Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2018 ARM Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 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" |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 31 | #include "arm_compute/core/NEON/NEAsymm.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 32 | #include "arm_compute/core/NEON/NEFixedPoint.h" |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 33 | #include "arm_compute/core/NEON/NEMath.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 34 | #include "arm_compute/core/TensorInfo.h" |
| 35 | #include "arm_compute/core/Utils.h" |
| 36 | #include "arm_compute/core/Validate.h" |
| 37 | #include "arm_compute/core/Window.h" |
| 38 | |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 39 | #include "support/ToolchainSupport.h" |
| 40 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 41 | #include <algorithm> |
| 42 | #include <arm_neon.h> |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 43 | #include <cmath> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 44 | #include <limits> |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 45 | #include <set> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 46 | #include <string> |
| 47 | #include <tuple> |
| 48 | |
| 49 | using namespace arm_compute; |
| 50 | |
| 51 | namespace |
| 52 | { |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 53 | void auto_init(const ITensorInfo *input, ITensorInfo *output, unsigned int pooled_w, unsigned int pooled_h) |
| 54 | { |
| 55 | TensorShape output_shape{ input->tensor_shape() }; |
| 56 | output_shape.set(0, pooled_w); |
| 57 | output_shape.set(1, pooled_h); |
| 58 | |
| 59 | auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape)); |
| 60 | } |
| 61 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 62 | template <bool exclude_padding> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 63 | inline float calculate_avg_scale(const Coordinates &id, const int pool_size, const int upper_bound_w, const int upper_bound_h, |
| 64 | const int pad_x, const int pad_y, const int stride_x, const int stride_y) |
| 65 | { |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 66 | int start_x = id.x() * stride_x - pad_x; |
| 67 | int start_y = id.y() * stride_y - pad_y; |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 68 | const int end_x = std::min(start_x + pool_size, upper_bound_w); |
| 69 | const int end_y = std::min(start_y + pool_size, upper_bound_h); |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 70 | if(exclude_padding) |
| 71 | { |
| 72 | start_x = std::max(0, start_x); |
| 73 | start_y = std::max(0, start_y); |
| 74 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 75 | return 1.f / ((end_y - start_y) * (end_x - start_x)); |
| 76 | } |
| 77 | |
| 78 | inline qint8_t calculate_avg_scale_q8(const Coordinates &id, int pool_size, int upper_bound_w, int upper_bound_h, |
| 79 | int pad_x, int pad_y, int stride_x, int stride_y, int fixed_point_position) |
| 80 | { |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 81 | static const std::array<qint8_t, 10> scale_values_q8 = |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 82 | { { 0x0, 0x0, 0x40, 0x2A, 0x20, 0x19, 0x15, 0x12, 0x10, 0xE } }; |
| 83 | const int start_x = id.x() * stride_x - pad_x; |
| 84 | const int start_y = id.y() * stride_y - pad_y; |
| 85 | const int end_x = std::min(start_x + pool_size, upper_bound_w); |
| 86 | const int end_y = std::min(start_y + pool_size, upper_bound_h); |
| 87 | const int val = ((end_y - start_y) * (end_x - start_x)); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 88 | return sshr_qs8(scale_values_q8[val], (7 - fixed_point_position)); |
| 89 | } |
| 90 | |
| 91 | inline qint16_t calculate_avg_scale_q16(const Coordinates &id, int pool_size, int upper_bound_w, int upper_bound_h, |
| 92 | int pad_x, int pad_y, int stride_x, int stride_y, int fixed_point_position) |
| 93 | { |
| 94 | static std::array<qint16_t, 10> scale_values_q16 = |
| 95 | { { 0x0, 0x0, 0x4000, 0x2AAB, 0x2000, 0x199A, 0x1555, 0x1249, 0x1000, 0xE38 } }; |
| 96 | const int start_x = id.x() * stride_x - pad_x; |
| 97 | const int start_y = id.y() * stride_y - pad_y; |
| 98 | const int end_x = std::min(start_x + pool_size, upper_bound_w); |
| 99 | const int end_y = std::min(start_y + pool_size, upper_bound_h); |
| 100 | const int val = ((end_y - start_y) * (end_x - start_x)); |
| 101 | return sshr_qs16(scale_values_q16[val], (15 - fixed_point_position)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 102 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 103 | |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 104 | template <bool exclude_padding> |
| 105 | inline void scale_vector_s16x8(uint16x8_t &v, const Coordinates &id, int id_offset, int step, |
| 106 | const int pool_size, const int upper_bound_w, const int upper_bound_h, |
| 107 | const int pad_x, const int pad_y, const int stride_x, const int stride_y) |
| 108 | { |
| 109 | int start_x = (id.x() + id_offset) * stride_x - pad_x; |
| 110 | int start_y = id.y() * stride_y - pad_y; |
| 111 | const int end_y = std::min(start_y + pool_size, upper_bound_h); |
| 112 | if(exclude_padding) |
| 113 | { |
| 114 | start_y = std::max(0, start_y); |
| 115 | } |
| 116 | |
| 117 | std::array<uint16_t, 8> elems = |
| 118 | { |
| 119 | { |
| 120 | vgetq_lane_u16(v, 0), |
| 121 | vgetq_lane_u16(v, 1), |
| 122 | vgetq_lane_u16(v, 2), |
| 123 | vgetq_lane_u16(v, 3), |
| 124 | vgetq_lane_u16(v, 4), |
| 125 | vgetq_lane_u16(v, 5), |
| 126 | vgetq_lane_u16(v, 6), |
| 127 | vgetq_lane_u16(v, 7), |
| 128 | } |
| 129 | }; |
| 130 | |
| 131 | for(auto &el : elems) |
| 132 | { |
| 133 | int c_start_x = start_x; |
| 134 | const int end_x = std::min(c_start_x + pool_size, upper_bound_w); |
| 135 | if(exclude_padding) |
| 136 | { |
| 137 | c_start_x = std::max(0, c_start_x); |
| 138 | } |
| 139 | float scale = 1.f / ((end_y - start_y) * (end_x - c_start_x)); |
| 140 | el *= scale; |
| 141 | start_x += step * stride_x; |
| 142 | } |
| 143 | |
| 144 | v = vsetq_lane_u16(elems[0], v, 0); |
| 145 | v = vsetq_lane_u16(elems[1], v, 1); |
| 146 | v = vsetq_lane_u16(elems[2], v, 2); |
| 147 | v = vsetq_lane_u16(elems[3], v, 3); |
| 148 | v = vsetq_lane_u16(elems[4], v, 4); |
| 149 | v = vsetq_lane_u16(elems[5], v, 5); |
| 150 | v = vsetq_lane_u16(elems[6], v, 6); |
| 151 | v = vsetq_lane_u16(elems[7], v, 7); |
| 152 | } |
| 153 | |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 154 | Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, unsigned int &pooled_w, unsigned int pooled_h, int pool_size) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 155 | { |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 156 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 157 | |
Georgios Pinitas | 4c2dd54 | 2017-11-13 12:58:41 +0000 | [diff] [blame] | 158 | int pool_pad_x = 0; |
| 159 | int pool_pad_y = 0; |
| 160 | int pool_stride_x = 0; |
| 161 | int pool_stride_y = 0; |
Georgios Pinitas | 4c2dd54 | 2017-11-13 12:58:41 +0000 | [diff] [blame] | 162 | PoolingType pool_type = pool_info.pool_type(); |
Georgios Pinitas | 4c2dd54 | 2017-11-13 12:58:41 +0000 | [diff] [blame] | 163 | const PadStrideInfo pad_stride_info = pool_info.pad_stride_info(); |
| 164 | const bool exclude_padding = pool_info.exclude_padding(); |
| 165 | const bool is_global_pooling = pool_info.is_global_pooling(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 166 | std::tie(pool_pad_x, pool_pad_y) = pad_stride_info.pad(); |
| 167 | std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame] | 168 | static const std::set<int> supported_pool_sizes = { 2, 3 }; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 169 | |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 170 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); |
| 171 | ARM_COMPUTE_RETURN_ERROR_ON(pool_type == PoolingType::L2 && is_data_type_quantized(input->data_type())); |
| 172 | ARM_COMPUTE_RETURN_ERROR_ON((supported_pool_sizes.find(pool_size) == supported_pool_sizes.end()) && ((input->data_type() != DataType::F32) && (input->data_type() != DataType::QASYMM8))); |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 173 | ARM_COMPUTE_RETURN_ERROR_ON(!is_global_pooling && (pool_pad_x >= pool_size || pool_pad_y >= pool_size)); |
| 174 | ARM_COMPUTE_RETURN_ERROR_ON(is_global_pooling && (input->tensor_shape().x() != input->tensor_shape().y())); |
| 175 | ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_fixed_point(input->data_type()) && pool_stride_x > 2); |
| 176 | ARM_COMPUTE_RETURN_ERROR_ON(exclude_padding && is_data_type_fixed_point(input->data_type())); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 177 | |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 178 | if(output->total_size() != 0) |
Georgios Pinitas | 1dad50e | 2017-07-03 17:51:34 +0100 | [diff] [blame] | 179 | { |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 180 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 181 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); |
| 182 | ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(0) != pooled_w) || (output->dimension(1) != pooled_h)); |
Georgios Pinitas | 1dad50e | 2017-07-03 17:51:34 +0100 | [diff] [blame] | 183 | } |
| 184 | |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 185 | return Status{}; |
| 186 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 187 | |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 188 | Status validate_arguments_pool_info(const ITensorInfo *input, const PoolingLayerInfo &pool_info, const unsigned int pool_size) |
| 189 | { |
| 190 | const bool is_global_pooling = pool_info.is_global_pooling(); |
| 191 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_global_pooling && (input->tensor_shape().x() != input->tensor_shape().y()), |
| 192 | "Global pooling is supported only with rectangular inputs!"); |
| 193 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_global_pooling && ((pool_info.pad_stride_info().pad().first >= pool_size) || (pool_info.pad_stride_info().pad().second >= pool_size)), |
| 194 | "Invalid pool size and pool pad combination!"); |
| 195 | |
| 196 | return Status{}; |
| 197 | } |
| 198 | |
| 199 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const PoolingLayerInfo &pool_info, unsigned int &num_elems_processed_per_iteration, |
