Moritz Pflanzer | 6db73ce | 2017-07-19 10:18:42 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "NormalizationLayer.h" |
| 25 | |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 26 | #include "arm_compute/core/Types.h" |
Moritz Pflanzer | a09de0c | 2017-09-01 20:41:12 +0100 | [diff] [blame] | 27 | #include "tests/validation/FixedPoint.h" |
Moritz Pflanzer | 6db73ce | 2017-07-19 10:18:42 +0100 | [diff] [blame] | 28 | |
| 29 | namespace arm_compute |
| 30 | { |
| 31 | namespace test |
| 32 | { |
| 33 | namespace validation |
| 34 | { |
| 35 | namespace reference |
| 36 | { |
| 37 | template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type> |
| 38 | SimpleTensor<T> normalization_layer(const SimpleTensor<T> &src, NormalizationLayerInfo info) |
| 39 | { |
| 40 | // Create reference |
| 41 | SimpleTensor<T> dst{ src.shape(), src.data_type(), 1, src.fixed_point_position() }; |
| 42 | |
| 43 | // Compute reference |
| 44 | const uint32_t norm_size = info.norm_size(); |
| 45 | NormType type = info.type(); |
| 46 | float beta = info.beta(); |
| 47 | uint32_t kappa = info.kappa(); |
| 48 | |
| 49 | const int cols = src.shape()[0]; |
| 50 | const int rows = src.shape()[1]; |
| 51 | const int depth = src.shape()[2]; |
| 52 | int upper_dims = src.shape().total_size() / (cols * rows); |
| 53 | |
| 54 | float coeff = info.scale_coeff(); |
| 55 | int radius_cols = norm_size / 2; |
| 56 | |
| 57 | // IN_MAP_1D and CROSS_MAP normalize over a single axis only |
| 58 | int radius_rows = (NormType::IN_MAP_2D == type) ? norm_size / 2 : 0; |
| 59 | |
| 60 | if(type == NormType::CROSS_MAP) |
| 61 | { |
| 62 | // Remove also depth from upper dimensions since it is the dimension we |
| 63 | // want to use for normalization |
| 64 | upper_dims /= depth; |
| 65 | |
| 66 | for(int r = 0; r < upper_dims; ++r) |
| 67 | { |
| 68 | for(int i = 0; i < rows; ++i) |
| 69 | { |
| 70 | for(int k = 0; k < cols; ++k) |
| 71 | { |
| 72 | for(int l = 0; l < depth; ++l) |
| 73 | { |
| 74 | float accumulated_scale = 0.f; |
| 75 | |
| 76 | for(int j = -radius_cols; j <= radius_cols; ++j) |
| 77 | { |
| 78 | const int z = l + j; |
| 79 | |
| 80 | if(z >= 0 && z < depth) |
| 81 | { |
| 82 | const T value = src[k + i * cols + z * rows * cols + r * cols * rows * depth]; |
| 83 | accumulated_scale += value * value; |
| 84 | } |
| 85 | } |
| 86 | |
| 87 | dst[k + i * cols + l * rows * cols + r * cols * rows * depth] = kappa + accumulated_scale * coeff; |
| 88 | } |
| 89 | } |
| 90 | } |
| 91 | } |
| 92 | } |
| 93 | else |
| 94 | { |
| 95 | for(int r = 0; r < upper_dims; ++r) |
| 96 | { |
| 97 | for(int i = 0; i < rows; ++i) |
| 98 | { |
| 99 | for(int k = 0; k < cols; ++k) |
| 100 | { |
| 101 | float accumulated_scale = 0.f; |
| 102 | |
| 103 | for(int j = -radius_rows; j <= radius_rows; ++j) |
| 104 | { |
| 105 | const int y = i + j; |
| 106 | for(int l = -radius_cols; l <= radius_cols; ++l) |
| 107 | { |
| 108 | const int x = k + l; |
| 109 | |
| 110 | if((x >= 0 && y >= 0) && (x < cols && y < rows)) |
