Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +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 "arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/Error.h" |
| 28 | #include "arm_compute/core/FixedPoint.h" |
| 29 | #include "arm_compute/core/Helpers.h" |
| 30 | #include "arm_compute/core/ITensor.h" |
| 31 | #include "arm_compute/core/NEON/NEFixedPoint.h" |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 32 | #include "arm_compute/core/NEON/NEMath.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 33 | #include "arm_compute/core/TensorInfo.h" |
| 34 | #include "arm_compute/core/Utils.h" |
| 35 | #include "arm_compute/core/Validate.h" |
| 36 | #include "arm_compute/core/Window.h" |
| 37 | |
| 38 | #include <algorithm> |
| 39 | #include <arm_neon.h> |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 40 | #include <cmath> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 41 | #include <limits> |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 42 | #include <set> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 43 | #include <string> |
| 44 | #include <tuple> |
| 45 | |
| 46 | using namespace arm_compute; |
| 47 | |
| 48 | namespace |
| 49 | { |
| 50 | inline float calculate_avg_scale(const Coordinates &id, const int pool_size, const int upper_bound_w, const int upper_bound_h, |
| 51 | const int pad_x, const int pad_y, const int stride_x, const int stride_y) |
| 52 | { |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 53 | const int start_x = id.x() * stride_x - pad_x; |
| 54 | const int start_y = id.y() * stride_y - pad_y; |
| 55 | const int end_x = std::min(start_x + pool_size, upper_bound_w); |
| 56 | const int end_y = std::min(start_y + pool_size, upper_bound_h); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 57 | return 1.f / ((end_y - start_y) * (end_x - start_x)); |
| 58 | } |
| 59 | |
| 60 | inline qint8_t calculate_avg_scale_q8(const Coordinates &id, int pool_size, int upper_bound_w, int upper_bound_h, |
| 61 | int pad_x, int pad_y, int stride_x, int stride_y, int fixed_point_position) |
| 62 | { |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 63 | static const std::array<qint8_t, 10> scale_values_q8 = |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 64 | { { 0x0, 0x0, 0x40, 0x2A, 0x20, 0x19, 0x15, 0x12, 0x10, 0xE } }; |
| 65 | const int start_x = id.x() * stride_x - pad_x; |
| 66 | const int start_y = id.y() * stride_y - pad_y; |
| 67 | const int end_x = std::min(start_x + pool_size, upper_bound_w); |
| 68 | const int end_y = std::min(start_y + pool_size, upper_bound_h); |
| 69 | const int val = ((end_y - start_y) * (end_x - start_x)); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 70 | return sshr_qs8(scale_values_q8[val], (7 - fixed_point_position)); |
| 71 | } |
| 72 | |
| 73 | inline qint16_t calculate_avg_scale_q16(const Coordinates &id, int pool_size, int upper_bound_w, int upper_bound_h, |
| 74 | int pad_x, int pad_y, int stride_x, int stride_y, int fixed_point_position) |
| 75 | { |
| 76 | static std::array<qint16_t, 10> scale_values_q16 = |
| 77 | { { 0x0, 0x0, 0x4000, 0x2AAB, 0x2000, 0x199A, 0x1555, 0x1249, 0x1000, 0xE38 } }; |
| 78 | const int start_x = id.x() * stride_x - pad_x; |
| 79 | const int start_y = id.y() * stride_y - pad_y; |
| 80 | const int end_x = std::min(start_x + pool_size, upper_bound_w); |
| 81 | const int end_y = std::min(start_y + pool_size, upper_bound_h); |
| 82 | const int val = ((end_y - start_y) * (end_x - start_x)); |
| 83 | return sshr_qs16(scale_values_q16[val], (15 - fixed_point_position)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 84 | } |
| 85 | } // namespace |
| 86 | |
| 87 | NEPoolingLayerKernel::NEPoolingLayerKernel() |
| 88 | : _func(nullptr), _input(nullptr), _output(nullptr), _pool_info(), _num_elems_processed_per_iteration(0), _border_size(0) |
| 89 | { |
| 90 | } |
| 91 | |
| 92 | BorderSize NEPoolingLayerKernel::border_size() const |
| 93 | { |
| 94 | return _border_size; |
| 95 | } |
| 96 | |
| 97 | void NEPoolingLayerKernel::configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info) |
| 98 | { |
Gian Marco Iodice | 4e28869 | 2017-06-27 11:41:59 +0100 | [diff] [blame] | 99 | int pool_pad_x = 0; |
| 100 | int pool_pad_y = 0; |
| 101 | int pool_stride_x = 0; |
| 102 | int pool_stride_y = 0; |
| 103 | unsigned int pooled_w = 0; |
| 104 | unsigned int pooled_h = 0; |
| 105 | PoolingType pool_type = pool_info.pool_type(); |
| 106 | int pool_size = pool_info.pool_size(); |
| 107 | const PadStrideInfo pad_stride_info = pool_info.pad_stride_info(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 108 | std::tie(pool_pad_x, pool_pad_y) = pad_stride_info.pad(); |
| 109 | std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); |
| 110 | |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame^] | 111 | static const std::set<int> supported_pool_sizes = { 2, 3 }; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 112 | ARM_COMPUTE_UNUSED(supported_pool_sizes); |
| 113 | |
Georgios Pinitas | 1dad50e | 2017-07-03 17:51:34 +0100 | [diff] [blame] | 114 | ARM_COMPUTE_ERROR_ON_NULLPTR(output); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 115 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 116 | ARM_COMPUTE_ERROR_ON(pool_type == PoolingType::L2 && is_data_type_fixed_point(input->info()->data_type())); |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame^] | 117 | ARM_COMPUTE_ERROR_ON((supported_pool_sizes.find(pool_size) == supported_pool_sizes.end()) && (input->info()->data_type() != DataType::F32)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 118 | ARM_COMPUTE_ERROR_ON(pool_pad_x >= pool_size || pool_pad_y >= pool_size); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 119 | ARM_COMPUTE_ERROR_ON(is_data_type_fixed_point(input->info()->data_type()) && pool_stride_x > 2); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 120 | |
| 121 | // Check output dimensions |
| 122 | std::tie(pooled_w, pooled_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), |
Gian Marco Iodice | 4e28869 | 2017-06-27 11:41:59 +0100 | [diff] [blame] | 123 | pool_size, pool_size, pool_info.pad_stride_info()); |
Georgios Pinitas | 1dad50e | 2017-07-03 17:51:34 +0100 | [diff] [blame] | 124 | |
| 125 | // Output auto initialization if not yet initialized |
| 126 | { |
| 127 | TensorShape output_shape{ input->info()->tensor_shape() }; |
| 128 | output_shape.set(0, pooled_w); |
| 129 | output_shape.set(1, pooled_h); |
| 130 | |
| 131 | auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); |
| 132 | } |
| 133 | |
| 134 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 135 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 136 | ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pooled_w) || (output->info()->dimension(1) != pooled_h)); |
| 137 | |
| 138 | unsigned int num_elems_read_per_iteration = 0; |
