Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h" |
| 25 | |
Gian Marco Iodice | 13edbff | 2017-06-26 17:20:16 +0100 | [diff] [blame] | 26 | #include "arm_compute/core/Size2D.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/Validate.h" |
| 28 | #include "arm_compute/runtime/CL/CLScheduler.h" |
| 29 | |
| 30 | #include <algorithm> |
| 31 | #include <cmath> |
| 32 | |
| 33 | using namespace arm_compute; |
| 34 | |
| 35 | CLFullyConnectedLayerReshapeWeights::CLFullyConnectedLayerReshapeWeights() |
| 36 | : _transpose_kernel(), _transpose1xW_kernel(), _transpose_output(), _transpose_weights(false), _is_batched_fc_layer(false) |
| 37 | { |
| 38 | } |
| 39 | |
| 40 | void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLTensor *output, bool transpose_weights, bool is_batched_fc_layer) |
| 41 | { |
Gian Marco Iodice | 368da83 | 2017-07-03 12:33:49 +0100 | [diff] [blame^] | 42 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 43 | ARM_COMPUTE_ERROR_ON(output == nullptr); |
| 44 | ARM_COMPUTE_ERROR_ON(input->info()->num_dimensions() != 2); |
| 45 | ARM_COMPUTE_ERROR_ON((transpose_weights == false) && (is_batched_fc_layer == false)); |
| 46 | |
| 47 | const DataType dt = input->info()->data_type(); |
| 48 | const int fixed_point_position = input->info()->fixed_point_position(); |
| 49 | |
| 50 | _transpose_weights = transpose_weights; |
| 51 | _is_batched_fc_layer = is_batched_fc_layer; |
| 52 | |
| 53 | // Check if we need to transpose the weights |
| 54 | if(_transpose_weights) |
| 55 | { |
| 56 | if(_is_batched_fc_layer) |
| 57 | { |
| 58 | // Initialize the output tensor for transpose |
| 59 | TensorShape shape_transposed(input->info()->dimension(1), input->info()->dimension(0)); |
| 60 | _transpose_output.allocator()->init(TensorInfo(shape_transposed, 1, dt, fixed_point_position)); |
| 61 | _transpose_kernel.configure(input, &_transpose_output); |
| 62 | |
| 63 | // Configure transpose 1xW kernel |
| 64 | _transpose1xW_kernel.configure(&_transpose_output, output); |
| 65 | |
| 66 | // Allocate temporary tensor used for transposing the weights |
| 67 | _transpose_output.allocator()->allocate(); |
| 68 | } |
| 69 | else |
| 70 | { |
| 71 | _transpose_kernel.configure(input, output); |
| 72 | } |
| 73 | } |
| 74 | else |
| 75 | { |
| 76 | if(_is_batched_fc_layer) |
| 77 | { |
| 78 | // Configure transpose 1xW kernel |
| 79 | _transpose1xW_kernel.configure(input, output); |
| 80 | } |
| 81 | else |
| 82 | { |
| 83 | ARM_COMPUTE_ERROR("Configuration transpose_weights=false & is_batched_fc_layer=false not supported"); |
| 84 | } |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | void CLFullyConnectedLayerReshapeWeights::run() |
| 89 | { |
| 90 | if(_transpose_weights) |
| 91 | { |
| 92 | CLScheduler::get().enqueue(_transpose_kernel, _is_batched_fc_layer); |
| 93 | } |
| 94 | if(_is_batched_fc_layer) |
| 95 | { |
| 96 | CLScheduler::get().enqueue(_transpose1xW_kernel); |
| 97 | } |
| 98 | } |
| 99 | |
| 100 | CLFullyConnectedLayer::CLFullyConnectedLayer() |
| 101 | : _im2col_kernel(), _reshape_weights_kernel(), _interleave4x4_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _interleave4x4_output(), _reshape_weights_output(), |
| 102 | _are_weights_reshaped(true), _is_fc_after_conv(true), _is_batched_fc_layer(false), _accumulate_biases(false) |
| 103 | { |
| 104 | } |
| 105 | |
| 106 | void CLFullyConnectedLayer::configure_conv_fc_wb(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output) |
| 107 | { |
