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 | |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 320 | const int upper_bound_w = ((pooled_w - 1) * pool_stride_x - pool_pad_x + num_elems_read_per_iteration) - input_width; |
| 321 | const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_y + pool_size) - input_height; |
| 322 | |
| 323 | border_size = BorderSize(pool_pad_y, pool_pad_x); |
| 324 | border_size.right = std::max(upper_bound_w, pool_pad_x); |
| 325 | border_size.bottom = std::max(upper_bound_h, pool_pad_y); |
| 326 | bool window_changed = false; |
| 327 | |
| 328 | TensorShape output_shape{ input->tensor_shape() }; |
| 329 | output_shape.set(0, pooled_w); |
| 330 | output_shape.set(1, pooled_h); |
| 331 | TensorInfo output_info(input->clone()->set_tensor_shape(output_shape)); |
| 332 | |
| 333 | Window win = calculate_max_window(output_info, Steps(num_elems_processed_per_iteration)); |
| 334 | AccessWindowStatic input_access(input, -pool_pad_x, -pool_pad_y, input_width + border_size.right, input_height + border_size.bottom); |
| 335 | |
| 336 | if(output->total_size() != 0) |
| 337 | { |
| 338 | AccessWindowHorizontal output_access(output, 0, num_elems_horizontal_window); |
| 339 | window_changed = update_window_and_padding(win, input_access, output_access); |
| 340 | output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); |
| 341 | } |
| 342 | else |
| 343 | { |
| 344 | window_changed = update_window_and_padding(win, input_access); |
| 345 | } |
| 346 | |
| 347 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 348 | return std::make_pair(err, win); |
| 349 | } |
| 350 | } // namespace |
| 351 | |
| 352 | NEPoolingLayerKernel::NEPoolingLayerKernel() |
| 353 | : _func(nullptr), _input(nullptr), _output(nullptr), _pool_info(), _num_elems_processed_per_iteration(0), _border_size(0) |
| 354 | { |
| 355 | } |
| 356 | |
| 357 | BorderSize NEPoolingLayerKernel::border_size() const |
| 358 | { |
| 359 | return _border_size; |
| 360 | } |
| 361 | |
| 362 | void NEPoolingLayerKernel::configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info) |
| 363 | { |
| 364 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 365 | |
| 366 | int pool_pad_x = 0; |
| 367 | int pool_pad_y = 0; |
| 368 | int pool_stride_x = 0; |
| 369 | int pool_stride_y = 0; |
| 370 | unsigned int pooled_w = 0; |
| 371 | unsigned int pooled_h = 0; |
| 372 | PoolingType pool_type = pool_info.pool_type(); |
| 373 | int pool_size = pool_info.pool_size(); |
| 374 | const PadStrideInfo pad_stride_info = pool_info.pad_stride_info(); |
| 375 | const bool exclude_padding = pool_info.exclude_padding(); |
| 376 | const bool is_global_pooling = pool_info.is_global_pooling(); |
| 377 | std::tie(pool_pad_x, pool_pad_y) = pad_stride_info.pad(); |
| 378 | std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); |
| 379 | |
| 380 | // Update pool size in case of global pooling |
| 381 | pool_size = is_global_pooling ? input->info()->dimension(0) : pool_size; |
| 382 | |
| 383 | // Validate pool info before calling scaled_dimensions |
| 384 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments_pool_info(input->info(), pool_info, pool_size)); |
| 385 | |
| 386 | // Check output dimensions |
| 387 | std::tie(pooled_w, pooled_h) = scaled_dimensions(input->info()->dimension(0), |
| 388 | input->info()->dimension(1), |
| 389 | pool_size, |
| 390 | pool_size, |
| 391 | pool_info.pad_stride_info()); |
| 392 | |
| 393 | // Output auto initialization if not yet initialized |
| 394 | auto_init(input->info(), output->info(), pooled_w, pooled_h); |
| 395 | |
| 396 | // Perform validation step |
| 397 | 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] | 398 | |
| 399 | // Set instance variables |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 400 | _input = input; |
| 401 | _output = output; |
| 402 | _pool_info = pool_info; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 403 | |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 404 | // Get data type |
| 405 | const DataType data_type = input->info()->data_type(); |
| 406 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 407 | // Select appropriate function |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 408 | if(data_type == DataType::QS8) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 409 | { |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 410 | switch(pool_size) |
| 411 | { |
| 412 | case 2: |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 413 | switch(pool_type) |
| 414 | { |
| 415 | case PoolingType::AVG: |
| 416 | _func = &NEPoolingLayerKernel::pooling2_q8<PoolingType::AVG>; |
| 417 | break; |
| 418 | case PoolingType::MAX: |
| 419 | _func = &NEPoolingLayerKernel::pooling2_q8<PoolingType::MAX>; |
| 420 | break; |
| 421 | default: |
| 422 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 423 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 424 | break; |
| 425 | case 3: |
| 426 | switch(pool_type) |
| 427 | { |
| 428 | case PoolingType::AVG: |
| 429 | _func = &NEPoolingLayerKernel::pooling3_q8<PoolingType::AVG>; |
| 430 | break; |
| 431 | case PoolingType::MAX: |
| 432 | _func = &NEPoolingLayerKernel::pooling3_q8<PoolingType::MAX>; |
| 433 | break; |
| 434 | default: |
| 435 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 436 | } |
| 437 | break; |
| 438 | default: |
| 439 | ARM_COMPUTE_ERROR("Unsupported pooling size!"); |
| 440 | } |
| 441 | } |
| 442 | else if(data_type == DataType::QASYMM8) |
| 443 | { |
| 444 | if(pool_size == 2 && pool_stride_x < 3) |
| 445 | { |
| 446 | switch(pool_type) |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 447 | { |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 448 | case PoolingType::AVG: |
| 449 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling2_qasymm8<PoolingType::AVG, true> : &NEPoolingLayerKernel::pooling2_qasymm8<PoolingType::AVG, false>; |
| 450 | break; |
| 451 | case PoolingType::MAX: |
| 452 | _func = &NEPoolingLayerKernel::pooling2_qasymm8<PoolingType::MAX>; |
| 453 | break; |
| 454 | default: |
| 455 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 456 | } |
| 457 | } |
| 458 | else if(pool_size == 3 && pool_stride_x < 3) |
| 459 | { |
| 460 | switch(pool_type) |
| 461 | { |
| 462 | case PoolingType::AVG: |
| 463 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling3_qasymm8<PoolingType::AVG, true> : &NEPoolingLayerKernel::pooling3_qasymm8<PoolingType::AVG, false>; |
| 464 | break; |
| 465 | case PoolingType::MAX: |
| 466 | _func = &NEPoolingLayerKernel::pooling3_qasymm8<PoolingType::MAX>; |
| 467 | break; |
| 468 | default: |
| 469 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 470 | } |
| 471 | } |
| 472 | else |
| 473 | { |
| 474 | switch(pool_type) |
| 475 | { |
| 476 | case PoolingType::AVG: |
| 477 | _func = (exclude_padding) ? &NEPoolingLayerKernel::poolingN_qasymm8<PoolingType::AVG, true> : &NEPoolingLayerKernel::poolingN_qasymm8<PoolingType::AVG, false>; |
| 478 | break; |
| 479 | case PoolingType::MAX: |
| 480 | _func = &NEPoolingLayerKernel::poolingN_qasymm8<PoolingType::MAX>; |
| 481 | break; |
| 482 | default: |
| 483 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 484 | } |
| 485 | } |
| 486 | } |
| 487 | else if(data_type == DataType::QS16) |
| 488 | { |
| 489 | switch(pool_size) |
| 490 | { |
| 491 | case 2: |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 492 | switch(pool_type) |
| 493 | { |
| 494 | case PoolingType::AVG: |
| 495 | _func = &NEPoolingLayerKernel::pooling2_q16<PoolingType::AVG>; |
| 496 | break; |
| 497 | case PoolingType::MAX: |
| 498 | _func = &NEPoolingLayerKernel::pooling2_q16<PoolingType::MAX>; |
| 499 | break; |
| 500 | default: |
| 501 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 502 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 503 | break; |
| 504 | case 3: |
| 505 | switch(pool_type) |
| 506 | { |
| 507 | case PoolingType::AVG: |
| 508 | _func = &NEPoolingLayerKernel::pooling3_q16<PoolingType::AVG>; |
| 509 | break; |
| 510 | case PoolingType::MAX: |
| 511 | _func = &NEPoolingLayerKernel::pooling3_q16<PoolingType::MAX>; |
| 512 | break; |
| 513 | default: |