| 139 | unsigned int num_elems_processed_per_iteration = 0; |
| 140 | unsigned int num_elems_horizontal_window = 0; |
| 141 | |
| 142 | // Select element size |
| 143 | switch(input->info()->data_type()) |
| 144 | { |
| 145 | case DataType::QS8: |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 146 | num_elems_read_per_iteration = 16; |
| 147 | switch(pool_size) |
| 148 | { |
| 149 | case 2: |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 150 | num_elems_processed_per_iteration = (pool_stride_x == 2) ? 8 : 15; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 151 | break; |
| 152 | case 3: |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 153 | num_elems_processed_per_iteration = (pool_stride_x == 2) ? 7 : 14; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 154 | break; |
| 155 | default: |
| 156 | ARM_COMPUTE_ERROR("Pooling size not supported"); |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 157 | break; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 158 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 159 | num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; |
| 160 | break; |
| 161 | case DataType::QS16: |
| 162 | num_elems_read_per_iteration = 8; |
| 163 | switch(pool_size) |
| 164 | { |
| 165 | case 2: |
| 166 | num_elems_processed_per_iteration = (pool_stride_x == 2) ? 4 : 7; |
| 167 | break; |
| 168 | case 3: |
| 169 | num_elems_processed_per_iteration = (pool_stride_x == 2) ? 3 : 6; |
| 170 | break; |
| 171 | default: |
| 172 | ARM_COMPUTE_ERROR("Pooling size not supported"); |
| 173 | } |
| 174 | num_elems_horizontal_window = (pool_stride_x == 2) ? 4 : 8; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 175 | break; |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 176 | #ifdef ARM_COMPUTE_ENABLE_FP16 |
| 177 | case DataType::F16: |
| 178 | switch(pool_size) |
| 179 | { |
| 180 | case 2: |
| 181 | num_elems_read_per_iteration = 16; |
| 182 | num_elems_processed_per_iteration = 8; |
| 183 | num_elems_horizontal_window = 8; |
| 184 | break; |
| 185 | case 3: |
| 186 | num_elems_read_per_iteration = 4; |
| 187 | num_elems_processed_per_iteration = 1; |
| 188 | num_elems_horizontal_window = 1; |
| 189 | break; |
| 190 | default: |
| 191 | ARM_COMPUTE_ERROR("Pooling size not supported"); |
| 192 | break; |
| 193 | } |
| 194 | break; |
| 195 | #endif /* ARM_COMPUTE_ENABLE_FP16 */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 196 | case DataType::F32: |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 197 | switch(pool_size) |
| 198 | { |
| 199 | case 2: |
| 200 | num_elems_read_per_iteration = 2; |
| 201 | break; |
| 202 | case 3: |
| 203 | num_elems_read_per_iteration = 4; // We use vload4 for pooling3 |
| 204 | break; |
| 205 | case 7: |
| 206 | num_elems_read_per_iteration = 8; // We use vload8 for pooling7 |
| 207 | break; |
| 208 | default: |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame^] | 209 | num_elems_read_per_iteration = 1; // We use vload4 for poolingN but with a leftover for loop |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 210 | break; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 211 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 212 | num_elems_processed_per_iteration = 1; |
| 213 | num_elems_horizontal_window = 1; |
| 214 | break; |
| 215 | default: |
| 216 | ARM_COMPUTE_ERROR("Element size not supported"); |
| 217 | break; |
| 218 | } |
| 219 | |
| 220 | _num_elems_processed_per_iteration = num_elems_processed_per_iteration; |
| 221 | const int input_width = input->info()->dimension(0); |
| 222 | const int input_height = input->info()->dimension(1); |
| 223 | const int upper_bound_w = ((pooled_w - 1) * pool_stride_x - pool_pad_x + num_elems_read_per_iteration) - input_width; |
| 224 | const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_y + pool_size) - input_height; |
| 225 | |
| 226 | // Set instance variables |
| 227 | _input = input; |
| 228 | _output = output; |
| 229 | _pool_info = pool_info; |
| 230 | _border_size = BorderSize(pool_pad_y, pool_pad_x); |
| 231 | _border_size.right = std::max(upper_bound_w, pool_pad_x); |
| 232 | _border_size.bottom = std::max(upper_bound_h, pool_pad_y); |
| 233 | |
| 234 | // Select appropriate function |
| 235 | switch(pool_size) |
| 236 | { |
| 237 | case 2: |
| 238 | if(input->info()->data_type() == DataType::QS8) |
| 239 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 240 | switch(pool_type) |
| 241 | { |
| 242 | case PoolingType::AVG: |
| 243 | _func = &NEPoolingLayerKernel::pooling2_q8<PoolingType::AVG>; |
| 244 | break; |
| 245 | case PoolingType::MAX: |
| 246 | _func = &NEPoolingLayerKernel::pooling2_q8<PoolingType::MAX>; |
| 247 | break; |
| 248 | default: |
| 249 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 250 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 251 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 252 | else if(input->info()->data_type() == DataType::QS16) |
| 253 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 254 | switch(pool_type) |
| 255 | { |
| 256 | case PoolingType::AVG: |
| 257 | _func = &NEPoolingLayerKernel::pooling2_q16<PoolingType::AVG>; |
| 258 | break; |
| 259 | case PoolingType::MAX: |
| 260 | _func = &NEPoolingLayerKernel::pooling2_q16<PoolingType::MAX>; |
| 261 | break; |
| 262 | default: |
| 263 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 264 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 265 | } |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 266 | else if(input->info()->data_type() == DataType::F16) |
| 267 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 268 | switch(pool_type) |
| 269 | { |
| 270 | case PoolingType::AVG: |
| 271 | _func = &NEPoolingLayerKernel::pooling2_f16<PoolingType::AVG>; |
| 272 | break; |
| 273 | case PoolingType::L2: |
| 274 | _func = &NEPoolingLayerKernel::pooling2_f16<PoolingType::L2>; |
| 275 | break; |
| 276 | case PoolingType::MAX: |
| 277 | _func = &NEPoolingLayerKernel::pooling2_f16<PoolingType::MAX>; |
| 278 | break; |
| 279 | default: |
| 280 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 281 | } |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 282 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 283 | else if(input->info()->data_type() == DataType::F32) |
| 284 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 285 | switch(pool_type) |
| 286 | { |
| 287 | case PoolingType::AVG: |
| 288 | _func = &NEPoolingLayerKernel::pooling2_f32<PoolingType::AVG>; |
| 289 | break; |
| 290 | case PoolingType::L2: |
| 291 | _func = &NEPoolingLayerKernel::pooling2_f32<PoolingType::L2>; |
| 292 | break; |
| 293 | case PoolingType::MAX: |
| 294 | _func = &NEPoolingLayerKernel::pooling2_f32<PoolingType::MAX>; |
| 295 | break; |
| 296 | default: |
| 297 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 298 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 299 | } |
