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
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
|
/*
* Copyright (c) 2017 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 "arm_compute/core/NEON/kernels/NENormalizationLayerKernel.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/NEON/NEFixedPoint.h"
#include "arm_compute/core/NEON/NEMath.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/core/Window.h"
using namespace arm_compute;
NENormalizationLayerKernel::NENormalizationLayerKernel()
: _func(nullptr), _input(nullptr), _input_squared(nullptr), _output(nullptr), _norm_info(NormType::IN_MAP_1D), _border_size()
{
}
BorderSize NENormalizationLayerKernel::border_size() const
{
return _border_size;
}
void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor *input_squared, ITensor *output, NormalizationLayerInfo norm_info)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::QS8);
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32, DataType::QS8);
ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, input_squared);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, input_squared, output);
ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, input_squared, output);
ARM_COMPUTE_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd");
ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.beta(), input);
ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.kappa(), input);
ARM_COMPUTE_ERROR_ON_VALUE_NOT_REPRESENTABLE_IN_FIXED_POINT(norm_info.scale_coeff(), input);
const unsigned int border_width = (norm_info.type() == NormType::CROSS_MAP) ? 0 : std::min(norm_info.norm_size() / 2, 3U);
_input = input;
_input_squared = input_squared;
_output = output;
_norm_info = norm_info;
_border_size = BorderSize(0, border_width);
const bool is_dt_f32 = _input->info()->data_type() == DataType::F32;
switch(norm_info.type())
{
case NormType::IN_MAP_1D:
_func = (is_dt_f32) ? &NENormalizationLayerKernel::normalize<0, false> : &NENormalizationLayerKernel::normalize_fixed_point<0, false>;
break;
case NormType::IN_MAP_2D:
// Normalize over X and Y
_func = (is_dt_f32) ? &NENormalizationLayerKernel::normalize<0, true> : &NENormalizationLayerKernel::normalize_fixed_point<0, true>;
break;
case NormType::CROSS_MAP:
_func = (is_dt_f32) ? &NENormalizationLayerKernel::normalize<2, false> : &NENormalizationLayerKernel::normalize_fixed_point<2, false>;
break;
default:
ARM_COMPUTE_ERROR("NOT SUPPORTED!");
}
const unsigned int num_elems_processed_per_iteration = (is_dt_f32) ? 4 : 16;
const unsigned int num_elems_read_per_iteration = num_elems_processed_per_iteration + 2 * (norm_info.norm_size() / 2);
const unsigned int num_rows = (norm_info.type() == NormType::IN_MAP_2D) ? norm_info.norm_size() : 1;
// Configure window
Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
AccessWindowRectangle input_access(input->info(), -_border_size.left, 0, num_elems_read_per_iteration, num_rows);
AccessWindowRectangle input_squared_access(input_squared->info(), -_border_size.left, 0, num_elems_read_per_iteration, num_rows);
AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
update_window_and_padding(win, input_access, input_squared_access, output_access);
output_access.set_valid_region(win, input->info()->valid_region());
INEKernel::configure(win);
}
template <unsigned int dim, bool do_2D_norm>
void NENormalizationLayerKernel::normalize(const Window &window)
{
Iterator input(_input, window);
Iterator input_squared(_input_squared, window);
Iterator output(_output, window);
const int dim_y = 1;
const int radius = _norm_info.norm_size() / 2;
const int total_size = _input->info()->dimension(dim) - 1;
const int input_squared_stride = _input_squared->info()->strides_in_bytes()[dim];
// We account padding across X only and we iterate over rows
const int min_left = (dim == 2) ? 0 : -static_cast<int>(border_size().left);
const int max_right = (dim == 2) ? total_size : total_size + border_size().left;
const int min_top = 0;
const int max_bottom = _input->info()->dimension(dim_y) - 1;
const float32x4_t coeff_vec = vdupq_n_f32(_norm_info.scale_coeff());
