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" |
| 32 | #include "arm_compute/core/TensorInfo.h" |
| 33 | #include "arm_compute/core/Utils.h" |
| 34 | #include "arm_compute/core/Validate.h" |
| 35 | #include "arm_compute/core/Window.h" |
| 36 | |
| 37 | #include <algorithm> |
| 38 | #include <arm_neon.h> |
| 39 | #include <limits> |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 40 | #include <set> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 41 | #include <string> |
| 42 | #include <tuple> |
| 43 | |
| 44 | using namespace arm_compute; |
| 45 | |
| 46 | namespace |
| 47 | { |
| 48 | inline float calculate_avg_scale(const Coordinates &id, const int pool_size, const int upper_bound_w, const int upper_bound_h, |
| 49 | const int pad_x, const int pad_y, const int stride_x, const int stride_y) |
| 50 | { |
| 51 | int start_x = id.x() * stride_x - pad_x; |
| 52 | int start_y = id.y() * stride_y - pad_y; |
| 53 | int end_x = std::min(start_x + pool_size, upper_bound_w); |
| 54 | int end_y = std::min(start_y + pool_size, upper_bound_h); |
| 55 | return 1.f / ((end_y - start_y) * (end_x - start_x)); |
| 56 | } |
| 57 | |
| 58 | inline qint8_t calculate_avg_scale_q8(const Coordinates &id, int pool_size, int upper_bound_w, int upper_bound_h, |
| 59 | int pad_x, int pad_y, int stride_x, int stride_y, int fixed_point_position) |
| 60 | { |
| 61 | static std::array<qint8_t, 10> scale_values_q8 = |
| 62 | { { 0x0, 0x0, 0x40, 0x2A, 0x20, 0x19, 0x15, 0x12, 0x10, 0xE } }; |
| 63 | const int start_x = id.x() * stride_x - pad_x; |
| 64 | const int start_y = id.y() * stride_y - pad_y; |
| 65 | const int end_x = std::min(start_x + pool_size, upper_bound_w); |
| 66 | const int end_y = std::min(start_y + pool_size, upper_bound_h); |
| 67 | const int val = ((end_y - start_y) * (end_x - start_x)); |
| 68 | return scale_values_q8[val] >> (7 - fixed_point_position); |
| 69 | } |
| 70 | } // namespace |
| 71 | |
| 72 | NEPoolingLayerKernel::NEPoolingLayerKernel() |
| 73 | : _func(nullptr), _input(nullptr), _output(nullptr), _pool_info(), _num_elems_processed_per_iteration(0), _border_size(0) |
| 74 | { |
| 75 | } |
| 76 | |
| 77 | BorderSize NEPoolingLayerKernel::border_size() const |
| 78 | { |
| 79 | return _border_size; |
| 80 | } |
| 81 | |
| 82 | void NEPoolingLayerKernel::configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info) |
| 83 | { |
Gian Marco Iodice | 4e28869 | 2017-06-27 11:41:59 +0100 | [diff] [blame] | 84 | int pool_pad_x = 0; |
| 85 | int pool_pad_y = 0; |
| 86 | int pool_stride_x = 0; |
| 87 | int pool_stride_y = 0; |
| 88 | unsigned int pooled_w = 0; |
| 89 | unsigned int pooled_h = 0; |
| 90 | PoolingType pool_type = pool_info.pool_type(); |
| 91 | int pool_size = pool_info.pool_size(); |
| 92 | const PadStrideInfo pad_stride_info = pool_info.pad_stride_info(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 93 | std::tie(pool_pad_x, pool_pad_y) = pad_stride_info.pad(); |
| 94 | std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); |
| 95 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 96 | static const std::set<int> supported_pool_sizes = { 2, 3, 7 }; |
| 97 | ARM_COMPUTE_UNUSED(supported_pool_sizes); |
| 98 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 99 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F32); |
Georgios Pinitas | 1dad50e | 2017-07-03 17:51:34 +0100 | [diff] [blame] | 100 | ARM_COMPUTE_ERROR_ON_NULLPTR(output); |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 101 | ARM_COMPUTE_ERROR_ON(supported_pool_sizes.find(pool_size) == supported_pool_sizes.end()); |
| 102 | ARM_COMPUTE_ERROR_ON(7 == pool_size && input->info()->data_type() != DataType::F32); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 103 | ARM_COMPUTE_ERROR_ON(pool_pad_x >= pool_size || pool_pad_y >= pool_size); |