| 108 | ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2) * (16 / weights->info()->element_size()))); |
| 109 | |
| 110 | const DataType dt = input->info()->data_type(); |
| 111 | const int fixed_point_position = input->info()->fixed_point_position(); |
| 112 | |
| 113 | // If the fully connected layer is called after a convolution layer, the input tensor must be linearized |
| 114 | |
| 115 | // Initialize output tensor for im2col |
| 116 | TensorShape shape_im2col; |
| 117 | shape_im2col.set(0, input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2)); |
| 118 | shape_im2col.set(1, input->info()->dimension(3)); |
| 119 | shape_im2col.set(2, input->info()->dimension(4)); |
| 120 | shape_im2col.set(3, input->info()->dimension(5)); |
| 121 | _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, dt, fixed_point_position)); |
| 122 | |
| 123 | // Initialize output tensor for interleave 4x4 |
| 124 | TensorShape shape_interleaved = _im2col_output.info()->tensor_shape(); |
| 125 | shape_interleaved.set(0, shape_interleaved.x() * 4); |
| 126 | shape_interleaved.set(1, std::ceil(static_cast<float>(shape_interleaved.y()) / 4)); |
| 127 | _interleave4x4_output.allocator()->init(TensorInfo(shape_interleaved, 1, dt, fixed_point_position)); |
| 128 | |
| 129 | // Configure im2col kernel |
Gian Marco Iodice | 13edbff | 2017-06-26 17:20:16 +0100 | [diff] [blame] | 130 | _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 131 | |
| 132 | // Configure interleave4x4 kernel |
| 133 | _interleave4x4_kernel.configure(&_im2col_output, &_interleave4x4_output); |
| 134 | |
| 135 | // Configure matrix multiply kernel |
| 136 | _mm_kernel.configure(&_interleave4x4_output, weights, output, 1.0f); |
| 137 | |
| 138 | // Allocate the tensors once all the configure methods have been called |
| 139 | _im2col_output.allocator()->allocate(); |
| 140 | _interleave4x4_output.allocator()->allocate(); |
| 141 | } |
| 142 | |
| 143 | void CLFullyConnectedLayer::configure_fc_fc_wb(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output) |
| 144 | { |
| 145 | const DataType dt = input->info()->data_type(); |
| 146 | const int fixed_point_position = input->info()->fixed_point_position(); |
| 147 | |
| 148 | // Initialize output tensor for interleave 4x4 |
| 149 | TensorShape shape_interleaved = input->info()->tensor_shape(); |
| 150 | shape_interleaved.set(0, shape_interleaved.x() * 4); |
| 151 | shape_interleaved.set(1, std::ceil(static_cast<float>(shape_interleaved.y()) / 4)); |
| 152 | _interleave4x4_output.allocator()->init(TensorInfo(shape_interleaved, 1, dt, fixed_point_position)); |
| 153 | |
| 154 | // Configure interleave4x4 kernel |
| 155 | _interleave4x4_kernel.configure(input, &_interleave4x4_output); |
| 156 | |
| 157 | // Configure matrix multiply kernel |
| 158 | _mm_kernel.configure(&_interleave4x4_output, weights, output, 1.0f); |
| 159 | |
| 160 | // Allocate the tensors once all the configure methods have been called |
| 161 | _interleave4x4_output.allocator()->allocate(); |
| 162 | } |
| 163 | |
| 164 | void CLFullyConnectedLayer::configure_conv_fc_nb(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output) |
| 165 | { |
| 166 | ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2)))); |
| 167 | |
| 168 | const DataType dt = input->info()->data_type(); |
| 169 | const int fixed_point_position = input->info()->fixed_point_position(); |
| 170 | |
| 171 | // If the fully connected layer is called after a convolution layer, the input tensor must be linearized |
| 172 | |
| 173 | // Initialize output tensor for im2col |
