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
|
/*
* Copyright (c) 2019 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 "AddLayer.h"
#include "OperationUtils.h"
#include "ncnn/layer/binaryop.h"
#include "cpp14/memory.h"
namespace
{
std::unique_ptr<nnfw::ncnn::Mat>
convertMatIgnoreLayout(neurun::backend::srcn::kernel::TensorDescriptor &desc, void *data)
{
if (desc.dimensions.size() == 1)
{
return nnfw::cpp14::make_unique<nnfw::ncnn::Mat>(desc.dimensions[0], data);
}
else if (desc.dimensions.size() == 2)
{
return nnfw::cpp14::make_unique<nnfw::ncnn::Mat>(desc.dimensions[1], desc.dimensions[0], data);
}
else if (desc.dimensions.size() == 3)
{
return nnfw::cpp14::make_unique<nnfw::ncnn::Mat>(desc.dimensions[2], desc.dimensions[1],
desc.dimensions[0], data);
}
else // rank == 4 and N == 1
{
return nnfw::cpp14::make_unique<nnfw::ncnn::Mat>(desc.dimensions[3], desc.dimensions[2],
desc.dimensions[1], data);
}
}
} // namespace
namespace neurun
{
namespace backend
{
namespace srcn
{
namespace kernel
{
void AddLayer::addFloat32()
{
assert(_activation == ir::Activation::NONE);
// ncnn kernel support
// 1. rank < 4
// 2. broadcasting
// 2-1 lhs, rhs have same rank, or
// 2-2 model layout and backend layout is same
// For safety, block all broadcasting (enable when ready)
assert(_lhsDescr.dimensions.size() < 4 ||
(_lhsDescr.dimensions.size() == 4 && _lhsDescr.dimensions[0] == 1));
assert(_rhsDescr.dimensions.size() < 4 ||
(_rhsDescr.dimensions.size() == 4 && _rhsDescr.dimensions[0] == 1));
assert((_lhsDescr.dimensions.size() == _rhsDescr.dimensions.size()));
nnfw::ncnn::BinaryOpParam param;
param.op_type = nnfw::ncnn::BinaryOp::Operation_ADD;
auto lhs_mat = convertMatIgnoreLayout(_lhsDescr, _lhsData.v);
auto rhs_mat = convertMatIgnoreLayout(_rhsDescr, _rhsData.v);
auto out_mat = convertMatIgnoreLayout(_outputDescr, _outputData.v);
::nnfw::ncnn::ncnn_binary_op(param, *lhs_mat.get(), *rhs_mat.get(), *out_mat.get());
}
void AddLayer::addQuant8()
{
// quant8 add is not implemented yet
throw std::runtime_error{"NYI"};
}
void AddLayer::configure(uint8_t *lhsData, const TensorDescriptor &lhsDescr, uint8_t *rhsData,
const TensorDescriptor &rhsDescr, const ir::Activation activation,
uint8_t *outputData, const TensorDescriptor &outputDescr,
const ir::Layout backendLayout)
{
_lhsData.u8 = lhsData;
_lhsDescr = lhsDescr;
_rhsData.u8 = rhsData;
_rhsDescr = rhsDescr;
_inputType = lhsDescr.type;
_activation = activation;
_outputData.u8 = outputData;
_outputDescr = outputDescr;
_backendLayout = backendLayout;
}
void AddLayer::run()
{
if (_inputType == OperandType::FLOAT32)
{
addFloat32();
}
else if (_inputType == OperandType::QUANT8_ASYMM)
{
addQuant8();
}
}
} // namespace kernel
} // namespace srcn
} // namespace backend
} // namespace neurun
|
