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
|
/*
* Copyright (c) 2017-2020 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "FullyConnectedLayer.h"
#include "arm_compute/core/Types.h"
#include "tests/validation/reference/UtilsQuantizedAsymm.h"
#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
#include <numeric>
namespace arm_compute
{
namespace test
{
namespace validation
{
namespace reference
{
namespace
{
// Vector matrix multiply for floating point
template < typename T, typename TB, typename std::enable_if < is_floating_point<T>::value &&is_floating_point<TB>::value, int >::type = 0 >
void vector_matrix_multiply(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &dst, int offset_src, int offset_dst, int cols_weights,
int rows_weights)
{
const T *src_ptr = src.data() + offset_src;
const T *weights_ptr = weights.data();
const TB *bias_ptr = bias.data();
T *dst_ptr = dst.data() + offset_dst;
#if defined(_OPENMP)
#pragma omp parallel for
#endif /* _OPENMP */
for(int y = 0; y < rows_weights; ++y)
{
dst_ptr[y] = std::inner_product(src_ptr, src_ptr + cols_weights, &weights_ptr[cols_weights * y], static_cast<T>(0)) + bias_ptr[y];
}
}
// Vector matrix multiply for quantized type
template < typename T, typename TB, typename std::enable_if < (std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value) &&std::is_same<TB, int32_t>::value, int >::type = 0 >
void vector_matrix_multiply(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &dst, int offset_src, int offset_dst,
int cols_weights, int rows_weights)
{
const T *src_ptr = src.data() + offset_src;
const T *weights_ptr = weights.data();
const TB *bias_ptr = bias.data();
T *dst_ptr = dst.data() + offset_dst;
const UniformQuantizationInfo iq_info = src.quantization_info().uniform();
const UniformQuantizationInfo wq_info = weights.quantization_info().uniform();
const UniformQuantizationInfo oq_info = dst.quantization_info().uniform();
const int input_offset = -iq_info.offset;
const float input_scale = iq_info.scale;
const int weights_offset = -wq_info.offset;
const float weights_scale = wq_info.scale;
const int output_offset = oq_info.offset;
const float output_scale = oq_info.scale;
int output_multiplier = 0;
int output_shift = 0;
const float multiplier = input_scale * weights_scale / output_scale;
arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
const int min = std::numeric_limits<T>::lowest();
const int max = std::numeric_limits<T>::max();
#if defined(_OPENMP)
#pragma omp parallel for
#endif /* _OPENMP */
for(int y = 0; y < rows_weights; ++y)
{
// Reset accumulator
int32_t acc = 0;
for(int x = 0; x < cols_weights; ++x)
{
acc += (src_ptr[x] + input_offset) * (weights_ptr[x + y * cols_weights] + weights_offset);
}
// Accumulate the bias
acc += bias_ptr[y];
// Quantize down
acc = quantize_down_scale_by_fixedpoint(acc, output_multiplier, output_shift, output_offset, min, max);
// Store the result
dst_ptr[y] = static_cast<T>(acc);
}
}
} // namespace
template <typename T, typename TB>
SimpleTensor<T> fully_connected_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, const TensorShape &dst_shape, QuantizationInfo out_quant_info)
{
// if no explicit quantization has been set you the same as src
if(out_quant_info == QuantizationInfo())
{
out_quant_info = src.quantization_info();
}
// Create reference
SimpleTensor<T> dst{ TensorShape{ dst_shape }, src.data_type(), 1, out_quant_info };
// Sanity checks
const int num_batch_dimensions = std::max(0, static_cast<int>(dst_shape.num_dimensions()) - 1);
const int num_input_dimensions = src.shape().num_dimensions() - num_batch_dimensions;
const unsigned int linear_input_size = src.shape().total_size_lower(num_input_dimensions);
ARM_COMPUTE_UNUSED(num_batch_dimensions);
ARM_COMPUTE_UNUSED(num_input_dimensions);
ARM_COMPUTE_UNUSED(linear_input_size);
ARM_COMPUTE_ERROR_ON(weights.shape().x() != linear_input_size);
ARM_COMPUTE_ERROR_ON(weights.shape().y() != bias.shape().x());
ARM_COMPUTE_ERROR_ON(weights.shape().y() != dst.shape().x());
// Compute reference
const int cols_weights = weights.shape().x();
const int rows_weights = weights.shape().y();
const int num_batches = dst_shape.total_size_upper(1);
for(int k = 0; k < num_batches; ++k)
{
const int offset_in = k * cols_weights;
const int offset_out = k * rows_weights;
vector_matrix_multiply<T>(src,
weights,
bias,
dst,
offset_in,
offset_out,
cols_weights,
rows_weights);
}
return dst;
}
template SimpleTensor<float> fully_connected_layer(const SimpleTensor<float> &src, const SimpleTensor<float> &weights, const SimpleTensor<float> &bias, const TensorShape &dst_shape,
QuantizationInfo out_quant_info);
template SimpleTensor<half> fully_connected_layer(const SimpleTensor<half> &src, const SimpleTensor<half> &weights, const SimpleTensor<half> &bias, const TensorShape &dst_shape,
QuantizationInfo out_quant_info);
template SimpleTensor<uint8_t> fully_connected_layer(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &bias, const TensorShape &dst_shape,
QuantizationInfo out_quant_info);
template SimpleTensor<int8_t> fully_connected_layer(const SimpleTensor<int8_t> &src, const SimpleTensor<int8_t> &weights, const SimpleTensor<int32_t> &bias, const TensorShape &dst_shape,
QuantizationInfo out_quant_info);
} // namespace reference
} // namespace validation
} // namespace test
} // namespace arm_compute
|