| 300 | break; |
| 301 | case 3: |
| 302 | if(input->info()->data_type() == DataType::QS8) |
| 303 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 304 | switch(pool_type) |
| 305 | { |
| 306 | case PoolingType::AVG: |
| 307 | _func = &NEPoolingLayerKernel::pooling3_q8<PoolingType::AVG>; |
| 308 | break; |
| 309 | case PoolingType::MAX: |
| 310 | _func = &NEPoolingLayerKernel::pooling3_q8<PoolingType::MAX>; |
| 311 | break; |
| 312 | default: |
| 313 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 314 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 315 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 316 | else if(input->info()->data_type() == DataType::QS16) |
| 317 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 318 | switch(pool_type) |
| 319 | { |
| 320 | case PoolingType::AVG: |
| 321 | _func = &NEPoolingLayerKernel::pooling3_q16<PoolingType::AVG>; |
| 322 | break; |
| 323 | case PoolingType::MAX: |
| 324 | _func = &NEPoolingLayerKernel::pooling3_q16<PoolingType::MAX>; |
| 325 | break; |
| 326 | default: |
| 327 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 328 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 329 | } |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 330 | else if(input->info()->data_type() == DataType::F16) |
| 331 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 332 | switch(pool_type) |
| 333 | { |
| 334 | case PoolingType::AVG: |
| 335 | _func = &NEPoolingLayerKernel::pooling3_f16<PoolingType::AVG>; |
| 336 | break; |
| 337 | case PoolingType::L2: |
| 338 | _func = &NEPoolingLayerKernel::pooling3_f16<PoolingType::L2>; |
| 339 | break; |
| 340 | case PoolingType::MAX: |
| 341 | _func = &NEPoolingLayerKernel::pooling3_f16<PoolingType::MAX>; |
| 342 | break; |
| 343 | default: |
| 344 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 345 | } |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 346 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 347 | else if(input->info()->data_type() == DataType::F32) |
| 348 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 349 | switch(pool_type) |
| 350 | { |
| 351 | case PoolingType::AVG: |
| 352 | _func = &NEPoolingLayerKernel::pooling3_f32<PoolingType::AVG>; |
| 353 | break; |
| 354 | case PoolingType::L2: |
| 355 | _func = &NEPoolingLayerKernel::pooling3_f32<PoolingType::L2>; |
| 356 | break; |
| 357 | case PoolingType::MAX: |
| 358 | _func = &NEPoolingLayerKernel::pooling3_f32<PoolingType::MAX>; |
| 359 | break; |
| 360 | default: |
| 361 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 362 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 363 | } |
| 364 | break; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 365 | case 7: |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 366 | switch(pool_type) |
| 367 | { |
| 368 | case PoolingType::AVG: |
| 369 | _func = &NEPoolingLayerKernel::pooling7_f32<PoolingType::AVG>; |
| 370 | break; |
| 371 | case PoolingType::L2: |
| 372 | _func = &NEPoolingLayerKernel::pooling7_f32<PoolingType::L2>; |
| 373 | break; |
| 374 | case PoolingType::MAX: |
| 375 | _func = &NEPoolingLayerKernel::pooling7_f32<PoolingType::MAX>; |
| 376 | break; |
| 377 | default: |
| 378 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 379 | } |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 380 | break; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 381 | default: |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame^] | 382 | switch(pool_type) |
| 383 | { |
| 384 | case PoolingType::AVG: |
| 385 | _func = &NEPoolingLayerKernel::poolingN_f32<PoolingType::AVG>; |
| 386 | break; |
| 387 | case PoolingType::L2: |
| 388 | _func = &NEPoolingLayerKernel::poolingN_f32<PoolingType::L2>; |
| 389 | break; |
| 390 | case PoolingType::MAX: |
| 391 | _func = &NEPoolingLayerKernel::poolingN_f32<PoolingType::MAX>; |
| 392 | break; |
| 393 | default: |
| 394 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 395 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 396 | break; |
| 397 | } |
| 398 | |
| 399 | // Configure kernel window |
| 400 | Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration)); |
| 401 | AccessWindowStatic input_access(input->info(), -pool_pad_x, -pool_pad_y, input_width + _border_size.right, input_height + _border_size.bottom); |
| 402 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_horizontal_window); |
| 403 | update_window_and_padding(win, input_access, output_access); |
| 404 | output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); |
| 405 | INEKernel::configure(win); |
| 406 | } |
| 407 | |
| 408 | template <PoolingType pooling_type> |
| 409 | void NEPoolingLayerKernel::pooling2_q8(const Window &window_input, const Window &window) |
| 410 | { |
| 411 | Iterator input(_input, window_input); |
| 412 | Iterator output(_output, window); |
| 413 | |
| 414 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 415 | constexpr int pool_size = 2; |
| 416 | int pool_pad_x = 0; |
| 417 | int pool_pad_y = 0; |
| 418 | int pool_stride_x = 0; |
| 419 | int pool_stride_y = 0; |
| 420 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 421 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 422 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 423 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 424 | |
| 425 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 426 | const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 1)); |
| 427 | |
| 428 | execute_window_loop(window, [&](const Coordinates & id) |
| 429 | { |
| 430 | const auto top_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_top_ptr + input.offset())); |
| 431 | const auto bottom_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_bottom_ptr + input.offset())); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 432 | qint8x8_t lower_res = {}; |
| 433 | qint8x8_t upper_res = {}; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 434 | if(pooling_type == PoolingType::AVG) |
| 435 | { |
| 436 | // Calculate scale |
| 437 | const qint8_t scale = calculate_avg_scale_q8(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y, fixed_point_position); |
| 438 | const qint8x8_t scale_vec = vdup_n_qs8(scale); |
| 439 | |
| 440 | // Perform pooling |
| 441 | const qint8x16_t sum_data = vqaddq_qs8(top_data, bottom_data); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 442 | lower_res = vqmul_qs8(vpadd_s8(vget_low_s8(sum_data), vget_high_s8(sum_data)), scale_vec, fixed_point_position); |
| 443 | if(pool_stride_x == 1) |
| 444 | { |
| 445 | const qint8x16_t sum_data_shifted = vextq_s8(sum_data, sum_data, 1); |
| 446 | upper_res = vqmul_qs8(vpadd_s8(vget_low_s8(sum_data_shifted), vget_high_s8(sum_data_shifted)), scale_vec, fixed_point_position); |
| 447 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 448 | } |
| 449 | else |
| 450 | { |
