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/NEON/functions/NEFullyConnectedLayer.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/NEON/NEScheduler.h" |
| 29 | |
| 30 | #include <algorithm> |
| 31 | #include <cmath> |
| 32 | |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 33 | namespace arm_compute |
| 34 | { |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame^] | 35 | NEFullyConnectedLayerReshapeWeights::NEFullyConnectedLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager) |
| 36 | : _memory_group(std::move(memory_manager)), _transpose_kernel(), _transpose1xW_kernel(), _transpose_output(), _transpose_weights(false), _is_batched_fc_layer(false) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 37 | { |
| 38 | } |
| 39 | |
| 40 | void NEFullyConnectedLayerReshapeWeights::configure(const ITensor *input, ITensor *output, bool transpose_weights, bool is_batched_fc_layer) |
| 41 | { |
Pablo Tello | dcdc85e | 2017-06-28 10:05:29 +0100 | [diff] [blame] | 42 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 43 | ARM_COMPUTE_ERROR_ON(input->info()->num_dimensions() > 2); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 44 | ARM_COMPUTE_ERROR_ON(output == nullptr); |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 45 | ARM_COMPUTE_ERROR_ON(!transpose_weights && !is_batched_fc_layer); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 46 | |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 47 | const DataType data_type = input->info()->data_type(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 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)); |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 60 | _transpose_output.allocator()->init(TensorInfo(shape_transposed, 1, data_type, fixed_point_position)); |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame^] | 61 | _memory_group.manage(&_transpose_output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 62 | _transpose_kernel.configure(input, &_transpose_output); |
| 63 | |
| 64 | // Configure transpose 1xW kernel |
| 65 | _transpose1xW_kernel.configure(&_transpose_output, output); |
| 66 | |
| 67 | // Allocate temporary tensor used for transposing the weights |
| 68 | _transpose_output.allocator()->allocate(); |
| 69 | } |
| 70 | else |
| 71 | { |
| 72 | _transpose_kernel.configure(input, output); |
| 73 | } |
| 74 | } |
| 75 | else |
| 76 | { |
| 77 | if(_is_batched_fc_layer) |
| 78 | { |
| 79 | // Configure transpose 1xW kernel |
| 80 | _transpose1xW_kernel.configure(input, output); |
| 81 | } |
| 82 | else |
| 83 | { |
| 84 | ARM_COMPUTE_ERROR("Configuration transpose_weights=false & is_batched_fc_layer=false not supported"); |
| 85 | } |
| 86 | } |
| 87 | } |
| 88 | |
| 89 | void NEFullyConnectedLayerReshapeWeights::run() |
| 90 | { |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame^] | 91 | _memory_group.acquire(); |
| 92 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 93 | if(_transpose_weights) |
| 94 | { |
| 95 | NEScheduler::get().schedule(&_transpose_kernel, Window::DimY); |
| 96 | } |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 97 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 98 | if(_is_batched_fc_layer) |
| 99 | { |
| 100 | NEScheduler::get().schedule(&_transpose1xW_kernel, Window::DimY); |
| 101 | } |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame^] | 102 | |
| 103 | _memory_group.release(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 104 | } |
| 105 | |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame^] | 106 | NEFullyConnectedLayer::NEFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager) |
| 107 | : _memory_group(std::move(memory_manager)), _im2col_kernel(), _reshape_weights_kernel(), _interleave4x4_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _interleave4x4_output(), |
| 108 | _reshape_weights_output(), _are_weights_reshaped(false), _is_batched_fc_layer(false), _linearize_input(false), _accumulate_biases(false) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 109 | { |
| 110 | } |
| 111 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 112 | void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose_weights, bool are_weights_reshaped) |
| 113 | { |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 114 | // With the Fully Connected layer we can have 4 different cases: |
| 115 | // 1) Convolution layer -> Fully Connected layer without batches |
| 116 | // 2) Fully Connected layer -> Fully Connected layer without batches |
