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
|
#pragma once
#include <torch/csrc/inductor/aoti_runtime/arrayref_tensor.h>
namespace torch::aot_inductor {
template <typename T>
struct ThreadLocalCachedOutputTensor;
template <>
struct ThreadLocalCachedOutputTensor<RAIIAtenTensorHandle> {
explicit ThreadLocalCachedOutputTensor(const RAIIAtenTensorHandle&) {}
void copy_data_from(const RAIIAtenTensorHandle& handle) {
throw std::runtime_error("can't happen");
}
AtenTensorHandle tensor() const {
throw std::runtime_error("can't happen");
}
};
template <>
struct ThreadLocalCachedOutputTensor<AtenTensorHandle> {
explicit ThreadLocalCachedOutputTensor(const AtenTensorHandle&) {}
void copy_data_from(const AtenTensorHandle& handle) {
throw std::runtime_error("can't happen");
}
AtenTensorHandle tensor() const {
throw std::runtime_error("can't happen");
}
};
template <>
struct ThreadLocalCachedOutputTensor<ConstantHandle> {
explicit ThreadLocalCachedOutputTensor(const ConstantHandle&) {}
void copy_data_from(const ConstantHandle& handle) {
throw std::runtime_error("can't happen");
}
AtenTensorHandle tensor() const {
throw std::runtime_error("can't happen");
}
};
template <typename T>
struct ThreadLocalCachedOutputTensor<ArrayRefTensor<T>> {
explicit ThreadLocalCachedOutputTensor(const ArrayRefTensor<T>& t) {
realloc(t);
}
void copy_data_from(const ArrayRefTensor<T>& t) {
if (t.numel() > capacity_) {
realloc(t);
}
std::copy(t.data(), t.data() + t.numel(), storage_.get());
}
AtenTensorHandle tensor() const {
return tensor_.get();
}
private:
void realloc(const ArrayRefTensor<T>& t) {
capacity_ = t.numel();
// NOLINTNEXTLINE(*arrays*)
storage_ = std::make_unique<T[]>(t.numel());
AtenTensorHandle handle = nullptr;
AOTI_TORCH_ERROR_CODE_CHECK(aoti_torch_create_tensor_from_blob(
storage_.get(),
t.sizes().size(),
t.sizes().data(),
t.strides().data(),
0,
aoti_torch_dtype<std::remove_const_t<T>>(),
t.device_type(),
t.device_idx(),
&handle));
tensor_ = handle;
}
// NOLINTNEXTLINE(*arrays*)
std::unique_ptr<T[]> storage_;
int64_t capacity_ = 0;
RAIIAtenTensorHandle tensor_;
};
template <typename T>
struct ThreadLocalCachedOutputArray;
// Just needs to compile, doesn't need to do anything.
template <>
struct ThreadLocalCachedOutputArray<RAIIAtenTensorHandle> {
explicit ThreadLocalCachedOutputArray(const RAIIAtenTensorHandle&) {
throw std::runtime_error("can't happen");
}
// Not supported yet! We would need to put contiguous() or
// expect_contiguous() into the ABI.
void copy_data_from(const RAIIAtenTensorHandle&) {
throw std::runtime_error("can't happen");
}
template <typename U>
ArrayRefTensor<U> arrayref_tensor() const {
throw std::runtime_error("can't happen");
}
};
// Just needs to compile, doesn't need to do anything.
template <>
struct ThreadLocalCachedOutputArray<ConstantHandle> {
explicit ThreadLocalCachedOutputArray(const ConstantHandle&) {
throw std::runtime_error("can't happen");
}
// Not supported yet! We would need to put contiguous() or
// expect_contiguous() into the ABI.
void copy_data_from(const ConstantHandle&) {
throw std::runtime_error("can't happen");
}
template <typename U>
ArrayRefTensor<U> arrayref_tensor() const {
throw std::runtime_error("can't happen");
}
};
template <typename T>
struct ThreadLocalCachedOutputArray<ArrayRefTensor<T>> {
explicit ThreadLocalCachedOutputArray(const ArrayRefTensor<T>& t) {}
template <
typename U,
std::enable_if_t<
std::is_same_v<std::remove_const_t<T>, std::remove_const_t<U>>,
bool> = true>
ArrayRefTensor<T> arrayref_tensor() const {
return tensor_;
}
void copy_data_from(const ArrayRefTensor<T>& t) {
if (t.numel() > capacity_) {
capacity_ = t.numel();
// NOLINTNEXTLINE(*arrays*)
storage_ = std::make_unique<T[]>(capacity_);
}
std::copy(t.data(), t.data() + t.numel(), storage_.get());
tensor_ = t;
tensor_.set_arrayref(MiniArrayRef<T>(storage_.get(), t.numel()));
}
private:
// NOLINTNEXTLINE(*arrays*)
std::unique_ptr<T[]> storage_;
uint32_t capacity_ = 0;
ArrayRefTensor<T> tensor_;
};
} // namespace torch::aot_inductor
|