| 451 | const qint8x16_t max_data = vmaxq_s8(top_data, bottom_data); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 452 | lower_res = vpmax_s8(vget_low_s8(max_data), vget_high_s8(max_data)); |
| 453 | if(pool_stride_x == 1) |
| 454 | { |
| 455 | const qint8x16_t max_data_shifted = vextq_s8(max_data, max_data, 1); |
| 456 | upper_res = vpmax_s8(vget_low_s8(max_data_shifted), vget_high_s8(max_data_shifted)); |
| 457 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 458 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 459 | if(pool_stride_x == 1) |
| 460 | { |
Georgios Pinitas | dc460f1 | 2017-08-24 19:02:44 +0100 | [diff] [blame] | 461 | const qint8x8x2_t res = { { lower_res, upper_res } }; |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 462 | vst2_s8(reinterpret_cast<qint8_t *>(output.ptr()), res); |
| 463 | } |
| 464 | else |
| 465 | { |
| 466 | vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), lower_res); |
| 467 | } |
| 468 | }, |
| 469 | input, output); |
| 470 | } |
| 471 | |
| 472 | template <PoolingType pooling_type> |
| 473 | void NEPoolingLayerKernel::pooling2_q16(const Window &window_input, const Window &window) |
| 474 | { |
| 475 | Iterator input(_input, window_input); |
| 476 | Iterator output(_output, window); |
| 477 | |
| 478 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 479 | constexpr int pool_size = 2; |
| 480 | int pool_pad_x = 0; |
| 481 | int pool_pad_y = 0; |
| 482 | int pool_stride_x = 0; |
| 483 | int pool_stride_y = 0; |
| 484 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 485 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 486 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 487 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 488 | |
| 489 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 490 | const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 1)); |
| 491 | |
| 492 | execute_window_loop(window, [&](const Coordinates & id) |
| 493 | { |
| 494 | const auto top_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_top_ptr + input.offset())); |
| 495 | const auto bottom_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_bottom_ptr + input.offset())); |
| 496 | qint16x4_t lower_res = {}; |
| 497 | qint16x4_t upper_res = {}; |
| 498 | if(pooling_type == PoolingType::AVG) |
| 499 | { |
| 500 | // Calculate scale |
| 501 | const qint16_t scale = calculate_avg_scale_q16(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y, fixed_point_position); |
| 502 | const qint16x4_t scale_vec = vdup_n_qs16(scale); |
| 503 | |
| 504 | // Perform pooling |
| 505 | const qint16x8_t sum_data = vqaddq_qs16(top_data, bottom_data); |
| 506 | lower_res = vqmul_qs16(vpadd_s16(vget_low_s16(sum_data), vget_high_s16(sum_data)), scale_vec, fixed_point_position); |
| 507 | if(pool_stride_x == 1) |
| 508 | { |
| 509 | const qint16x8_t sum_data_shifted = vextq_s16(sum_data, sum_data, 1); |
| 510 | upper_res = vqmul_qs16(vpadd_s16(vget_low_s16(sum_data_shifted), vget_high_s16(sum_data_shifted)), scale_vec, fixed_point_position); |
| 511 | } |
| 512 | } |
| 513 | else |
| 514 | { |
| 515 | const qint16x8_t max_data = vmaxq_s16(top_data, bottom_data); |
| 516 | lower_res = vpmax_s16(vget_low_s16(max_data), vget_high_s16(max_data)); |
| 517 | if(pool_stride_x == 1) |
| 518 | { |
| 519 | const qint16x8_t max_data_shifted = vextq_s16(max_data, max_data, 1); |
| 520 | upper_res = vpmax_s16(vget_low_s16(max_data_shifted), vget_high_s16(max_data_shifted)); |
| 521 | } |
| 522 | } |
| 523 | if(pool_stride_x == 1) |
| 524 | { |
Georgios Pinitas | dc460f1 | 2017-08-24 19:02:44 +0100 | [diff] [blame] | 525 | const qint16x4x2_t res = { { lower_res, upper_res } }; |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 526 | vst2_s16(reinterpret_cast<qint16_t *>(output.ptr()), res); |
| 527 | } |
| 528 | else |
| 529 | { |
| 530 | vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), lower_res); |
| 531 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 532 | }, |
| 533 | input, output); |
| 534 | } |
| 535 | |
| 536 | template <PoolingType pooling_type> |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 537 | void NEPoolingLayerKernel::pooling3_f16(const Window &window_input, const Window &window) |
| 538 | { |
| 539 | #ifdef ARM_COMPUTE_ENABLE_FP16 |
| 540 | Iterator input(_input, window_input); |
| 541 | Iterator output(_output, window); |
| 542 | |
| 543 | constexpr const int pool_size = 3; |
| 544 | int pool_pad_x = 0; |
| 545 | int pool_pad_y = 0; |
| 546 | int pool_stride_x = 0; |
| 547 | int pool_stride_y = 0; |
| 548 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 549 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 550 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 551 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 552 | |
| 553 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 554 | const unsigned char *const input_middle_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 1)); |
| 555 | const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 2)); |
| 556 | |
| 557 | execute_window_loop(window, [&](const Coordinates & id) |
| 558 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 559 | float16x4_t top_data = vld1_f16(reinterpret_cast<const float16_t *>(input_top_ptr + input.offset())); |
| 560 | float16x4_t middle_data = vld1_f16(reinterpret_cast<const float16_t *>(input_middle_ptr + input.offset())); |
| 561 | float16x4_t bottom_data = vld1_f16(reinterpret_cast<const float16_t *>(input_bottom_ptr + input.offset())); |
| 562 | float16x4_t res = {}; |
| 563 | |
| 564 | // Get power of 2 in case of l2 pooling |
| 565 | if(pooling_type == PoolingType::L2) |
| 566 | { |
| 567 | top_data = vmul_f16(top_data, top_data); |
| 568 | middle_data = vmul_f16(middle_data, middle_data); |
| 569 | bottom_data = vmul_f16(bottom_data, bottom_data); |
| 570 | } |
| 571 | |
| 572 | if(pooling_type != PoolingType::MAX) |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 573 | { |
| 574 | // Calculate scale |
| 575 | const float scale = calculate_avg_scale(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 576 | const float16x4_t scale_v = vdup_n_f16(scale); |
| 577 | // Perform pooling |
| 578 | const float16x4_t sum_data = vadd_f16(vadd_f16(top_data, bottom_data), middle_data); |
| 579 | res = vpadd_f16(vset_lane_f16(0.f, sum_data, 3), sum_data); |
| 580 | res = vmul_f16(vpadd_f16(res, res), scale_v); |
| 581 | } |
| 582 | else |
| 583 | { |
| 584 | const float16x4_t max_data = vmax_f16(vmax_f16(top_data, bottom_data), middle_data); |
| 585 | res = vpmax_f16(vset_lane_f16(-std::numeric_limits<float>::max(), max_data, 3), max_data); |
| 586 | res = vpmax_f16(res, res); |
| 587 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 588 | |
| 589 | // Calculate square-root in case of l2 pooling |
| 590 | if(pooling_type == PoolingType::L2) |
| 591 | { |
| 592 | res = vinv_f16(vinvsqrt_f16(res)); |
