO: Missing test for 5D+ tensors - verify works for higher ranks.
// TODO: Missing test for negative values in perm.
// ONNX spec doesn't support negative indices in perm, should be rejected.
}