| 200 | BorderSize &border_size, |
| 201 | unsigned int pooled_w, unsigned int pooled_h, int pool_size) |
| 202 | { |
| 203 | unsigned int num_elems_read_per_iteration = 0; |
| 204 | unsigned int num_elems_horizontal_window = 0; |
| 205 | int pool_pad_x = 0; |
| 206 | int pool_pad_y = 0; |
| 207 | int pool_stride_x = 0; |
| 208 | int pool_stride_y = 0; |
| 209 | const int input_width = input->dimension(0); |
| 210 | const int input_height = input->dimension(1); |
| 211 | const PadStrideInfo pad_stride_info = pool_info.pad_stride_info(); |
| 212 | std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); |
| 213 | std::tie(pool_pad_x, pool_pad_y) = pad_stride_info.pad(); |
| 214 | |
| 215 | // Check output dimensions |
| 216 | std::tie(pooled_w, pooled_h) = scaled_dimensions(input->dimension(0), |
| 217 | input->dimension(1), |
| 218 | pool_size, |
| 219 | pool_size, |
| 220 | pad_stride_info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 221 | |
| 222 | // Select element size |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 223 | switch(input->data_type()) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 224 | { |
| 225 | case DataType::QS8: |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 226 | num_elems_read_per_iteration = 16; |
| 227 | switch(pool_size) |
| 228 | { |
| 229 | case 2: |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 230 | num_elems_processed_per_iteration = (pool_stride_x == 2) ? 8 : 15; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 231 | break; |
| 232 | case 3: |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 233 | num_elems_processed_per_iteration = (pool_stride_x == 2) ? 7 : 14; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 234 | break; |
| 235 | default: |
| 236 | ARM_COMPUTE_ERROR("Pooling size not supported"); |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 237 | break; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 238 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 239 | num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; |
| 240 | break; |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 241 | case DataType::QASYMM8: |
| 242 | switch(pool_size) |
| 243 | { |
| 244 | case 2: |
| 245 | num_elems_read_per_iteration = 16; |
| 246 | num_elems_processed_per_iteration = (pool_stride_x == 2) ? 8 : 15; |
| 247 | num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; |
| 248 | break; |
| 249 | case 3: |
| 250 | num_elems_read_per_iteration = 16; |
| 251 | num_elems_processed_per_iteration = (pool_stride_x == 2) ? 7 : 14; |
| 252 | num_elems_horizontal_window = (pool_stride_x == 2) ? 8 : 16; |
| 253 | break; |
| 254 | default: |
| 255 | num_elems_read_per_iteration = 1; |
| 256 | num_elems_processed_per_iteration = 1; |
| 257 | num_elems_horizontal_window = 1; |
| 258 | break; |
| 259 | } |
| 260 | break; |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 261 | case DataType::QS16: |
| 262 | num_elems_read_per_iteration = 8; |
| 263 | switch(pool_size) |
| 264 | { |
| 265 | case 2: |
| 266 | num_elems_processed_per_iteration = (pool_stride_x == 2) ? 4 : 7; |
| 267 | break; |
| 268 | case 3: |
| 269 | num_elems_processed_per_iteration = (pool_stride_x == 2) ? 3 : 6; |
| 270 | break; |
| 271 | default: |
| 272 | ARM_COMPUTE_ERROR("Pooling size not supported"); |
| 273 | } |
| 274 | num_elems_horizontal_window = (pool_stride_x == 2) ? 4 : 8; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 275 | break; |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 276 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 277 | case DataType::F16: |
| 278 | switch(pool_size) |
| 279 | { |
| 280 | case 2: |
| 281 | num_elems_read_per_iteration = 16; |
| 282 | num_elems_processed_per_iteration = 8; |
| 283 | num_elems_horizontal_window = 8; |
| 284 | break; |
| 285 | case 3: |
| 286 | num_elems_read_per_iteration = 4; |
| 287 | num_elems_processed_per_iteration = 1; |
| 288 | num_elems_horizontal_window = 1; |
| 289 | break; |
| 290 | default: |
| 291 | ARM_COMPUTE_ERROR("Pooling size not supported"); |
| 292 | break; |
| 293 | } |
| 294 | break; |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 295 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 296 | case DataType::F32: |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 297 | switch(pool_size) |
| 298 | { |
| 299 | case 2: |
| 300 | num_elems_read_per_iteration = 2; |
| 301 | break; |
| 302 | case 3: |
| 303 | num_elems_read_per_iteration = 4; // We use vload4 for pooling3 |
| 304 | break; |
| 305 | case 7: |
| 306 | num_elems_read_per_iteration = 8; // We use vload8 for pooling7 |
| 307 | break; |
| 308 | default: |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame] | 309 | 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] | 310 | break; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 311 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 312 | num_elems_processed_per_iteration = 1; |
| 313 | num_elems_horizontal_window = 1; |
| 314 | break; |
| 315 | default: |
| 316 | ARM_COMPUTE_ERROR("Element size not supported"); |
| 317 | break; |
| 318 | } |
| 319 | |
Diego Lopez Recas | 61ef5bf | 2017-12-11 12:36:55 +0000 | [diff] [blame] | 320 | // Number of iterations in X dimension |
| 321 | const int num_iterations_x = (pooled_w + num_elems_processed_per_iteration - 1) / num_elems_processed_per_iteration; |
| 322 | |
| 323 | // Upper limit for the number of right/bottom border elements that are accessed |
| 324 | const int upper_bound_w = ((num_iterations_x - 1) * num_elems_processed_per_iteration * pool_stride_x - pool_pad_x + num_elems_read_per_iteration) - input_width; |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 325 | const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_y + pool_size) - input_height; |
| 326 | |
| 327 | border_size = BorderSize(pool_pad_y, pool_pad_x); |
| 328 | border_size.right = std::max(upper_bound_w, pool_pad_x); |
| 329 | border_size.bottom = std::max(upper_bound_h, pool_pad_y); |
| 330 | bool window_changed = false; |
| 331 | |
| 332 | TensorShape output_shape{ input->tensor_shape() }; |
| 333 | output_shape.set(0, pooled_w); |
| 334 | output_shape.set(1, pooled_h); |
| 335 | TensorInfo output_info(input->clone()->set_tensor_shape(output_shape)); |
| 336 | |
| 337 | Window win = calculate_max_window(output_info, Steps(num_elems_processed_per_iteration)); |
| 338 | AccessWindowStatic input_access(input, -pool_pad_x, -pool_pad_y, input_width + border_size.right, input_height + border_size.bottom); |
| 339 | |
| 340 | if(output->total_size() != 0) |
| 341 | { |
| 342 | AccessWindowHorizontal output_access(output, 0, num_elems_horizontal_window); |
| 343 | window_changed = update_window_and_padding(win, input_access, output_access); |
| 344 | output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); |
| 345 | } |
| 346 | else |
| 347 | { |
| 348 | window_changed = update_window_and_padding(win, input_access); |
| 349 | } |
| 350 | |
| 351 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 352 | return std::make_pair(err, win); |
| 353 | } |
| 354 | } // namespace |
| 355 | |
| 356 | NEPoolingLayerKernel::NEPoolingLayerKernel() |
| 357 | : _func(nullptr), _input(nullptr), _output(nullptr), _pool_info(), _num_elems_processed_per_iteration(0), _border_size(0) |
| 358 | { |
| 359 | } |
| 360 | |
| 361 | BorderSize NEPoolingLayerKernel::border_size() const |
| 362 | { |
| 363 | return _border_size; |
| 364 | } |
| 365 | |
| 366 | void NEPoolingLayerKernel::configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info) |
| 367 | { |
| 368 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 369 | |
Diego Lopez Recas | 61ef5bf | 2017-12-11 12:36:55 +0000 | [diff] [blame] | 370 | const PoolingType pool_type = pool_info.pool_type(); |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 371 | const PadStrideInfo pad_stride_info = pool_info.pad_stride_info(); |
| 372 | const bool exclude_padding = pool_info.exclude_padding(); |
| 373 | const bool is_global_pooling = pool_info.is_global_pooling(); |
Diego Lopez Recas | 61ef5bf | 2017-12-11 12:36:55 +0000 | [diff] [blame] | 374 | const int pool_stride_x = pad_stride_info.stride().first; |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 375 | |
| 376 | // Update pool size in case of global pooling |
Diego Lopez Recas | 61ef5bf | 2017-12-11 12:36:55 +0000 | [diff] [blame] | 377 | const int pool_size = is_global_pooling ? input->info()->dimension(0) : pool_info.pool_size(); |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 378 | |
| 379 | // Validate pool info before calling scaled_dimensions |
| 380 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_pool_info(input->info(), pool_info, pool_size)); |
| 381 | |
| 382 | // Check output dimensions |
Diego Lopez Recas | 61ef5bf | 2017-12-11 12:36:55 +0000 | [diff] [blame] | 383 | unsigned int pooled_w, pooled_h; |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 384 | std::tie(pooled_w, pooled_h) = scaled_dimensions(input->info()->dimension(0), |
| 385 | input->info()->dimension(1), |
| 386 | pool_size, |
| 387 | pool_size, |
Diego Lopez Recas | 61ef5bf | 2017-12-11 12:36:55 +0000 | [diff] [blame] | 388 | pad_stride_info); |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 389 | |
| 390 | // Output auto initialization if not yet initialized |
| 391 | auto_init(input->info(), output->info(), pooled_w, pooled_h); |
| 392 | |
| 393 | // Perform validation step |
| 394 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), pool_info, pooled_w, pooled_h, pool_size)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 395 | |
| 396 | // Set instance variables |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 397 | _input = input; |
| 398 | _output = output; |
| 399 | _pool_info = pool_info; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 400 | |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 401 | // Get data type |
| 402 | const DataType data_type = input->info()->data_type(); |
| 403 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 404 | // Select appropriate function |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 405 | if(data_type == DataType::QS8) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 406 | { |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 407 | switch(pool_size) |
| 408 | { |
| 409 | case 2: |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 410 | switch(pool_type) |