| 593 | } |
| 594 | |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 595 | *(reinterpret_cast<float16_t *>(output.ptr())) = vget_lane_f16(res, 0); |
| 596 | }, |
| 597 | input, output); |
| 598 | #else /* ARM_COMPUTE_ENABLE_FP16 */ |
| 599 | ARM_COMPUTE_UNUSED(window_input); |
| 600 | ARM_COMPUTE_UNUSED(window); |
| 601 | ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); |
| 602 | #endif /* ARM_COMPUTE_ENABLE_FP16 */ |
| 603 | } |
| 604 | |
| 605 | template <PoolingType pooling_type> |
| 606 | void NEPoolingLayerKernel::pooling2_f16(const Window &window_input, const Window &window) |
| 607 | { |
| 608 | #ifdef ARM_COMPUTE_ENABLE_FP16 |
| 609 | Iterator input(_input, window_input); |
| 610 | Iterator output(_output, window); |
| 611 | constexpr int pool_size = 2; |
| 612 | int pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y = 0; |
| 613 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 614 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 615 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 616 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 617 | |
| 618 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 619 | const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 1)); |
| 620 | |
| 621 | execute_window_loop(window, [&](const Coordinates & id) |
| 622 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 623 | auto top_data = vld2q_f16(reinterpret_cast<const float16_t *>(input_top_ptr + input.offset())); |
| 624 | auto bottom_data = vld2q_f16(reinterpret_cast<const float16_t *>(input_bottom_ptr + input.offset())); |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 625 | float16x8_t res = {}; |
| 626 | |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 627 | // Get power of 2 in case of l2 pooling |
| 628 | if(pooling_type == PoolingType::L2) |
| 629 | { |
| 630 | top_data.val[0] = vmulq_f16(top_data.val[0], top_data.val[0]); |
| 631 | top_data.val[1] = vmulq_f16(top_data.val[1], top_data.val[1]); |
| 632 | bottom_data.val[0] = vmulq_f16(bottom_data.val[0], bottom_data.val[0]); |
| 633 | bottom_data.val[1] = vmulq_f16(bottom_data.val[1], bottom_data.val[1]); |
| 634 | } |
| 635 | |
| 636 | if(pooling_type != PoolingType::MAX) |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 637 | { |
| 638 | const float scale = calculate_avg_scale(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 639 | const float16x8_t scale_v = vdupq_n_f16(scale); |
| 640 | res = vmulq_f16(scale_v, vaddq_f16(bottom_data.val[1], vaddq_f16(bottom_data.val[0], vaddq_f16(top_data.val[0], top_data.val[1])))); |
| 641 | } |
| 642 | else |
| 643 | { |
| 644 | res = vmaxq_f16(bottom_data.val[1], vmaxq_f16(bottom_data.val[0], vmaxq_f16(top_data.val[0], top_data.val[1]))); |
| 645 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 646 | |
| 647 | // Calculate square-root in case of l2 pooling |
| 648 | if(pooling_type == PoolingType::L2) |
| 649 | { |
| 650 | res = vinvq_f16(vinvsqrtq_f16(res)); |
| 651 | } |
| 652 | |
| 653 | // Store result |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 654 | vst1q_f16(reinterpret_cast<float16_t *>(output.ptr()), res); |
| 655 | }, |
| 656 | input, output); |
| 657 | #else /* ARM_COMPUTE_ENABLE_FP16 */ |
| 658 | ARM_COMPUTE_UNUSED(window_input); |
| 659 | ARM_COMPUTE_UNUSED(window); |
| 660 | ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); |
| 661 | #endif /* ARM_COMPUTE_ENABLE_FP16 */ |
| 662 | } |
| 663 | |
| 664 | template <PoolingType pooling_type> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 665 | void NEPoolingLayerKernel::pooling2_f32(const Window &window_input, const Window &window) |
| 666 | { |
| 667 | Iterator input(_input, window_input); |
| 668 | Iterator output(_output, window); |
| 669 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 670 | constexpr int pool_size = 2; |
| 671 | int pool_pad_x = 0; |
| 672 | int pool_pad_y = 0; |
| 673 | int pool_stride_x = 0; |
| 674 | int pool_stride_y = 0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 675 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 676 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 677 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 678 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 679 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 680 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 681 | const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 682 | |
| 683 | execute_window_loop(window, [&](const Coordinates & id) |
| 684 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 685 | float32x2_t top_data = vld1_f32(reinterpret_cast<const float *>(input_top_ptr + input.offset())); |
| 686 | float32x2_t bottom_data = vld1_f32(reinterpret_cast<const float *>(input_bottom_ptr + input.offset())); |
| 687 | float32x2_t res = {}; |
| 688 | float final_res = 0; |
| 689 | |
| 690 | // Get power of 2 in case of l2 pooling |
| 691 | if(pooling_type == PoolingType::L2) |
| 692 | { |
| 693 | top_data = vmul_f32(top_data, top_data); |
| 694 | bottom_data = vmul_f32(bottom_data, bottom_data); |
| 695 | } |
| 696 | |
| 697 | if(pooling_type != PoolingType::MAX) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 698 | { |
| 699 | // Calculate scale |
| 700 | float scale = calculate_avg_scale(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 701 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 702 | |
| 703 | // Perform pooling |
| 704 | const float32x2_t sum_data = vadd_f32(top_data, bottom_data); |
| 705 | res = vmul_f32(vpadd_f32(sum_data, sum_data), scale_v); |
| 706 | } |
| 707 | else |
| 708 | { |
| 709 | const float32x2_t max_data = vmax_f32(top_data, bottom_data); |
| 710 | res = vpmax_f32(max_data, max_data); |
| 711 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 712 | final_res = vget_lane_f32(res, 0); |
| 713 | |
| 714 | // Calculate square-root in case of l2 pooling |
| 715 | if(pooling_type == PoolingType::L2) |
| 716 | { |
| 717 | final_res = sqrt(final_res); |
| 718 | } |
| 719 | |
| 720 | // Store result |
| 721 | *(reinterpret_cast<float *>(output.ptr())) = final_res; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 722 | }, |
| 723 | input, output); |
| 724 | } |
| 725 | |
| 726 | template <PoolingType pooling_type> |
| 727 | void NEPoolingLayerKernel::pooling3_q8(const Window &window_input, const Window &window) |
| 728 | { |
| 729 | Iterator input(_input, window_input); |
| 730 | Iterator output(_output, window); |
| 731 | |
| 732 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 733 | constexpr int pool_size = 3; |
| 734 | int pool_pad_x = 0; |
| 735 | int pool_pad_y = 0; |
| 736 | int pool_stride_x = 0; |
| 737 | int pool_stride_y = 0; |
| 738 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 739 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 740 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 741 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 742 | |