| 111 | { |
| 112 | const T value = src[x + y * cols + r * cols * rows]; |
| 113 | accumulated_scale += value * value; |
| 114 | } |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | dst[k + i * cols + r * cols * rows] = kappa + accumulated_scale * coeff; |
| 119 | } |
| 120 | } |
| 121 | } |
| 122 | } |
| 123 | |
| 124 | if(beta == 1.f) |
| 125 | { |
| 126 | for(int i = 0; i < dst.num_elements(); ++i) |
| 127 | { |
| 128 | dst[i] = src[i] / dst[i]; |
| 129 | } |
| 130 | } |
| 131 | else if(beta == 0.5f) |
| 132 | { |
| 133 | for(int i = 0; i < dst.num_elements(); ++i) |
| 134 | { |
| 135 | dst[i] = src[i] / std::sqrt(dst[i]); |
| 136 | } |
| 137 | } |
| 138 | else |
| 139 | { |
| 140 | for(int i = 0; i < dst.num_elements(); ++i) |
| 141 | { |
| 142 | dst[i] = src[i] * std::exp(std::log(dst[i]) * -beta); |
| 143 | } |
| 144 | } |
| 145 | |
| 146 | return dst; |
| 147 | } |
| 148 | |
| 149 | template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type> |
| 150 | SimpleTensor<T> normalization_layer(const SimpleTensor<T> &src, NormalizationLayerInfo info) |
| 151 | { |
| 152 | using namespace fixed_point_arithmetic; |
| 153 | |
| 154 | // Create reference |
| 155 | SimpleTensor<T> dst{ src.shape(), src.data_type(), 1, src.fixed_point_position() }; |
| 156 | |
| 157 | // Compute reference |
| 158 | const int fixed_point_position = src.fixed_point_position(); |
| 159 | |
| 160 | const uint32_t norm_size = info.norm_size(); |
| 161 | NormType type = info.type(); |
| 162 | fixed_point<T> beta(info.beta(), fixed_point_position); |
| 163 | fixed_point<T> kappa(info.kappa(), fixed_point_position); |
| 164 | |
| 165 | const int cols = src.shape()[0]; |
| 166 | const int rows = src.shape()[1]; |
| 167 | const int depth = src.shape()[2]; |
| 168 | int upper_dims = src.shape().total_size() / (cols * rows); |
| 169 | |
| 170 | fixed_point<T> coeff(info.scale_coeff(), fixed_point_position); |
| 171 | int radius_cols = norm_size / 2; |
| 172 | |
| 173 | // IN_MAP_1D and CROSS_MAP normalize over a single axis only |
| 174 | int radius_rows = (NormType::IN_MAP_2D == type) ? norm_size / 2 : 0; |
| 175 | |
| 176 | if(type == NormType::CROSS_MAP) |
| 177 | { |
| 178 | // Remove also depth from upper dimensions since it is the dimension we |
| 179 | // want to use for normalization |
| 180 | upper_dims /= depth; |
| 181 | |
| 182 | for(int r = 0; r < upper_dims; ++r) |
| 183 | { |
| 184 | for(int i = 0; i < rows; ++i) |
| 185 | { |
| 186 | for(int k = 0; k < cols; ++k) |
| 187 | { |
| 188 | for(int l = 0; l < depth; ++l) |
| 189 | { |
| 190 | fixed_point<T> accumulated_scale(0.f, fixed_point_position); |
| 191 | |
| 192 | for(int j = -radius_cols; j <= radius_cols; ++j) |
| 193 | { |
| 194 | const int z = l + j; |
| 195 | |
| 196 | if(z >= 0 && z < depth) |
| 197 | { |
| 198 | const T value = src[k + i * cols + z * rows * cols + r * cols * rows * depth]; |
| 199 | const fixed_point<T> fp_value(value, fixed_point_position, true); |