| 411 | { |
| 412 | case PoolingType::AVG: |
| 413 | _func = &NEPoolingLayerKernel::pooling2_q8<PoolingType::AVG>; |
| 414 | break; |
| 415 | case PoolingType::MAX: |
| 416 | _func = &NEPoolingLayerKernel::pooling2_q8<PoolingType::MAX>; |
| 417 | break; |
| 418 | default: |
| 419 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 420 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 421 | break; |
| 422 | case 3: |
| 423 | switch(pool_type) |
| 424 | { |
| 425 | case PoolingType::AVG: |
| 426 | _func = &NEPoolingLayerKernel::pooling3_q8<PoolingType::AVG>; |
| 427 | break; |
| 428 | case PoolingType::MAX: |
| 429 | _func = &NEPoolingLayerKernel::pooling3_q8<PoolingType::MAX>; |
| 430 | break; |
| 431 | default: |
| 432 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 433 | } |
| 434 | break; |
| 435 | default: |
| 436 | ARM_COMPUTE_ERROR("Unsupported pooling size!"); |
| 437 | } |
| 438 | } |
| 439 | else if(data_type == DataType::QASYMM8) |
| 440 | { |
| 441 | if(pool_size == 2 && pool_stride_x < 3) |
| 442 | { |
| 443 | switch(pool_type) |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 444 | { |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 445 | case PoolingType::AVG: |
| 446 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling2_qasymm8<PoolingType::AVG, true> : &NEPoolingLayerKernel::pooling2_qasymm8<PoolingType::AVG, false>; |
| 447 | break; |
| 448 | case PoolingType::MAX: |
| 449 | _func = &NEPoolingLayerKernel::pooling2_qasymm8<PoolingType::MAX>; |
| 450 | break; |
| 451 | default: |
| 452 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 453 | } |
| 454 | } |
| 455 | else if(pool_size == 3 && pool_stride_x < 3) |
| 456 | { |
| 457 | switch(pool_type) |
| 458 | { |
| 459 | case PoolingType::AVG: |
| 460 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling3_qasymm8<PoolingType::AVG, true> : &NEPoolingLayerKernel::pooling3_qasymm8<PoolingType::AVG, false>; |
| 461 | break; |
| 462 | case PoolingType::MAX: |
| 463 | _func = &NEPoolingLayerKernel::pooling3_qasymm8<PoolingType::MAX>; |
| 464 | break; |
| 465 | default: |
| 466 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 467 | } |
| 468 | } |
| 469 | else |
| 470 | { |
| 471 | switch(pool_type) |
| 472 | { |
| 473 | case PoolingType::AVG: |
| 474 | _func = (exclude_padding) ? &NEPoolingLayerKernel::poolingN_qasymm8<PoolingType::AVG, true> : &NEPoolingLayerKernel::poolingN_qasymm8<PoolingType::AVG, false>; |
| 475 | break; |
| 476 | case PoolingType::MAX: |
| 477 | _func = &NEPoolingLayerKernel::poolingN_qasymm8<PoolingType::MAX>; |
| 478 | break; |
| 479 | default: |
| 480 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 481 | } |
| 482 | } |
| 483 | } |
| 484 | else if(data_type == DataType::QS16) |
| 485 | { |
| 486 | switch(pool_size) |
| 487 | { |
| 488 | case 2: |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 489 | switch(pool_type) |
| 490 | { |
| 491 | case PoolingType::AVG: |
| 492 | _func = &NEPoolingLayerKernel::pooling2_q16<PoolingType::AVG>; |
| 493 | break; |
| 494 | case PoolingType::MAX: |
| 495 | _func = &NEPoolingLayerKernel::pooling2_q16<PoolingType::MAX>; |
| 496 | break; |
| 497 | default: |
| 498 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 499 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 500 | break; |
| 501 | case 3: |
| 502 | switch(pool_type) |
| 503 | { |
| 504 | case PoolingType::AVG: |
| 505 | _func = &NEPoolingLayerKernel::pooling3_q16<PoolingType::AVG>; |
| 506 | break; |
| 507 | case PoolingType::MAX: |
| 508 | _func = &NEPoolingLayerKernel::pooling3_q16<PoolingType::MAX>; |
| 509 | break; |
| 510 | default: |
| 511 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 512 | } |
| 513 | break; |
| 514 | default: |
| 515 | ARM_COMPUTE_ERROR("Unsupported pooling size!"); |
| 516 | } |
| 517 | } |
| 518 | else if(data_type == DataType::F16) |
| 519 | { |
| 520 | switch(pool_size) |
| 521 | { |
| 522 | case 2: |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 523 | switch(pool_type) |
| 524 | { |
| 525 | case PoolingType::AVG: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 526 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling2_f16<PoolingType::AVG, true> : &NEPoolingLayerKernel::pooling2_f16<PoolingType::AVG, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 527 | break; |
| 528 | case PoolingType::L2: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 529 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling2_f16<PoolingType::L2, true> : &NEPoolingLayerKernel::pooling2_f16<PoolingType::L2, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 530 | break; |
| 531 | case PoolingType::MAX: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 532 | _func = &NEPoolingLayerKernel::pooling2_f16<PoolingType::MAX, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 533 | break; |
| 534 | default: |
| 535 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 536 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 537 | break; |
| 538 | case 3: |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 539 | switch(pool_type) |
| 540 | { |
| 541 | case PoolingType::AVG: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 542 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling3_f16<PoolingType::AVG, true> : &NEPoolingLayerKernel::pooling3_f16<PoolingType::AVG, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 543 | break; |
| 544 | case PoolingType::L2: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 545 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling3_f16<PoolingType::L2, true> : &NEPoolingLayerKernel::pooling3_f16<PoolingType::L2, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 546 | break; |
| 547 | case PoolingType::MAX: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 548 | _func = &NEPoolingLayerKernel::pooling3_f16<PoolingType::MAX, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 549 | break; |
| 550 | default: |
| 551 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 552 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 553 | break; |
| 554 | default: |
| 555 | ARM_COMPUTE_ERROR("Unsupported pooling size!"); |
| 556 | } |
| 557 | } |
| 558 | else if(data_type == DataType::F32) |
| 559 | { |
| 560 | switch(pool_size) |
| 561 | { |
| 562 | case 2: |
| 563 | switch(pool_type) |
| 564 | { |
| 565 | case PoolingType::AVG: |
| 566 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling2_f32<PoolingType::AVG, true> : &NEPoolingLayerKernel::pooling2_f32<PoolingType::AVG, false>; |
| 567 | break; |
| 568 | case PoolingType::L2: |
| 569 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling2_f32<PoolingType::L2, true> : &NEPoolingLayerKernel::pooling2_f32<PoolingType::L2, false>; |
| 570 | break; |
| 571 | case PoolingType::MAX: |
| 572 | _func = &NEPoolingLayerKernel::pooling2_f32<PoolingType::MAX, false>; |
| 573 | break; |
| 574 | default: |
| 575 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 576 | } |
| 577 | break; |
| 578 | case 3: |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 579 | switch(pool_type) |
| 580 | { |
| 581 | case PoolingType::AVG: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 582 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling3_f32<PoolingType::AVG, true> : &NEPoolingLayerKernel::pooling3_f32<PoolingType::AVG, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 583 | break; |
| 584 | case PoolingType::L2: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 585 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling3_f32<PoolingType::L2, true> : &NEPoolingLayerKernel::pooling3_f32<PoolingType::L2, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 586 | break; |
| 587 | case PoolingType::MAX: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 588 | _func = &NEPoolingLayerKernel::pooling3_f32<PoolingType::MAX, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 589 | break; |
| 590 | default: |
| 591 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 592 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 593 | break; |
| 594 | case 7: |
| 595 | switch(pool_type) |
| 596 | { |
| 597 | case PoolingType::AVG: |
| 598 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling7_f32<PoolingType::AVG, true> : &NEPoolingLayerKernel::pooling7_f32<PoolingType::AVG, false>; |
| 599 | break; |
| 600 | case PoolingType::L2: |
| 601 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling7_f32<PoolingType::L2, true> : &NEPoolingLayerKernel::pooling7_f32<PoolingType::L2, false>; |
| 602 | break; |
| 603 | case PoolingType::MAX: |
| 604 | _func = &NEPoolingLayerKernel::pooling7_f32<PoolingType::MAX, false>; |
| 605 | break; |
| 606 | default: |
| 607 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 608 | } |
| 609 | break; |
| 610 | default: |
| 611 | switch(pool_type) |
| 612 | { |
| 613 | case PoolingType::AVG: |
| 614 | _func = (exclude_padding) ? &NEPoolingLayerKernel::poolingN_f32<PoolingType::AVG, true> : &NEPoolingLayerKernel::poolingN_f32<PoolingType::AVG, false>; |
| 615 | break; |
| 616 | case PoolingType::L2: |
| 617 | _func = (exclude_padding) ? &NEPoolingLayerKernel::poolingN_f32<PoolingType::L2, true> : &NEPoolingLayerKernel::poolingN_f32<PoolingType::L2, false>; |
| 618 | break; |
| 619 | case PoolingType::MAX: |
| 620 | _func = &NEPoolingLayerKernel::poolingN_f32<PoolingType::MAX, false>; |
| 621 | break; |
| 622 | default: |
| 623 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 624 | } |
| 625 | break; |
| 626 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 627 | } |
| 628 | |
| 629 | // Configure kernel window |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 630 | auto win_config = validate_and_configure_window(input->info(), output->info(), pool_info, _num_elems_processed_per_iteration, _border_size, pooled_w, pooled_h, pool_size); |
| 631 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 632 | INEKernel::configure(win_config.second); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 633 | } |
| 634 | |
| 635 | template <PoolingType pooling_type> |
| 636 | void NEPoolingLayerKernel::pooling2_q8(const Window &window_input, const Window &window) |
| 637 | { |
| 638 | Iterator input(_input, window_input); |
| 639 | Iterator output(_output, window); |
| 640 | |
| 641 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 642 | constexpr int pool_size = 2; |