| 174 | TensorShape shape_im2col; |
| 175 | shape_im2col.set(0, input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2)); |
| 176 | shape_im2col.set(1, 1); |
| 177 | _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, dt, fixed_point_position)); |
| 178 | |
| 179 | // Configure im2col kernel |
Gian Marco Iodice | 13edbff | 2017-06-26 17:20:16 +0100 | [diff] [blame] | 180 | _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 181 | |
| 182 | // Configure matrix multiply kernel |
| 183 | _mm_kernel.configure(&_im2col_output, weights, output, 1.0f); |
| 184 | |
| 185 | // Allocate the output tensor for im2col once all the configure methods have been called |
| 186 | _im2col_output.allocator()->allocate(); |
| 187 | } |
| 188 | |
| 189 | void CLFullyConnectedLayer::configure_fc_fc_nb(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output) |
| 190 | { |
| 191 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1)); |
| 192 | |
| 193 | // Configure matrix multiply kernel |
| 194 | _mm_kernel.configure(input, weights, output, 1.0f); |
| 195 | } |
| 196 | |
| 197 | void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights, bool are_weights_reshaped) |
| 198 | { |
Gian Marco Iodice | 368da83 | 2017-07-03 12:33:49 +0100 | [diff] [blame^] | 199 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 200 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); |
| 201 | ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() != 2); |
| 202 | |
| 203 | const DataType dt = input->info()->data_type(); |
| 204 | const int fixed_point_position = input->info()->fixed_point_position(); |
| 205 | |
| 206 | _are_weights_reshaped = are_weights_reshaped; |
| 207 | _is_fc_after_conv = true; |
| 208 | _is_batched_fc_layer = false; |
| 209 | _accumulate_biases = false; |
| 210 | |
| 211 | if(biases != nullptr) |
| 212 | { |
| 213 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); |
| 214 | |
| 215 | _accumulate_biases = true; |
| 216 | |
| 217 | // Configure accumulate biases kernel |
| 218 | _accumulate_biases_kernel.configure(output, biases); |
| 219 | } |
| 220 | |
| 221 | // With the Fully Connected layer we can have 4 different cases: |
| 222 | // 1) Convolution layer -> Fully Connected layer without batches |
| 223 | // 2) Fully Connected layer -> Fully Connected layer without batches |
| 224 | // 3) Convolution layer -> Fully Connected layer with batches |
| 225 | // 4) Fully Connected layer -> Fully Connected layer with batches |
| 226 | |
| 227 | // Check if we have a fully connected layer with batches |
| 228 | _is_batched_fc_layer = (output->info()->dimension(1) > 1); |
| 229 | |
| 230 | const ICLTensor *weights_to_use = weights; |
| 231 | |
| 232 | if(!are_weights_reshaped) |
| 233 | { |
| 234 | if((transpose_weights || _is_batched_fc_layer)) |
| 235 | { |
| 236 | weights_to_use = &_reshape_weights_output; |
| 237 | |
| 238 | if(transpose_weights) |
| 239 | { |
| 240 | if(_is_batched_fc_layer) |
| 241 | { |
| 242 | const float transpose_width = 16.0f / input->info()->element_size(); |
| 243 | TensorShape shape_wt(weights->info()->dimension(0) * static_cast<unsigned int>(transpose_width), static_cast<unsigned int>(std::ceil(weights->info()->dimension(1) / transpose_width))); |
| 244 | TensorInfo info_wt(shape_wt, 1, dt, fixed_point_position); |
| 245 | _reshape_weights_output.allocator()->init(info_wt); |
| 246 | } |
| 247 | else |
| 248 | { |
| 249 | TensorShape shape_wt(weights->info()->dimension(1), weights->info()->dimension(0)); |