| 743 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 744 | const uint8_t *const input_middle_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 1)); |
| 745 | const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 2)); |
| 746 | |
| 747 | execute_window_loop(window, [&](const Coordinates & id) |
| 748 | { |
| 749 | const auto top_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_top_ptr + input.offset())); |
| 750 | const auto middle_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_middle_ptr + input.offset())); |
| 751 | const auto bottom_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_bottom_ptr + input.offset())); |
| 752 | qint8x8_t res = {}; |
| 753 | if(pooling_type == PoolingType::AVG) |
| 754 | { |
| 755 | // Calculate scale |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 756 | const qint8_t scale = calculate_avg_scale_q8(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y, fixed_point_position); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 757 | |
| 758 | // Perform pooling for stride 2 |
| 759 | const qint8x16_t sum_data = vqaddq_qs8(vqaddq_qs8(top_data, bottom_data), middle_data); |
| 760 | const qint8x16_t sum_data2 = vextq_s8(sum_data, sum_data, 1); |
| 761 | const qint8x16_t sum_data3 = vextq_s8(sum_data, sum_data, 2); |
| 762 | const qint8x16_t final_sum = vqaddq_qs8(vqaddq_qs8(sum_data, sum_data2), sum_data3); |
| 763 | if(pool_stride_x == 2) |
| 764 | { |
| 765 | const qint8x8x2_t table = { { vget_low_s8(final_sum), vget_high_s8(final_sum) } }; |
| 766 | static const qint8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 767 | const qint8x8_t scale_vec = vdup_n_qs8(scale); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 768 | res = vtbl2_s8(table, lookup_val); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 769 | res = vqmul_qs8(res, scale_vec, fixed_point_position); |
| 770 | vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), res); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 771 | } |
| 772 | else |
| 773 | { |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 774 | const qint8x16_t scale_vec = vdupq_n_qs8(scale); |
| 775 | vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), vqmulq_qs8(final_sum, scale_vec, fixed_point_position)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 776 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 777 | } |
| 778 | else |
| 779 | { |
| 780 | const qint8x16_t max_data = vmaxq_s8(vmaxq_s8(top_data, bottom_data), middle_data); |
| 781 | const qint8x16_t max_data2 = vextq_s8(max_data, max_data, 1); |
| 782 | const qint8x16_t max_data3 = vextq_s8(max_data, max_data, 2); |
| 783 | const qint8x16_t final_max = vmaxq_s8(vmaxq_s8(max_data, max_data2), max_data3); |
| 784 | |
| 785 | if(pool_stride_x == 2) |
| 786 | { |
| 787 | const qint8x8x2_t table = { { vget_low_s8(final_max), vget_high_s8(final_max) } }; |
| 788 | static const qint8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; |
| 789 | res = vtbl2_s8(table, lookup_val); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 790 | vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), res); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 791 | } |
| 792 | else |
| 793 | { |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 794 | vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), final_max); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 795 | } |
| 796 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 797 | }, |
| 798 | input, output); |
| 799 | } |
| 800 | |
| 801 | template <PoolingType pooling_type> |
| 802 | void NEPoolingLayerKernel::pooling3_q16(const Window &window_input, const Window &window) |
| 803 | { |
| 804 | Iterator input(_input, window_input); |
| 805 | Iterator output(_output, window); |
| 806 | |
| 807 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 808 | constexpr int pool_size = 3; |
| 809 | int pool_pad_x = 0; |
| 810 | int pool_pad_y = 0; |
| 811 | int pool_stride_x = 0; |
| 812 | int pool_stride_y = 0; |
| 813 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 814 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 815 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 816 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 817 | |
| 818 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 819 | const unsigned char *const input_middle_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 1)); |
| 820 | const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 2)); |
| 821 | |
| 822 | execute_window_loop(window, [&](const Coordinates & id) |
| 823 | { |
| 824 | const auto top_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_top_ptr + input.offset())); |
| 825 | const auto middle_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_middle_ptr + input.offset())); |
| 826 | const auto bottom_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_bottom_ptr + input.offset())); |
| 827 | |
| 828 | if(pooling_type == PoolingType::AVG) |
| 829 | { |
| 830 | // Calculate scale |
| 831 | const qint16_t scale = calculate_avg_scale_q16(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y, fixed_point_position); |
| 832 | |
| 833 | // Perform pooling for stride 2 |
| 834 | const qint16x8_t sum_data = vqaddq_qs16(vqaddq_qs16(top_data, bottom_data), middle_data); |
| 835 | const qint16x8_t sum_data2 = vextq_s16(sum_data, sum_data, 1); |
| 836 | const qint16x8_t sum_data3 = vextq_s16(sum_data, sum_data, 2); |
| 837 | const qint16x8_t final_sum = vqaddq_qs16(vqaddq_qs16(sum_data, sum_data2), sum_data3); |
| 838 | if(pool_stride_x == 2) |
| 839 | { |
| 840 | const qint16x4_t tmp = { vgetq_lane_s16(final_sum, 0), vgetq_lane_s16(final_sum, 2), vgetq_lane_s16(final_sum, 4), vgetq_lane_s16(final_sum, 6) }; |
| 841 | const qint16x4_t scale_vec = vdup_n_qs16(scale); |
| 842 | vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), vqmul_qs16(tmp, scale_vec, fixed_point_position)); |
| 843 | } |
| 844 | else |
| 845 | { |
| 846 | const qint16x8_t scale_vec = vdupq_n_qs16(scale); |
| 847 | vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()), vqmulq_qs16(final_sum, scale_vec, fixed_point_position)); |
| 848 | } |
| 849 | } |
| 850 | else |
| 851 | { |
| 852 | const qint16x8_t max_data = vmaxq_s16(vmaxq_s16(top_data, bottom_data), middle_data); |
| 853 | const qint16x8_t max_data2 = vextq_s16(max_data, max_data, 1); |
| 854 | const qint16x8_t max_data3 = vextq_s16(max_data, max_data, 2); |
| 855 | const qint16x8_t final_max = vmaxq_s16(vmaxq_s16(max_data, max_data2), max_data3); |
| 856 | |
| 857 | if(pool_stride_x == 2) |
| 858 | { |
| 859 | const qint16x4_t tmp = { vgetq_lane_s16(final_max, 0), vgetq_lane_s16(final_max, 2), vgetq_lane_s16(final_max, 4), vgetq_lane_s16(final_max, 6) }; |
| 860 | vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), tmp); |