| 514 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 515 | } |
| 516 | break; |
| 517 | default: |
| 518 | ARM_COMPUTE_ERROR("Unsupported pooling size!"); |
| 519 | } |
| 520 | } |
| 521 | else if(data_type == DataType::F16) |
| 522 | { |
| 523 | switch(pool_size) |
| 524 | { |
| 525 | case 2: |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 526 | switch(pool_type) |
| 527 | { |
| 528 | case PoolingType::AVG: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 529 | _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] | 530 | break; |
| 531 | case PoolingType::L2: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 532 | _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] | 533 | break; |
| 534 | case PoolingType::MAX: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 535 | _func = &NEPoolingLayerKernel::pooling2_f16<PoolingType::MAX, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 536 | break; |
| 537 | default: |
| 538 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 539 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 540 | break; |
| 541 | case 3: |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 542 | switch(pool_type) |
| 543 | { |
| 544 | case PoolingType::AVG: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 545 | _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] | 546 | break; |
| 547 | case PoolingType::L2: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 548 | _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] | 549 | break; |
| 550 | case PoolingType::MAX: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 551 | _func = &NEPoolingLayerKernel::pooling3_f16<PoolingType::MAX, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 552 | break; |
| 553 | default: |
| 554 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 555 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 556 | break; |
| 557 | default: |
| 558 | ARM_COMPUTE_ERROR("Unsupported pooling size!"); |
| 559 | } |
| 560 | } |
| 561 | else if(data_type == DataType::F32) |
| 562 | { |
| 563 | switch(pool_size) |
| 564 | { |
| 565 | case 2: |
| 566 | switch(pool_type) |
| 567 | { |
| 568 | case PoolingType::AVG: |
| 569 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling2_f32<PoolingType::AVG, true> : &NEPoolingLayerKernel::pooling2_f32<PoolingType::AVG, false>; |
| 570 | break; |
| 571 | case PoolingType::L2: |
| 572 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling2_f32<PoolingType::L2, true> : &NEPoolingLayerKernel::pooling2_f32<PoolingType::L2, false>; |
| 573 | break; |
| 574 | case PoolingType::MAX: |
| 575 | _func = &NEPoolingLayerKernel::pooling2_f32<PoolingType::MAX, false>; |
| 576 | break; |
| 577 | default: |
| 578 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 579 | } |
| 580 | break; |
| 581 | case 3: |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 582 | switch(pool_type) |
| 583 | { |
| 584 | case PoolingType::AVG: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 585 | _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] | 586 | break; |
| 587 | case PoolingType::L2: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 588 | _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] | 589 | break; |
| 590 | case PoolingType::MAX: |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 591 | _func = &NEPoolingLayerKernel::pooling3_f32<PoolingType::MAX, false>; |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 592 | break; |
| 593 | default: |
| 594 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 595 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 596 | break; |
| 597 | case 7: |
| 598 | switch(pool_type) |
| 599 | { |
| 600 | case PoolingType::AVG: |
| 601 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling7_f32<PoolingType::AVG, true> : &NEPoolingLayerKernel::pooling7_f32<PoolingType::AVG, false>; |
| 602 | break; |
| 603 | case PoolingType::L2: |
| 604 | _func = (exclude_padding) ? &NEPoolingLayerKernel::pooling7_f32<PoolingType::L2, true> : &NEPoolingLayerKernel::pooling7_f32<PoolingType::L2, false>; |
| 605 | break; |
| 606 | case PoolingType::MAX: |
| 607 | _func = &NEPoolingLayerKernel::pooling7_f32<PoolingType::MAX, false>; |
| 608 | break; |
| 609 | default: |
| 610 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 611 | } |
| 612 | break; |
| 613 | default: |
| 614 | switch(pool_type) |
| 615 | { |
| 616 | case PoolingType::AVG: |
| 617 | _func = (exclude_padding) ? &NEPoolingLayerKernel::poolingN_f32<PoolingType::AVG, true> : &NEPoolingLayerKernel::poolingN_f32<PoolingType::AVG, false>; |
| 618 | break; |
| 619 | case PoolingType::L2: |
| 620 | _func = (exclude_padding) ? &NEPoolingLayerKernel::poolingN_f32<PoolingType::L2, true> : &NEPoolingLayerKernel::poolingN_f32<PoolingType::L2, false>; |
| 621 | break; |
| 622 | case PoolingType::MAX: |
| 623 | _func = &NEPoolingLayerKernel::poolingN_f32<PoolingType::MAX, false>; |
| 624 | break; |
| 625 | default: |
| 626 | ARM_COMPUTE_ERROR("Unsupported pooling type!"); |
| 627 | } |
| 628 | break; |
| 629 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 630 | } |
| 631 | |
| 632 | // Configure kernel window |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 633 | 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); |
| 634 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 635 | INEKernel::configure(win_config.second); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 636 | } |
| 637 | |
| 638 | template <PoolingType pooling_type> |
| 639 | void NEPoolingLayerKernel::pooling2_q8(const Window &window_input, const Window &window) |
| 640 | { |
| 641 | Iterator input(_input, window_input); |
| 642 | Iterator output(_output, window); |
| 643 | |
| 644 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 645 | constexpr int pool_size = 2; |
| 646 | int pool_pad_x = 0; |
| 647 | int pool_pad_y = 0; |
| 648 | int pool_stride_x = 0; |
| 649 | int pool_stride_y = 0; |
| 650 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 651 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 652 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 653 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 654 | |
| 655 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 656 | 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)); |
| 657 | |
| 658 | execute_window_loop(window, [&](const Coordinates & id) |
| 659 | { |
| 660 | const auto top_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_top_ptr + input.offset())); |
| 661 | 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] | 662 | qint8x8_t lower_res = {}; |
| 663 | qint8x8_t upper_res = {}; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 664 | if(pooling_type == PoolingType::AVG) |
| 665 | { |
| 666 | // Calculate scale |
| 667 | 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); |
| 668 | const qint8x8_t scale_vec = vdup_n_qs8(scale); |
| 669 | |
| 670 | // Perform pooling |
| 671 | const qint8x16_t sum_data = vqaddq_qs8(top_data, bottom_data); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 672 | lower_res = vqmul_qs8(vpadd_s8(vget_low_s8(sum_data), vget_high_s8(sum_data)), scale_vec, fixed_point_position); |
| 673 | if(pool_stride_x == 1) |
| 674 | { |
| 675 | const qint8x16_t sum_data_shifted = vextq_s8(sum_data, sum_data, 1); |
| 676 | upper_res = vqmul_qs8(vpadd_s8(vget_low_s8(sum_data_shifted), vget_high_s8(sum_data_shifted)), scale_vec, fixed_point_position); |
| 677 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 678 | } |
| 679 | else |
| 680 | { |
| 681 | const qint8x16_t max_data = vmaxq_s8(top_data, bottom_data); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 682 | lower_res = vpmax_s8(vget_low_s8(max_data), vget_high_s8(max_data)); |
| 683 | if(pool_stride_x == 1) |
| 684 | { |
| 685 | const qint8x16_t max_data_shifted = vextq_s8(max_data, max_data, 1); |
| 686 | upper_res = vpmax_s8(vget_low_s8(max_data_shifted), vget_high_s8(max_data_shifted)); |
| 687 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 688 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 689 | if(pool_stride_x == 1) |
| 690 | { |
