giuros01 | 154bc1c | 2019-03-26 17:44:40 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2019 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/NEFFTConvolutionLayer.h" |
| 25 | |
| 26 | #include "arm_compute/core/ITensor.h" |
| 27 | #include "arm_compute/core/Utils.h" |
| 28 | #include "arm_compute/core/Validate.h" |
| 29 | #include "arm_compute/core/utils/helpers/fft.h" |
| 30 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 31 | |
| 32 | namespace arm_compute |
| 33 | { |
| 34 | namespace |
| 35 | { |
| 36 | int pad_decomposable(int N) |
| 37 | { |
| 38 | const auto supported_radix = NEFFTRadixStageKernel::supported_radix(); |
| 39 | |
| 40 | int pad = 0; |
| 41 | bool is_decomposed = false; |
| 42 | while(!is_decomposed) |
| 43 | { |
| 44 | const auto decomposed_vector = arm_compute::helpers::fft::decompose_stages(N++, supported_radix); |
| 45 | is_decomposed = !decomposed_vector.empty(); |
| 46 | if(!is_decomposed) |
| 47 | { |
| 48 | ++pad; |
| 49 | } |
| 50 | } |
| 51 | return pad; |
| 52 | } |
| 53 | } // namespace |
| 54 | |
| 55 | NEFFTConvolutionLayer::NEFFTConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) |
| 56 | : _memory_group(memory_manager), |
| 57 | _flip_weights_func(), |
| 58 | _permute_input_func(), |
| 59 | _permute_output_func(), |
| 60 | _permute_weights_func(), |
| 61 | _permute_bias_func(), |
| 62 | _pad_input_func(), |
| 63 | _pad_weights_func(), |
| 64 | _transform_input_func(memory_manager), |
| 65 | _transform_weights_func(), |
| 66 | _itransform_output_func(memory_manager), |
| 67 | _prod_func(), |
| 68 | _reduce_func(), |
| 69 | _extract_output_func(), |
| 70 | _bias_add_func(), |
| 71 | _activation_layer_func(), |
| 72 | _permuted_input(), |
| 73 | _permuted_weights(), |
| 74 | _permuted_bias(), |
| 75 | _permuted_output(), |
| 76 | _padded_input(), |
| 77 | _padded_weights(), |
| 78 | _flip_axis(), |
| 79 | _flipped_weights(), |
| 80 | _transformed_input(), |
| 81 | _transformed_weights(), |
| 82 | _input_weights_product(), |
| 83 | _output_product(), |
| 84 | _output_reduced(), |
| 85 | _itransformed_output(), |
| 86 | _reshaped_output(), |
| 87 | _bias_output(), |
| 88 | _original_weights(nullptr), |
| 89 | _original_bias(nullptr), |
| 90 | _is_activationlayer_enabled(false), |
| 91 | _needs_permute(false), |
| 92 | _has_bias(false), |
| 93 | _is_prepared(false) |
| 94 | { |
| 95 | } |
| 96 | |
| 97 | void NEFFTConvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, |
| 98 | const ActivationLayerInfo &act_info) |
| 99 | { |
| 100 | _original_weights = weights; |
| 101 | _original_bias = biases; |
| 102 | |
| 103 | // Flat if bias addition is required |
| 104 | _has_bias = biases != nullptr; |
| 105 | |
| 106 | // Get indices for the width and height |
| 107 | const size_t idx_width = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH); |
| 108 | const size_t idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT); |
| 109 | |
| 110 | // Input shape, kernel size and output tile |
| 111 | const Size2D input_dims = Size2D(input->info()->tensor_shape()[idx_width], input->info()->tensor_shape()[idx_height]); |
| 112 | const Size2D kernel_size = Size2D(weights->info()->tensor_shape()[idx_width], weights->info()->tensor_shape()[idx_height]); |
| 113 | const Size2D pad_valid = Size2D(pad_decomposable(input_dims.x() + kernel_size.x() - 1), |
| 114 | pad_decomposable(input_dims.y() + kernel_size.y() - 1)); |
| 115 | // Tensors to use |
| 116 | ITensor *input_to_use = input; |
| 117 | const ITensor *weights_to_use = weights; |
| 118 | ITensor *output_to_use = _has_bias ? &_bias_output : output; |
| 119 | |
| 120 | // Permute bias |