| 104 | ARM_COMPUTE_ERROR_ON(input->info()->data_type() == DataType::QS8 && pool_type == PoolingType::AVG && input->info()->fixed_point_position() > 6); |
| 105 | ARM_COMPUTE_ERROR_ON(input->info()->data_type() == DataType::QS8 && pool_stride_x > 2); |
| 106 | |
| 107 | // Check output dimensions |
| 108 | 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] | 109 | pool_size, pool_size, pool_info.pad_stride_info()); |
Georgios Pinitas | 1dad50e | 2017-07-03 17:51:34 +0100 | [diff] [blame] | 110 | |
| 111 | // Output auto initialization if not yet initialized |
| 112 | { |
| 113 | TensorShape output_shape{ input->info()->tensor_shape() }; |
| 114 | output_shape.set(0, pooled_w); |
| 115 | output_shape.set(1, pooled_h); |
| 116 | |
| 117 | auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); |
| 118 | } |
| 119 | |
| 120 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 121 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 122 | ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pooled_w) || (output->info()->dimension(1) != pooled_h)); |
| 123 | |
| 124 | unsigned int num_elems_read_per_iteration = 0; |
| 125 | unsigned int num_elems_processed_per_iteration = 0; |
| 126 | unsigned int num_elems_horizontal_window = 0; |
| 127 | |
| 128 | // Select element size |
| 129 | switch(input->info()->data_type()) |
| 130 | { |
| 131 | case DataType::QS8: |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 132 | num_elems_read_per_iteration = 16; |
| 133 | switch(pool_size) |
| 134 | { |
| 135 | case 2: |
| 136 | num_elems_processed_per_iteration = 8; |
| 137 | break; |
| 138 | case 3: |
| 139 | num_elems_processed_per_iteration = 7; |
| 140 | break; |
| 141 | default: |
| 142 | ARM_COMPUTE_ERROR("Pooling size not supported"); |
| 143 | } |
| 144 | num_elems_horizontal_window = 8; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 145 | break; |
| 146 | case DataType::F32: |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 147 | switch(pool_size) |
| 148 | { |
| 149 | case 2: |
| 150 | num_elems_read_per_iteration = 2; |
| 151 | break; |
| 152 | case 3: |
| 153 | num_elems_read_per_iteration = 4; // We use vload4 for pooling3 |
| 154 | break; |
| 155 | case 7: |
| 156 | num_elems_read_per_iteration = 8; // We use vload8 for pooling7 |
| 157 | break; |
| 158 | default: |
| 159 | ARM_COMPUTE_ERROR("Pooling size not supported"); |
| 160 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 161 | num_elems_processed_per_iteration = 1; |
| 162 | num_elems_horizontal_window = 1; |
| 163 | break; |
| 164 | default: |
| 165 | ARM_COMPUTE_ERROR("Element size not supported"); |
| 166 | break; |
| 167 | } |
| 168 | |
| 169 | _num_elems_processed_per_iteration = num_elems_processed_per_iteration; |
| 170 | const int input_width = input->info()->dimension(0); |
| 171 | const int input_height = input->info()->dimension(1); |
| 172 | const int upper_bound_w = ((pooled_w - 1) * pool_stride_x - pool_pad_x + num_elems_read_per_iteration) - input_width; |
| 173 | const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_y + pool_size) - input_height; |
| 174 | |
| 175 | // Set instance variables |
| 176 | _input = input; |
| 177 | _output = output; |
| 178 | _pool_info = pool_info; |
| 179 | _border_size = BorderSize(pool_pad_y, pool_pad_x); |
| 180 | _border_size.right = std::max(upper_bound_w, pool_pad_x); |
| 181 | _border_size.bottom = std::max(upper_bound_h, pool_pad_y); |
| 182 | |
| 183 | // Select appropriate function |
| 184 | switch(pool_size) |
| 185 | { |
| 186 | case 2: |
| 187 | if(input->info()->data_type() == DataType::QS8) |
| 188 | { |
| 189 | _func = (PoolingType::AVG == pool_type) ? &NEPoolingLayerKernel::pooling2_q8<PoolingType::AVG> : &NEPoolingLayerKernel::pooling2_q8<PoolingType::MAX>; |
| 190 | } |
