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
|
#include <torch/csrc/jit/passes/onnx/preprocess_for_onnx.h>
#include <torch/csrc/jit/jit_log.h>
#include <torch/csrc/jit/passes/onnx/helper.h>
namespace torch {
namespace jit {
namespace onnx {
using namespace ::c10::onnx;
}
namespace {
at::optional<Node*> FindFusibleListUnpack(Node* n) {
// 1. number of outputs is restricted to 1.
// 2. output is only used by prim::ListUnpack.
if (n->outputs().size() != 1) {
return at::nullopt;
}
if (n->output()->uses().size() != 1) {
return at::nullopt;
}
auto listUnpackNode = n->output()->uses()[0].user;
if (listUnpackNode->kind() != prim::ListUnpack) {
return at::nullopt;
}
return listUnpackNode;
}
// Fuse node + ListUnpack
// Node such as split/unbind produces tensor[] of static size,
// that is later unpacked by ListUnpack.
// This pass fuses the two nodes, and adds an additional input "_outputs" such
// that the symbolic function is aware of the number of outputs.
//
// Example IR
// split.Tensor(Tensor(a) self, int split_size, int dim=0) -> Tensor(a)[]
// split_with_sizes(Tensor self, int[] split_sizes, int dim=0) -> Tensor[]
//
// graph(%input : Float(5:12, 4:3, 3:1)):
// %13 : int[] = prim::Constant[value=[2, 1, 2]]()
// %7 : int = prim::Constant[value=0]()
// %8 : Tensor[] = aten::split_with_sizes(%input, %13, %7)
// %9 : Float(2:12, 4:3, 3:1), %10 : Float(1:12, 4:3, 3:1), %11 : Float(2:12,
// 4:3, 3:1) = prim::ListUnpack(%8) return (%9, %10, %11)
//
// After fusion
// graph(%input : Float(5:12, 4:3, 3:1)):
// %13 : int[] = prim::Constant[value=[2, 1, 2]]()
// %7 : int = prim::Constant[value=0]()
// %8 : int = prim::Constant[value=3]() # Adding addtional input of value 3
// representing the number of outputs.
// %14 : Float(2:12, 4:3, 3:1), %15 : Float(1:12, 4:3, 3:1), %16 : Float(2:12,
// 4:3, 3:1) = aten::split_with_sizes(%input, %13, %7, %8)
// return (%14, %15, %16)
void FuseWithListUnpack(Node* n) {
auto found_listUnpack = FindFusibleListUnpack(n);
if (!found_listUnpack) {
return;
}
auto listUnpack_node = found_listUnpack.value();
TORCH_INTERNAL_ASSERT(n->outputs().size() == 1);
// 1. Add internal input "_outputs" to node, so that later symbolic function
// conversion
// is aware of the number of outputs.
// 2. Add the exact number of outputs to n, copy metadata and replace uses of
// listUnpack outputs.
n->i_(
Symbol::fromQualString("attr::_outputs"),
static_cast<int64_t>(listUnpack_node->outputs().size()));
for (auto i = 0; i < listUnpack_node->outputs().size(); ++i) {
auto new_output = n->addOutput();
new_output->copyMetadata(listUnpack_node->output(i));
}
listUnpack_node->removeAllInputs();
// remove original output, which is input to listUnpack node.
n->eraseOutput(0);
listUnpack_node->replaceAllUsesWith(n);
}
static void FuseWithListUnpack(Block* b) {
for (auto it = b->nodes().begin(), end = b->nodes().end(); it != end; ++it) {
for (auto* child_block : it->blocks()) {
FuseWithListUnpack(child_block);
}
auto n_kind = it->kind();
switch (n_kind) {
case aten::split:
case aten::split_with_sizes:
case aten::unsafe_split:
case aten::unsafe_split_with_sizes:
case aten::unbind:
case aten::unsafe_chunk:
case aten::where:
FuseWithListUnpack(*it);
break;
default:
break;
}
}
}
// Replace aten::add with onnx::Concat
// when inputs to the add node are two int lists
//
// before the pass:
// graph(%x.1 : Float(2:12, 3:4, 4:1, requires_grad=0, device=cpu),
// %y.1 : Float(1:6, 2:3, 3:1, requires_grad=0, device=cpu)):
// %2 : None = prim::Constant()
// %3 : int[] = aten::size(%x.1)
// %l1.1 : int[] = aten::list(%3
// %5 : int[] = aten::size(%y.1)
// %l2.1 : int[] = aten::list(%5)
// %7 : int[] = aten::add(%l1.1, %l2.1)
// %8 : Tensor = aten::new_zeros(%x.1, %7, %2, %2, %2, %2)
// return (%8)
//
// after the pass:
// graph(%x.1 : Float(2:12, 3:4, 4:1, requires_grad=0, device=cpu),
// %y.1 : Float(1:6, 2:3, 3:1, requires_grad=0, device=cpu)):
// %2 : None = prim::Constant()
// %3 : int[] = aten::size(%x.1)
// %l1.1 : int[] = aten::list(%3)
// %5 : int[] = aten::size(%y.1)
// %l2.1 : int[] = aten::list(%5)
// %9 : Tensor = onnx::Concat[axis=0](%l1.1, %l2.1)
// %8 : Tensor = aten::new_zeros(%x.1, %9, %2, %2, %2, %2)
// return (%8)
static void ReplaceAddWithConcat(Block* b) {
for (auto it = b->nodes().begin(), end = b->nodes().end(); it != end; ++it) {
for (auto* child_block : it->blocks()) {
ReplaceAddWithConcat(child_block);
}
if (it->kind() == aten::add) {
if (!it->input(0)->type()->cast<ListType>() ||
!it->input(1)->type()->cast<ListType>()) {
continue;
}
TypePtr elem = it->input(0)->type()->cast<ListType>()->getElementType();
if (elem->cast<IntType>()) {
Node* concat_node = b->owningGraph()->create(onnx::Concat, 1);
concat_node->i_(attr::axis, 0);
concat_node->insertBefore(*it);
concat_node->addInput(it->input(0));
concat_node->addInput(it->input(1));
concat_node->outputs()[0]->setType(
TensorType::fromNumberType(std::move(elem)));
it->replaceAllUsesWith(concat_node);
it->removeAllInputs();
it.destroyCurrent();
}
}
}
}
} // namespace
void PreprocessForONNX(std::shared_ptr<Graph>& graph) {
FuseWithListUnpack(graph->block());
ReplaceAddWithConcat(graph->block());
}
} // namespace jit
} // namespace torch
|
