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
|
#ifndef CAFFE2_OPERATORS_NUMPY_TILE_OP_H_
#define CAFFE2_OPERATORS_NUMPY_TILE_OP_H_
#include "caffe2/core/common_omp.h"
#include "caffe2/core/context.h"
#include "caffe2/core/logging.h"
#include "caffe2/core/operator.h"
#include "caffe2/utils/math.h"
#include "c10/util/irange.h"
namespace caffe2 {
// Copy a Blob n times along a specified axis.
template <class Context>
class NumpyTileOp : public Operator<Context> {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
template <class... Args>
explicit NumpyTileOp(Args&&... args)
: Operator<Context>(std::forward<Args>(args)...) {}
~NumpyTileOp() {}
bool RunOnDevice() override {
const auto& input = Input(0);
const auto& repeats = Input(1);
// Check that the `repeats` tensor has the correct rank, has a number of
// elements equal to the number of axes of `input`.
CAFFE_ENFORCE_EQ(repeats.dim(), 1, "repeats input must be a 1-d tensor");
CAFFE_ENFORCE_EQ(
repeats.numel(),
input.dim(),
"repeats input have the same"
" number of elements as `inputs` has dimensions.");
const int64_t* repeats_data = repeats.template data<int64_t>();
// NOLINTNEXTLINE(clang-diagnostic-sign-compare)
for (const auto i : c10::irange(repeats.numel())) {
CAFFE_ENFORCE_GE(repeats_data[i], 0);
}
auto* output = Output(0);
// Alternate inputs and outputs between two buffers. Repeatedly apply the
// Tile kernel along each axis. Then copy out the resulting data into the
// output tensor.
Tensor *src = &buffer, *dst = output;
src->CopyFrom(input);
vector<int64_t> output_dims(input.sizes().vec());
for (const auto i : c10::irange(repeats.numel())) {
if (repeats_data[i] == 1) {
continue;
}
// size up to (and not including) axis
const auto outer_dim = src->size_to_dim(i);
// size from axis up
const auto inner_dim = src->size_from_dim(i);
dst->Resize(outer_dim, inner_dim * repeats_data[i]);
/**
* How this works:
* Imagine a 2D tensor (matrix) of size 3x10, tiled 2 times.
* - Tiling along axis 0 (row) means copying the entire 3x10 Matrix 2
* times. outer_dim = 0, inner_dim = 30.
* - Tiling along axis 1 (column) means copying each row 2 times, then
* proceed to the next row, until the end. outer_dim = 3, inner_dim = 10.
*/
const char* src_data = static_cast<const char*>(src->raw_data());
char* dst_data = static_cast<char*>(dst->raw_mutable_data(src->dtype()));
DoTile(
src->dtype(),
src->itemsize(),
outer_dim,
inner_dim,
repeats_data[i],
src_data,
dst_data);
output_dims[i] *= repeats_data[i];
dst->Reshape(output_dims);
std::swap(src, dst);
}
// NB: because we have the swap at the end of the above loop, our real
// result tensor is going to live in *src when we reach this line
// whether we entered the loop or not :)
if (output != src)
output->CopyFrom(*src);
return true;
}
private:
void DoTile(
const TypeMeta meta,
int item_size,
int outer_dim,
int inner_dim,
int64_t num_tiles,
const char* input_data,
char* output_data) {
for (const auto i : c10::irange(outer_dim)) {
(void)i; // Suppress unused variable warning
for (const auto t : c10::irange(num_tiles)) {
(void)t; // Suppress unused variable warning
context_.CopyItemsSameDevice(meta, inner_dim, input_data, output_data);
output_data += inner_dim * item_size;
}
input_data += inner_dim * item_size;
}
}
Tensor buffer{Context::GetDeviceType()};
};
} // namespace caffe2
#endif // CAFFE2_OPERATORS_NUMPY_TILE_OP_H_
|