| 191 | else if(input->info()->data_type() == DataType::F32) |
| 192 | { |
| 193 | _func = (PoolingType::AVG == pool_type) ? &NEPoolingLayerKernel::pooling2_f32<PoolingType::AVG> : &NEPoolingLayerKernel::pooling2_f32<PoolingType::MAX>; |
| 194 | } |
| 195 | break; |
| 196 | case 3: |
| 197 | if(input->info()->data_type() == DataType::QS8) |
| 198 | { |
| 199 | _func = (PoolingType::AVG == pool_type) ? &NEPoolingLayerKernel::pooling3_q8<PoolingType::AVG> : &NEPoolingLayerKernel::pooling3_q8<PoolingType::MAX>; |
| 200 | } |
| 201 | else if(input->info()->data_type() == DataType::F32) |
| 202 | { |
| 203 | _func = (PoolingType::AVG == pool_type) ? &NEPoolingLayerKernel::pooling3_f32<PoolingType::AVG> : &NEPoolingLayerKernel::pooling3_f32<PoolingType::MAX>; |
| 204 | } |
| 205 | break; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 206 | case 7: |
| 207 | _func = (PoolingType::AVG == pool_type) ? &NEPoolingLayerKernel::pooling7_f32<PoolingType::AVG> : &NEPoolingLayerKernel::pooling7_f32<PoolingType::MAX>; |
| 208 | break; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 209 | default: |
| 210 | ARM_COMPUTE_ERROR("Unsupported pooling size"); |
| 211 | break; |
| 212 | } |
| 213 | |
| 214 | // Configure kernel window |
| 215 | Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration)); |
| 216 | AccessWindowStatic input_access(input->info(), -pool_pad_x, -pool_pad_y, input_width + _border_size.right, input_height + _border_size.bottom); |
| 217 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_horizontal_window); |
| 218 | update_window_and_padding(win, input_access, output_access); |
| 219 | output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); |
| 220 | INEKernel::configure(win); |
| 221 | } |
| 222 | |
| 223 | template <PoolingType pooling_type> |
| 224 | void NEPoolingLayerKernel::pooling2_q8(const Window &window_input, const Window &window) |
| 225 | { |
| 226 | Iterator input(_input, window_input); |
| 227 | Iterator output(_output, window); |
| 228 | |
| 229 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 230 | constexpr int pool_size = 2; |
| 231 | int pool_pad_x = 0; |
| 232 | int pool_pad_y = 0; |
| 233 | int pool_stride_x = 0; |
| 234 | int pool_stride_y = 0; |
| 235 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 236 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 237 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 238 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 239 | |
| 240 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 241 | 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)); |
| 242 | |
| 243 | execute_window_loop(window, [&](const Coordinates & id) |
| 244 | { |
| 245 | const auto top_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_top_ptr + input.offset())); |
| 246 | const auto bottom_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_bottom_ptr + input.offset())); |
| 247 | qint8x8_t res = {}; |
| 248 | if(pooling_type == PoolingType::AVG) |
| 249 | { |
| 250 | // Calculate scale |
| 251 | 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); |
| 252 | const qint8x8_t scale_vec = vdup_n_qs8(scale); |
| 253 | |
| 254 | // Perform pooling |
| 255 | const qint8x16_t sum_data = vqaddq_qs8(top_data, bottom_data); |
| 256 | res = vqmul_qs8(vpadd_s8(vget_low_s8(sum_data), vget_high_s8(sum_data)), scale_vec, fixed_point_position); |
| 257 | } |
| 258 | else |
| 259 | { |
| 260 | const qint8x16_t max_data = vmaxq_s8(top_data, bottom_data); |
| 261 | res = vpmax_s8(vget_low_s8(max_data), vget_high_s8(max_data)); |
| 262 | } |
| 263 | vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), res); |
| 264 | }, |
| 265 | input, output); |
| 266 | } |
| 267 | |
| 268 | template <PoolingType pooling_type> |
| 269 | void NEPoolingLayerKernel::pooling2_f32(const Window &window_input, const Window &window) |
