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
//! # SpaceToDepth
//!
//! Rearranges blocks of spatial data into depth. This operation moves values from the
//! height and width dimensions into the depth/channel dimension. It is the reverse
//! transformation of DepthToSpace.
//!
//! More specifically, this operator outputs a copy of the input tensor where values from
//! the height and width dimensions are moved to the depth dimension. The spatial dimensions
//! are reduced by a factor of `blocksize`, while the depth is increased by `blocksize^2`.
//!
//! **ONNX Spec**: <https://onnx.ai/onnx/operators/onnx__SpaceToDepth.html>
//!
//! ## Opset Versions
//! - **Opset 1-12**: Initial version with blocksize attribute
//! - **Opset 13+**: Extended type support (added bfloat16, uint types)
use derive_new::new;
use onnx_ir_derive::NodeBuilder;
use crate::processor::{
InputSpec, NodeProcessor, NodeSpec, OutputPreferences, OutputSpec, ProcessError,
};
use crate::ir::{ArgType, Argument, Node, RawNode, TensorType};
/// Configuration for SpaceToDepth operations
#[derive(Debug, Clone, new)]
pub struct SpaceToDepthConfig {
/// Block size for space-to-depth transformation
pub block_size: usize,
}
/// Node representation for SpaceToDepth operation
#[derive(Debug, Clone, NodeBuilder)]
pub struct SpaceToDepthNode {
pub name: String,
pub inputs: Vec<Argument>,
pub outputs: Vec<Argument>,
pub config: SpaceToDepthConfig,
}
pub(crate) struct SpaceToDepthProcessor;
impl NodeProcessor for SpaceToDepthProcessor {
type Config = SpaceToDepthConfig;
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> {
// Validate unexpected attributes before config extraction
for (key, _value) in node.attrs.iter() {
match key.as_str() {
"blocksize" => {}
_ => {
return Err(ProcessError::InvalidAttribute {
name: key.clone(),
reason: format!("Unexpected attribute for SpaceToDepth: {}", key),
});
}
}
}
// Get reference to config for type inference
let config = self
.extract_config(node, opset)
.expect("Config extraction failed");
let block_size = config.block_size;
// Validate block_size
if block_size == 0 {
return Err(ProcessError::InvalidAttribute {
name: "blocksize".to_string(),
reason: "block_size must be greater than 0".to_string(),
});
}
// Extract the input tensor type to determine rank and shape
let tensor = match &node.inputs[0].ty {
ArgType::Tensor(tensor) => tensor,
_ => {
return Err(ProcessError::TypeMismatch {
expected: "Tensor".to_string(),
actual: format!("{:?}", node.inputs[0].ty),
});
}
};
// TODO: Missing validation that input is rank 4 with NCHW format.
// ONNX spec requires input to be 4D [N, C, H, W] but only rank is checked, not semantics.
if tensor.rank != 4 {
return Err(ProcessError::Custom(
"SpaceToDepth: only rank 4 tensors are supported".to_string(),
));
}
// TODO: Missing validation that H and W are divisible by blocksize.
// ONNX spec requires H % blocksize == 0 and W % blocksize == 0, but this isn't validated.
// Should check when static_shape is available to catch errors early.
// Infer static shape based on rank and block size
let static_shape = tensor.static_shape.clone().map(|shape| {
let [b, c, h, w] = shape
.try_into()
.expect("SpaceToDepth: input tensor rank is not 4");
vec![
b,
c.map(|v| v * block_size * block_size),
h.map(|v| v / block_size),
w.map(|v| v / block_size),
]
});
node.outputs[0].ty = ArgType::Tensor(TensorType {
dtype: tensor.dtype,
rank: tensor.rank,
static_shape,
});
Ok(())
}
fn extract_config(&self, node: &RawNode, _opset: usize) -> Result<Self::Config, ProcessError> {
let mut block_size: Option<usize> = None;
for (key, value) in node.attrs.iter() {
if key.as_str() == "blocksize" {
block_size = Some(value.clone().into_i64() as usize)
}
}
let block_size =
block_size.ok_or_else(|| ProcessError::MissingAttribute("blocksize".to_string()))?;
let config = SpaceToDepthConfig { block_size };
Ok(config)
}
fn build_node(&self, builder: RawNode, opset: usize) -> Node {
let config = self
.extract_config(&builder, opset)
.expect("Config extraction failed");
Node::SpaceToDepth(SpaceToDepthNode {
name: builder.name,
inputs: builder.inputs,
outputs: builder.outputs,
config,
})
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::ir::DType;
use crate::ir::NodeType;
use crate::node::test_utils::TestNodeBuilder;
/// Helper function to create test nodes with different repeat values
fn create_test_node(rank: usize, static_shape: Option<Vec<usize>>, block_size: i64) -> RawNode {
let builder = TestNodeBuilder::new(NodeType::DepthToSpace, "test_space_to_depth")
.input_tensor_f32("input", rank, static_shape)
.output_tensor_f32("output", rank, None) // Same rank as input
.attr_int("blocksize", block_size);
builder.build()
}
#[test]
fn test_basic_config() {
let node = create_test_node(4, None, 2);
let mut node = node;
let processor = SpaceToDepthProcessor;
let prefs = OutputPreferences::new();
let config = processor.extract_config(&node, 16).unwrap();
processor.infer_types(&mut node, 16, &prefs).unwrap();
assert_eq!(config.block_size, 2);
}
#[test]
fn test_static_shape_update_outputs() {
let mut node = create_test_node(4, Some(vec![2, 1, 4, 6]), 2);
let processor = SpaceToDepthProcessor;
let prefs = OutputPreferences::new();
let _config = processor.extract_config(&node, 16).unwrap();
processor.infer_types(&mut node, 16, &prefs).unwrap();
match &node.outputs[0].ty {
ArgType::Tensor(tensor) => {
assert_eq!(
tensor.static_shape,
Some(vec![Some(2), Some(4), Some(2), Some(3)])
);
assert_eq!(tensor.dtype, DType::F32);
assert_eq!(tensor.rank, 4);
}
_ => panic!("Expected tensor output"),
}
}
// TODO: Missing test for blocksize validation - blocksize must be > 0.
// Currently validated but not explicitly tested.
// TODO: Missing test for non-divisible dimensions - H or W not divisible by blocksize.
// E.g., input [1, 1, 5, 6], blocksize=2 should fail (5 % 2 != 0).
// TODO: Missing test for blocksize=1 edge case - should be no-op transformation.
// TODO: Missing test for large blocksize - e.g., blocksize > H or blocksize > W.
// Should be rejected as dimensions would become negative.
// TODO: Missing test for different data types - verify works with int8, float16, etc.
// Implementation should support all types per ONNX spec (opset 13+).
}