Georgios Pinitas | 68c6a79 | 2019-05-15 13:24:00 +0100 | [diff] [blame] | 121 | if(biases != nullptr) |
| 122 | { |
| 123 | _permute_bias_func.configure(biases, &_permuted_bias, PermutationVector(1U, 2U, 0U)); |
| 124 | _permuted_bias.info()->set_data_layout(DataLayout::NCHW); |
| 125 | } |
giuros01 | 154bc1c | 2019-03-26 17:44:40 +0000 | [diff] [blame] | 126 | |
| 127 | // Permute input if needed |
| 128 | _needs_permute = input->info()->data_layout() == DataLayout::NHWC; |
| 129 | if(_needs_permute) |
| 130 | { |
| 131 | _memory_group.manage(&_permuted_input); |
| 132 | // Configure the function to transform the input tensor from NHWC -> NCHW |
| 133 | _permute_input_func.configure(input, &_permuted_input, PermutationVector(1U, 2U, 0U)); |
| 134 | _permuted_input.info()->set_data_layout(DataLayout::NCHW); |
| 135 | |
| 136 | // Configure the function to transform the weights tensor from HWI -> IHW |
| 137 | _permute_weights_func.configure(weights, &_permuted_weights, PermutationVector(1U, 2U, 0U)); |
| 138 | _permuted_weights.info()->set_data_layout(DataLayout::NCHW); |
| 139 | |
| 140 | input_to_use = &_permuted_input; |
| 141 | weights_to_use = &_permuted_weights; |
| 142 | } |
| 143 | |
| 144 | // Flip weights |
| 145 | _flipped_weights.allocator()->init(weights_to_use->info()->clone()->set_is_resizable(true).reset_padding()); |
| 146 | _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32)); |
| 147 | _flip_weights_func.configure(weights_to_use, &_flipped_weights, &_flip_axis); |
| 148 | |
| 149 | // Pad weights |
| 150 | const PaddingList padding_w = { { 0, input_dims.x() + pad_valid.x() - 1 }, { 0, input_dims.y() + pad_valid.y() - 1 } }; |
| 151 | _pad_weights_func.configure(&_flipped_weights, &_padded_weights, padding_w); |
| 152 | |
| 153 | // Transform weights |
| 154 | _transform_weights_func = support::cpp14::make_unique<NEFFT2D>(); |
| 155 | _transform_weights_func->configure(&_padded_weights, &_transformed_weights, FFT2DInfo()); |
| 156 | |
| 157 | // Pad input |
| 158 | const PaddingList padding_in = { { 0, kernel_size.x() + pad_valid.x() - 1 }, { 0, kernel_size.y() + pad_valid.y() - 1 } }; |
| 159 | _memory_group.manage(&_padded_input); |
| 160 | _pad_input_func.configure(input_to_use, &_padded_input, padding_in); |
| 161 | if(_needs_permute) |
| 162 | { |
| 163 | _permuted_input.allocator()->allocate(); |
| 164 | } |
| 165 | |
| 166 | // Transform input |
| 167 | _memory_group.manage(&_transformed_input); |
| 168 | _transform_input_func.configure(&_padded_input, &_transformed_input, FFT2DInfo()); |
| 169 | _padded_input.allocator()->allocate(); |
| 170 | |
| 171 | // Perform product |
| 172 | _memory_group.manage(&_output_product); |
| 173 | _prod_func.configure(&_transformed_input, &_transformed_weights, &_output_product); |
| 174 | _transformed_input.allocator()->allocate(); |
| 175 | |
| 176 | // Perform reduction |
| 177 | _memory_group.manage(&_output_reduced); |
| 178 | _reduce_func.configure(&_output_product, &_output_reduced, 2, ReductionOperation::SUM); |
| 179 | _output_product.allocator()->allocate(); |
| 180 | |
| 181 | // Transform output |
| 182 | _memory_group.manage(&_itransformed_output); |
| 183 | FFT2DInfo itranform_info; |
| 184 | itranform_info.direction = FFTDirection::Inverse; |
| 185 | _itransformed_output.allocator()->init(_output_reduced.info()->clone()->set_is_resizable(true).set_num_channels(1).reset_padding()); |
| 186 | _itransform_output_func.configure(&_output_reduced, &_itransformed_output, itranform_info); |
| 187 | _output_reduced.allocator()->allocate(); |
| 188 | |
| 189 | // Reshape output |
| 190 | TensorShape reshaped_shape = _itransformed_output.info()->tensor_shape(); |