| 270 | { |
| 271 | Iterator input(_input, window_input); |
| 272 | Iterator output(_output, window); |
| 273 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 274 | constexpr int pool_size = 2; |
| 275 | int pool_pad_x = 0; |
| 276 | int pool_pad_y = 0; |
| 277 | int pool_stride_x = 0; |
| 278 | int pool_stride_y = 0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 279 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 280 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 281 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 282 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 283 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 284 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 285 | 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] | 286 | |
| 287 | execute_window_loop(window, [&](const Coordinates & id) |
| 288 | { |
| 289 | const float32x2_t top_data = vld1_f32(reinterpret_cast<const float *>(input_top_ptr + input.offset())); |
| 290 | const float32x2_t bottom_data = vld1_f32(reinterpret_cast<const float *>(input_bottom_ptr + input.offset())); |
| 291 | float32x2_t res = {}; |
| 292 | if(pooling_type == PoolingType::AVG) |
| 293 | { |
| 294 | // Calculate scale |
| 295 | 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); |
| 296 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 297 | |
| 298 | // Perform pooling |
| 299 | const float32x2_t sum_data = vadd_f32(top_data, bottom_data); |
| 300 | res = vmul_f32(vpadd_f32(sum_data, sum_data), scale_v); |
| 301 | } |
| 302 | else |
| 303 | { |
| 304 | const float32x2_t max_data = vmax_f32(top_data, bottom_data); |
| 305 | res = vpmax_f32(max_data, max_data); |
| 306 | } |
| 307 | *(reinterpret_cast<float *>(output.ptr())) = vget_lane_f32(res, 0); |
| 308 | }, |
| 309 | input, output); |
| 310 | } |
| 311 | |
| 312 | template <PoolingType pooling_type> |
| 313 | void NEPoolingLayerKernel::pooling3_q8(const Window &window_input, const Window &window) |
| 314 | { |
| 315 | Iterator input(_input, window_input); |
| 316 | Iterator output(_output, window); |
| 317 | |
| 318 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 319 | constexpr int pool_size = 3; |
| 320 | int pool_pad_x = 0; |
| 321 | int pool_pad_y = 0; |
| 322 | int pool_stride_x = 0; |
| 323 | int pool_stride_y = 0; |
| 324 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 325 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 326 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 327 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 328 | |
| 329 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 330 | 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)); |
| 331 | 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)); |
| 332 | |
| 333 | execute_window_loop(window, [&](const Coordinates & id) |
| 334 | { |
| 335 | const auto top_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_top_ptr + input.offset())); |
| 336 | const auto middle_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_middle_ptr + input.offset())); |
| 337 | const auto bottom_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_bottom_ptr + input.offset())); |
| 338 | qint8x8_t res = {}; |
| 339 | if(pooling_type == PoolingType::AVG) |
| 340 | { |
| 341 | // Calculate scale |
| 342 | 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); |
| 343 | const qint8x8_t scale_vec = vdup_n_qs8(scale); |
| 344 | |
| 345 | // Perform pooling for stride 2 |
| 346 | const qint8x16_t sum_data = vqaddq_qs8(vqaddq_qs8(top_data, bottom_data), middle_data); |
| 347 | const qint8x16_t sum_data2 = vextq_s8(sum_data, sum_data, 1); |
| 348 | const qint8x16_t sum_data3 = vextq_s8(sum_data, sum_data, 2); |
| 349 | const qint8x16_t final_sum = vqaddq_qs8(vqaddq_qs8(sum_data, sum_data2), sum_data3); |