| 117 | // 3) Convolution layer -> Fully Connected layer with batches |
| 118 | // 4) Fully Connected layer -> Fully Connected layer with batches |
| 119 | |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 120 | // Expected shape before transpose and reshaping |
| 121 | // Input: In x B (In and B can be multi-dimensional) |
| 122 | // Weights: flat(In) x Out |
| 123 | // Biases: Out |
| 124 | // Output: Out x B (B can be multi-dimensional) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 125 | |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 126 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); |
| 127 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); |
| 128 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, weights, output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 129 | |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 130 | const DataType data_type = input->info()->data_type(); |
| 131 | const int fixed_point_position = input->info()->fixed_point_position(); |
| 132 | const int num_batch_dimensions = std::max(0, static_cast<int>(output->info()->tensor_shape().num_dimensions()) - 1); |
| 133 | const int num_input_dimensions = input->info()->tensor_shape().num_dimensions() - num_batch_dimensions; |
| 134 | const size_t linear_input_size = input->info()->tensor_shape().total_size_lower(num_input_dimensions); |
| 135 | |
| 136 | _linearize_input = input->info()->tensor_shape().x() != linear_input_size; |
| 137 | _are_weights_reshaped = are_weights_reshaped; |
| 138 | _accumulate_biases = biases != nullptr; |
| 139 | _is_batched_fc_layer = num_batch_dimensions > 0; |
| 140 | |
| 141 | // Check if number of batches match |
| 142 | ARM_COMPUTE_ERROR_ON(input->info()->tensor_shape().total_size_upper(num_input_dimensions) != output->info()->tensor_shape().total_size_upper(1)); |
| 143 | ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 2); |
| 144 | |
| 145 | const size_t interleave_width = 16 / input->info()->element_size(); |
| 146 | const ITensor *weights_to_use = weights; |
| 147 | |
| 148 | if(!are_weights_reshaped && (transpose_weights || _is_batched_fc_layer)) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 149 | { |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 150 | weights_to_use = &_reshape_weights_output; |
| 151 | |
| 152 | TensorShape reshaped_weights_shape(weights->info()->tensor_shape()); |
| 153 | |
| 154 | // Transpose weights if the user hasn't done it |
| 155 | if(transpose_weights) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 156 | { |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 157 | const size_t shape_x = reshaped_weights_shape.x(); |
| 158 | reshaped_weights_shape.set(0, reshaped_weights_shape.y()); |
| 159 | reshaped_weights_shape.set(1, shape_x); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 160 | } |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 161 | |
| 162 | // If the we run multiple batches we need 1xW transpose, too. |
| 163 | if(_is_batched_fc_layer) |
| 164 | { |
| 165 | const float shape_x = reshaped_weights_shape.x(); |
| 166 | reshaped_weights_shape.set(0, reshaped_weights_shape.y() * interleave_width); |
| 167 | reshaped_weights_shape.set(1, static_cast<unsigned int>(std::ceil(shape_x / interleave_width))); |
| 168 | } |
| 169 | |
| 170 | _reshape_weights_output.allocator()->init(TensorInfo(reshaped_weights_shape, 1, data_type, fixed_point_position)); |
| 171 | |
| 172 | // Reshape the weights |
| 173 | _reshape_weights_kernel.configure(weights, &_reshape_weights_output, transpose_weights, _is_batched_fc_layer); |
| 174 | } |
| 175 | |
| 176 | // Check correct shape of weights |
| 177 | if(_is_batched_fc_layer) |
| 178 | { |
| 179 | // Transpose + Transpose1xW |
| 180 | ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().x() != linear_input_size * interleave_width); |
| 181 | ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().y() != static_cast<unsigned int>(std::ceil(static_cast<float>(output->info()->tensor_shape().x()) / interleave_width))); |
| 182 | } |
| 183 | else |
| 184 | { |
| 185 | // Transpose |
| 186 | ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().x() != output->info()->tensor_shape().x()); |
| 187 | ARM_COMPUTE_ERROR_ON(weights_to_use->info()->tensor_shape().y() != linear_input_size); |
| 188 | } |
| 189 | |
| 190 | const ITensor *multiply_input = input; |
| 191 | |
| 192 | if(_linearize_input) |
| 193 | { |
