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
|
#include <torch/csrc/jit/runtime/static/ops.h>
#include <ATen/NativeFunctions.h>
#include <torch/csrc/jit/ir/ir.h>
namespace torch {
namespace jit {
bool canRunOutOfPlace(Node* n) {
auto str = std::string(n->kind().toQualString());
if ((str == "aten::add") || (str == "aten::mul") || (str == "aten::addmm") ||
(str == "aten::bmm") || (str == "aten::sigmoid") ||
(str == "aten::cat") || (str == "aten::transpose") ||
(str == "aten::flatten")) {
return true;
}
return false;
}
std::function<void(StaticRuntime::ConstantMap&)> getOutOfPlaceOperation(
Node* n) {
auto create_empty_from = [](const at::Tensor& t) {
return at::empty({0}, t.options());
};
if (n->kind() == c10::Symbol::fromQualString("aten::add")) {
auto out = n->outputs().at(0);
auto in0 = n->inputs().at(0);
auto in1 = n->inputs().at(1);
auto in2 = n->inputs().at(2);
return [=](StaticRuntime::ConstantMap& ws) {
auto in0_t = ws.at(in0).toTensor();
auto in1_t = ws.at(in1).toTensor();
auto in2_s = ws.at(in2).toScalar();
if (!ws.count(out)) {
ws.emplace(out, create_empty_from(in0_t));
}
auto out_t = ws.at(out).toTensor();
at::native::add_out(out_t, in0_t, in1_t, in2_s);
};
} else if (n->kind() == c10::Symbol::fromQualString("aten::mul")) {
auto out = n->outputs().at(0);
auto in0 = n->inputs().at(0);
auto in1 = n->inputs().at(1);
return [=](StaticRuntime::ConstantMap& ws) {
auto in0_t = ws.at(in0).toTensor();
auto in1_t = ws.at(in1).toTensor();
if (!ws.count(out)) {
ws.emplace(out, create_empty_from(in0_t));
}
auto out_t = ws.at(out).toTensor();
at::native::mul_out(out_t, in0_t, in1_t);
};
} else if (n->kind() == c10::Symbol::fromQualString("aten::addmm")) {
auto out = n->outputs().at(0);
auto in0 = n->inputs().at(0);
auto in1 = n->inputs().at(1);
auto in2 = n->inputs().at(2);
auto in3 = n->inputs().at(3);
auto in4 = n->inputs().at(4);
return [=](StaticRuntime::ConstantMap& ws) {
auto in0_t = ws.at(in0).toTensor();
auto in1_t = ws.at(in1).toTensor();
auto in2_t = ws.at(in2).toTensor();
auto in3_s = ws.at(in3).toScalar();
auto in4_s = ws.at(in3).toScalar();
if (!ws.count(out)) {
ws.emplace(out, create_empty_from(in0_t));
}
auto out_t = ws.at(out).toTensor();
at::native::addmm_cpu_out(out_t, in0_t, in1_t, in2_t, in3_s, in4_s);
};
} else if (n->kind() == c10::Symbol::fromQualString("aten::clamp")) {
auto out = n->outputs().at(0);
auto in0 = n->inputs().at(0);
auto in1 = n->inputs().at(1);
auto in2 = n->inputs().at(2);
return [=](StaticRuntime::ConstantMap& ws) {
auto in0_t = ws.at(in0).toTensor();
auto in1_s = ws.at(in1).toScalar();
auto in2_s = ws.at(in2).toScalar();
if (!ws.count(out)) {
ws.emplace(out, create_empty_from(in0_t));
}
auto out_t = ws.at(out).toTensor();
at::native::clamp_out(out_t, in0_t, in1_s, in2_s);
};
} else if (n->kind() == c10::Symbol::fromQualString("aten::bmm")) {
auto out = n->outputs().at(0);
auto in0 = n->inputs().at(0);
auto in1 = n->inputs().at(1);
return [=](StaticRuntime::ConstantMap& ws) {
auto in0_t = ws.at(in0).toTensor();
auto in1_t = ws.at(in1).toTensor();
if (!ws.count(out)) {
ws.emplace(out, create_empty_from(in0_t));
}
auto out_t = ws.at(out).toTensor();
at::native::bmm_out_cpu(out_t, in0_t, in1_t);
};
} else if (n->kind() == c10::Symbol::fromQualString("aten::cat")) {
auto out = n->outputs().at(0);
auto in0 = n->inputs().at(0);
auto in1 = n->inputs().at(1);
return [=](StaticRuntime::ConstantMap& ws) {
auto in0_tl = ws.at(in0).toTensorVector();
auto in1_i = ws.at(in1).toInt();
if (!ws.count(out)) {
ws.emplace(out, create_empty_from(in0_tl[0]));
}
auto out_t = ws.at(out).toTensor();
at::native::_cat_out_cpu(out_t, in0_tl, in1_i);
};
} else if (n->kind() == c10::Symbol::fromQualString("aten::sigmoid")) {
auto out = n->outputs().at(0);
auto in0 = n->inputs().at(0);
return [=](StaticRuntime::ConstantMap& ws) {
auto in0_t = ws.at(in0).toTensor();
if (!ws.count(out)) {
ws.emplace(out, create_empty_from(in0_t));
}
auto out_t = ws.at(out).toTensor();
at::native::sigmoid_out(out_t, in0_t);
};
} else if (n->kind() == c10::Symbol::fromQualString("aten::transpose")) {
auto out = n->outputs().at(0);
auto in0 = n->inputs().at(0);
auto in1 = n->inputs().at(1);
auto in2 = n->inputs().at(2);
return [=](StaticRuntime::ConstantMap& ws) {
auto in0_t = ws.at(in0).toTensor();
auto in1_i = ws.at(in1).toInt();
auto in2_i = ws.at(in2).toInt();
ws[out] = at::native::transpose(in0_t, in1_i, in2_i);
};
} else if (n->kind() == c10::Symbol::fromQualString("aten::flatten")) {
auto out = n->outputs().at(0);
auto in0 = n->inputs().at(0);
auto in1 = n->inputs().at(1);
auto in2 = n->inputs().at(2);
return [=](StaticRuntime::ConstantMap& ws) {
auto in0_t = ws.at(in0).toTensor();
auto in1_i = ws.at(in1).toInt();
auto in2_i = ws.at(in2).toInt();
ws[out] = at::native::flatten(in0_t, in1_i, in2_i);
};
}
return [](StaticRuntime::ConstantMap&) { TORCH_CHECK(0); };
}
} // namespace jit
} // namespace torch
|
