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
|
/*
* 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 "FullyConnectedLayer.h"
#include <cker/operation/FullyConnected.h>
namespace onert
{
namespace backend
{
namespace cpu
{
namespace kernel
{
FullyConnectedLayer::FullyConnectedLayer()
: _input(nullptr), _weights(nullptr), _bias(nullptr), _output(nullptr),
_activation(ir::Activation::NONE), _temp_arena(new nnfw::cker::FCTempArena())
{
// DO NOTHING
}
FullyConnectedLayer::~FullyConnectedLayer() = default;
void FullyConnectedLayer::fullyConnectedFloat32()
{
float output_activation_min, output_activation_max;
CalculateActivationRangeFloat(_activation, &output_activation_min, &output_activation_max);
nnfw::cker::FullyConnectedParams op_params;
op_params.float_activation_min = output_activation_min;
op_params.float_activation_max = output_activation_max;
op_params.activation = convertActivationType(_activation);
nnfw::cker::FullyConnected(
op_params, convertTensorToCkerShape(_input),
reinterpret_cast<const float *>(_input->buffer()), convertTensorToCkerShape(_weights),
reinterpret_cast<const float *>(_weights->buffer()), convertTensorToCkerShape(_bias),
reinterpret_cast<const float *>(_bias->buffer()), convertTensorToCkerShape(_output),
reinterpret_cast<float *>(_output->buffer()));
}
// executionMutex is used to protect concurrent access of non-threadsafe resources
// like gemmlowp::GemmContext.
void FullyConnectedLayer::fullyConnectedQuant8()
{
double real_multiplier = 0.0;
int32_t output_multiplier = 0;
int32_t output_shift = 0;
int32_t output_activation_min = 0;
int32_t output_activation_max = 0;
GetQuantizedConvolutionMultiplier(_input, _weights, _bias, _output, &real_multiplier);
QuantizeMultiplier(real_multiplier, &output_multiplier, &output_shift);
CalculateActivationRangeUint8(_activation, _output, &output_activation_min,
&output_activation_max);
nnfw::cker::FullyConnectedParams op_params;
op_params.input_offset = -_input->offset();
op_params.weights_offset = -_weights->offset();
op_params.output_offset = _output->offset();
op_params.output_multiplier = output_multiplier;
op_params.output_shift = output_shift;
op_params.quantized_activation_min = output_activation_min;
op_params.quantized_activation_max = output_activation_max;
nnfw::cker::FullyConnected(
op_params, convertTensorToCkerShape(_input),
reinterpret_cast<const uint8_t *>(_input->buffer()), convertTensorToCkerShape(_weights),
reinterpret_cast<const uint8_t *>(_weights->buffer()), convertTensorToCkerShape(_bias),
reinterpret_cast<const int32_t *>(_bias->buffer()), convertTensorToCkerShape(_output),
reinterpret_cast<uint8_t *>(_output->buffer()));
}
void FullyConnectedLayer::fullyConnectedHybrid()
{
nnfw::cker::FCTempArena &temp_arena = *_temp_arena;
if (!temp_arena.prepared)
{
temp_arena.prepare(convertTensorToCkerShape(_input), convertTensorToCkerShape(_weights));
}
nnfw::cker::FullyConnectedParams op_params;
op_params.activation = convertActivationType(_activation);
op_params.weights_scale = _weights->scale();
nnfw::cker::FullyConnectedHybrid(
op_params, convertTensorToCkerShape(_input),
reinterpret_cast<const float *>(_input->buffer()), convertTensorToCkerShape(_weights),
reinterpret_cast<const int8_t *>(_weights->buffer()), convertTensorToCkerShape(_bias),
reinterpret_cast<const float *>(_bias->buffer()), convertTensorToCkerShape(_output),
reinterpret_cast<float *>(_output->buffer()), temp_arena);
}
void FullyConnectedLayer::configure(const operand::Tensor *input, const operand::Tensor *weights,
const operand::Tensor *bias, ir::Activation activation,
operand::Tensor *output)
{
_input = input;
_weights = weights;
_bias = bias;
_activation = activation;
_output = output;
}
void FullyConnectedLayer::run()
{
if (_input->data_type() == OperandType::FLOAT32)
{
if (_weights->data_type() == OperandType::QUANT8_SYMM)
{
fullyConnectedHybrid();
}
else
{
fullyConnectedFloat32();
}
}
else if (_input->data_type() == OperandType::QUANT8_ASYMM)
{
fullyConnectedQuant8();
}
}
} // namespace kernel
} // namespace cpu
} // namespace backend
} // namespace onert
|