const float32x4_t beta_vec = vdupq_n_f32(_norm_info.beta());
const float32x4_t kappa_vec = vdupq_n_f32(_norm_info.kappa());
execute_window_loop(window, [&](const Coordinates & id)
{
// Get range to normalize
const int current_row = do_2D_norm ? id[dim_y] : 0;
const int current_slice = id[dim];
const int first_row = do_2D_norm ? std::max(current_row - radius, min_top) : 0;
const int last_row = do_2D_norm ? std::min(current_row + radius, max_bottom) : 0;
const int first_slice = std::max(current_slice - radius, min_left);
const int last_slice = std::min(current_slice + radius, max_right);
// Accumulate 2D In-Map values
float32x4_t accu = vdupq_n_f32(0.f);
for(int j = first_row; j <= last_row; j++)
{
// Compute row displacement
const int row = (j - current_row) * _input_squared->info()->strides_in_bytes()[dim_y];
const uint8_t *const input_squared_ptr = input_squared.ptr() + row - (current_slice * input_squared_stride);
for(int i = first_slice; i <= last_slice; ++i)
{
accu = vaddq_f32(accu, vld1q_f32(reinterpret_cast<const float *>(input_squared_ptr + i * input_squared_stride)));
}
}
// Normalize
const float32x4_t normalized = vpowq_f32(vmlaq_f32(kappa_vec, coeff_vec, accu), beta_vec);
const float32x4_t normalized_pixel = vmulq_f32(vld1q_f32(reinterpret_cast<const float *>(input.ptr())), vinvq_f32(normalized));
vst1q_f32(reinterpret_cast<float *>(output.ptr()), normalized_pixel);
},
input, input_squared, output);
}
template <unsigned int dim, bool do_2D_norm>
void NENormalizationLayerKernel::normalize_fixed_point(const Window &window)
{
Iterator input(_input, window);
Iterator input_squared(_input_squared, window);
Iterator output(_output, window);
const int dim_y = 1;
const int radius = _norm_info.norm_size() / 2;
const int total_size = _input->info()->dimension(dim) - 1;
const int input_squared_stride = _input_squared->info()->strides_in_bytes()[dim];
// We account padding across X only and we iterate over rows
const int min_left = (dim == 2) ? 0 : -static_cast<int>(border_size().left);
const int max_right = (dim == 2) ? total_size : total_size + border_size().left;
const int min_top = 0;
const int max_bottom = _input->info()->dimension(dim_y) - 1;
const int fixed_point_position = _input->info()->fixed_point_position();
const qint8x16_t coeff_vec = vdupq_n_qs8_f32(_norm_info.scale_coeff(), fixed_point_position);
const qint8x16_t beta_vec = vdupq_n_qs8_f32(_norm_info.beta(), fixed_point_position);
const qint8x16_t kappa_vec = vdupq_n_qs8_f32(_norm_info.kappa(), fixed_point_position);
execute_window_loop(window, [&](const Coordinates & id)
{
// Get range to normalize
const int current_row = do_2D_norm ? id[dim_y] : 0;
const int current_slice = id[dim];
const int first_row = do_2D_norm ? std::max(current_row - radius, min_top) : 0;
const int last_row = do_2D_norm ? std::min(current_row + radius, max_bottom) : 0;
const int first_slice = std::max(current_slice - radius, min_left);
const int last_slice = std::min(current_slice + radius, max_right);
// Accumulate 2D In-Map values
qint8x16_t accu = vdupq_n_qs8(0);
for(int j = first_row; j <= last_row; ++j)
{
// Compute row displacement
const int row = (j - current_row) * _input_squared->info()->strides_in_bytes()[dim_y];
const uint8_t *const input_squared_ptr = input_squared.ptr() + row - (current_slice * input_squared_stride);
for(int i = first_slice; i <= last_slice; ++i)
{
accu = vqaddq_qs8(accu, vld1q_qs8(reinterpret_cast<const qint8_t *>(input_squared_ptr + i * input_squared_stride)));
}
}
// Normalize
const qint8x16_t accu_scale = vqmlaq_qs8(kappa_vec, coeff_vec, accu, fixed_point_position);
const qint8x16_t normalized = vqpowq_qs8(accu_scale, beta_vec, fixed_point_position);
const qint8x16_t normalized_pixel = vdivq_qs8(vld1q_qs8(reinterpret_cast<const qint8_t *>(input.ptr())), normalized, fixed_point_position);
vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), normalized_pixel);
},
input, input_squared, output);
}
void NENormalizationLayerKernel::run(const Window &window)
{
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
ARM_COMPUTE_ERROR_ON(_func == nullptr);
// Run function
(this->*_func)(window);
}
|