Georgios Pinitas | dc460f1 | 2017-08-24 19:02:44 +0100 | [diff] [blame] | 691 | const qint8x8x2_t res = { { lower_res, upper_res } }; |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 692 | vst2_s8(reinterpret_cast<qint8_t *>(output.ptr()), res); |
| 693 | } |
| 694 | else |
| 695 | { |
| 696 | vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), lower_res); |
| 697 | } |
| 698 | }, |
| 699 | input, output); |
| 700 | } |
| 701 | |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 702 | template <PoolingType pooling_type, bool exclude_padding> |
| 703 | void NEPoolingLayerKernel::pooling2_qasymm8(const Window &window_input, const Window &window) |
| 704 | { |
| 705 | Iterator input(_input, window_input); |
| 706 | Iterator output(_output, window); |
| 707 | |
| 708 | constexpr int pool_size = 2; |
| 709 | int pool_pad_x = 0; |
| 710 | int pool_pad_y = 0; |
| 711 | int pool_stride_x = 0; |
| 712 | int pool_stride_y = 0; |
| 713 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 714 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 715 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 716 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
| 717 | |
| 718 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 719 | 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)); |
| 720 | |
| 721 | const int scale_step_x = (pool_stride_x == 1) ? 2 : 1; |
| 722 | |
| 723 | execute_window_loop(window, [&](const Coordinates & id) |
| 724 | { |
| 725 | const auto top_data = vld1q_u8(reinterpret_cast<const uint8_t *>(input_top_ptr + input.offset())); |
| 726 | const auto bottom_data = vld1q_u8(reinterpret_cast<const uint8_t *>(input_bottom_ptr + input.offset())); |
| 727 | uint8x8_t lower_res = {}; |
| 728 | uint8x8_t upper_res = {}; |
| 729 | |
| 730 | if(pooling_type != PoolingType::MAX) |
| 731 | { |
| 732 | const uint16x8x2_t top_data_u16 = { { vmovl_u8(vget_low_u8(top_data)), vmovl_u8(vget_high_u8(top_data)) } }; |
| 733 | const uint16x8x2_t bottom_data_u16 = { { vmovl_u8(vget_low_u8(bottom_data)), vmovl_u8(vget_high_u8(bottom_data)) } }; |
| 734 | |
| 735 | // Add rows |
| 736 | const uint16x8x2_t vrsum = |
| 737 | { |
| 738 | { |
| 739 | vaddq_u16(top_data_u16.val[0], bottom_data_u16.val[0]), |
| 740 | vaddq_u16(top_data_u16.val[1], bottom_data_u16.val[1]), |
| 741 | } |
| 742 | }; |
| 743 | |
| 744 | // Pair-wise add row data |
| 745 | const uint16x4x2_t vpsum = |
| 746 | { |
| 747 | { |
| 748 | vpadd_u16(vget_low_u16(vrsum.val[0]), vget_high_u16(vrsum.val[0])), |
| 749 | vpadd_u16(vget_low_u16(vrsum.val[1]), vget_high_u16(vrsum.val[1])), |
| 750 | } |
| 751 | }; |
| 752 | |
| 753 | uint16x8_t res_lower = vcombine_u16(vpsum.val[0], vpsum.val[1]); |
| 754 | |
| 755 | // Scale lower result |
| 756 | scale_vector_s16x8<exclude_padding>(res_lower, id, 0, scale_step_x, |
| 757 | pool_size, upper_bound_w, upper_bound_h, |
| 758 | pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 759 | lower_res = vmovn_u16(res_lower); |
| 760 | |
| 761 | // Compute upper result for stride_x == 1 |
| 762 | if(pool_stride_x == 1) |
| 763 | { |
| 764 | // Shifted row sum |
| 765 | const uint16x8x2_t vrsum_shifted = |
| 766 | { |
| 767 | { |
| 768 | vextq_u16(vrsum.val[0], vrsum.val[1], 1), |
| 769 | vextq_u16(vrsum.val[1], vrsum.val[1], 1) |
| 770 | } |
| 771 | }; |
| 772 | |
| 773 | // Pair-wise add shifted row |
| 774 | const uint16x4x2_t vpsum_shifted = |
| 775 | { |
| 776 | { |
| 777 | vpadd_u16(vget_low_u16(vrsum_shifted.val[0]), vget_high_u16(vrsum_shifted.val[0])), |
| 778 | vpadd_u16(vget_low_u16(vrsum_shifted.val[1]), vget_high_u16(vrsum_shifted.val[1])), |
| 779 | } |
| 780 | }; |
| 781 | uint16x8_t res_upper = vcombine_u16(vpsum_shifted.val[0], vpsum_shifted.val[1]); |
| 782 | |
| 783 | // Scale lower result |
| 784 | scale_vector_s16x8<exclude_padding>(res_upper, id, 1, 2, |
| 785 | pool_size, upper_bound_w, upper_bound_h, |
| 786 | pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 787 | upper_res = vmovn_u16(res_upper); |
| 788 | } |
| 789 | } |
| 790 | else |
| 791 | { |
| 792 | const uint8x16_t max_data = vmaxq_u8(top_data, bottom_data); |
| 793 | lower_res = vpmax_u8(vget_low_u8(max_data), vget_high_u8(max_data)); |
| 794 | if(pool_stride_x == 1) |
| 795 | { |
| 796 | const uint8x16_t max_data_shifted = vextq_u8(max_data, max_data, 1); |
| 797 | upper_res = vpmax_u8(vget_low_u8(max_data_shifted), vget_high_u8(max_data_shifted)); |
| 798 | } |
| 799 | } |
| 800 | |
| 801 | // Store result |
| 802 | if(pool_stride_x == 1) |
| 803 | { |
| 804 | const uint8x8x2_t res = { { lower_res, upper_res } }; |
| 805 | vst2_u8(reinterpret_cast<uint8_t *>(output.ptr()), res); |
| 806 | } |
| 807 | else |
| 808 | { |
| 809 | vst1_u8(reinterpret_cast<uint8_t *>(output.ptr()), lower_res); |
| 810 | } |
| 811 | }, |
| 812 | input, output); |
| 813 | } |
| 814 | |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 815 | template <PoolingType pooling_type> |
| 816 | void NEPoolingLayerKernel::pooling2_q16(const Window &window_input, const Window &window) |
| 817 | { |
| 818 | Iterator input(_input, window_input); |
| 819 | Iterator output(_output, window); |
| 820 | |
| 821 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 822 | constexpr int pool_size = 2; |
| 823 | int pool_pad_x = 0; |
| 824 | int pool_pad_y = 0; |
| 825 | int pool_stride_x = 0; |
| 826 | int pool_stride_y = 0; |
| 827 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 828 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 829 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 830 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 831 | |
| 832 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 833 | 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)); |
| 834 | |
| 835 | execute_window_loop(window, [&](const Coordinates & id) |
| 836 | { |
| 837 | const auto top_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_top_ptr + input.offset())); |
| 838 | const auto bottom_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_bottom_ptr + input.offset())); |
| 839 | qint16x4_t lower_res = {}; |
| 840 | qint16x4_t upper_res = {}; |
| 841 | if(pooling_type == PoolingType::AVG) |
| 842 | { |
| 843 | // Calculate scale |
| 844 | 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); |
| 845 | const qint16x4_t scale_vec = vdup_n_qs16(scale); |
| 846 | |
| 847 | // Perform pooling |
| 848 | const qint16x8_t sum_data = vqaddq_qs16(top_data, bottom_data); |
| 849 | lower_res = vqmul_qs16(vpadd_s16(vget_low_s16(sum_data), vget_high_s16(sum_data)), scale_vec, fixed_point_position); |
| 850 | if(pool_stride_x == 1) |
| 851 | { |
| 852 | const qint16x8_t sum_data_shifted = vextq_s16(sum_data, sum_data, 1); |
| 853 | upper_res = vqmul_qs16(vpadd_s16(vget_low_s16(sum_data_shifted), vget_high_s16(sum_data_shifted)), scale_vec, fixed_point_position); |
| 854 | } |
| 855 | } |
| 856 | else |
| 857 | { |
| 858 | const qint16x8_t max_data = vmaxq_s16(top_data, bottom_data); |
| 859 | lower_res = vpmax_s16(vget_low_s16(max_data), vget_high_s16(max_data)); |
| 860 | if(pool_stride_x == 1) |
| 861 | { |
| 862 | const qint16x8_t max_data_shifted = vextq_s16(max_data, max_data, 1); |
| 863 | upper_res = vpmax_s16(vget_low_s16(max_data_shifted), vget_high_s16(max_data_shifted)); |
| 864 | } |
| 865 | } |
| 866 | if(pool_stride_x == 1) |
| 867 | { |
Georgios Pinitas | dc460f1 | 2017-08-24 19:02:44 +0100 | [diff] [blame] | 868 | const qint16x4x2_t res = { { lower_res, upper_res } }; |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 869 | vst2_s16(reinterpret_cast<qint16_t *>(output.ptr()), res); |
| 870 | } |
| 871 | else |
| 872 | { |
| 873 | vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), lower_res); |
| 874 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 875 | }, |
| 876 | input, output); |
| 877 | } |
| 878 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 879 | template <PoolingType pooling_type, bool exclude_padding> |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 880 | void NEPoolingLayerKernel::pooling3_f16(const Window &window_input, const Window &window) |
| 881 | { |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 882 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 883 | Iterator input(_input, window_input); |
