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
|
/*
* 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/Types.h"
#include "cker/Utils.h"
#include "cker/TensorUtils.h"
namespace nnfw
{
namespace cker
{
class FCTempArena
{
public:
FCTempArena(void) : prepared(false), input_quantized(), scaling_factors()
{
// DO NOTHING
}
void prepare(const Shape &input_shape, const Shape &weights_shape)
{
auto input_size = input_shape.FlatSize();
input_quantized.resize(input_size);
assert(weights_shape.DimensionsCount() == 2);
int batch_size = input_size / weights_shape.Dims(1);
scaling_factors.resize(batch_size);
prepared = true;
}
public:
bool prepared;
std::vector<int8_t> input_quantized;
std::vector<float> scaling_factors;
};
inline void FullyConnected(const FullyConnectedParams ¶ms, const Shape &input_shape,
const float *input_data, const Shape &weights_shape,
const float *weights_data, const Shape &, const float *bias_data,
const Shape &, float *output_data)
{
int total_input_size = input_shape.FlatSize();
int input_size = weights_shape.Dims(1);
const int batch_size = total_input_size / input_size;
const int num_units = weights_shape.Dims(0);
// Output = bias if bias tensor exists.
if (bias_data)
{
VectorBatchVectorAssign(bias_data, num_units, batch_size, output_data);
}
else
{
ZeroVector(output_data, batch_size * num_units);
}
// Compute output += weight * input
MatrixBatchVectorMultiplyAccumulate(weights_data, num_units, input_size, input_data, batch_size,
output_data, /*result_stride=*/1);
// Apply activation function
ApplyActivationToVector(output_data, batch_size * num_units, params.activation, output_data);
}
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);
}
}
}
inline void FullyConnectedHybrid(const FullyConnectedParams ¶ms, const Shape &input_shape,
const float *input_data, const Shape &filter_shape,
const int8_t *filter_data, const Shape &, const float *bias_data,
const Shape &, float *output_data, FCTempArena &temp_arena)
{
int total_input_size = input_shape.FlatSize();
const int input_size = filter_shape.Dims(1);
const int batch_size = total_input_size / input_size;
const int num_units = filter_shape.Dims(0);
// Output = bias if bias tensor exists.
VectorBatchVectorAssign(bias_data, num_units, batch_size, output_data);
// Save matrix multiplication computation for all zero input.
if (IsZeroVector(input_data, total_input_size))
{
ApplyActivationToVector(output_data, batch_size * num_units, params.activation, output_data);
return;
}
// Quantize input from float to uint8 + quantization params (scaling factor).
float unused_min, unused_max;
float *scaling_factors_ptr = temp_arena.scaling_factors.data();
int8_t *quant_data = temp_arena.input_quantized.data();
// Quantize each batch independently.
for (int b = 0; b < batch_size; ++b)
{
const int offset = b * input_size;
SymmetricQuantizeFloats(input_data + offset, input_size, quant_data + offset, &unused_min,
&unused_max, &scaling_factors_ptr[b]);
// Incorporate scaling of the filter.
scaling_factors_ptr[b] *= params.weights_scale;
}
// Compute output += weight * quantized_input
MatrixBatchVectorMultiplyAccumulate(filter_data, num_units, input_size, quant_data,
scaling_factors_ptr, batch_size, output_data,
/*result_stride=*/1);
// Apply activation function to floats.
ApplyActivationToVector(output_data, batch_size * num_units, params.activation, output_data);
return;
}
} // namespace cker
} // namespace nnfw
#endif // __NNFW_CKER_FULLY_CONNECTED_H__
|