| 861 | } |
| 862 | else |
| 863 | { |
| 864 | vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()), final_max); |
| 865 | } |
| 866 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 867 | }, |
| 868 | input, output); |
| 869 | } |
| 870 | |
| 871 | template <PoolingType pooling_type> |
| 872 | void NEPoolingLayerKernel::pooling3_f32(const Window &window_input, const Window &window) |
| 873 | { |
| 874 | Iterator input(_input, window_input); |
| 875 | Iterator output(_output, window); |
| 876 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 877 | constexpr const int pool_size = 3; |
| 878 | int pool_pad_x = 0; |
| 879 | int pool_pad_y = 0; |
| 880 | int pool_stride_x = 0; |
| 881 | int pool_stride_y = 0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 882 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 883 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 884 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 885 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 886 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 887 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 888 | const uint8_t *const input_middle_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 1)); |
| 889 | const uint8_t *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 2)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 890 | |
| 891 | execute_window_loop(window, [&](const Coordinates & id) |
| 892 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 893 | float32x4_t top_data = vld1q_f32(reinterpret_cast<const float *>(input_top_ptr + input.offset())); |
| 894 | float32x4_t middle_data = vld1q_f32(reinterpret_cast<const float *>(input_middle_ptr + input.offset())); |
| 895 | float32x4_t bottom_data = vld1q_f32(reinterpret_cast<const float *>(input_bottom_ptr + input.offset())); |
| 896 | float32x2_t res = {}; |
| 897 | float final_res = 0; |
| 898 | |
| 899 | // Get power of 2 in case of l2 pooling |
| 900 | if(pooling_type == PoolingType::L2) |
| 901 | { |
| 902 | top_data = vmulq_f32(top_data, top_data); |
| 903 | middle_data = vmulq_f32(middle_data, middle_data); |
| 904 | bottom_data = vmulq_f32(bottom_data, bottom_data); |
| 905 | } |
| 906 | |
| 907 | if(pooling_type != PoolingType::MAX) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 908 | { |
| 909 | // Calculate scale |
| 910 | float scale = calculate_avg_scale(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 911 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 912 | |
| 913 | // Perform pooling |
| 914 | const float32x4_t sum_data = vaddq_f32(vaddq_f32(top_data, bottom_data), middle_data); |
| 915 | res = vpadd_f32(vget_high_f32(vsetq_lane_f32(0.f, sum_data, 3)), vget_low_f32(sum_data)); |
| 916 | res = vmul_f32(vpadd_f32(res, res), scale_v); |
| 917 | } |
| 918 | else |
| 919 | { |
| 920 | const float32x4_t max_data = vmaxq_f32(vmaxq_f32(top_data, bottom_data), middle_data); |
| 921 | res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits<float>::max(), max_data, 3)), vget_low_f32(max_data)); |
| 922 | res = vpmax_f32(res, res); |
| 923 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 924 | final_res = vget_lane_f32(res, 0); |
| 925 | |
| 926 | // Calculate square-root in case of l2 pooling |
| 927 | if(pooling_type == PoolingType::L2) |
| 928 | { |
| 929 | final_res = sqrt(final_res); |
| 930 | } |
| 931 | |
| 932 | // Store result |
| 933 | *(reinterpret_cast<float *>(output.ptr())) = final_res; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 934 | }, |
| 935 | input, output); |
| 936 | } |
| 937 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 938 | template <PoolingType pooling_type> |
| 939 | void NEPoolingLayerKernel::pooling7_f32(const Window &window_input, const Window &window) |
| 940 | { |
| 941 | Iterator input(_input, window_input); |
| 942 | Iterator output(_output, window); |
| 943 | |
| 944 | constexpr const int pool_size = 7; |
| 945 | int pool_pad_x = 0; |
| 946 | int pool_pad_y = 0; |
| 947 | int pool_stride_x = 0; |
| 948 | int pool_stride_y = 0; |
| 949 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 950 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 951 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 952 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 953 | |
| 954 | std::array<const uint8_t *, pool_size> input_ptrs{ {} }; |
| 955 | for(int i = 0; i < pool_size; ++i) |
| 956 | { |
| 957 | input_ptrs[i] = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + i)); |
| 958 | } |
| 959 | |
| 960 | execute_window_loop(window, [&](const Coordinates & id) |
| 961 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 962 | float32x2_t res = {}; |
| 963 | float final_res = 0.f; |
| 964 | if(pooling_type != PoolingType::MAX) |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 965 | { |
| 966 | // Calculate scale |
| 967 | float scale = calculate_avg_scale(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 968 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 969 | |
| 970 | // Perform pooling |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 971 | float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[0] + input.offset())); |
| 972 | // Get power of 2 in case of l2 pooling |
| 973 | if(pooling_type == PoolingType::L2) |
| 974 | { |
| 975 | data.val[0] = vmulq_f32(data.val[0], data.val[0]); |
| 976 | data.val[1] = vmulq_f32(data.val[1], data.val[1]); |
| 977 | } |
| 978 | float32x4_t sum_data = vaddq_f32(data.val[0], vsetq_lane_f32(0.f, data.val[1], 3)); |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 979 | for(int i = 1; i < pool_size; ++i) |
| 980 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 981 | data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[i] + input.offset())); |
| 982 | // Get power of 2 in case of l2 pooling |
| 983 | if(pooling_type == PoolingType::L2) |
| 984 | { |
| 985 | data.val[0] = vmulq_f32(data.val[0], data.val[0]); |
| 986 | data.val[1] = vmulq_f32(data.val[1], data.val[1]); |
| 987 | } |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 988 | sum_data = vaddq_f32(sum_data, data.val[0]); |
| 989 | sum_data = vaddq_f32(sum_data, vsetq_lane_f32(0.f, data.val[1], 3)); |
| 990 | } |
| 991 | res = vpadd_f32(vget_high_f32(sum_data), vget_low_f32(sum_data)); |
| 992 | res = vmul_f32(vpadd_f32(res, res), scale_v); |
| 993 | } |
| 994 | else |
| 995 | { |
| 996 | float32x4x2_t max_data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[0] + input.offset())); |
| 997 | for(int i = 1; i < pool_size; ++i) |
| 998 | { |
| 999 | const float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[i] + input.offset())); |
| 1000 | max_data = vmax2q_f32(max_data, data); |
| 1001 | } |
| 1002 | res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits<float>::max(), max_data.val[1], 3)), vget_low_f32(max_data.val[1])); |
