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
|
/*
* Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
* Copyright 2017 The TensorFlow Authors. 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 __NNFW_CKER_FULLY_CONNECTED_H__
#define __NNFW_CKER_FULLY_CONNECTED_H__
#include "cker/Shape.h"
#include "cker/Utils.h"
namespace nnfw
{
namespace cker
{
struct FullyConnectedParams
{
// uint8 inference params.
// TODO(b/65838351): Use smaller types if appropriate.
int32_t input_offset;
int32_t weights_offset;
int32_t output_offset;
int32_t output_multiplier;
int output_shift;
// uint8, etc, activation params.
int32_t quantized_activation_min;
int32_t quantized_activation_max;
// float activation params.
float float_activation_min;
float float_activation_max;
// FullyConnectedWeightsFormat weights_format;
};
inline void FullyConnected(const FullyConnectedParams ¶ms, const Shape &input_shape,
const float *input_data, const Shape &weights_shape,
const float *weights_data, const Shape &bias_shape,
const float *bias_data, const Shape &output_shape, float *output_data)
{
UNUSED_RELEASE(input_shape);
UNUSED_RELEASE(bias_shape);
const float output_activation_min = params.float_activation_min;
const float output_activation_max = params.float_activation_max;
// TODO(benoitjacob): This really should be:
// const int batches = ArraySize(output_dims, 1);
// but the current --variable_batch hack consists in overwriting the 3rd
// dimension with the runtime batch size, as we don't keep track for each
// array of which dimension is the batch dimension in it.
const int output_dims_count = output_shape.DimensionsCount();
const int weights_dims_count = weights_shape.DimensionsCount();
const int batches = FlatSizeSkipDim(output_shape, output_dims_count - 1);
const int output_depth =
MatchingDim(weights_shape, weights_dims_count - 2, output_shape, output_dims_count - 1);
const int accum_depth = weights_shape.Dims(weights_dims_count - 1);
for (int b = 0; b < batches; ++b)
{
for (int out_c = 0; out_c < output_depth; ++out_c)
{
float total = 0.f;
for (int d = 0; d < accum_depth; ++d)
{
total += input_data[b * accum_depth + d] * weights_data[out_c * accum_depth + d];
}
float bias_value = 0.0f;
if (bias_data)
{
bias_value = bias_data[out_c];
}
output_data[out_c + output_depth * b] = ActivationFunctionWithMinMax(
total + bias_value, output_activation_min, output_activation_max);
}
}
}
inline void FullyConnected(const FullyConnectedParams ¶ms, const Shape &input_shape,
const uint8_t *input_data, const Shape &filter_shape,
const uint8_t *filter_data, const Shape &bias_shape,
const int32_t *bias_data, const Shape &output_shape,
uint8_t *output_data)
{
UNUSED_RELEASE(input_shape);
UNUSED_RELEASE(bias_shape);
const int32_t input_offset = params.input_offset;
const int32_t filter_offset = params.weights_offset;
const int32_t output_offset = params.output_offset;
const int32_t output_multiplier = params.output_multiplier;
const int output_shift = params.output_shift;
const int32_t output_activation_min = params.quantized_activation_min;
const int32_t output_activation_max = params.quantized_activation_max;
assert(filter_shape.DimensionsCount() >= 2);
assert(output_shape.DimensionsCount() >= 1);
assert(output_activation_min <= output_activation_max);
// TODO(benoitjacob): This really should be:
// const int batches = ArraySize(output_dims, 1);
// but the current --variable_batch hack consists in overwriting the 3rd
// dimension with the runtime batch size, as we don't keep track for each
// array of which dimension is the batch dimension in it.
const int output_dim_count = output_shape.DimensionsCount();
const int filter_dim_count = filter_shape.DimensionsCount();
const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1);
const int output_depth =
MatchingDim(filter_shape, filter_dim_count - 2, output_shape, output_dim_count - 1);
const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
for (int b = 0; b < batches; ++b)
{
for (int out_c = 0; out_c < output_depth; ++out_c)
{
int32_t acc = 0;
for (int d = 0; d < accum_depth; ++d)
{
int32_t input_val = input_data[b * accum_depth + d];
int32_t filter_val = filter_data[out_c * accum_depth + d];
acc += (filter_val + filter_offset) * (input_val + input_offset);
}
if (bias_data)
{
acc += bias_data[out_c];
}
acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
acc += output_offset;
acc = std::max(acc, output_activation_min);
acc = std::min(acc, output_activation_max);
output_data[out_c + output_depth * b] = static_cast<uint8_t>(acc);
}
}
}
} // namespace cker
} // namespace nnfw
#endif // __NNFW_CKER_FULLY_CONNECTED_H__
|