| 643 | int pool_pad_x = 0; |
| 644 | int pool_pad_y = 0; |
| 645 | int pool_stride_x = 0; |
| 646 | int pool_stride_y = 0; |
| 647 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 648 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 649 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 650 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 651 | |
| 652 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 653 | 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)); |
| 654 | |
| 655 | execute_window_loop(window, [&](const Coordinates & id) |
| 656 | { |
| 657 | const auto top_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_top_ptr + input.offset())); |
| 658 | 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] | 659 | qint8x8_t lower_res = {}; |
| 660 | qint8x8_t upper_res = {}; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 661 | if(pooling_type == PoolingType::AVG) |
| 662 | { |
| 663 | // Calculate scale |
| 664 | 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); |
| 665 | const qint8x8_t scale_vec = vdup_n_qs8(scale); |
| 666 | |
| 667 | // Perform pooling |
| 668 | const qint8x16_t sum_data = vqaddq_qs8(top_data, bottom_data); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 669 | lower_res = vqmul_qs8(vpadd_s8(vget_low_s8(sum_data), vget_high_s8(sum_data)), scale_vec, fixed_point_position); |
| 670 | if(pool_stride_x == 1) |
| 671 | { |
| 672 | const qint8x16_t sum_data_shifted = vextq_s8(sum_data, sum_data, 1); |
| 673 | upper_res = vqmul_qs8(vpadd_s8(vget_low_s8(sum_data_shifted), vget_high_s8(sum_data_shifted)), scale_vec, fixed_point_position); |
| 674 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 675 | } |
| 676 | else |
| 677 | { |
| 678 | const qint8x16_t max_data = vmaxq_s8(top_data, bottom_data); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 679 | lower_res = vpmax_s8(vget_low_s8(max_data), vget_high_s8(max_data)); |
| 680 | if(pool_stride_x == 1) |
| 681 | { |
| 682 | const qint8x16_t max_data_shifted = vextq_s8(max_data, max_data, 1); |
| 683 | upper_res = vpmax_s8(vget_low_s8(max_data_shifted), vget_high_s8(max_data_shifted)); |
| 684 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 685 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 686 | if(pool_stride_x == 1) |
| 687 | { |
Georgios Pinitas | dc460f1 | 2017-08-24 19:02:44 +0100 | [diff] [blame] | 688 | const qint8x8x2_t res = { { lower_res, upper_res } }; |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 689 | vst2_s8(reinterpret_cast<qint8_t *>(output.ptr()), res); |
| 690 | } |
| 691 | else |
| 692 | { |
| 693 | vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), lower_res); |
| 694 | } |
| 695 | }, |
| 696 | input, output); |
| 697 | } |
| 698 | |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 699 | template <PoolingType pooling_type, bool exclude_padding> |
| 700 | void NEPoolingLayerKernel::pooling2_qasymm8(const Window &window_input, const Window &window) |
| 701 | { |
| 702 | Iterator input(_input, window_input); |
| 703 | Iterator output(_output, window); |
| 704 | |
| 705 | constexpr int pool_size = 2; |
| 706 | int pool_pad_x = 0; |
| 707 | int pool_pad_y = 0; |
| 708 | int pool_stride_x = 0; |
| 709 | int pool_stride_y = 0; |
| 710 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 711 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 712 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 713 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
| 714 | |
| 715 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 716 | 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)); |
| 717 | |
| 718 | const int scale_step_x = (pool_stride_x == 1) ? 2 : 1; |
| 719 | |
| 720 | execute_window_loop(window, [&](const Coordinates & id) |
| 721 | { |
| 722 | const auto top_data = vld1q_u8(reinterpret_cast<const uint8_t *>(input_top_ptr + input.offset())); |
| 723 | const auto bottom_data = vld1q_u8(reinterpret_cast<const uint8_t *>(input_bottom_ptr + input.offset())); |
| 724 | uint8x8_t lower_res = {}; |
| 725 | uint8x8_t upper_res = {}; |
| 726 | |
| 727 | if(pooling_type != PoolingType::MAX) |
| 728 | { |
| 729 | const uint16x8x2_t top_data_u16 = { { vmovl_u8(vget_low_u8(top_data)), vmovl_u8(vget_high_u8(top_data)) } }; |
| 730 | const uint16x8x2_t bottom_data_u16 = { { vmovl_u8(vget_low_u8(bottom_data)), vmovl_u8(vget_high_u8(bottom_data)) } }; |
| 731 | |
| 732 | // Add rows |
| 733 | const uint16x8x2_t vrsum = |
| 734 | { |
| 735 | { |
| 736 | vaddq_u16(top_data_u16.val[0], bottom_data_u16.val[0]), |
| 737 | vaddq_u16(top_data_u16.val[1], bottom_data_u16.val[1]), |
| 738 | } |
| 739 | }; |
| 740 | |
| 741 | // Pair-wise add row data |
| 742 | const uint16x4x2_t vpsum = |
| 743 | { |
| 744 | { |
| 745 | vpadd_u16(vget_low_u16(vrsum.val[0]), vget_high_u16(vrsum.val[0])), |
| 746 | vpadd_u16(vget_low_u16(vrsum.val[1]), vget_high_u16(vrsum.val[1])), |
| 747 | } |
| 748 | }; |
| 749 | |
| 750 | uint16x8_t res_lower = vcombine_u16(vpsum.val[0], vpsum.val[1]); |
| 751 | |
| 752 | // Scale lower result |
| 753 | scale_vector_s16x8<exclude_padding>(res_lower, id, 0, scale_step_x, |
| 754 | pool_size, upper_bound_w, upper_bound_h, |
| 755 | pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 756 | lower_res = vmovn_u16(res_lower); |
| 757 | |
| 758 | // Compute upper result for stride_x == 1 |
| 759 | if(pool_stride_x == 1) |
| 760 | { |
| 761 | // Shifted row sum |
| 762 | const uint16x8x2_t vrsum_shifted = |
| 763 | { |
| 764 | { |
| 765 | vextq_u16(vrsum.val[0], vrsum.val[1], 1), |
| 766 | vextq_u16(vrsum.val[1], vrsum.val[1], 1) |
| 767 | } |
| 768 | }; |
| 769 | |
| 770 | // Pair-wise add shifted row |
| 771 | const uint16x4x2_t vpsum_shifted = |
| 772 | { |
| 773 | { |
| 774 | vpadd_u16(vget_low_u16(vrsum_shifted.val[0]), vget_high_u16(vrsum_shifted.val[0])), |
| 775 | vpadd_u16(vget_low_u16(vrsum_shifted.val[1]), vget_high_u16(vrsum_shifted.val[1])), |
| 776 | } |
| 777 | }; |
| 778 | uint16x8_t res_upper = vcombine_u16(vpsum_shifted.val[0], vpsum_shifted.val[1]); |
| 779 | |
| 780 | // Scale lower result |
| 781 | scale_vector_s16x8<exclude_padding>(res_upper, id, 1, 2, |
| 782 | pool_size, upper_bound_w, upper_bound_h, |
| 783 | pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 784 | upper_res = vmovn_u16(res_upper); |
| 785 | } |
| 786 | } |
| 787 | else |
| 788 | { |
| 789 | const uint8x16_t max_data = vmaxq_u8(top_data, bottom_data); |
| 790 | lower_res = vpmax_u8(vget_low_u8(max_data), vget_high_u8(max_data)); |
| 791 | if(pool_stride_x == 1) |
| 792 | { |
| 793 | const uint8x16_t max_data_shifted = vextq_u8(max_data, max_data, 1); |
| 794 | upper_res = vpmax_u8(vget_low_u8(max_data_shifted), vget_high_u8(max_data_shifted)); |
| 795 | } |
| 796 | } |
| 797 | |
| 798 | // Store result |
| 799 | if(pool_stride_x == 1) |
| 800 | { |
| 801 | const uint8x8x2_t res = { { lower_res, upper_res } }; |
| 802 | vst2_u8(reinterpret_cast<uint8_t *>(output.ptr()), res); |
| 803 | } |
| 804 | else |
| 805 | { |
| 806 | vst1_u8(reinterpret_cast<uint8_t *>(output.ptr()), lower_res); |
| 807 | } |
| 808 | }, |
| 809 | input, output); |
| 810 | } |
| 811 | |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 812 | template <PoolingType pooling_type> |
| 813 | void NEPoolingLayerKernel::pooling2_q16(const Window &window_input, const Window &window) |
| 814 | { |
| 815 | Iterator input(_input, window_input); |
| 816 | Iterator output(_output, window); |
| 817 | |
| 818 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 819 | constexpr int pool_size = 2; |
| 820 | int pool_pad_x = 0; |
| 821 | int pool_pad_y = 0; |
| 822 | int pool_stride_x = 0; |
| 823 | int pool_stride_y = 0; |
| 824 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 825 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 826 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 827 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 828 | |
| 829 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 830 | 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)); |
| 831 | |
| 832 | execute_window_loop(window, [&](const Coordinates & id) |
| 833 | { |
| 834 | const auto top_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_top_ptr + input.offset())); |
| 835 | const auto bottom_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_bottom_ptr + input.offset())); |
| 836 | qint16x4_t lower_res = {}; |
| 837 | qint16x4_t upper_res = {}; |
| 838 | if(pooling_type == PoolingType::AVG) |
| 839 | { |
| 840 | // Calculate scale |
| 841 | 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); |
| 842 | const qint16x4_t scale_vec = vdup_n_qs16(scale); |
| 843 | |
| 844 | // Perform pooling |
| 845 | const qint16x8_t sum_data = vqaddq_qs16(top_data, bottom_data); |
| 846 | lower_res = vqmul_qs16(vpadd_s16(vget_low_s16(sum_data), vget_high_s16(sum_data)), scale_vec, fixed_point_position); |
| 847 | if(pool_stride_x == 1) |
| 848 | { |
| 849 | const qint16x8_t sum_data_shifted = vextq_s16(sum_data, sum_data, 1); |
| 850 | upper_res = vqmul_qs16(vpadd_s16(vget_low_s16(sum_data_shifted), vget_high_s16(sum_data_shifted)), scale_vec, fixed_point_position); |
| 851 | } |
| 852 | } |
| 853 | else |
| 854 | { |
| 855 | const qint16x8_t max_data = vmaxq_s16(top_data, bottom_data); |
| 856 | lower_res = vpmax_s16(vget_low_s16(max_data), vget_high_s16(max_data)); |
| 857 | if(pool_stride_x == 1) |
| 858 | { |
| 859 | const qint16x8_t max_data_shifted = vextq_s16(max_data, max_data, 1); |
| 860 | upper_res = vpmax_s16(vget_low_s16(max_data_shifted), vget_high_s16(max_data_shifted)); |
| 861 | } |
| 862 | } |
| 863 | if(pool_stride_x == 1) |
| 864 | { |
Georgios Pinitas | dc460f1 | 2017-08-24 19:02:44 +0100 | [diff] [blame] | 865 | const qint16x4x2_t res = { { lower_res, upper_res } }; |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 866 | vst2_s16(reinterpret_cast<qint16_t *>(output.ptr()), res); |
| 867 | } |
| 868 | else |
| 869 | { |
| 870 | vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), lower_res); |
| 871 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 872 | }, |
| 873 | input, output); |
| 874 | } |
| 875 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 876 | template <PoolingType pooling_type, bool exclude_padding> |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 877 | void NEPoolingLayerKernel::pooling3_f16(const Window &window_input, const Window &window) |