| 884 | Iterator output(_output, window); |
| 885 | |
| 886 | constexpr const int pool_size = 3; |
| 887 | int pool_pad_x = 0; |
| 888 | int pool_pad_y = 0; |
| 889 | int pool_stride_x = 0; |
| 890 | int pool_stride_y = 0; |
| 891 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 892 | 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] | 893 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 894 | 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] | 895 | |
| 896 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 897 | 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)); |
| 898 | 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)); |
| 899 | |
| 900 | execute_window_loop(window, [&](const Coordinates & id) |
| 901 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 902 | float16x4_t top_data = vld1_f16(reinterpret_cast<const float16_t *>(input_top_ptr + input.offset())); |
| 903 | float16x4_t middle_data = vld1_f16(reinterpret_cast<const float16_t *>(input_middle_ptr + input.offset())); |
| 904 | float16x4_t bottom_data = vld1_f16(reinterpret_cast<const float16_t *>(input_bottom_ptr + input.offset())); |
| 905 | float16x4_t res = {}; |
| 906 | |
| 907 | // Get power of 2 in case of l2 pooling |
| 908 | if(pooling_type == PoolingType::L2) |
| 909 | { |
| 910 | top_data = vmul_f16(top_data, top_data); |
| 911 | middle_data = vmul_f16(middle_data, middle_data); |
| 912 | bottom_data = vmul_f16(bottom_data, bottom_data); |
| 913 | } |
| 914 | |
| 915 | if(pooling_type != PoolingType::MAX) |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 916 | { |
| 917 | // Calculate scale |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 918 | 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] | 919 | const float16x4_t scale_v = vdup_n_f16(scale); |
| 920 | // Perform pooling |
| 921 | const float16x4_t sum_data = vadd_f16(vadd_f16(top_data, bottom_data), middle_data); |
| 922 | res = vpadd_f16(vset_lane_f16(0.f, sum_data, 3), sum_data); |
| 923 | res = vmul_f16(vpadd_f16(res, res), scale_v); |
| 924 | } |
| 925 | else |
| 926 | { |
| 927 | const float16x4_t max_data = vmax_f16(vmax_f16(top_data, bottom_data), middle_data); |
| 928 | res = vpmax_f16(vset_lane_f16(-std::numeric_limits<float>::max(), max_data, 3), max_data); |
| 929 | res = vpmax_f16(res, res); |
| 930 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 931 | |
| 932 | // Calculate square-root in case of l2 pooling |
| 933 | if(pooling_type == PoolingType::L2) |
| 934 | { |
| 935 | res = vinv_f16(vinvsqrt_f16(res)); |
| 936 | } |
| 937 | |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 938 | *(reinterpret_cast<float16_t *>(output.ptr())) = vget_lane_f16(res, 0); |
| 939 | }, |
| 940 | input, output); |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 941 | #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 942 | ARM_COMPUTE_UNUSED(window_input); |
| 943 | ARM_COMPUTE_UNUSED(window); |
| 944 | 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] | 945 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 946 | } |
| 947 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 948 | template <PoolingType pooling_type, bool exclude_padding> |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 949 | void NEPoolingLayerKernel::pooling2_f16(const Window &window_input, const Window &window) |
| 950 | { |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 951 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 952 | Iterator input(_input, window_input); |
| 953 | Iterator output(_output, window); |
| 954 | constexpr int pool_size = 2; |
| 955 | int pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y = 0; |
| 956 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 957 | 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] | 958 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 959 | 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] | 960 | |
| 961 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 962 | 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)); |
| 963 | |
| 964 | execute_window_loop(window, [&](const Coordinates & id) |
| 965 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 966 | auto top_data = vld2q_f16(reinterpret_cast<const float16_t *>(input_top_ptr + input.offset())); |
| 967 | 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] | 968 | float16x8_t res = {}; |
| 969 | |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 970 | // Get power of 2 in case of l2 pooling |
| 971 | if(pooling_type == PoolingType::L2) |
| 972 | { |
| 973 | top_data.val[0] = vmulq_f16(top_data.val[0], top_data.val[0]); |
| 974 | top_data.val[1] = vmulq_f16(top_data.val[1], top_data.val[1]); |
| 975 | bottom_data.val[0] = vmulq_f16(bottom_data.val[0], bottom_data.val[0]); |
| 976 | bottom_data.val[1] = vmulq_f16(bottom_data.val[1], bottom_data.val[1]); |
| 977 | } |
| 978 | |
| 979 | if(pooling_type != PoolingType::MAX) |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 980 | { |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 981 | 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] | 982 | const float16x8_t scale_v = vdupq_n_f16(scale); |
| 983 | 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])))); |
| 984 | } |
| 985 | else |
| 986 | { |
| 987 | res = vmaxq_f16(bottom_data.val[1], vmaxq_f16(bottom_data.val[0], vmaxq_f16(top_data.val[0], top_data.val[1]))); |
| 988 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 989 | |
| 990 | // Calculate square-root in case of l2 pooling |
| 991 | if(pooling_type == PoolingType::L2) |
| 992 | { |
| 993 | res = vinvq_f16(vinvsqrtq_f16(res)); |
| 994 | } |
| 995 | |
| 996 | // Store result |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 997 | vst1q_f16(reinterpret_cast<float16_t *>(output.ptr()), res); |
| 998 | }, |
| 999 | input, output); |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 1000 | #else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1001 | ARM_COMPUTE_UNUSED(window_input); |
| 1002 | ARM_COMPUTE_UNUSED(window); |
| 1003 | 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] | 1004 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1005 | } |
| 1006 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1007 | template <PoolingType pooling_type, bool exclude_padding> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1008 | void NEPoolingLayerKernel::pooling2_f32(const Window &window_input, const Window &window) |
| 1009 | { |
| 1010 | Iterator input(_input, window_input); |
| 1011 | Iterator output(_output, window); |
| 1012 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1013 | constexpr int pool_size = 2; |
| 1014 | int pool_pad_x = 0; |
| 1015 | int pool_pad_y = 0; |
| 1016 | int pool_stride_x = 0; |
| 1017 | int pool_stride_y = 0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1018 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1019 | 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] | 1020 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1021 | 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] | 1022 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1023 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 1024 | 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] | 1025 | |
| 1026 | execute_window_loop(window, [&](const Coordinates & id) |
| 1027 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1028 | float32x2_t top_data = vld1_f32(reinterpret_cast<const float *>(input_top_ptr + input.offset())); |
| 1029 | float32x2_t bottom_data = vld1_f32(reinterpret_cast<const float *>(input_bottom_ptr + input.offset())); |
| 1030 | float32x2_t res = {}; |
| 1031 | float final_res = 0; |
| 1032 | |
| 1033 | // Get power of 2 in case of l2 pooling |
| 1034 | if(pooling_type == PoolingType::L2) |
| 1035 | { |
| 1036 | top_data = vmul_f32(top_data, top_data); |
| 1037 | bottom_data = vmul_f32(bottom_data, bottom_data); |
| 1038 | } |
| 1039 | |
| 1040 | if(pooling_type != PoolingType::MAX) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1041 | { |
| 1042 | // Calculate scale |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1043 | 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] | 1044 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 1045 | |