| 200 | accumulated_scale = add(accumulated_scale, mul(fp_value, fp_value)); |
| 201 | } |
| 202 | } |
| 203 | |
| 204 | accumulated_scale = add(kappa, mul(accumulated_scale, coeff)); |
| 205 | dst[k + i * cols + l * rows * cols + r * cols * rows * depth] = accumulated_scale.raw(); |
| 206 | } |
| 207 | } |
| 208 | } |
| 209 | } |
| 210 | } |
| 211 | else |
| 212 | { |
| 213 | for(int r = 0; r < upper_dims; ++r) |
| 214 | { |
| 215 | for(int i = 0; i < rows; ++i) |
| 216 | { |
| 217 | for(int k = 0; k < cols; ++k) |
| 218 | { |
| 219 | fixed_point<T> accumulated_scale(0.f, fixed_point_position); |
| 220 | |
| 221 | for(int j = -radius_rows; j <= radius_rows; ++j) |
| 222 | { |
| 223 | const int y = i + j; |
| 224 | |
| 225 | for(int l = -radius_cols; l <= radius_cols; ++l) |
| 226 | { |
| 227 | const int x = k + l; |
| 228 | |
| 229 | if((x >= 0 && y >= 0) && (x < cols && y < rows)) |
| 230 | { |
| 231 | const T value = src[x + y * cols + r * cols * rows]; |
| 232 | const fixed_point<T> fp_value(value, fixed_point_position, true); |
| 233 | accumulated_scale = add(accumulated_scale, mul(fp_value, fp_value)); |
| 234 | } |
| 235 | } |
| 236 | } |
| 237 | |
| 238 | accumulated_scale = add(kappa, mul(accumulated_scale, coeff)); |
| 239 | dst[k + i * cols + r * cols * rows] = accumulated_scale.raw(); |
| 240 | } |
| 241 | } |
| 242 | } |
| 243 | } |
| 244 | |
| 245 | if(info.beta() == 1.f) |
| 246 | { |
| 247 | for(int i = 0; i < dst.num_elements(); ++i) |
| 248 | { |
| 249 | fixed_point<T> res = div(fixed_point<T>(src[i], fixed_point_position, true), fixed_point<T>(dst[i], fixed_point_position, true)); |
| 250 | dst[i] = res.raw(); |
| 251 | } |
| 252 | } |
| 253 | else |
| 254 | { |
| 255 | const fixed_point<T> beta(info.beta(), fixed_point_position); |
| 256 | |
| 257 | for(int i = 0; i < dst.num_elements(); ++i) |
| 258 | { |
| 259 | fixed_point<T> res = pow(fixed_point<T>(dst[i], fixed_point_position, true), beta); |
| 260 | res = div(fixed_point<T>(src[i], fixed_point_position, true), res); |
| 261 | dst[i] = res.raw(); |
| 262 | } |
| 263 | } |
| 264 | |
| 265 | return dst; |
| 266 | } |
| 267 | |
| 268 | template SimpleTensor<float> normalization_layer(const SimpleTensor<float> &src, NormalizationLayerInfo info); |
Georgios Pinitas | 583137c | 2017-08-31 18:12:42 +0100 | [diff] [blame] | 269 | template SimpleTensor<half> normalization_layer(const SimpleTensor<half> &src, NormalizationLayerInfo info); |
Moritz Pflanzer | 6db73ce | 2017-07-19 10:18:42 +0100 | [diff] [blame] | 270 | template SimpleTensor<qint8_t> normalization_layer(const SimpleTensor<qint8_t> &src, NormalizationLayerInfo info); |
Michele Di Giorgio | d5e65c7 | 2017-07-26 17:09:17 +0100 | [diff] [blame] | 271 | template SimpleTensor<qint16_t> normalization_layer(const SimpleTensor<qint16_t> &src, NormalizationLayerInfo info); |
Moritz Pflanzer | 6db73ce | 2017-07-19 10:18:42 +0100 | [diff] [blame] | 272 | } // namespace reference |
| 273 | } // namespace validation |
| 274 | } // namespace test |
| 275 | } // namespace arm_compute |