| 878 | { |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 879 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 880 | Iterator input(_input, window_input); |
| 881 | Iterator output(_output, window); |
| 882 | |
| 883 | constexpr const int pool_size = 3; |
| 884 | int pool_pad_x = 0; |
| 885 | int pool_pad_y = 0; |
| 886 | int pool_stride_x = 0; |
| 887 | int pool_stride_y = 0; |
| 888 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 889 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 890 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 891 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 892 | |
| 893 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 894 | 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)); |
| 895 | 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)); |
| 896 | |
| 897 | execute_window_loop(window, [&](const Coordinates & id) |
| 898 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 899 | float16x4_t top_data = vld1_f16(reinterpret_cast<const float16_t *>(input_top_ptr + input.offset())); |
| 900 | float16x4_t middle_data = vld1_f16(reinterpret_cast<const float16_t *>(input_middle_ptr + input.offset())); |
| 901 | float16x4_t bottom_data = vld1_f16(reinterpret_cast<const float16_t *>(input_bottom_ptr + input.offset())); |
| 902 | float16x4_t res = {}; |
| 903 | |
| 904 | // Get power of 2 in case of l2 pooling |
| 905 | if(pooling_type == PoolingType::L2) |
| 906 | { |
| 907 | top_data = vmul_f16(top_data, top_data); |
| 908 | middle_data = vmul_f16(middle_data, middle_data); |
| 909 | bottom_data = vmul_f16(bottom_data, bottom_data); |
| 910 | } |
| 911 | |
| 912 | if(pooling_type != PoolingType::MAX) |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 913 | { |
| 914 | // Calculate scale |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 915 | const float scale = calculate_avg_scale<exclude_padding>(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 916 | const float16x4_t scale_v = vdup_n_f16(scale); |
| 917 | // Perform pooling |
| 918 | const float16x4_t sum_data = vadd_f16(vadd_f16(top_data, bottom_data), middle_data); |
| 919 | res = vpadd_f16(vset_lane_f16(0.f, sum_data, 3), sum_data); |
| 920 | res = vmul_f16(vpadd_f16(res, res), scale_v); |
| 921 | } |
| 922 | else |
| 923 | { |
| 924 | const float16x4_t max_data = vmax_f16(vmax_f16(top_data, bottom_data), middle_data); |
| 925 | res = vpmax_f16(vset_lane_f16(-std::numeric_limits<float>::max(), max_data, 3), max_data); |
| 926 | res = vpmax_f16(res, res); |
| 927 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 928 | |
| 929 | // Calculate square-root in case of l2 pooling |
| 930 | if(pooling_type == PoolingType::L2) |
| 931 | { |
| 932 | res = vinv_f16(vinvsqrt_f16(res)); |
| 933 | } |
| 934 | |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 935 | *(reinterpret_cast<float16_t *>(output.ptr())) = vget_lane_f16(res, 0); |
| 936 | }, |
| 937 | input, output); |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 938 | #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 939 | ARM_COMPUTE_UNUSED(window_input); |
| 940 | ARM_COMPUTE_UNUSED(window); |
| 941 | ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 942 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 943 | } |
| 944 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 945 | template <PoolingType pooling_type, bool exclude_padding> |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 946 | void NEPoolingLayerKernel::pooling2_f16(const Window &window_input, const Window &window) |
| 947 | { |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 948 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 949 | Iterator input(_input, window_input); |
| 950 | Iterator output(_output, window); |
| 951 | constexpr int pool_size = 2; |
| 952 | int pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y = 0; |
| 953 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 954 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 955 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 956 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 957 | |
| 958 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 959 | 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)); |
| 960 | |
| 961 | execute_window_loop(window, [&](const Coordinates & id) |
| 962 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 963 | auto top_data = vld2q_f16(reinterpret_cast<const float16_t *>(input_top_ptr + input.offset())); |
| 964 | 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] | 965 | float16x8_t res = {}; |
| 966 | |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 967 | // Get power of 2 in case of l2 pooling |
| 968 | if(pooling_type == PoolingType::L2) |
| 969 | { |
| 970 | top_data.val[0] = vmulq_f16(top_data.val[0], top_data.val[0]); |
| 971 | top_data.val[1] = vmulq_f16(top_data.val[1], top_data.val[1]); |
| 972 | bottom_data.val[0] = vmulq_f16(bottom_data.val[0], bottom_data.val[0]); |
| 973 | bottom_data.val[1] = vmulq_f16(bottom_data.val[1], bottom_data.val[1]); |
| 974 | } |
| 975 | |
| 976 | if(pooling_type != PoolingType::MAX) |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 977 | { |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 978 | const float scale = calculate_avg_scale<exclude_padding>(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 979 | const float16x8_t scale_v = vdupq_n_f16(scale); |
| 980 | 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])))); |
| 981 | } |
| 982 | else |
| 983 | { |
| 984 | res = vmaxq_f16(bottom_data.val[1], vmaxq_f16(bottom_data.val[0], vmaxq_f16(top_data.val[0], top_data.val[1]))); |
| 985 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 986 | |
| 987 | // Calculate square-root in case of l2 pooling |
| 988 | if(pooling_type == PoolingType::L2) |
| 989 | { |
| 990 | res = vinvq_f16(vinvsqrtq_f16(res)); |
| 991 | } |
| 992 | |
| 993 | // Store result |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 994 | vst1q_f16(reinterpret_cast<float16_t *>(output.ptr()), res); |
| 995 | }, |
| 996 | input, output); |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 997 | #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 998 | ARM_COMPUTE_UNUSED(window_input); |
| 999 | ARM_COMPUTE_UNUSED(window); |
| 1000 | ARM_COMPUTE_ERROR("FP16 Not supported! Recompile the library with arch=arm64-v8.2-a"); |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 1001 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1002 | } |
| 1003 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1004 | template <PoolingType pooling_type, bool exclude_padding> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1005 | void NEPoolingLayerKernel::pooling2_f32(const Window &window_input, const Window &window) |
| 1006 | { |
| 1007 | Iterator input(_input, window_input); |
| 1008 | Iterator output(_output, window); |
| 1009 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1010 | constexpr int pool_size = 2; |
| 1011 | int pool_pad_x = 0; |
| 1012 | int pool_pad_y = 0; |
| 1013 | int pool_stride_x = 0; |
| 1014 | int pool_stride_y = 0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1015 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1016 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1017 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1018 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1019 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1020 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 1021 | 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] | 1022 | |
| 1023 | execute_window_loop(window, [&](const Coordinates & id) |
| 1024 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1025 | float32x2_t top_data = vld1_f32(reinterpret_cast<const float *>(input_top_ptr + input.offset())); |
| 1026 | float32x2_t bottom_data = vld1_f32(reinterpret_cast<const float *>(input_bottom_ptr + input.offset())); |
| 1027 | float32x2_t res = {}; |
| 1028 | float final_res = 0; |
| 1029 | |
| 1030 | // Get power of 2 in case of l2 pooling |
| 1031 | if(pooling_type == PoolingType::L2) |
| 1032 | { |
| 1033 | top_data = vmul_f32(top_data, top_data); |
| 1034 | bottom_data = vmul_f32(bottom_data, bottom_data); |
| 1035 | } |
| 1036 | |
| 1037 | if(pooling_type != PoolingType::MAX) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1038 | { |
| 1039 | // Calculate scale |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1040 | float scale = calculate_avg_scale<exclude_padding>(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1041 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 1042 | |
| 1043 | // Perform pooling |
| 1044 | const float32x2_t sum_data = vadd_f32(top_data, bottom_data); |
| 1045 | res = vmul_f32(vpadd_f32(sum_data, sum_data), scale_v); |
| 1046 | } |
| 1047 | else |
| 1048 | { |
| 1049 | const float32x2_t max_data = vmax_f32(top_data, bottom_data); |
| 1050 | res = vpmax_f32(max_data, max_data); |
| 1051 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1052 | final_res = vget_lane_f32(res, 0); |
| 1053 | |
| 1054 | // Calculate square-root in case of l2 pooling |
| 1055 | if(pooling_type == PoolingType::L2) |
| 1056 | { |
| 1057 | final_res = sqrt(final_res); |
| 1058 | } |
| 1059 | |
| 1060 | // Store result |
| 1061 | *(reinterpret_cast<float *>(output.ptr())) = final_res; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1062 | }, |
| 1063 | input, output); |
| 1064 | } |
| 1065 | |
| 1066 | template <PoolingType pooling_type> |
| 1067 | void NEPoolingLayerKernel::pooling3_q8(const Window &window_input, const Window &window) |
| 1068 | { |
| 1069 | Iterator input(_input, window_input); |
| 1070 | Iterator output(_output, window); |
| 1071 | |
| 1072 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 1073 | constexpr int pool_size = 3; |
| 1074 | int pool_pad_x = 0; |
| 1075 | int pool_pad_y = 0; |
| 1076 | int pool_stride_x = 0; |
| 1077 | int pool_stride_y = 0; |
| 1078 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1079 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 1080 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 1081 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 1082 | |
| 1083 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 1084 | 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)); |