| 1046 | // Perform pooling |
| 1047 | const float32x2_t sum_data = vadd_f32(top_data, bottom_data); |
| 1048 | res = vmul_f32(vpadd_f32(sum_data, sum_data), scale_v); |
| 1049 | } |
| 1050 | else |
| 1051 | { |
| 1052 | const float32x2_t max_data = vmax_f32(top_data, bottom_data); |
| 1053 | res = vpmax_f32(max_data, max_data); |
| 1054 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1055 | final_res = vget_lane_f32(res, 0); |
| 1056 | |
| 1057 | // Calculate square-root in case of l2 pooling |
| 1058 | if(pooling_type == PoolingType::L2) |
| 1059 | { |
| 1060 | final_res = sqrt(final_res); |
| 1061 | } |
| 1062 | |
| 1063 | // Store result |
| 1064 | *(reinterpret_cast<float *>(output.ptr())) = final_res; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1065 | }, |
| 1066 | input, output); |
| 1067 | } |
| 1068 | |
| 1069 | template <PoolingType pooling_type> |
| 1070 | void NEPoolingLayerKernel::pooling3_q8(const Window &window_input, const Window &window) |
| 1071 | { |
| 1072 | Iterator input(_input, window_input); |
| 1073 | Iterator output(_output, window); |
| 1074 | |
| 1075 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 1076 | constexpr int pool_size = 3; |
| 1077 | int pool_pad_x = 0; |
| 1078 | int pool_pad_y = 0; |
| 1079 | int pool_stride_x = 0; |
| 1080 | int pool_stride_y = 0; |
| 1081 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1082 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 1083 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 1084 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 1085 | |
| 1086 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 1087 | 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)); |
| 1088 | 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)); |
| 1089 | |
| 1090 | execute_window_loop(window, [&](const Coordinates & id) |
| 1091 | { |
| 1092 | const auto top_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_top_ptr + input.offset())); |
| 1093 | const auto middle_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_middle_ptr + input.offset())); |
| 1094 | const auto bottom_data = vld1q_qs8(reinterpret_cast<const qint8_t *>(input_bottom_ptr + input.offset())); |
| 1095 | qint8x8_t res = {}; |
| 1096 | if(pooling_type == PoolingType::AVG) |
| 1097 | { |
| 1098 | // Calculate scale |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1099 | 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] | 1100 | |
| 1101 | // Perform pooling for stride 2 |
| 1102 | const qint8x16_t sum_data = vqaddq_qs8(vqaddq_qs8(top_data, bottom_data), middle_data); |
| 1103 | const qint8x16_t sum_data2 = vextq_s8(sum_data, sum_data, 1); |
| 1104 | const qint8x16_t sum_data3 = vextq_s8(sum_data, sum_data, 2); |
| 1105 | const qint8x16_t final_sum = vqaddq_qs8(vqaddq_qs8(sum_data, sum_data2), sum_data3); |
| 1106 | if(pool_stride_x == 2) |
| 1107 | { |
| 1108 | const qint8x8x2_t table = { { vget_low_s8(final_sum), vget_high_s8(final_sum) } }; |
| 1109 | 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] | 1110 | const qint8x8_t scale_vec = vdup_n_qs8(scale); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1111 | res = vtbl2_s8(table, lookup_val); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1112 | res = vqmul_qs8(res, scale_vec, fixed_point_position); |
| 1113 | vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), res); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1114 | } |
| 1115 | else |
| 1116 | { |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1117 | const qint8x16_t scale_vec = vdupq_n_qs8(scale); |
| 1118 | 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] | 1119 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1120 | } |
| 1121 | else |
| 1122 | { |
| 1123 | const qint8x16_t max_data = vmaxq_s8(vmaxq_s8(top_data, bottom_data), middle_data); |
| 1124 | const qint8x16_t max_data2 = vextq_s8(max_data, max_data, 1); |
| 1125 | const qint8x16_t max_data3 = vextq_s8(max_data, max_data, 2); |
| 1126 | const qint8x16_t final_max = vmaxq_s8(vmaxq_s8(max_data, max_data2), max_data3); |
| 1127 | |
| 1128 | if(pool_stride_x == 2) |
| 1129 | { |
| 1130 | const qint8x8x2_t table = { { vget_low_s8(final_max), vget_high_s8(final_max) } }; |
| 1131 | static const qint8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; |
| 1132 | res = vtbl2_s8(table, lookup_val); |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1133 | vst1_qs8(reinterpret_cast<qint8_t *>(output.ptr()), res); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1134 | } |
| 1135 | else |
| 1136 | { |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1137 | vst1q_qs8(reinterpret_cast<qint8_t *>(output.ptr()), final_max); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1138 | } |
| 1139 | } |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1140 | }, |
| 1141 | input, output); |
| 1142 | } |
| 1143 | |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 1144 | template <PoolingType pooling_type, bool exclude_padding> |
| 1145 | void NEPoolingLayerKernel::pooling3_qasymm8(const Window &window_input, const Window &window) |
| 1146 | { |
| 1147 | Iterator input(_input, window_input); |
| 1148 | Iterator output(_output, window); |
| 1149 | |
| 1150 | constexpr int pool_size = 3; |
| 1151 | int pool_pad_x = 0; |
| 1152 | int pool_pad_y = 0; |
| 1153 | int pool_stride_x = 0; |
| 1154 | int pool_stride_y = 0; |
| 1155 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1156 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 1157 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1158 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
| 1159 | |
| 1160 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 1161 | 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)); |
| 1162 | 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)); |
| 1163 | |
| 1164 | execute_window_loop(window, [&](const Coordinates & id) |
| 1165 | { |
| 1166 | const auto top_data = vld1q_u8(reinterpret_cast<const uint8_t *>(input_top_ptr + input.offset())); |
| 1167 | const auto middle_data = vld1q_u8(reinterpret_cast<const uint8_t *>(input_middle_ptr + input.offset())); |
| 1168 | const auto bottom_data = vld1q_u8(reinterpret_cast<const uint8_t *>(input_bottom_ptr + input.offset())); |
| 1169 | |
| 1170 | if(pooling_type == PoolingType::AVG) |
| 1171 | { |
| 1172 | // Convert data to u16 |
| 1173 | const uint16x8x2_t top_data_u16 = { { vmovl_u8(vget_low_u8(top_data)), vmovl_u8(vget_high_u8(top_data)) } }; |
| 1174 | const uint16x8x2_t middle_data_u16 = { { vmovl_u8(vget_low_u8(middle_data)), vmovl_u8(vget_high_u8(middle_data)) } }; |
| 1175 | const uint16x8x2_t bottom_data_u16 = { { vmovl_u8(vget_low_u8(bottom_data)), vmovl_u8(vget_high_u8(bottom_data)) } }; |
| 1176 | |
| 1177 | // Calculate row sums |
| 1178 | const uint16x8x2_t vrsum = |
| 1179 | { |
| 1180 | { |
| 1181 | vaddq_u16(vaddq_u16(top_data_u16.val[0], bottom_data_u16.val[0]), middle_data_u16.val[0]), |
| 1182 | vaddq_u16(vaddq_u16(top_data_u16.val[1], bottom_data_u16.val[1]), middle_data_u16.val[1]), |
| 1183 | } |
| 1184 | }; |
| 1185 | const uint16x8x2_t vrsum_shifted_1 = |
| 1186 | { |
| 1187 | { |
| 1188 | vextq_u16(vrsum.val[0], vrsum.val[1], 1), |
| 1189 | vextq_u16(vrsum.val[1], vrsum.val[1], 1) |
| 1190 | } |
| 1191 | }; |
| 1192 | const uint16x8x2_t vrsum_shifted_2 = |
| 1193 | { |
| 1194 | { |
| 1195 | vextq_u16(vrsum.val[0], vrsum.val[1], 2), |
| 1196 | vextq_u16(vrsum.val[1], vrsum.val[1], 2) |
| 1197 | } |
| 1198 | }; |
| 1199 | // Calculate final sum |
| 1200 | uint16x8x2_t final_sum = |
| 1201 | { |
| 1202 | { |
| 1203 | vaddq_u16(vaddq_u16(vrsum.val[0], vrsum_shifted_1.val[0]), vrsum_shifted_2.val[0]), |
| 1204 | vaddq_u16(vaddq_u16(vrsum.val[1], vrsum_shifted_1.val[1]), vrsum_shifted_2.val[1]), |
| 1205 | } |
| 1206 | }; |
| 1207 | if(pool_stride_x == 2) |
| 1208 | { |
| 1209 | uint16x8_t res = |
| 1210 | { |
| 1211 | vgetq_lane_u16(final_sum.val[0], 0), |
| 1212 | vgetq_lane_u16(final_sum.val[0], 2), |
| 1213 | vgetq_lane_u16(final_sum.val[0], 4), |
| 1214 | vgetq_lane_u16(final_sum.val[0], 6), |
| 1215 | vgetq_lane_u16(final_sum.val[1], 0), |