| 191 | reshaped_shape.remove_dimension(2); |
| 192 | _reshaped_output.allocator()->init(_itransformed_output.info()->clone()->set_tensor_shape(reshaped_shape)); |
| 193 | |
| 194 | // Extract correct region |
| 195 | const int start_left = kernel_size.x() - conv_info.pad_left() - 1; |
| 196 | const int start_top = kernel_size.y() - conv_info.pad_top() - 1; |
| 197 | const int end_right = _reshaped_output.info()->tensor_shape().x() - (kernel_size.x() - conv_info.pad_right() - 1) - pad_valid.x(); |
| 198 | const int end_botton = _reshaped_output.info()->tensor_shape().y() - (kernel_size.y() - conv_info.pad_bottom() - 1) - pad_valid.y(); |
| 199 | if(_has_bias) |
| 200 | { |
| 201 | _memory_group.manage(&_bias_output); |
| 202 | } |
| 203 | else if(_needs_permute) |
| 204 | { |
| 205 | output_to_use = &_permuted_output; |
| 206 | _memory_group.manage(&_permuted_output); |
| 207 | } |
| 208 | _extract_output_func.configure(&_reshaped_output, output_to_use, Coordinates(start_left, start_top), Coordinates(end_right, end_botton)); |
| 209 | _reshaped_output.allocator()->allocate(); |
| 210 | _itransformed_output.allocator()->allocate(); |
| 211 | |
| 212 | // Add bias |
| 213 | if(biases != nullptr) |
| 214 | { |
| 215 | output_to_use = output; |
| 216 | if(_needs_permute) |
| 217 | { |
| 218 | output_to_use = &_permuted_output; |
| 219 | _memory_group.manage(&_permuted_output); |
| 220 | } |
| 221 | auto_init_if_empty(*output_to_use->info(), *_bias_output.info()); |
| 222 | _bias_add_func.configure(&_bias_output, &_permuted_bias, output_to_use, ConvertPolicy::WRAP); |
| 223 | _bias_output.allocator()->allocate(); |
| 224 | } |
| 225 | |
| 226 | // Permute output |
| 227 | if(_needs_permute) |
| 228 | { |
| 229 | // Configure the function to transform the convoluted output to ACL's native ordering format NCHW |
| 230 | _permuted_output.info()->set_data_layout(DataLayout::NCHW); |
| 231 | _permute_output_func.configure(&_permuted_output, output, PermutationVector(2U, 0U, 1U)); |
| 232 | |
| 233 | // Allocate tensors |
| 234 | _permuted_output.allocator()->allocate(); |
| 235 | } |
| 236 | |
| 237 | // Configure Activation Layer |
| 238 | _is_activationlayer_enabled = act_info.enabled(); |
| 239 | if(_is_activationlayer_enabled) |
| 240 | { |
| 241 | _activation_layer_func.configure(output, nullptr, act_info); |
| 242 | } |
| 243 | |
| 244 | // Setup flip axis data |
| 245 | _flip_axis.allocator()->allocate(); |
| 246 | |
| 247 | auto axis_data = reinterpret_cast<uint32_t *>(_flip_axis.buffer()); |
| 248 | axis_data[0] = 0; |
| 249 | axis_data[1] = 1; |
| 250 | } |
| 251 | |
| 252 | Status NEFFTConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, |
| 253 | const ActivationLayerInfo &act_info) |
| 254 | { |
| 255 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); |
| 256 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); |
| 257 | |
| 258 | // Get indices for the width and height |
| 259 | const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); |
| 260 | const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); |
| 261 | |
| 262 | // Input shape, kernel size and output tile |
| 263 | const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]); |
| 264 | |
| 265 | // Strides |
| 266 | const auto strides = conv_info.stride(); |
| 267 | ARM_COMPUTE_RETURN_ERROR_ON(strides.first != strides.second && strides.first != 1); |
| 268 | ARM_COMPUTE_RETURN_ERROR_ON(kernel_size.x() != kernel_size.y()); |
| 269 | ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_left() != (kernel_size.x() / 2) || conv_info.pad_right() != (kernel_size.x() / 2)); |