| 1085 | 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)); |
| 1086 | |
| 1087 | execute_window_loop(window, [&](const Coordinates & id) |
| 1088 | { |
| 1089 | const auto top_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_top_ptr + input.offset())); |
| 1090 | const auto middle_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_middle_ptr + input.offset())); |
| 1091 | const auto bottom_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_bottom_ptr + input.offset())); |
| 1092 | qint8x8_t res = {}; |
| 1093 | if(pooling_type == PoolingType::AVG) |
| 1094 | { |
| 1095 | // Calculate scale |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1096 | 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] | 1097 | |
| 1098 | // Perform pooling for stride 2 |
| 1099 | const qint8x16_t sum_data = vqaddq_qs8(vqaddq_qs8(top_data, bottom_data), middle_data); |
| 1100 | const qint8x16_t sum_data2 = vextq_s8(sum_data, sum_data, 1); |
| 1101 | const qint8x16_t sum_data3 = vextq_s8(sum_data, sum_data, 2); |
| 1102 | const qint8x16_t final_sum = vqaddq_qs8(vqaddq_qs8(sum_data, sum_data2), sum_data3); |
| 1103 | if(pool_stride_x == 2) |
| 1104 | { |
| 1105 | const qint8x8x2_t table = { { vget_low_s8(final_sum), vget_high_s8(final_sum) } }; |
| 1106 | 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] | 1107 | const qint8x8_t scale_vec = vdup_n_qs8(scale); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1108 | res = vtbl2_s8(table, lookup_val); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1109 | res = vqmul_qs8(res, scale_vec, fixed_point_position); |
| 1110 | vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), res); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1111 | } |
| 1112 | else |
| 1113 | { |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1114 | const qint8x16_t scale_vec = vdupq_n_qs8(scale); |
| 1115 | 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] | 1116 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1117 | } |
| 1118 | else |
| 1119 | { |
| 1120 | const qint8x16_t max_data = vmaxq_s8(vmaxq_s8(top_data, bottom_data), middle_data); |
| 1121 | const qint8x16_t max_data2 = vextq_s8(max_data, max_data, 1); |
| 1122 | const qint8x16_t max_data3 = vextq_s8(max_data, max_data, 2); |
| 1123 | const qint8x16_t final_max = vmaxq_s8(vmaxq_s8(max_data, max_data2), max_data3); |
| 1124 | |
| 1125 | if(pool_stride_x == 2) |
| 1126 | { |
| 1127 | const qint8x8x2_t table = { { vget_low_s8(final_max), vget_high_s8(final_max) } }; |
| 1128 | static const qint8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; |
| 1129 | res = vtbl2_s8(table, lookup_val); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1130 | vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), res); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1131 | } |
| 1132 | else |
| 1133 | { |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1134 | vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), final_max); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1135 | } |
| 1136 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1137 | }, |
| 1138 | input, output); |
| 1139 | } |
| 1140 | |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 1141 | template <PoolingType pooling_type, bool exclude_padding> |
| 1142 | void NEPoolingLayerKernel::pooling3_qasymm8(const Window &window_input, const Window &window) |
| 1143 | { |
| 1144 | Iterator input(_input, window_input); |
| 1145 | Iterator output(_output, window); |
| 1146 | |
| 1147 | constexpr int pool_size = 3; |
| 1148 | int pool_pad_x = 0; |
| 1149 | int pool_pad_y = 0; |
| 1150 | int pool_stride_x = 0; |
| 1151 | int pool_stride_y = 0; |
| 1152 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1153 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 1154 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1155 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
| 1156 | |
| 1157 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 1158 | 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)); |
| 1159 | 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)); |
| 1160 | |
| 1161 | execute_window_loop(window, [&](const Coordinates & id) |
| 1162 | { |
| 1163 | const auto top_data = vld1q_u8(reinterpret_cast<const uint8_t *>(input_top_ptr + input.offset())); |
| 1164 | const auto middle_data = vld1q_u8(reinterpret_cast<const uint8_t *>(input_middle_ptr + input.offset())); |
| 1165 | const auto bottom_data = vld1q_u8(reinterpret_cast<const uint8_t *>(input_bottom_ptr + input.offset())); |
| 1166 | |
| 1167 | if(pooling_type == PoolingType::AVG) |
| 1168 | { |
| 1169 | // Convert data to u16 |
| 1170 | const uint16x8x2_t top_data_u16 = { { vmovl_u8(vget_low_u8(top_data)), vmovl_u8(vget_high_u8(top_data)) } }; |
| 1171 | const uint16x8x2_t middle_data_u16 = { { vmovl_u8(vget_low_u8(middle_data)), vmovl_u8(vget_high_u8(middle_data)) } }; |
| 1172 | const uint16x8x2_t bottom_data_u16 = { { vmovl_u8(vget_low_u8(bottom_data)), vmovl_u8(vget_high_u8(bottom_data)) } }; |
| 1173 | |
| 1174 | // Calculate row sums |
| 1175 | const uint16x8x2_t vrsum = |
| 1176 | { |
| 1177 | { |
| 1178 | vaddq_u16(vaddq_u16(top_data_u16.val[0], bottom_data_u16.val[0]), middle_data_u16.val[0]), |
| 1179 | vaddq_u16(vaddq_u16(top_data_u16.val[1], bottom_data_u16.val[1]), middle_data_u16.val[1]), |
| 1180 | } |
| 1181 | }; |
| 1182 | const uint16x8x2_t vrsum_shifted_1 = |
| 1183 | { |
| 1184 | { |
| 1185 | vextq_u16(vrsum.val[0], vrsum.val[1], 1), |
| 1186 | vextq_u16(vrsum.val[1], vrsum.val[1], 1) |
| 1187 | } |
| 1188 | }; |
| 1189 | const uint16x8x2_t vrsum_shifted_2 = |
| 1190 | { |
| 1191 | { |
| 1192 | vextq_u16(vrsum.val[0], vrsum.val[1], 2), |
| 1193 | vextq_u16(vrsum.val[1], vrsum.val[1], 2) |
| 1194 | } |
| 1195 | }; |
| 1196 | // Calculate final sum |
| 1197 | uint16x8x2_t final_sum = |
| 1198 | { |
| 1199 | { |
| 1200 | vaddq_u16(vaddq_u16(vrsum.val[0], vrsum_shifted_1.val[0]), vrsum_shifted_2.val[0]), |
| 1201 | vaddq_u16(vaddq_u16(vrsum.val[1], vrsum_shifted_1.val[1]), vrsum_shifted_2.val[1]), |
| 1202 | } |
| 1203 | }; |
| 1204 | if(pool_stride_x == 2) |
| 1205 | { |
| 1206 | uint16x8_t res = |
| 1207 | { |
| 1208 | vgetq_lane_u16(final_sum.val[0], 0), |
| 1209 | vgetq_lane_u16(final_sum.val[0], 2), |
| 1210 | vgetq_lane_u16(final_sum.val[0], 4), |
| 1211 | vgetq_lane_u16(final_sum.val[0], 6), |
| 1212 | vgetq_lane_u16(final_sum.val[1], 0), |
| 1213 | vgetq_lane_u16(final_sum.val[1], 2), |
| 1214 | vgetq_lane_u16(final_sum.val[1], 4), |
| 1215 | vgetq_lane_u16(final_sum.val[1], 6), |
| 1216 | }; |
| 1217 | |
| 1218 | scale_vector_s16x8<exclude_padding>(res, id, 0, 1, |
| 1219 | pool_size, upper_bound_w, upper_bound_h, |
| 1220 | pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 1221 | vst1_u8(reinterpret_cast<uint8_t *>(output.ptr()), vmovn_u16(res)); |
| 1222 | } |
| 1223 | else |
| 1224 | { |
| 1225 | // Scale lower result |
| 1226 | scale_vector_s16x8<exclude_padding>(final_sum.val[0], id, 0, 1, |
| 1227 | pool_size, upper_bound_w, upper_bound_h, |
| 1228 | pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 1229 | // Scale lower result |
| 1230 | scale_vector_s16x8<exclude_padding>(final_sum.val[1], id, 8, 1, |
| 1231 | pool_size, upper_bound_w, upper_bound_h, |
| 1232 | pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 1233 | const uint8x16_t res = vcombine_u8(vmovn_u16(final_sum.val[0]), vmovn_u16(final_sum.val[1])); |
| 1234 | vst1q_u8(reinterpret_cast<uint8_t *>(output.ptr()), res); |
| 1235 | } |
| 1236 | } |
| 1237 | else |
| 1238 | { |
| 1239 | const uint8x16_t max_data = vmaxq_u8(vmaxq_u8(top_data, bottom_data), middle_data); |
| 1240 | const uint8x16_t max_data_shift1 = vextq_u8(max_data, max_data, 1); |
| 1241 | const uint8x16_t max_data_shift2 = vextq_u8(max_data, max_data, 2); |
| 1242 | const uint8x16_t final_max = vmaxq_u8(vmaxq_u8(max_data, max_data_shift1), max_data_shift2); |
| 1243 | |
| 1244 | if(pool_stride_x == 2) |
| 1245 | { |
| 1246 | const uint8x8x2_t table = { { vget_low_u8(final_max), vget_high_u8(final_max) } }; |
| 1247 | static const uint8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; |
| 1248 | const uint8x8_t res = vtbl2_u8(table, lookup_val); |
| 1249 | vst1_u8(reinterpret_cast<uint8_t *>(output.ptr()), res); |
| 1250 | } |
| 1251 | else |
| 1252 | { |
| 1253 | vst1q_u8(reinterpret_cast<uint8_t *>(output.ptr()), final_max); |
| 1254 | } |
| 1255 | } |
| 1256 | }, |
| 1257 | input, output); |
| 1258 | } |
| 1259 | |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1260 | template <PoolingType pooling_type> |
| 1261 | void NEPoolingLayerKernel::pooling3_q16(const Window &window_input, const Window &window) |
| 1262 | { |
| 1263 | Iterator input(_input, window_input); |
| 1264 | Iterator output(_output, window); |
| 1265 | |
| 1266 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 1267 | constexpr int pool_size = 3; |
| 1268 | int pool_pad_x = 0; |
| 1269 | int pool_pad_y = 0; |
| 1270 | int pool_stride_x = 0; |
| 1271 | int pool_stride_y = 0; |
| 1272 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1273 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 1274 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 1275 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 1276 | |
| 1277 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 1278 | 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)); |
| 1279 | 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)); |
| 1280 | |
| 1281 | execute_window_loop(window, [&](const Coordinates & id) |
| 1282 | { |
| 1283 | const auto top_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_top_ptr + input.offset())); |
| 1284 | const auto middle_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_middle_ptr + input.offset())); |
| 1285 | const auto bottom_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_bottom_ptr + input.offset())); |
| 1286 | |
| 1287 | if(pooling_type == PoolingType::AVG) |
| 1288 | { |
| 1289 | // Calculate scale |
| 1290 | 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); |
| 1291 | |
| 1292 | // Perform pooling for stride 2 |
| 1293 | const qint16x8_t sum_data = vqaddq_qs16(vqaddq_qs16(top_data, bottom_data), middle_data); |
| 1294 | const qint16x8_t sum_data2 = vextq_s16(sum_data, sum_data, 1); |
| 1295 | const qint16x8_t sum_data3 = vextq_s16(sum_data, sum_data, 2); |