| 1003 | res = vpmax_f32(res, vpmax_f32(vget_high_f32(max_data.val[0]), vget_low_f32(max_data.val[0]))); |
| 1004 | res = vpmax_f32(res, res); |
| 1005 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1006 | final_res = vget_lane_f32(res, 0); |
| 1007 | |
| 1008 | // Calculate square-root in case of l2 pooling |
| 1009 | if(pooling_type == PoolingType::L2) |
| 1010 | { |
| 1011 | final_res = sqrt(final_res); |
| 1012 | } |
| 1013 | |
| 1014 | // Store result |
| 1015 | *(reinterpret_cast<float *>(output.ptr())) = final_res; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1016 | }, |
| 1017 | input, output); |
| 1018 | } |
| 1019 | |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame^] | 1020 | template <PoolingType pooling_type> |
| 1021 | void NEPoolingLayerKernel::poolingN_f32(const Window &window_input, const Window &window) |
| 1022 | { |
| 1023 | Iterator input(_input, window_input); |
| 1024 | Iterator output(_output, window); |
| 1025 | |
| 1026 | const int pool_size = _pool_info.pool_size(); |
| 1027 | int pool_pad_x = 0; |
| 1028 | int pool_pad_y = 0; |
| 1029 | int pool_stride_x = 0; |
| 1030 | int pool_stride_y = 0; |
| 1031 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1032 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 1033 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 1034 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 1035 | |
| 1036 | execute_window_loop(window, [&](const Coordinates & id) |
| 1037 | { |
| 1038 | float res = 0.0f; |
| 1039 | |
| 1040 | if(pooling_type != PoolingType::MAX) |
| 1041 | { |
| 1042 | // Calculate scale |
| 1043 | const float scale = calculate_avg_scale(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 1044 | |
| 1045 | // Perform pooling |
| 1046 | float32x4_t vres = vdupq_n_f32(0.0f); |
| 1047 | |
| 1048 | for(int y = 0; y < pool_size; ++y) |
| 1049 | { |
| 1050 | int x = 0; |
| 1051 | for(; x <= (pool_size - 4); x += 4) |
| 1052 | { |
| 1053 | const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + |
| 1054 | (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1055 | |
| 1056 | // Get power of 2 in case of l2 pooling and accumulate |
| 1057 | if(pooling_type == PoolingType::L2) |
| 1058 | { |
| 1059 | vres = vmlaq_f32(vres, data, data); |
| 1060 | } |
| 1061 | else |
| 1062 | { |
| 1063 | vres = vaddq_f32(vres, data); |
| 1064 | } |
| 1065 | } |
| 1066 | |
| 1067 | // Leftover for loop |
| 1068 | for(; x < pool_size; ++x) |
| 1069 | { |
| 1070 | float data = *(reinterpret_cast<const float *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1071 | |
| 1072 | // Get power of 2 in case of l2 pooling |
| 1073 | if(pooling_type == PoolingType::L2) |
| 1074 | { |
| 1075 | data *= data; |
| 1076 | } |
| 1077 | |
| 1078 | res += data; |
| 1079 | } |
| 1080 | } |
| 1081 | |
| 1082 | #if defined(__aarch64__) |
| 1083 | // Reduction operation available on 64 bit architectures only |
| 1084 | res += vaddvq_f32(vres); |
| 1085 | #else // __aarch64__ |
| 1086 | // Reduction |
| 1087 | float32x2_t tmp = vpadd_f32(vget_high_f32(vres), vget_low_f32(vres)); |
| 1088 | tmp = vpadd_f32(tmp, tmp); |
| 1089 | |
| 1090 | res += vget_lane_f32(tmp, 0); |
| 1091 | #endif // __aarch64__ |
| 1092 | // Divide by scale |
| 1093 | res *= scale; |
| 1094 | } |
| 1095 | else |
| 1096 | { |
| 1097 | float32x4_t vres = vdupq_n_f32(std::numeric_limits<float>::min()); |
| 1098 | res = std::numeric_limits<float>::min(); |
| 1099 | |
| 1100 | for(int y = 0; y < pool_size; ++y) |
| 1101 | { |
| 1102 | int x = 0; |
| 1103 | for(; x <= (pool_size - 4); x += 4) |
| 1104 | { |
| 1105 | const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + |
| 1106 | (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1107 | vres = vmaxq_f32(vres, data); |
| 1108 | } |
| 1109 | |
| 1110 | // Leftover for loop |
| 1111 | for(; x < pool_size; ++x) |
| 1112 | { |
| 1113 | const float data = *(reinterpret_cast<const float *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1114 | res = std::max(res, data); |
| 1115 | } |
| 1116 | } |
| 1117 | |
| 1118 | #if defined(__aarch64__) |
| 1119 | // Reduction operation available on 64 bit architectures only |
| 1120 | res = std::max(vmaxvq_f32(vres), res); |
| 1121 | #else // __aarch64__ |
| 1122 | float32x2_t tmp = vpmax_f32(vget_high_f32(vres), vget_low_f32(vres)); |
| 1123 | tmp = vpmax_f32(tmp, tmp); |
| 1124 | |
| 1125 | res = std::max(res, vget_lane_f32(tmp, 0)); |
| 1126 | #endif // __aarch64__ |
| 1127 | } |
| 1128 | |
| 1129 | // Calculate square-root in case of l2 pooling |
| 1130 | if(pooling_type == PoolingType::L2) |
| 1131 | { |
| 1132 | res = std::sqrt(res); |
| 1133 | } |
| 1134 | |
| 1135 | // Store result |
| 1136 | *(reinterpret_cast<float *>(output.ptr())) = res; |
| 1137 | }, |
| 1138 | input, output); |
| 1139 | } |
| 1140 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1141 | void NEPoolingLayerKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1142 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1143 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1144 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 1145 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 1146 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 1147 | |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1148 | const unsigned int pool_stride_x = _pool_info.pad_stride_info().stride().first; |
| 1149 | const unsigned int pool_stride_y = _pool_info.pad_stride_info().stride().second; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1150 | |
| 1151 | // Set step for input in x and y direction for the input |
| 1152 | Window window_input(window); |
| 1153 | unsigned int window_x_inc = 0; |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1154 | switch(_input->info()->data_type()) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1155 | { |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1156 | case DataType::QS8: |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1157 | case DataType::QS16: |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1158 | case DataType::F16: |
| 1159 | { |
| 1160 | window_x_inc = (pool_stride_x == 2) ? _num_elems_processed_per_iteration * 2 : _num_elems_processed_per_iteration; |
| 1161 | break; |
| 1162 | } |
| 1163 | case DataType::F32: |
| 1164 | { |
| 1165 | window_x_inc = pool_stride_x; |
| 1166 | break; |
| 1167 | } |
| 1168 | default: |
| 1169 | { |
| 1170 | ARM_COMPUTE_ERROR("Not supported"); |
| 1171 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1172 | } |
| 1173 | window_input.set(Window::DimX, Window::Dimension(window.x().start() * pool_stride_x, window.x().end() * pool_stride_x, window_x_inc)); |
| 1174 | window_input.set(Window::DimY, Window::Dimension(window.y().start() * pool_stride_y, window.y().end() * pool_stride_y, pool_stride_y)); |
| 1175 | |
| 1176 | // Run function |
| 1177 | (this->*_func)(window_input, window); |
| 1178 | } |