| 270 | ARM_COMPUTE_RETURN_ERROR_ON(conv_info.pad_top() != (kernel_size.y() / 2) || conv_info.pad_bottom() != (kernel_size.y() / 2)); |
| 271 | |
| 272 | // Validate biases |
| 273 | if(biases != nullptr) |
| 274 | { |
| 275 | const size_t idx_channels = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL); |
| 276 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); |
| 277 | ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_channels] != biases->tensor_shape().x()); |
| 278 | } |
| 279 | |
| 280 | // Checks performed when output is configured |
| 281 | if((output != nullptr) && (output->total_size() != 0)) |
| 282 | { |
| 283 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 284 | ARM_COMPUTE_RETURN_ERROR_ON((input->tensor_shape()[idx_height] != output->tensor_shape()[idx_height]) || (input->tensor_shape()[idx_width] != output->tensor_shape()[idx_width])); |
| 285 | |
| 286 | // Validate Activation Layer |
| 287 | if(act_info.enabled()) |
| 288 | { |
| 289 | ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(output, nullptr, act_info)); |
| 290 | } |
| 291 | } |
| 292 | |
| 293 | return Status{}; |
| 294 | } |
| 295 | |
| 296 | void NEFFTConvolutionLayer::run() |
| 297 | { |
| 298 | prepare(); |
| 299 | |
| 300 | MemoryGroupResourceScope scope_mg(_memory_group); |
| 301 | |
| 302 | // Transform input |
| 303 | if(_needs_permute) |
| 304 | { |
| 305 | _permute_input_func.run(); |
| 306 | } |
| 307 | _pad_input_func.run(); |
| 308 | _transform_input_func.run(); |
| 309 | |
| 310 | // Perform operations to frequency domain |
| 311 | _prod_func.run(); |
| 312 | |
| 313 | _reduce_func.run(); |
| 314 | |
| 315 | // Transform output |
| 316 | _itransform_output_func.run(); |
| 317 | _reshaped_output.allocator()->import_memory(_itransformed_output.buffer()); |
| 318 | _extract_output_func.run(); |
| 319 | |
| 320 | // Add bias |
| 321 | if(_has_bias) |
| 322 | { |
| 323 | _bias_add_func.run(); |
| 324 | } |
| 325 | if(_needs_permute) |
| 326 | { |
| 327 | _permute_output_func.run(); |
| 328 | } |
| 329 | |
| 330 | // Run activation layer |
| 331 | if(_is_activationlayer_enabled) |
| 332 | { |
| 333 | _activation_layer_func.run(); |
| 334 | } |
| 335 | } |
| 336 | |
| 337 | void NEFFTConvolutionLayer::prepare() |
| 338 | { |
| 339 | if(!_is_prepared) |
| 340 | { |
| 341 | // Permute bias to NCHW |
| 342 | if(_original_bias != nullptr) |
| 343 | { |
| 344 | _permuted_bias.allocator()->allocate(); |
| 345 | _permute_bias_func.run(); |
| 346 | _original_bias->mark_as_unused(); |
| 347 | } |
| 348 | |
| 349 | const ITensor *cur_weights = _original_weights; |
| 350 | |
| 351 | // Permute weights |
| 352 | if(_needs_permute) |
| 353 | { |
| 354 | ARM_COMPUTE_ERROR_ON(!cur_weights->is_used()); |
| 355 | |
| 356 | _permuted_weights.allocator()->allocate(); |
| 357 | _permute_weights_func.run(); |
| 358 | cur_weights->mark_as_unused(); |
| 359 | cur_weights = &_permuted_weights; |
| 360 | } |
| 361 | |
| 362 | // Flip weights |
| 363 | _flipped_weights.allocator()->allocate(); |
| 364 | _flip_weights_func.run(); |
| 365 | cur_weights->mark_as_unused(); |
| 366 | |
| 367 | // Pad weights |
| 368 | _padded_weights.allocator()->allocate(); |
| 369 | _pad_weights_func.run(); |
| 370 | _flipped_weights.mark_as_unused(); |
| 371 | _flipped_weights.allocator()->free(); |
| 372 | |
| 373 | // Transform weights to frequency domain |
| 374 | _transformed_weights.allocator()->allocate(); |
| 375 | _transform_weights_func->run(); |
| 376 | _transform_weights_func.reset(); |
| 377 | |
| 378 | _padded_weights.mark_as_unused(); |
| 379 | _padded_weights.allocator()->free(); |
| 380 | |
| 381 | _is_prepared = true; |
| 382 | } |
| 383 | } |
| 384 | } // namespace arm_compute |