| 1216 | vgetq_lane_u16(final_sum.val[1], 2), |
| 1217 | vgetq_lane_u16(final_sum.val[1], 4), |
| 1218 | vgetq_lane_u16(final_sum.val[1], 6), |
| 1219 | }; |
| 1220 | |
| 1221 | scale_vector_s16x8<exclude_padding>(res, id, 0, 1, |
| 1222 | pool_size, upper_bound_w, upper_bound_h, |
| 1223 | pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 1224 | vst1_u8(reinterpret_cast<uint8_t *>(output.ptr()), vmovn_u16(res)); |
| 1225 | } |
| 1226 | else |
| 1227 | { |
| 1228 | // Scale lower result |
| 1229 | scale_vector_s16x8<exclude_padding>(final_sum.val[0], id, 0, 1, |
| 1230 | pool_size, upper_bound_w, upper_bound_h, |
| 1231 | pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 1232 | // Scale lower result |
| 1233 | scale_vector_s16x8<exclude_padding>(final_sum.val[1], id, 8, 1, |
| 1234 | pool_size, upper_bound_w, upper_bound_h, |
| 1235 | pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y); |
| 1236 | const uint8x16_t res = vcombine_u8(vmovn_u16(final_sum.val[0]), vmovn_u16(final_sum.val[1])); |
| 1237 | vst1q_u8(reinterpret_cast<uint8_t *>(output.ptr()), res); |
| 1238 | } |
| 1239 | } |
| 1240 | else |
| 1241 | { |
| 1242 | const uint8x16_t max_data = vmaxq_u8(vmaxq_u8(top_data, bottom_data), middle_data); |
| 1243 | const uint8x16_t max_data_shift1 = vextq_u8(max_data, max_data, 1); |
| 1244 | const uint8x16_t max_data_shift2 = vextq_u8(max_data, max_data, 2); |
| 1245 | const uint8x16_t final_max = vmaxq_u8(vmaxq_u8(max_data, max_data_shift1), max_data_shift2); |
| 1246 | |
| 1247 | if(pool_stride_x == 2) |
| 1248 | { |
| 1249 | const uint8x8x2_t table = { { vget_low_u8(final_max), vget_high_u8(final_max) } }; |
| 1250 | static const uint8x8_t lookup_val = { 0, 2, 4, 6, 8, 10, 12, 14 }; |
| 1251 | const uint8x8_t res = vtbl2_u8(table, lookup_val); |
| 1252 | vst1_u8(reinterpret_cast<uint8_t *>(output.ptr()), res); |
| 1253 | } |
| 1254 | else |
| 1255 | { |
| 1256 | vst1q_u8(reinterpret_cast<uint8_t *>(output.ptr()), final_max); |
| 1257 | } |
| 1258 | } |
| 1259 | }, |
| 1260 | input, output); |
| 1261 | } |
| 1262 | |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1263 | template <PoolingType pooling_type> |
| 1264 | void NEPoolingLayerKernel::pooling3_q16(const Window &window_input, const Window &window) |
| 1265 | { |
| 1266 | Iterator input(_input, window_input); |
| 1267 | Iterator output(_output, window); |
| 1268 | |
| 1269 | const int fixed_point_position = _input->info()->fixed_point_position(); |
| 1270 | constexpr int pool_size = 3; |
| 1271 | int pool_pad_x = 0; |
| 1272 | int pool_pad_y = 0; |
| 1273 | int pool_stride_x = 0; |
| 1274 | int pool_stride_y = 0; |
| 1275 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1276 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 1277 | const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x; |
| 1278 | const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y; |
| 1279 | |
| 1280 | const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 1281 | 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)); |
| 1282 | 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)); |
| 1283 | |
| 1284 | execute_window_loop(window, [&](const Coordinates & id) |
| 1285 | { |
| 1286 | const auto top_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_top_ptr + input.offset())); |
| 1287 | const auto middle_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_middle_ptr + input.offset())); |
| 1288 | const auto bottom_data = vld1q_qs16(reinterpret_cast<const qint16_t *>(input_bottom_ptr + input.offset())); |
| 1289 | |
| 1290 | if(pooling_type == PoolingType::AVG) |
| 1291 | { |
| 1292 | // Calculate scale |
| 1293 | 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); |
| 1294 | |
| 1295 | // Perform pooling for stride 2 |
| 1296 | const qint16x8_t sum_data = vqaddq_qs16(vqaddq_qs16(top_data, bottom_data), middle_data); |
| 1297 | const qint16x8_t sum_data2 = vextq_s16(sum_data, sum_data, 1); |
| 1298 | const qint16x8_t sum_data3 = vextq_s16(sum_data, sum_data, 2); |
| 1299 | const qint16x8_t final_sum = vqaddq_qs16(vqaddq_qs16(sum_data, sum_data2), sum_data3); |
| 1300 | if(pool_stride_x == 2) |
| 1301 | { |
| 1302 | 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) }; |
| 1303 | const qint16x4_t scale_vec = vdup_n_qs16(scale); |
| 1304 | vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), vqmul_qs16(tmp, scale_vec, fixed_point_position)); |
| 1305 | } |
| 1306 | else |
| 1307 | { |
| 1308 | const qint16x8_t scale_vec = vdupq_n_qs16(scale); |
| 1309 | vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()), vqmulq_qs16(final_sum, scale_vec, fixed_point_position)); |
| 1310 | } |
| 1311 | } |
| 1312 | else |
| 1313 | { |
| 1314 | const qint16x8_t max_data = vmaxq_s16(vmaxq_s16(top_data, bottom_data), middle_data); |
| 1315 | const qint16x8_t max_data2 = vextq_s16(max_data, max_data, 1); |
| 1316 | const qint16x8_t max_data3 = vextq_s16(max_data, max_data, 2); |
| 1317 | const qint16x8_t final_max = vmaxq_s16(vmaxq_s16(max_data, max_data2), max_data3); |
| 1318 | |
| 1319 | if(pool_stride_x == 2) |
| 1320 | { |
| 1321 | 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) }; |
| 1322 | vst1_qs16(reinterpret_cast<qint16_t *>(output.ptr()), tmp); |
| 1323 | } |
| 1324 | else |
| 1325 | { |
| 1326 | vst1q_qs16(reinterpret_cast<qint16_t *>(output.ptr()), final_max); |
| 1327 | } |
| 1328 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1329 | }, |
| 1330 | input, output); |
| 1331 | } |
| 1332 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1333 | template <PoolingType pooling_type, bool exclude_padding> |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1334 | void NEPoolingLayerKernel::pooling3_f32(const Window &window_input, const Window &window) |
| 1335 | { |
| 1336 | Iterator input(_input, window_input); |
| 1337 | Iterator output(_output, window); |
| 1338 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1339 | constexpr const int pool_size = 3; |
| 1340 | int pool_pad_x = 0; |
| 1341 | int pool_pad_y = 0; |
| 1342 | int pool_stride_x = 0; |
| 1343 | int pool_stride_y = 0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1344 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1345 | 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] | 1346 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1347 | 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] | 1348 | |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1349 | const uint8_t *const input_top_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y))); |
| 1350 | 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)); |
| 1351 | 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] | 1352 | |
| 1353 | execute_window_loop(window, [&](const Coordinates & id) |
| 1354 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1355 | float32x4_t top_data = vld1q_f32(reinterpret_cast<const float *>(input_top_ptr + input.offset())); |
| 1356 | float32x4_t middle_data = vld1q_f32(reinterpret_cast<const float *>(input_middle_ptr + input.offset())); |
| 1357 | float32x4_t bottom_data = vld1q_f32(reinterpret_cast<const float *>(input_bottom_ptr + input.offset())); |
| 1358 | float32x2_t res = {}; |
| 1359 | float final_res = 0; |
| 1360 | |
| 1361 | // Get power of 2 in case of l2 pooling |
| 1362 | if(pooling_type == PoolingType::L2) |
| 1363 | { |
| 1364 | top_data = vmulq_f32(top_data, top_data); |
| 1365 | middle_data = vmulq_f32(middle_data, middle_data); |
| 1366 | bottom_data = vmulq_f32(bottom_data, bottom_data); |
| 1367 | } |
| 1368 | |
| 1369 | if(pooling_type != PoolingType::MAX) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1370 | { |
| 1371 | // Calculate scale |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1372 | 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] | 1373 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 1374 | |
| 1375 | // Perform pooling |
| 1376 | const float32x4_t sum_data = vaddq_f32(vaddq_f32(top_data, bottom_data), middle_data); |
| 1377 | res = vpadd_f32(vget_high_f32(vsetq_lane_f32(0.f, sum_data, 3)), vget_low_f32(sum_data)); |
| 1378 | res = vmul_f32(vpadd_f32(res, res), scale_v); |
| 1379 | } |
| 1380 | else |
| 1381 | { |