| 1296 | const qint16x8_t final_sum = vqaddq_qs16(vqaddq_qs16(sum_data, sum_data2), sum_data3); |
| 1297 | if(pool_stride_x == 2) |
| 1298 | { |
| 1299 | 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) }; |
| 1300 | const qint16x4_t scale_vec = vdup_n_qs16(scale); |
| 1301 | vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), vqmul_qs16(tmp, scale_vec, fixed_point_position)); |
| 1302 | } |
| 1303 | else |
| 1304 | { |
| 1305 | const qint16x8_t scale_vec = vdupq_n_qs16(scale); |
| 1306 | vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()), vqmulq_qs16(final_sum, scale_vec, fixed_point_position)); |
| 1307 | } |
| 1308 | } |
| 1309 | else |
| 1310 | { |
| 1311 | const qint16x8_t max_data = vmaxq_s16(vmaxq_s16(top_data, bottom_data), middle_data); |
| 1312 | const qint16x8_t max_data2 = vextq_s16(max_data, max_data, 1); |
| 1313 | const qint16x8_t max_data3 = vextq_s16(max_data, max_data, 2); |
| 1314 | const qint16x8_t final_max = vmaxq_s16(vmaxq_s16(max_data, max_data2), max_data3); |
| 1315 | |
| 1316 | if(pool_stride_x == 2) |
| 1317 | { |
| 1318 | 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) }; |
| 1319 | vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), tmp); |
| 1320 | } |
| 1321 | else |
| 1322 | { |
| 1323 | vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()), final_max); |
| 1324 | } |
| 1325 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1326 | }, |
| 1327 | input, output); |
| 1328 | } |
| 1329 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1330 | template <PoolingType pooling_type, bool exclude_padding> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1331 | void NEPoolingLayerKernel::pooling3_f32(const Window &window_input, const Window &window) |
| 1332 | { |
| 1333 | Iterator input(_input, window_input); |
| 1334 | Iterator output(_output, window); |
| 1335 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1336 | constexpr const int pool_size = 3; |
| 1337 | int pool_pad_x = 0; |
| 1338 | int pool_pad_y = 0; |
| 1339 | int pool_stride_x = 0; |
| 1340 | int pool_stride_y = 0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1341 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1342 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1343 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1344 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1345 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1346 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 1347 | 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)); |
| 1348 | 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] | 1349 | |
| 1350 | execute_window_loop(window, [&](const Coordinates & id) |
| 1351 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1352 | float32x4_t top_data = vld1q_f32(reinterpret_cast<const float *>(input_top_ptr + input.offset())); |
| 1353 | float32x4_t middle_data = vld1q_f32(reinterpret_cast<const float *>(input_middle_ptr + input.offset())); |
| 1354 | float32x4_t bottom_data = vld1q_f32(reinterpret_cast<const float *>(input_bottom_ptr + input.offset())); |
| 1355 | float32x2_t res = {}; |
| 1356 | float final_res = 0; |
| 1357 | |
| 1358 | // Get power of 2 in case of l2 pooling |
| 1359 | if(pooling_type == PoolingType::L2) |
| 1360 | { |
| 1361 | top_data = vmulq_f32(top_data, top_data); |
| 1362 | middle_data = vmulq_f32(middle_data, middle_data); |
| 1363 | bottom_data = vmulq_f32(bottom_data, bottom_data); |
| 1364 | } |
| 1365 | |
| 1366 | if(pooling_type != PoolingType::MAX) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1367 | { |
| 1368 | // Calculate scale |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1369 | float scale = calculate_avg_scale<exclude_padding>(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1370 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 1371 | |
| 1372 | // Perform pooling |
| 1373 | const float32x4_t sum_data = vaddq_f32(vaddq_f32(top_data, bottom_data), middle_data); |
| 1374 | res = vpadd_f32(vget_high_f32(vsetq_lane_f32(0.f, sum_data, 3)), vget_low_f32(sum_data)); |
| 1375 | res = vmul_f32(vpadd_f32(res, res), scale_v); |
| 1376 | } |
| 1377 | else |
| 1378 | { |
| 1379 | const float32x4_t max_data = vmaxq_f32(vmaxq_f32(top_data, bottom_data), middle_data); |
| 1380 | res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits<float>::max(), max_data, 3)), vget_low_f32(max_data)); |
| 1381 | res = vpmax_f32(res, res); |
| 1382 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1383 | final_res = vget_lane_f32(res, 0); |
| 1384 | |
| 1385 | // Calculate square-root in case of l2 pooling |
| 1386 | if(pooling_type == PoolingType::L2) |
| 1387 | { |
| 1388 | final_res = sqrt(final_res); |
| 1389 | } |
| 1390 | |
| 1391 | // Store result |
| 1392 | *(reinterpret_cast<float *>(output.ptr())) = final_res; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1393 | }, |
| 1394 | input, output); |
| 1395 | } |
| 1396 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1397 | template <PoolingType pooling_type, bool exclude_padding> |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1398 | void NEPoolingLayerKernel::pooling7_f32(const Window &window_input, const Window &window) |
| 1399 | { |
| 1400 | Iterator input(_input, window_input); |
| 1401 | Iterator output(_output, window); |
| 1402 | |
| 1403 | constexpr const int pool_size = 7; |
| 1404 | int pool_pad_x = 0; |
| 1405 | int pool_pad_y = 0; |
| 1406 | int pool_stride_x = 0; |
| 1407 | int pool_stride_y = 0; |
| 1408 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1409 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1410 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1411 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1412 | |
| 1413 | std::array<const uint8_t *, pool_size> input_ptrs{ {} }; |
| 1414 | for(int i = 0; i < pool_size; ++i) |
| 1415 | { |
| 1416 | input_ptrs[i] = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + i)); |
| 1417 | } |
| 1418 | |
| 1419 | execute_window_loop(window, [&](const Coordinates & id) |
| 1420 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1421 | float32x2_t res = {}; |
| 1422 | float final_res = 0.f; |
| 1423 | if(pooling_type != PoolingType::MAX) |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1424 | { |
| 1425 | // Calculate scale |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1426 | float scale = calculate_avg_scale<exclude_padding>(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1427 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 1428 | |
| 1429 | // Perform pooling |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1430 | float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[0] + input.offset())); |
| 1431 | // Get power of 2 in case of l2 pooling |
| 1432 | if(pooling_type == PoolingType::L2) |
| 1433 | { |
| 1434 | data.val[0] = vmulq_f32(data.val[0], data.val[0]); |
| 1435 | data.val[1] = vmulq_f32(data.val[1], data.val[1]); |
| 1436 | } |
| 1437 | 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] | 1438 | for(int i = 1; i < pool_size; ++i) |
| 1439 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1440 | data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[i] + input.offset())); |
| 1441 | // Get power of 2 in case of l2 pooling |
| 1442 | if(pooling_type == PoolingType::L2) |
| 1443 | { |
| 1444 | data.val[0] = vmulq_f32(data.val[0], data.val[0]); |
| 1445 | data.val[1] = vmulq_f32(data.val[1], data.val[1]); |
| 1446 | } |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1447 | sum_data = vaddq_f32(sum_data, data.val[0]); |
| 1448 | sum_data = vaddq_f32(sum_data, vsetq_lane_f32(0.f, data.val[1], 3)); |
| 1449 | } |
| 1450 | res = vpadd_f32(vget_high_f32(sum_data), vget_low_f32(sum_data)); |
| 1451 | res = vmul_f32(vpadd_f32(res, res), scale_v); |
| 1452 | } |
| 1453 | else |
| 1454 | { |
| 1455 | float32x4x2_t max_data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[0] + input.offset())); |
| 1456 | for(int i = 1; i < pool_size; ++i) |
| 1457 | { |
| 1458 | const float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[i] + input.offset())); |
| 1459 | max_data = vmax2q_f32(max_data, data); |
| 1460 | } |
| 1461 | 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])); |
| 1462 | res = vpmax_f32(res, vpmax_f32(vget_high_f32(max_data.val[0]), vget_low_f32(max_data.val[0]))); |
| 1463 | res = vpmax_f32(res, res); |
| 1464 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1465 | final_res = vget_lane_f32(res, 0); |
| 1466 | |
| 1467 | // Calculate square-root in case of l2 pooling |
| 1468 | if(pooling_type == PoolingType::L2) |
| 1469 | { |
| 1470 | final_res = sqrt(final_res); |
| 1471 | } |
| 1472 | |
| 1473 | // Store result |
| 1474 | *(reinterpret_cast<float *>(output.ptr())) = final_res; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1475 | }, |
| 1476 | input, output); |
| 1477 | } |
| 1478 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1479 | template <PoolingType pooling_type, bool exclude_padding> |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame] | 1480 | void NEPoolingLayerKernel::poolingN_f32(const Window &window_input, const Window &window) |
| 1481 | { |
| 1482 | Iterator input(_input, window_input); |
| 1483 | Iterator output(_output, window); |
| 1484 | |
Georgios Pinitas | 4c2dd54 | 2017-11-13 12:58:41 +0000 | [diff] [blame] | 1485 | const int pool_size = _pool_info.is_global_pooling() ? _input->info()->tensor_shape().x() : _pool_info.pool_size(); |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame] | 1486 | int pool_pad_x = 0; |
| 1487 | int pool_pad_y = 0; |
| 1488 | int pool_stride_x = 0; |
| 1489 | int pool_stride_y = 0; |
| 1490 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1491 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1492 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1493 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame] | 1494 | |
| 1495 | execute_window_loop(window, [&](const Coordinates & id) |
| 1496 | { |
| 1497 | float res = 0.0f; |
| 1498 | |
| 1499 | if(pooling_type != PoolingType::MAX) |
| 1500 | { |
| 1501 | // Calculate scale |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1502 | const float scale = calculate_avg_scale<exclude_padding>(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame] | 1503 | |
| 1504 | // Perform pooling |
| 1505 | float32x4_t vres = vdupq_n_f32(0.0f); |
| 1506 | |
