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
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
|
/*
* Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#ifndef __NEURUN_BACKEND_CPU_KERNEL_PERMUTE_LAYER_H__
#define __NEURUN_BACKEND_CPU_KERNEL_PERMUTE_LAYER_H__
#include <exec/IFunction.h>
#include "util/feature/nhwc/View.h"
#include "OperationUtils.h"
#include <backend/operand/IObject.h>
#include "model/operation/PermuteNode.h"
#include "util/feature/nhwc/Reader.h"
#include "util/feature/nchw/View.h"
#include "util/Coordinates.h"
#include <misc/feature/IndexIterator.h>
namespace neurun
{
namespace backend
{
namespace cpu
{
namespace kernel
{
class PermuteLayer : public ::neurun::exec::IFunction
{
public:
PermuteLayer() = default;
public:
void configure(std::shared_ptr<::neurun::backend::operand::IObject> input,
std::shared_ptr<::neurun::backend::operand::IObject> output,
const model::Shape &output_shape, model::operation::PermuteNode::Type type,
model::DataType dataType);
void run();
void runSync()
{
// this abstract method is used just for profiling and called for
// backend::acl_common::AclFunction
run();
}
private:
template <class T> void runTempl()
{
auto rank = _output_shape.rank();
auto fn = [&](::neurun::backend::operand::ITensor &in_tensor) {
_output->access([&](::neurun::backend::operand::ITensor &out_tensor) {
auto input_buffer = in_tensor.buffer();
auto input_size = in_tensor.total_size();
auto output_buffer = out_tensor.buffer();
if (_type == model::operation::PermuteNode::Type::COPY)
{
assert(in_tensor.layout() == out_tensor.layout());
if (!in_tensor.has_padding() && !out_tensor.has_padding())
{
assert(input_size == out_tensor.total_size());
memcpy(output_buffer, input_buffer, input_size);
return;
}
}
switch (rank)
{
case 0:
case 1:
{
const int32_t copy_len = _output_shape.dim(0);
memcpy(output_buffer, input_buffer, copy_len);
break;
}
case 2:
{
const int32_t copy_len = _output_shape.dim(1);
for (auto i = 0; i < _output_shape.dim(0); ++i)
{
neurun::util::Coordinates coords{i, 0};
memcpy(output_buffer + out_tensor.calcOffset(coords),
input_buffer + in_tensor.calcOffset(coords), copy_len * sizeof(T));
}
break;
}
case 3:
{
const int32_t copy_len = _output_shape.dim(2);
for (auto i = 0; i < _output_shape.dim(0); ++i)
{
for (auto j = 0; j < _output_shape.dim(1); ++j)
{
neurun::util::Coordinates coords{i, j, 0};
memcpy(output_buffer + out_tensor.calcOffset(coords),
input_buffer + in_tensor.calcOffset(coords), copy_len * sizeof(T));
}
}
break;
}
case 4:
{
// TODO Unify permute type and remove switch case
switch (_type)
{
case model::operation::PermuteNode::Type::NHWC_TO_NCHW:
{
for (auto n = 0; n < _output_shape.dim(0); ++n)
{
for (auto c = 0; c < _output_shape.dim(1); ++c)
{
for (auto h = 0; h < _output_shape.dim(2); ++h)
{
for (auto w = 0; w < _output_shape.dim(3); ++w)
{
const neurun::util::Coordinates in_coords{n, h, w, c};
const auto out_coords =
convertCoordinates(in_coords, in_tensor.layout(), out_tensor.layout());
const auto value =
*reinterpret_cast<T *>(input_buffer + in_tensor.calcOffset(in_coords));
*reinterpret_cast<T *>(output_buffer + out_tensor.calcOffset(out_coords)) =
value;
}
}
}
}
break;
}
case model::operation::PermuteNode::Type::NCHW_TO_NHWC:
{
for (auto n = 0; n < _output_shape.dim(0); ++n)
{
for (auto h = 0; h < _output_shape.dim(1); ++h)
{
for (auto w = 0; w < _output_shape.dim(2); ++w)
{
for (auto c = 0; c < _output_shape.dim(3); ++c)
{
const neurun::util::Coordinates in_coords{n, c, h, w};
const auto out_coords =
convertCoordinates(in_coords, in_tensor.layout(), out_tensor.layout());
const auto value =
*reinterpret_cast<T *>(input_buffer + in_tensor.calcOffset(in_coords));
*reinterpret_cast<T *>(output_buffer + out_tensor.calcOffset(out_coords)) =
value;
}
}
}
}
break;
}
case model::operation::PermuteNode::Type::COPY:
{
const int32_t copy_len = _output_shape.dim(3);
for (auto i = 0; i < _output_shape.dim(0); ++i)
{
for (auto j = 0; j < _output_shape.dim(1); ++j)
{
for (auto k = 0; k < _output_shape.dim(2); ++k)
{
neurun::util::Coordinates coords{i, j, k, 0};
memcpy(output_buffer + out_tensor.calcOffset(coords),
input_buffer + in_tensor.calcOffset(coords), copy_len * sizeof(T));
}
}
}
break;
}
default:
throw std::runtime_error("NYI");
break;
}
break;
}
default:
throw std::runtime_error("NYI");
break;
}
});
};
_input->access(fn);
}
private:
std::shared_ptr<::neurun::backend::operand::IObject> _input{nullptr};
std::shared_ptr<::neurun::backend::operand::IObject> _output{nullptr};
model::Shape _output_shape{};
model::operation::PermuteNode::Type _type{model::operation::PermuteNode::Type::COPY};
model::DataType _dataType{model::DataType::FLOAT32};
};
} // namespace kernel
} // namespace cpu
} // namespace backend
} // namespace neurun
#endif // __NEURUN_BACKEND_CPU_KERNEL_PERMUTE_LAYER_H__
|