| 1382 | const float32x4_t max_data = vmaxq_f32(vmaxq_f32(top_data, bottom_data), middle_data); |
| 1383 | res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits<float>::max(), max_data, 3)), vget_low_f32(max_data)); |
| 1384 | res = vpmax_f32(res, res); |
| 1385 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1386 | final_res = vget_lane_f32(res, 0); |
| 1387 | |
| 1388 | // Calculate square-root in case of l2 pooling |
| 1389 | if(pooling_type == PoolingType::L2) |
| 1390 | { |
| 1391 | final_res = sqrt(final_res); |
| 1392 | } |
| 1393 | |
| 1394 | // Store result |
| 1395 | *(reinterpret_cast<float *>(output.ptr())) = final_res; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1396 | }, |
| 1397 | input, output); |
| 1398 | } |
| 1399 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1400 | template <PoolingType pooling_type, bool exclude_padding> |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1401 | void NEPoolingLayerKernel::pooling7_f32(const Window &window_input, const Window &window) |
| 1402 | { |
| 1403 | Iterator input(_input, window_input); |
| 1404 | Iterator output(_output, window); |
| 1405 | |
| 1406 | constexpr const int pool_size = 7; |
| 1407 | int pool_pad_x = 0; |
| 1408 | int pool_pad_y = 0; |
| 1409 | int pool_stride_x = 0; |
| 1410 | int pool_stride_y = 0; |
| 1411 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1412 | 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] | 1413 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1414 | 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] | 1415 | |
| 1416 | std::array<const uint8_t *, pool_size> input_ptrs{ {} }; |
| 1417 | for(int i = 0; i < pool_size; ++i) |
| 1418 | { |
| 1419 | input_ptrs[i] = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + i)); |
| 1420 | } |
| 1421 | |
| 1422 | execute_window_loop(window, [&](const Coordinates & id) |
| 1423 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1424 | float32x2_t res = {}; |
| 1425 | float final_res = 0.f; |
| 1426 | if(pooling_type != PoolingType::MAX) |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1427 | { |
| 1428 | // Calculate scale |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1429 | 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] | 1430 | const float32x2_t scale_v = vdup_n_f32(scale); |
| 1431 | |
| 1432 | // Perform pooling |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1433 | float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[0] + input.offset())); |
| 1434 | // Get power of 2 in case of l2 pooling |
| 1435 | if(pooling_type == PoolingType::L2) |
| 1436 | { |
| 1437 | data.val[0] = vmulq_f32(data.val[0], data.val[0]); |
| 1438 | data.val[1] = vmulq_f32(data.val[1], data.val[1]); |
| 1439 | } |
| 1440 | 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] | 1441 | for(int i = 1; i < pool_size; ++i) |
| 1442 | { |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1443 | data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[i] + input.offset())); |
| 1444 | // Get power of 2 in case of l2 pooling |
| 1445 | if(pooling_type == PoolingType::L2) |
| 1446 | { |
| 1447 | data.val[0] = vmulq_f32(data.val[0], data.val[0]); |
| 1448 | data.val[1] = vmulq_f32(data.val[1], data.val[1]); |
| 1449 | } |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1450 | sum_data = vaddq_f32(sum_data, data.val[0]); |
| 1451 | sum_data = vaddq_f32(sum_data, vsetq_lane_f32(0.f, data.val[1], 3)); |
| 1452 | } |
| 1453 | res = vpadd_f32(vget_high_f32(sum_data), vget_low_f32(sum_data)); |
| 1454 | res = vmul_f32(vpadd_f32(res, res), scale_v); |
| 1455 | } |
| 1456 | else |
| 1457 | { |
| 1458 | float32x4x2_t max_data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[0] + input.offset())); |
| 1459 | for(int i = 1; i < pool_size; ++i) |
| 1460 | { |
| 1461 | const float32x4x2_t data = vld2q_f32(reinterpret_cast<const float *>(input_ptrs[i] + input.offset())); |
| 1462 | max_data = vmax2q_f32(max_data, data); |
| 1463 | } |
| 1464 | 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])); |
| 1465 | res = vpmax_f32(res, vpmax_f32(vget_high_f32(max_data.val[0]), vget_low_f32(max_data.val[0]))); |
| 1466 | res = vpmax_f32(res, res); |
| 1467 | } |
Georgios Pinitas | cdf5145 | 2017-08-31 14:21:36 +0100 | [diff] [blame] | 1468 | final_res = vget_lane_f32(res, 0); |
| 1469 | |
| 1470 | // Calculate square-root in case of l2 pooling |
| 1471 | if(pooling_type == PoolingType::L2) |
| 1472 | { |
| 1473 | final_res = sqrt(final_res); |
| 1474 | } |
| 1475 | |
| 1476 | // Store result |
| 1477 | *(reinterpret_cast<float *>(output.ptr())) = final_res; |
Michele Di Giorgio | 8af2dd6 | 2017-06-19 15:19:29 +0100 | [diff] [blame] | 1478 | }, |
| 1479 | input, output); |
| 1480 | } |
| 1481 | |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1482 | template <PoolingType pooling_type, bool exclude_padding> |
Gian Marco Iodice | 1682430 | 2017-09-28 15:41:37 +0100 | [diff] [blame] | 1483 | void NEPoolingLayerKernel::poolingN_f32(const Window &window_input, const Window &window) |
| 1484 | { |
| 1485 | Iterator input(_input, window_input); |
| 1486 | Iterator output(_output, window); |
| 1487 | |
Georgios Pinitas | 4c2dd54 | 2017-11-13 12:58:41 +0000 | [diff] [blame] | 1488 | 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] | 1489 | int pool_pad_x = 0; |
| 1490 | int pool_pad_y = 0; |
| 1491 | int pool_stride_x = 0; |
| 1492 | int pool_stride_y = 0; |
| 1493 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1494 | 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] | 1495 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1496 | 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] | 1497 | |
| 1498 | execute_window_loop(window, [&](const Coordinates & id) |
| 1499 | { |
| 1500 | float res = 0.0f; |
| 1501 | |
| 1502 | if(pooling_type != PoolingType::MAX) |
| 1503 | { |
| 1504 | // Calculate scale |
Georgios Pinitas | adaae7e | 2017-10-30 15:56:32 +0000 | [diff] [blame] | 1505 | 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] | 1506 | |
| 1507 | // Perform pooling |
| 1508 | float32x4_t vres = vdupq_n_f32(0.0f); |
| 1509 | |
| 1510 | for(int y = 0; y < pool_size; ++y) |
| 1511 | { |
| 1512 | int x = 0; |
| 1513 | for(; x <= (pool_size - 4); x += 4) |
| 1514 | { |
| 1515 | const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + |
| 1516 | (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1517 | |
| 1518 | // Get power of 2 in case of l2 pooling and accumulate |
| 1519 | if(pooling_type == PoolingType::L2) |
| 1520 | { |
| 1521 | vres = vmlaq_f32(vres, data, data); |
| 1522 | } |
| 1523 | else |
| 1524 | { |
| 1525 | vres = vaddq_f32(vres, data); |
| 1526 | } |
| 1527 | } |
| 1528 | |
| 1529 | // Leftover for loop |
| 1530 | for(; x < pool_size; ++x) |
| 1531 | { |
| 1532 | 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())); |
| 1533 | |
| 1534 | // Get power of 2 in case of l2 pooling |
| 1535 | if(pooling_type == PoolingType::L2) |
| 1536 | { |
| 1537 | data *= data; |
| 1538 | } |
| 1539 | |
| 1540 | res += data; |
| 1541 | } |
| 1542 | } |
| 1543 | |
| 1544 | #if defined(__aarch64__) |
| 1545 | // Reduction operation available on 64 bit architectures only |
| 1546 | res += vaddvq_f32(vres); |
| 1547 | #else // __aarch64__ |
| 1548 | // Reduction |
| 1549 | float32x2_t tmp = vpadd_f32(vget_high_f32(vres), vget_low_f32(vres)); |
| 1550 | tmp = vpadd_f32(tmp, tmp); |
| 1551 | |
| 1552 | res += vget_lane_f32(tmp, 0); |
| 1553 | #endif // __aarch64__ |
| 1554 | // Divide by scale |
| 1555 | res *= scale; |
| 1556 | } |
| 1557 | else |
| 1558 | { |
| 1559 | float32x4_t vres = vdupq_n_f32(std::numeric_limits<float>::min()); |
| 1560 | res = std::numeric_limits<float>::min(); |
| 1561 | |
| 1562 | for(int y = 0; y < pool_size; ++y) |
| 1563 | { |
| 1564 | int x = 0; |
| 1565 | for(; x <= (pool_size - 4); x += 4) |
| 1566 | { |
| 1567 | const float32x4_t data = vld1q_f32(reinterpret_cast<const float *>(input.ptr() + (x - pool_pad_x) * _input->info()->strides_in_bytes().x() + |
| 1568 | (y - pool_pad_y) * _input->info()->strides_in_bytes().y())); |
| 1569 | vres = vmaxq_f32(vres, data); |
| 1570 | } |
| 1571 | |
| 1572 | // Leftover for loop |
