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
|
/*
* 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.
*/
#include "PermuteLayer.h"
#include "util/feature/nhwc/Reader.h"
#include "util/feature/nhwc/View.h"
#include "util/feature/nchw/View.h"
#include "util/feature/Coordinate4D.h"
#include <misc/feature/IndexIterator.h>
namespace neurun
{
namespace kernel
{
namespace cpu
{
using Type = model::operation::PermuteNode::Type;
void PermuteLayer::configure(std::shared_ptr<::neurun::backend::operand::IObject> input,
std::shared_ptr<::neurun::backend::operand::IObject> output,
const model::operand::Shape &shape, Type type)
{
_input = input;
_output = output;
_shape = shape;
_type = type;
}
void PermuteLayer::run()
{
auto rank = _shape.rank();
switch (_type)
{
case Type::NHWC_TO_NCHW:
{
auto fn = [&](::neurun::backend::operand::ITensor &tensor) {
auto input_tensor = _input->ptr();
auto input_buffer = input_tensor->buffer();
auto input_size = input_tensor->total_size();
auto output_buffer = tensor.buffer();
auto output_size = tensor.total_size();
switch (rank)
{
case 0:
case 1:
{
memcpy(output_buffer, input_buffer, input_size);
break;
}
case 2:
{
auto matrix_shape = _shape.asMatrix();
for (auto h = 0; h < matrix_shape.H; ++h)
{
neurun::util::feature::Coordinate4D coord{0, h, 0, 0};
memcpy(output_buffer + tensor.calcOffset(coord), input_buffer + h * matrix_shape.W,
matrix_shape.W * sizeof(float));
}
break;
}
case 3:
{
const int32_t depth = _shape.dim(0);
const int32_t height = _shape.dim(1);
const int32_t width = _shape.dim(2);
for (auto c = 0; c < depth; ++c)
{
for (auto h = 0; h < height; ++h)
{
neurun::util::feature::Coordinate4D coord{0, h, 0, c};
memcpy(output_buffer + tensor.calcOffset(coord),
input_buffer + c * height * width + h * width, width * sizeof(float));
}
}
break;
}
case 4:
{
auto feature = _shape.asFeature();
const util::feature::nhwc::Reader<float> from{
feature, reinterpret_cast<const float *>(input_buffer), input_size};
util::feature::nchw::View<float> into{&tensor};
::nnfw::misc::feature::iterate(feature)
<< [&](uint32_t batch, uint32_t ch, uint32_t row, uint32_t col) {
const auto value = from.at(batch, ch, row, col);
into.at(batch, ch, row, col) = value;
};
break;
}
default:
throw "NYI";
break;
}
};
_output->access(fn);
break;
}
case Type::NCHW_TO_NHWC:
{
auto fn = [&](::neurun::backend::operand::ITensor &tensor) {
auto input_buffer = tensor.buffer();
auto input_size = tensor.total_size();
auto output_tensor = _output->ptr();
auto output_buffer = output_tensor->buffer();
auto output_size = output_tensor->total_size();
switch (rank)
{
case 0:
case 1:
{
memcpy(output_buffer, input_buffer, output_size);
break;
}
case 2:
{
auto matrix_shape = _shape.asMatrix();
for (auto h = 0; h < matrix_shape.H; ++h)
{
neurun::util::feature::Coordinate4D coord{0, h, 0, 0};
memcpy(output_buffer + h * matrix_shape.W, input_buffer + tensor.calcOffset(coord),
matrix_shape.W * sizeof(float));
}
break;
}
case 3:
{
const int32_t depth = _shape.dim(0);
const int32_t height = _shape.dim(1);
const int32_t width = _shape.dim(2);
for (auto c = 0; c < depth; ++c)
{
for (auto h = 0; h < height; ++h)
{
neurun::util::feature::Coordinate4D coord{0, h, 0, c};
memcpy(output_buffer + c * height * width + h * width,
input_buffer + tensor.calcOffset(coord), width * sizeof(float));
}
}
break;
}
case 4:
{
auto feature = _shape.asFeature();
const util::feature::nchw::View<float> from{&tensor};
util::feature::nhwc::View<float> into{feature, reinterpret_cast<float *>(output_buffer),
output_size};
::nnfw::misc::feature::iterate(feature)
<< [&](uint32_t batch, uint32_t ch, uint32_t row, uint32_t col) {
const auto value = from.at(batch, ch, row, col);
into.at(batch, ch, row, col) = value;
};
break;
}
default:
throw "NYI";
break;
}
};
_input->access(fn);
break;
}
case Type::COPY:
// If two different backends using same tensor layout, we need this.
throw "NYI";
break;
}
}
} // namespace cpu
} // namespace kernel
} // namespace neurun
|