| 250 | TensorInfo info_wt(shape_wt, 1, dt, fixed_point_position); |
| 251 | _reshape_weights_output.allocator()->init(info_wt); |
| 252 | } |
| 253 | } |
| 254 | else |
| 255 | { |
| 256 | ARM_COMPUTE_ERROR_ON(!_is_batched_fc_layer); |
| 257 | |
| 258 | const float transpose_width = 16.0f / input->info()->element_size(); |
| 259 | TensorShape shape_wt(weights->info()->dimension(1) * static_cast<unsigned int>(transpose_width), static_cast<unsigned int>(std::ceil(weights->info()->dimension(0) / transpose_width))); |
| 260 | TensorInfo info_wt(shape_wt, 1, dt, fixed_point_position); |
| 261 | _reshape_weights_output.allocator()->init(info_wt); |
| 262 | } |
| 263 | |
| 264 | // Reshape the weights |
| 265 | _reshape_weights_kernel.configure(weights, &_reshape_weights_output, transpose_weights, _is_batched_fc_layer); |
| 266 | } |
| 267 | } |
| 268 | |
| 269 | if(_is_batched_fc_layer) |
| 270 | { |
| 271 | _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->info()->tensor_shape().cbegin() + 3, |
| 272 | input->info()->tensor_shape().cend(), |
| 273 | output->info()->tensor_shape().cbegin() + 1)); |
| 274 | |
| 275 | if(_is_fc_after_conv) |
| 276 | { |
| 277 | // Fully Connected layer after a Convolution Layer with batches |
| 278 | configure_conv_fc_wb(input, weights_to_use, output); |
| 279 | } |
| 280 | else |
| 281 | { |
| 282 | // Fully Connected layer after a Fully Connected Layer with batches |
| 283 | configure_fc_fc_wb(input, weights_to_use, output); |
| 284 | } |
| 285 | } |
| 286 | else |
| 287 | { |
| 288 | // In case of not batched fully connected layer, the weights will not be reshaped using transposed1xW |
| 289 | _is_fc_after_conv = ((weights_to_use->info()->dimension(1)) == (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))); |
| 290 | |
| 291 | if(_is_fc_after_conv) |
| 292 | { |
| 293 | // Fully Connected layer after a Convolution Layer without batches |
| 294 | configure_conv_fc_nb(input, weights_to_use, output); |
| 295 | } |
| 296 | else |
| 297 | { |
| 298 | // Fully Connected layer after a Fully Connected Layer without batches |
| 299 | configure_fc_fc_nb(input, weights_to_use, output); |
| 300 | } |
| 301 | } |
| 302 | |
| 303 | // Allocate the transpose tensor if the are_weights_reshaped flag is false and once all the configure methods have been called |
| 304 | if(!are_weights_reshaped) |
| 305 | { |
| 306 | if(transpose_weights || _is_batched_fc_layer) |
| 307 | { |
| 308 | // Allocate the tensor for the weights reshaped |
| 309 | _reshape_weights_output.allocator()->allocate(); |
| 310 | } |
| 311 | } |
| 312 | } |
| 313 | |
| 314 | void CLFullyConnectedLayer::run() |
| 315 | { |
| 316 | // Reshape of the weights (happens only once) |
| 317 | if(!_are_weights_reshaped) |
| 318 | { |
| 319 | _are_weights_reshaped = true; |
| 320 | _reshape_weights_kernel.run(); |
| 321 | } |
| 322 | |
| 323 | // Linearize input if it comes from a convolutional layer |
| 324 | if(_is_fc_after_conv) |
| 325 | { |
| 326 | CLScheduler::get().enqueue(_im2col_kernel, false); |
| 327 | } |
| 328 | |
| 329 | // Interleave input |
| 330 | if(_is_batched_fc_layer) |
| 331 | { |
| 332 | CLScheduler::get().enqueue(_interleave4x4_kernel, false); |
| 333 | } |
| 334 | |
| 335 | // Run matrix multiply |
| 336 | CLScheduler::get().enqueue(_mm_kernel, !_accumulate_biases); |
| 337 | |
| 338 | // Accumulate biases if provided |
| 339 | if(_accumulate_biases) |
| 340 | { |
| 341 | CLScheduler::get().enqueue(_accumulate_biases_kernel); |
| 342 | } |
| 343 | } |