| 350 | if(pool_stride_x == 2) |
| 351 | { |
| 352 | const qint8x8x2_t table = { { vget_low_s8(final_sum), vget_high_s8(final_sum) } }; |
| 353 | static const qint8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; |
| 354 | res = vtbl2_s8(table, lookup_val); |
| 355 | } |
| 356 | else |
| 357 | { |
| 358 | res = vget_low_s8(final_sum); |
| 359 | } |
| 360 | res = vqmul_qs8(res, scale_vec, fixed_point_position); |
| 361 | } |
| 362 | else |
| 363 | { |
| 364 | const qint8x16_t max_data = vmaxq_s8(vmaxq_s8(top_data, bottom_data), middle_data); |
| 365 | const qint8x16_t max_data2 = vextq_s8(max_data, max_data, 1); |
| 366 | const qint8x16_t max_data3 = vextq_s8(max_data, max_data, 2); |
| 367 | const qint8x16_t final_max = vmaxq_s8(vmaxq_s8(max_data, max_data2), max_data3); |
| 368 | |
| 369 | if(pool_stride_x == 2) |
| 370 | { |
| 371 | const qint8x8x2_t table = { { vget_low_s8(final_max), vget_high_s8(final_max) } }; |
| 372 | static const qint8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; |
| 373 | res = vtbl2_s8(table, lookup_val); |
| 374 | } |
| 375 | else |
| 376 | { |
| 377 | res = vget_low_s8(final_max); |
| 378 | } |
| 379 | } |
| 380 | vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), res); |
| 381 | }, |
| 382 | input, output); |
| 383 | } |
| 384 | |
| 385 | template <PoolingType pooling_type> |
| 386 | void NEPoolingLayerKernel::pooling3_f32(const Window &window_input, const Window &window) |
| 387 | { |
| 388 | Iterator input(_input, window_input); |
| 389 | Iterator output(_output, window); |
| 390 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 391 | constexpr const int pool_size = 3; |
| 392 | int pool_pad_x = 0; |
| 393 | int pool_pad_y = 0; |
| 394 | int pool_stride_x = 0; |
| 395 | int pool_stride_y = 0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 396 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 397 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 398 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 399 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 400 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 401 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 402 | 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)); |
| 403 | 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] | 404 | |
| 405 | execute_window_loop(window, [&](const Coordinates & id) |
| 406 | { |
| 407 | const float32x4_t top_data = vld1q_f32(reinterpret_cast<const float *>(input_top_ptr + input.offset())); |
| 408 | const float32x4_t middle_data = vld1q_f32(reinterpret_cast<const float *>(input_middle_ptr + input.offset())); |
| 409 | const float32x4_t bottom_data = vld1q_f32(reinterpret_cast<const float *>(input_bottom_ptr + input.offset())); |
| 410 | float32x2_t res = {}; |
| 411 | if(pooling_type == PoolingType::AVG) |
| 412 | { |
| 413 | // Calculate scale |
| 414 | 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); |
| 415 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 416 | |
| 417 | // Perform pooling |
| 418 | const float32x4_t sum_data = vaddq_f32(vaddq_f32(top_data, bottom_data), middle_data); |
| 419 | res = vpadd_f32(vget_high_f32(vsetq_lane_f32(0.f, sum_data, 3)), vget_low_f32(sum_data)); |
| 420 | res = vmul_f32(vpadd_f32(res, res), scale_v); |
| 421 | } |
| 422 | else |
| 423 | { |
| 424 | const float32x4_t max_data = vmaxq_f32(vmaxq_f32(top_data, bottom_data), middle_data); |
| 425 | res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits<float>::max(), max_data, 3)), vget_low_f32(max_data)); |
| 426 | res = vpmax_f32(res, res); |
| 427 | } |
| 428 | *(reinterpret_cast<float *>(output.ptr())) = vget_lane_f32(res, 0); |
| 429 | }, |
| 430 | input, output); |
| 431 | } |
| 432 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 433 | template <PoolingType pooling_type> |