| 1573 | for(; x < pool_size; ++x) |
| 1574 | { |
| 1575 | 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())); |
| 1576 | res = std::max(res, data); |
| 1577 | } |
| 1578 | } |
| 1579 | |
| 1580 | #if defined(__aarch64__) |
| 1581 | // Reduction operation available on 64 bit architectures only |
| 1582 | res = std::max(vmaxvq_f32(vres), res); |
| 1583 | #else // __aarch64__ |
| 1584 | float32x2_t tmp = vpmax_f32(vget_high_f32(vres), vget_low_f32(vres)); |
| 1585 | tmp = vpmax_f32(tmp, tmp); |
| 1586 | |
| 1587 | res = std::max(res, vget_lane_f32(tmp, 0)); |
| 1588 | #endif // __aarch64__ |
| 1589 | } |
| 1590 | |
| 1591 | // Calculate square-root in case of l2 pooling |
| 1592 | if(pooling_type == PoolingType::L2) |
| 1593 | { |
| 1594 | res = std::sqrt(res); |
| 1595 | } |
| 1596 | |
| 1597 | // Store result |
| 1598 | *(reinterpret_cast<float *>(output.ptr())) = res; |
| 1599 | }, |
| 1600 | input, output); |
| 1601 | } |
| 1602 | |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 1603 | template <PoolingType pooling_type, bool exclude_padding> |
| 1604 | void NEPoolingLayerKernel::poolingN_qasymm8(const Window &window_input, const Window &window) |
| 1605 | { |
| 1606 | Iterator input(_input, window_input); |
| 1607 | Iterator output(_output, window); |
| 1608 | |
| 1609 | const int pool_size = _pool_info.is_global_pooling() ? _input->info()->tensor_shape().x() : _pool_info.pool_size(); |
| 1610 | int pool_pad_x = 0; |
| 1611 | int pool_pad_y = 0; |
| 1612 | int pool_stride_x = 0; |
| 1613 | int pool_stride_y = 0; |
| 1614 | std::tie(pool_pad_x, pool_pad_y) = _pool_info.pad_stride_info().pad(); |
| 1615 | std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride(); |
| 1616 | const int upper_bound_w = _input->info()->dimension(0) + (exclude_padding ? 0 : pool_pad_x); |
| 1617 | const int upper_bound_h = _input->info()->dimension(1) + (exclude_padding ? 0 : pool_pad_y); |
| 1618 | |
| 1619 | execute_window_loop(window, [&](const Coordinates & id) |
| 1620 | { |
| 1621 | uint8_t res = 0; |
| 1622 | |
| 1623 | if(pooling_type != PoolingType::MAX) |
| 1624 | { |
| 1625 | uint32x4_t vres = vdupq_n_u32(0); |
| 1626 | uint32_t sres = 0; |
| 1627 | |
| 1628 | // Calculate scale |
| 1629 | 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); |
| 1630 | |
| 1631 | // Perform pooling |
| 1632 | for(int y = 0; y < pool_size; ++y) |
| 1633 | { |
| 1634 | int x = 0; |
| 1635 | for(; x <= (pool_size - 8); x += 8) |
| 1636 | { |
| 1637 | 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())); |
| 1638 | |
| 1639 | const uint16x8_t data_u16 = vmovl_u8(data); |
| 1640 | vres = vaddq_u32(vres, vaddl_u16(vget_high_u16(data_u16), vget_low_u16(data_u16))); |
| 1641 | } |
| 1642 | |
| 1643 | // Leftover for loop |
| 1644 | for(; x < pool_size; ++x) |
| 1645 | { |
| 1646 | 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())); |
| 1647 | sres += data; |
| 1648 | } |
| 1649 | } |
| 1650 | |
| 1651 | // Reduction |
| 1652 | const auto tmp = vpadd_u32(vget_high_u32(vres), vget_low_u32(vres)); |
| 1653 | sres += vget_lane_u32(tmp, 0) + vget_lane_u32(tmp, 1); |
| 1654 | |
| 1655 | // Divide by scale |
| 1656 | res = static_cast<uint8_t>(support::cpp11::round(sres * scale)); |
| 1657 | } |
| 1658 | else |
| 1659 | { |
| 1660 | uint8x8_t vres = vdup_n_u8(0); |
| 1661 | res = 0; |
| 1662 | |
| 1663 | for(int y = 0; y < pool_size; ++y) |
| 1664 | { |
| 1665 | int x = 0; |
| 1666 | for(; x <= (pool_size - 8); x += 8) |
| 1667 | { |
| 1668 | 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())); |
| 1669 | vres = vmax_u8(vres, data); |
| 1670 | } |
| 1671 | |
| 1672 | // Leftover for loop |
| 1673 | for(; x < pool_size; ++x) |
| 1674 | { |
| 1675 | 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())); |
| 1676 | res = std::max(res, data); |
| 1677 | } |
| 1678 | } |
| 1679 | |
| 1680 | // Reduce max |
| 1681 | vres = vpmax_u8(vres, vres); |
| 1682 | vres = vpmax_u8(vres, vres); |
| 1683 | vres = vpmax_u8(vres, vres); |
| 1684 | |
| 1685 | // Get max value |
| 1686 | res = std::max(res, vget_lane_u8(vres, 0)); |
| 1687 | } |
| 1688 | |
| 1689 | // Store result |
| 1690 | *(reinterpret_cast<uint8_t *>(output.ptr())) = res; |
| 1691 | }, |
| 1692 | input, output); |
| 1693 | } |
| 1694 | |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 1695 | Status NEPoolingLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info) |
| 1696 | { |
| 1697 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); |
| 1698 | |
| 1699 | unsigned int pooled_w = 0; |
| 1700 | unsigned int pooled_h = 0; |
| 1701 | unsigned int num_elems_processed_per_iteration = 0; |
| 1702 | BorderSize border_size(0); |
| 1703 | |
| 1704 | const bool is_global_pooling = pool_info.is_global_pooling(); |
| 1705 | const unsigned int pool_size = is_global_pooling ? input->tensor_shape().x() : pool_info.pool_size(); |
| 1706 | |
| 1707 | // Validate pool info befor calling scaled_dimensions |
| 1708 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_pool_info(input, pool_info, pool_size)); |
| 1709 | |
| 1710 | // Check output dimensions |
| 1711 | std::tie(pooled_w, pooled_h) = scaled_dimensions(input->dimension(0), |
| 1712 | input->dimension(1), |
| 1713 | pool_size, |
| 1714 | pool_size, |
| 1715 | pool_info.pad_stride_info()); |
| 1716 | |
| 1717 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, pool_info, pooled_w, pooled_h, pool_size)); |
| 1718 | 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); |
| 1719 | |
| 1720 | return Status{}; |
| 1721 | } |
| 1722 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1723 | void NEPoolingLayerKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1724 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1725 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1726 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 1727 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 1728 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 1729 | |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1730 | const unsigned int pool_stride_x = _pool_info.pad_stride_info().stride().first; |
| 1731 | 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^] | 1732 | const unsigned int pool_size = _pool_info.pool_size(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1733 | |
| 1734 | // Set step for input in x and y direction for the input |
| 1735 | Window window_input(window); |
| 1736 | unsigned int window_x_inc = 0; |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1737 | switch(_input->info()->data_type()) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1738 | { |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1739 | case DataType::QS8: |
Michalis Spyrou | bbd9fb9 | 2017-06-22 12:57:51 +0100 | [diff] [blame] | 1740 | case DataType::QS16: |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1741 | case DataType::F16: |
| 1742 | { |
| 1743 | window_x_inc = (pool_stride_x == 2) ? _num_elems_processed_per_iteration * 2 : _num_elems_processed_per_iteration; |
| 1744 | break; |
| 1745 | } |
Georgios Pinitas | 5518671 | 2018-01-08 17:37:12 +0000 | [diff] [blame^] | 1746 | case DataType::QASYMM8: |
| 1747 | { |
| 1748 | window_x_inc = pool_stride_x; |
| 1749 | if((pool_size == 2 || pool_size == 3) && pool_stride_x < 3) |
| 1750 | { |
| 1751 | window_x_inc = (pool_stride_x == 2) ? _num_elems_processed_per_iteration * 2 : _num_elems_processed_per_iteration; |
| 1752 | } |
| 1753 | break; |
| 1754 | } |
Pablo Tello | 0c34fe2 | 2017-06-26 17:17:42 +0100 | [diff] [blame] | 1755 | case DataType::F32: |
| 1756 | { |
| 1757 | window_x_inc = pool_stride_x; |
| 1758 | break; |
| 1759 | } |
| 1760 | default: |
| 1761 | { |
| 1762 | ARM_COMPUTE_ERROR("Not supported"); |
| 1763 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1764 | } |
| 1765 | window_input.set(Window::DimX, Window::Dimension(window.x().start() * pool_stride_x, window.x().end() * pool_stride_x, window_x_inc)); |
| 1766 | window_input.set(Window::DimY, Window::Dimension(window.y().start() * pool_stride_y, window.y().end() * pool_stride_y, pool_stride_y)); |
| 1767 | |
| 1768 | // Run function |
| 1769 | (this->*_func)(window_input, window); |
| 1770 | } |