forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
tensor_cuda.cpp
134 lines (104 loc) · 5.49 KB
/
tensor_cuda.cpp
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
#include <gtest/gtest.h>
#include <ATen/ATen.h>
#include <cmath>
#define REQUIRE_TENSOR_OPTIONS(device_, index_, type_, layout_) \
ASSERT_TRUE( \
tensor.device().type() == at::Device((device_), (index_)).type()); \
ASSERT_TRUE( \
tensor.device().index() == at::Device((device_), (index_)).index()); \
ASSERT_EQ(tensor.dtype(), (type_)); \
ASSERT_TRUE(tensor.layout() == (layout_))
// NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
TEST(TensorTest, AllocatesTensorOnTheCorrectDevice_MultiCUDA) {
auto tensor = at::tensor({1, 2, 3}, at::device({at::kCUDA, 1}));
ASSERT_EQ(tensor.device().type(), at::Device::Type::CUDA);
ASSERT_EQ(tensor.device().index(), 1);
}
// NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
TEST(TensorTest, ToDevice_MultiCUDA) {
auto tensor = at::empty({3, 4});
REQUIRE_TENSOR_OPTIONS(at::kCPU, -1, at::kFloat, at::kStrided);
tensor = tensor.to({at::kCUDA, 1});
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 1, at::kFloat, at::kStrided);
tensor = tensor.to({at::kCUDA, 0});
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 0, at::kFloat, at::kStrided);
tensor = tensor.to({at::kCUDA, 1});
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 1, at::kFloat, at::kStrided);
tensor = tensor.to(at::Device(at::kCPU));
REQUIRE_TENSOR_OPTIONS(at::kCPU, -1, at::kFloat, at::kStrided);
tensor = tensor.to(at::kCUDA);
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 0, at::kFloat, at::kStrided);
tensor = tensor.to(at::TensorOptions({at::kCUDA, 1}));
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 1, at::kFloat, at::kStrided);
tensor = tensor.to(at::TensorOptions({at::kCUDA, 0}));
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 0, at::kFloat, at::kStrided);
tensor = tensor.to(at::TensorOptions(at::kDouble));
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 0, at::kDouble, at::kStrided);
tensor = tensor.to(at::TensorOptions({at::kCUDA, 1}));
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 1, at::kDouble, at::kStrided);
tensor = tensor.to(at::TensorOptions(at::kInt));
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 1, at::kInt, at::kStrided);
tensor = tensor.to(at::TensorOptions(at::Device(at::kCPU)));
REQUIRE_TENSOR_OPTIONS(at::kCPU, -1, at::kInt, at::kStrided);
tensor = tensor.to(at::TensorOptions(at::kCUDA));
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 0, at::kInt, at::kStrided);
}
// NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
TEST(TensorTest, ToTensorAndTensorAttributes_MultiCUDA) {
auto tensor = at::empty({3, 4});
REQUIRE_TENSOR_OPTIONS(at::kCPU, -1, at::kFloat, at::kStrided);
auto other = at::empty({3, 4}, at::kFloat);
tensor = tensor.to(other);
REQUIRE_TENSOR_OPTIONS(at::kCPU, -1, at::kFloat, at::kStrided);
other = at::empty({3, 4}, at::TensorOptions(at::kCUDA).dtype(at::kDouble));
tensor = tensor.to(other.dtype());
REQUIRE_TENSOR_OPTIONS(at::kCPU, -1, at::kDouble, at::kStrided);
tensor = tensor.to(other.device());
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 0, at::kDouble, at::kStrided);
other = at::empty({3, 4}, at::TensorOptions({at::kCUDA, 1}).dtype(at::kLong));
tensor = tensor.to(other.device(), other.dtype());
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 1, at::kLong, at::kStrided);
other = at::empty({3, 4}, at::kFloat);
tensor = tensor.to(other.options());
REQUIRE_TENSOR_OPTIONS(at::kCPU, -1, at::kFloat, at::kStrided);
}
// NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
TEST(TensorTest, ToDoesNotCopyWhenOptionsAreAllTheSame_CUDA) {
auto tensor = at::empty({3, 4}, at::TensorOptions(at::kFloat).device(at::Device("cuda")));
auto hopefully_not_copy = tensor.to(tensor.options());
ASSERT_EQ(hopefully_not_copy.data_ptr<float>(), tensor.data_ptr<float>());
hopefully_not_copy = tensor.to(at::kFloat);
ASSERT_EQ(hopefully_not_copy.data_ptr<float>(), tensor.data_ptr<float>());
hopefully_not_copy = tensor.to("cuda");
ASSERT_EQ(hopefully_not_copy.data_ptr<float>(), tensor.data_ptr<float>());
hopefully_not_copy = tensor.to(at::TensorOptions("cuda"));
ASSERT_EQ(hopefully_not_copy.data_ptr<float>(), tensor.data_ptr<float>());
hopefully_not_copy = tensor.to(at::TensorOptions(at::kFloat));
ASSERT_EQ(hopefully_not_copy.data_ptr<float>(), tensor.data_ptr<float>());
}
// NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
TEST(TensorTest, ToDeviceAndDtype_MultiCUDA) {
auto tensor = at::empty({3, 4});
REQUIRE_TENSOR_OPTIONS(at::kCPU, -1, at::kFloat, at::kStrided);
tensor = tensor.to({at::kCUDA, 1}, at::kInt);
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 1, at::kInt, at::kStrided);
tensor = tensor.to(at::TensorOptions({at::kCUDA, 0}).dtype(at::kLong));
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 0, at::kLong, at::kStrided);
tensor = tensor.to(at::TensorOptions({at::kCUDA, 1}).dtype(at::kDouble));
REQUIRE_TENSOR_OPTIONS(at::kCUDA, 1, at::kDouble, at::kStrided);
tensor = tensor.to(at::kCPU, at::kInt);
REQUIRE_TENSOR_OPTIONS(at::kCPU, -1, at::kInt, at::kStrided);
}
// NOLINTNEXTLINE(cppcoreguidelines-avoid-non-const-global-variables)
TEST(TensorTest, MagmaInitializesCorrectly_CUDA) {
// Any tensor will work here as long as it's invertible
// NOLINTNEXTLINE(cppcoreguidelines-avoid-c-arrays,modernize-avoid-c-arrays)
float data[] = { 1, 1, 1, 0,
0, 3, 1, 2,
2, 3, 1, 0,
1, 0, 2, 1 };
auto tensor = at::from_blob(data, {4, 4}, at::TensorOptions(at::kFloat)).cuda();
if (at::hasMAGMA()) {
at::inverse(tensor);
}
}