| 434 | void NEPoolingLayerKernel::pooling7_f32(const Window &window_input, const Window &window) |
| 435 | { |
| 436 | Iterator input(_input, window_input); |
| 437 | Iterator output(_output, window); |
| 438 | |
| 439 | constexpr const int pool_size = 7; |
| 440 | int pool_pad_x = 0; |
| 441 | int pool_pad_y = 0; |
| 442 | int pool_stride_x = 0; |
| 443 | int pool_stride_y = 0; |
| 444 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 445 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 446 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 447 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 448 | |
| 449 | std::array<const uint8_t *, pool_size> input_ptrs{ {} }; |
| 450 | for(int i = 0; i < pool_size; ++i) |
| 451 | { |
| 452 | input_ptrs[i] = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + i)); |
| 453 | } |
| 454 | |
| 455 | execute_window_loop(window, [&](const Coordinates & id) |
| 456 | { |
| 457 | float32x2_t res = {}; |
| 458 | if(pooling_type == PoolingType::AVG) |
| 459 | { |
| 460 | // Calculate scale |
| 461 | 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); |
| 462 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 463 | |
| 464 | // Perform pooling |
| 465 | float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[0] + input.offset())); |
| 466 | float32x4_t sum_data = vaddq_f32(data.val[0], vsetq_lane_f32(0.f, data.val[1], 3)); |
| 467 | for(int i = 1; i < pool_size; ++i) |
| 468 | { |
| 469 | data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[i] + input.offset())); |
| 470 | sum_data = vaddq_f32(sum_data, data.val[0]); |
| 471 | sum_data = vaddq_f32(sum_data, vsetq_lane_f32(0.f, data.val[1], 3)); |
| 472 | } |
| 473 | res = vpadd_f32(vget_high_f32(sum_data), vget_low_f32(sum_data)); |
| 474 | res = vmul_f32(vpadd_f32(res, res), scale_v); |
| 475 | } |
| 476 | else |
| 477 | { |
| 478 | float32x4x2_t max_data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[0] + input.offset())); |
| 479 | for(int i = 1; i < pool_size; ++i) |
| 480 | { |
| 481 | const float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[i] + input.offset())); |
| 482 | max_data = vmax2q_f32(max_data, data); |
| 483 | } |
| 484 | 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])); |
| 485 | res = vpmax_f32(res, vpmax_f32(vget_high_f32(max_data.val[0]), vget_low_f32(max_data.val[0]))); |
| 486 | res = vpmax_f32(res, res); |
| 487 | } |
| 488 | *(reinterpret_cast<float *>(output.ptr())) = vget_lane_f32(res, 0); |
| 489 | }, |
| 490 | input, output); |
| 491 | } |
| 492 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 493 | void NEPoolingLayerKernel::run(const Window &window) |
| 494 | { |
| 495 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 496 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 497 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 498 | |
| 499 | unsigned int pool_stride_x, pool_stride_y = 0; |
| 500 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 501 | |
| 502 | // Set step for input in x and y direction for the input |
| 503 | Window window_input(window); |
| 504 | unsigned int window_x_inc = 0; |
| 505 | if(_input->info()->data_type() == DataType::QS8) |
| 506 | { |
| 507 | window_x_inc = (pool_stride_x == 2) ? _num_elems_processed_per_iteration * 2 : _num_elems_processed_per_iteration; |
| 508 | } |
| 509 | else |
| 510 | { |
| 511 | window_x_inc = pool_stride_x; |
| 512 | } |
| 513 | window_input.set(Window::DimX, Window::Dimension(window.x().start() * pool_stride_x, window.x().end() * pool_stride_x, window_x_inc)); |
| 514 | window_input.set(Window::DimY, Window::Dimension(window.y().start() * pool_stride_y, window.y().end() * pool_stride_y, pool_stride_y)); |
| 515 | |
| 516 | // Run function |
| 517 | (this->*_func)(window_input, window); |
| 518 | } |