| 1507 | for(int y = 0; y < pool_size; ++y) |
| 1508 | { |
| 1509 | int x = 0; |
| 1510 | for(; x <= (pool_size - 4); x += 4) |
| 1511 | { |
| 1512 | const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + |
| 1513 | (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1514 | |
| 1515 | // Get power of 2 in case of l2 pooling and accumulate |
| 1516 | if(pooling_type == PoolingType::L2) |
| 1517 | { |
| 1518 | vres = vmlaq_f32(vres, data, data); |
| 1519 | } |
| 1520 | else |
| 1521 | { |
| 1522 | vres = vaddq_f32(vres, data); |
| 1523 | } |
| 1524 | } |
| 1525 | |
| 1526 | // Leftover for loop |
| 1527 | for(; x < pool_size; ++x) |
| 1528 | { |
| 1529 | 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())); |
| 1530 | |
| 1531 | // Get power of 2 in case of l2 pooling |
| 1532 | if(pooling_type == PoolingType::L2) |
| 1533 | { |
| 1534 | data *= data; |
| 1535 | } |
| 1536 | |
| 1537 | res += data; |
| 1538 | } |
| 1539 | } |
| 1540 | |
| 1541 | #if defined(__aarch64__) |
| 1542 | // Reduction operation available on 64 bit architectures only |
| 1543 | res += vaddvq_f32(vres); |
| 1544 | #else // __aarch64__ |
| 1545 | // Reduction |
| 1546 | float32x2_t tmp = vpadd_f32(vget_high_f32(vres), vget_low_f32(vres)); |
| 1547 | tmp = vpadd_f32(tmp, tmp); |
| 1548 | |
| 1549 | res += vget_lane_f32(tmp, 0); |
| 1550 | #endif // __aarch64__ |
| 1551 | // Divide by scale |
| 1552 | res *= scale; |
| 1553 | } |
| 1554 | else |
| 1555 | { |
| 1556 | float32x4_t vres = vdupq_n_f32(std::numeric_limits<float>::min()); |
| 1557 | res = std::numeric_limits<float>::min(); |
| 1558 | |
| 1559 | for(int y = 0; y < pool_size; ++y) |
| 1560 | { |
| 1561 | int x = 0; |
| 1562 | for(; x <= (pool_size - 4); x += 4) |
| 1563 | { |
| 1564 | const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + |
| 1565 | (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1566 | vres = vmaxq_f32(vres, data); |
| 1567 | } |
| 1568 | |
| 1569 | // Leftover for loop |
| 1570 | for(; x < pool_size; ++x) |
| 1571 | { |
| 1572 | 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())); |
| 1573 | res = std::max(res, data); |
| 1574 | } |
| 1575 | } |
| 1576 | |
| 1577 | #if defined(__aarch64__) |
| 1578 | // Reduction operation available on 64 bit architectures only |
| 1579 | res = std::max(vmaxvq_f32(vres), res); |
| 1580 | #else // __aarch64__ |
| 1581 | float32x2_t tmp = vpmax_f32(vget_high_f32(vres), vget_low_f32(vres)); |
| 1582 | tmp = vpmax_f32(tmp, tmp); |
| 1583 | |
| 1584 | res = std::max(res, vget_lane_f32(tmp, 0)); |
| 1585 | #endif // __aarch64__ |
| 1586 | } |
| 1587 | |
| 1588 | // Calculate square-root in case of l2 pooling |
| 1589 | if(pooling_type == PoolingType::L2) |
| 1590 | { |
| 1591 | res = std::sqrt(res); |
| 1592 | } |
| 1593 | |
| 1594 | // Store result |
| 1595 | *(reinterpret_cast<float *>(output.ptr())) = res; |
| 1596 | }, |
| 1597 | input, output); |
| 1598 | } |
| 1599 | |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 1600 | template <PoolingType pooling_type, bool exclude_padding> |
| 1601 | void NEPoolingLayerKernel::poolingN_qasymm8(const Window &window_input, const Window &window) |
| 1602 | { |
| 1603 | Iterator input(_input, window_input); |
| 1604 | Iterator output(_output, window); |
| 1605 | |
| 1606 | const int pool_size = _pool_info.is_global_pooling() ? _input->info()->tensor_shape().x() : _pool_info.pool_size(); |
| 1607 | int pool_pad_x = 0; |
| 1608 | int pool_pad_y = 0; |
| 1609 | int pool_stride_x = 0; |
| 1610 | int pool_stride_y = 0; |
| 1611 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1612 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 1613 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1614 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
| 1615 | |
| 1616 | execute_window_loop(window, [&](const Coordinates & id) |
| 1617 | { |
| 1618 | uint8_t res = 0; |
| 1619 | |
| 1620 | if(pooling_type != PoolingType::MAX) |
| 1621 | { |
| 1622 | uint32x4_t vres = vdupq_n_u32(0); |
| 1623 | uint32_t sres = 0; |
| 1624 | |
| 1625 | // Calculate scale |
| 1626 | const float scale = calculate_avg_scale<exclude_padding>(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 1627 | |
| 1628 | // Perform pooling |
| 1629 | for(int y = 0; y < pool_size; ++y) |
| 1630 | { |
| 1631 | int x = 0; |
| 1632 | for(; x <= (pool_size - 8); x += 8) |
| 1633 | { |
| 1634 | const uint8x8_t data = vld1_u8(reinterpret_cast<const uint8_t *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1635 | |
| 1636 | const uint16x8_t data_u16 = vmovl_u8(data); |
| 1637 | vres = vaddq_u32(vres, vaddl_u16(vget_high_u16(data_u16), vget_low_u16(data_u16))); |
| 1638 | } |
| 1639 | |
| 1640 | // Leftover for loop |
| 1641 | for(; x < pool_size; ++x) |
| 1642 | { |
| 1643 | uint8_t data = *(reinterpret_cast<const uint8_t *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1644 | sres += data; |
| 1645 | } |
| 1646 | } |
| 1647 | |
| 1648 | // Reduction |
| 1649 | const auto tmp = vpadd_u32(vget_high_u32(vres), vget_low_u32(vres)); |
| 1650 | sres += vget_lane_u32(tmp, 0) + vget_lane_u32(tmp, 1); |
| 1651 | |
| 1652 | // Divide by scale |
| 1653 | res = static_cast<uint8_t>(support::cpp11::round(sres * scale)); |
| 1654 | } |
| 1655 | else |
| 1656 | { |
| 1657 | uint8x8_t vres = vdup_n_u8(0); |
| 1658 | res = 0; |
| 1659 | |
| 1660 | for(int y = 0; y < pool_size; ++y) |
| 1661 | { |
| 1662 | int x = 0; |
| 1663 | for(; x <= (pool_size - 8); x += 8) |
| 1664 | { |
| 1665 | const uint8x8_t data = vld1_u8(reinterpret_cast<const uint8_t *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1666 | vres = vmax_u8(vres, data); |
| 1667 | } |
| 1668 | |
| 1669 | // Leftover for loop |
| 1670 | for(; x < pool_size; ++x) |
| 1671 | { |
| 1672 | const uint8_t data = *(reinterpret_cast<const uint8_t *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1673 | res = std::max(res, data); |
| 1674 | } |
| 1675 | } |
| 1676 | |
| 1677 | // Reduce max |
| 1678 | vres = vpmax_u8(vres, vres); |
| 1679 | vres = vpmax_u8(vres, vres); |
| 1680 | vres = vpmax_u8(vres, vres); |
| 1681 | |
| 1682 | // Get max value |
| 1683 | res = std::max(res, vget_lane_u8(vres, 0)); |
| 1684 | } |
| 1685 | |
| 1686 | // Store result |
| 1687 | *(reinterpret_cast<uint8_t *>(output.ptr())) = res; |
| 1688 | }, |
| 1689 | input, output); |
| 1690 | } |
| 1691 | |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 1692 | Status NEPoolingLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info) |
| 1693 | { |
| 1694 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); |
| 1695 | |
| 1696 | unsigned int pooled_w = 0; |
| 1697 | unsigned int pooled_h = 0; |
| 1698 | unsigned int num_elems_processed_per_iteration = 0; |
| 1699 | BorderSize border_size(0); |
| 1700 | |
| 1701 | const bool is_global_pooling = pool_info.is_global_pooling(); |
| 1702 | const unsigned int pool_size = is_global_pooling ? input->tensor_shape().x() : pool_info.pool_size(); |
| 1703 | |
| 1704 | // Validate pool info befor calling scaled_dimensions |
| 1705 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_pool_info(input, pool_info, pool_size)); |
| 1706 | |
| 1707 | // Check output dimensions |
| 1708 | std::tie(pooled_w, pooled_h) = scaled_dimensions(input->dimension(0), |
| 1709 | input->dimension(1), |
| 1710 | pool_size, |
| 1711 | pool_size, |
| 1712 | pool_info.pad_stride_info()); |
| 1713 | |
| 1714 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, pool_info, pooled_w, pooled_h, pool_size)); |
| 1715 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), pool_info, num_elems_processed_per_iteration, border_size, pooled_w, pooled_h, pool_size).first); |
| 1716 | |
| 1717 | return Status{}; |
| 1718 | } |
| 1719 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1720 | void NEPoolingLayerKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1721 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1722 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1723 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 1724 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 1725 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 1726 | |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1727 | const unsigned int pool_stride_x = _pool_info.pad_stride_info().stride().first; |
| 1728 | const unsigned int pool_stride_y = _pool_info.pad_stride_info().stride().second; |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 1729 | const unsigned int pool_size = _pool_info.pool_size(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1730 | |
| 1731 | // Set step for input in x and y direction for the input |
| 1732 | Window window_input(window); |
| 1733 | unsigned int window_x_inc = 0; |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1734 | switch(_input->info()->data_type()) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1735 | { |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1736 | case DataType::QS8: |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1737 | case DataType::QS16: |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1738 | case DataType::F16: |
| 1739 | { |
| 1740 | window_x_inc = (pool_stride_x == 2) ? _num_elems_processed_per_iteration * 2 : _num_elems_processed_per_iteration; |
| 1741 | break; |
| 1742 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame] | 1743 | case DataType::QASYMM8: |
| 1744 | { |
| 1745 | window_x_inc = pool_stride_x; |
| 1746 | if((pool_size == 2 || pool_size == 3) && pool_stride_x < 3) |
| 1747 | { |
| 1748 | window_x_inc = (pool_stride_x == 2) ? _num_elems_processed_per_iteration * 2 : _num_elems_processed_per_iteration; |
| 1749 | } |
| 1750 | break; |
| 1751 | } |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1752 | case DataType::F32: |
| 1753 | { |
| 1754 | window_x_inc = pool_stride_x; |
| 1755 | break; |
| 1756 | } |
| 1757 | default: |
| 1758 | { |
| 1759 | ARM_COMPUTE_ERROR("Not supported"); |
| 1760 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1761 | } |
| 1762 | window_input.set(Window::DimX, Window::Dimension(window.x().start() * pool_stride_x, window.x().end() * pool_stride_x, window_x_inc)); |
| 1763 | window_input.set(Window::DimY, Window::Dimension(window.y().start() * pool_stride_y, window.y().end() * pool_stride_y, pool_stride_y)); |
| 1764 | |
| 1765 | // Run function |
| 1766 | (this->*_func)(window_input, window); |
| 1767 | } |