| 194 | TensorShape shape_im2col(input->info()->tensor_shape()); |
| 195 | shape_im2col.collapse(num_input_dimensions); |
| 196 | _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, data_type, fixed_point_position)); |
| 197 | |
| 198 | // Configure im2col kernel |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame^] | 199 | _memory_group.manage(&_im2col_output); |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 200 | _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false); |
| 201 | |
| 202 | multiply_input = &_im2col_output; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 203 | } |
| 204 | |
| 205 | if(_is_batched_fc_layer) |
| 206 | { |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 207 | TensorShape shape_interleaved(multiply_input->info()->tensor_shape()); |
| 208 | shape_interleaved.set(0, shape_interleaved.x() * 4); |
| 209 | shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); |
| 210 | _interleave4x4_output.allocator()->init(TensorInfo(shape_interleaved, 1, data_type, fixed_point_position)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 211 | |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 212 | // Configure interleave4x4 kernel |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame^] | 213 | _memory_group.manage(&_interleave4x4_output); |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 214 | _interleave4x4_kernel.configure(multiply_input, &_interleave4x4_output); |
| 215 | |
| 216 | multiply_input = &_interleave4x4_output; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 217 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 218 | |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 219 | // Configure matrix multiply kernel |
| 220 | _mm_kernel.configure(multiply_input, weights_to_use, output, 1.0f); |
| 221 | |
| 222 | if(_accumulate_biases) |
| 223 | { |
| 224 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); |
| 225 | ARM_COMPUTE_ERROR_ON(biases->info()->tensor_shape().x() != output->info()->tensor_shape().x()); |
| 226 | |
| 227 | // Configure accumulate biases kernel |
| 228 | _accumulate_biases_kernel.configure(output, biases); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 229 | } |
| 230 | |
| 231 | // Allocate the transpose tensor if the are_weights_reshaped flag is false and once all the configure methods have been called |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 232 | if(!are_weights_reshaped && (transpose_weights || _is_batched_fc_layer)) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 233 | { |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 234 | // Allocate the tensor for the weights reshaped |
| 235 | _reshape_weights_output.allocator()->allocate(); |
| 236 | } |
| 237 | |
| 238 | if(_linearize_input) |
| 239 | { |
| 240 | _im2col_output.allocator()->allocate(); |
| 241 | } |
| 242 | |
| 243 | if(_is_batched_fc_layer) |
| 244 | { |
| 245 | _interleave4x4_output.allocator()->allocate(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 246 | } |
| 247 | } |
| 248 | |
| 249 | void NEFullyConnectedLayer::run() |
| 250 | { |
| 251 | // Reshape of the weights (happens only once) |
| 252 | if(!_are_weights_reshaped) |
| 253 | { |
| 254 | _are_weights_reshaped = true; |
| 255 | _reshape_weights_kernel.run(); |
| 256 | } |
| 257 | |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame^] | 258 | _memory_group.acquire(); |
| 259 | |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 260 | // Linearize input if it comes from a convolutional layer |
| 261 | if(_linearize_input) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 262 | { |
| 263 | NEScheduler::get().schedule(&_im2col_kernel, Window::DimY); |
| 264 | } |
| 265 | |
| 266 | // Interleave input |
| 267 | if(_is_batched_fc_layer) |
| 268 | { |
| 269 | NEScheduler::get().schedule(&_interleave4x4_kernel, Window::DimY); |
| 270 | } |
| 271 | |
| 272 | // Run matrix multiply |
| 273 | NEScheduler::get().schedule(&_mm_kernel, _is_batched_fc_layer ? Window::DimY : Window::DimX); |
| 274 | |
| 275 | // Accumulate biases if provided |
| 276 | if(_accumulate_biases) |
| 277 | { |
| 278 | NEScheduler::get().schedule(&_accumulate_biases_kernel, Window::DimY); |
| 279 | } |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame^] | 280 | |
| 281 | _memory_group.release(); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 282 | } |
Moritz Pflanzer | 484e7b3 | 2017-08-09 11:43:18 +0100 | [diff] [blame] | 283 | } // namespace arm_compute |