Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 1 | /* |
Georgios Pinitas | 2481d46 | 2019-02-19 18:47:46 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2019 ARM Limited. |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +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 | */ |
Giorgio Arena | 04a8f8c | 2017-11-23 11:45:24 +0000 | [diff] [blame] | 24 | #include "arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h" |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 25 | |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 26 | #include "arm_compute/core/utils/misc/InfoHelpers.h" |
Georgios Pinitas | d05dce4 | 2018-01-22 16:29:17 +0000 | [diff] [blame] | 27 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Georgios Pinitas | f72f936 | 2018-01-12 16:29:45 +0000 | [diff] [blame] | 28 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 29 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 30 | |
Georgios Pinitas | d05dce4 | 2018-01-22 16:29:17 +0000 | [diff] [blame] | 31 | using namespace arm_compute::misc; |
Georgios Pinitas | 4074c99 | 2018-01-30 18:13:46 +0000 | [diff] [blame] | 32 | using namespace arm_compute::misc::shape_calculator; |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 33 | |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 34 | namespace arm_compute |
| 35 | { |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 36 | namespace |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 37 | { |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 38 | Status validate_arguments_optimized(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, |
| 39 | unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation) |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 40 | { |
| 41 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); |
Michele Di Giorgio | ff27192 | 2019-07-17 15:59:32 +0100 | [diff] [blame] | 42 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); |
| 43 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 44 | ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN); |
| 45 | ARM_COMPUTE_RETURN_ERROR_ON(dilation.x() < 1 || dilation.y() < 1); |
| 46 | const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); |
| 47 | const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); |
| 48 | ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) + (weights->dimension(idx_w) - 1) * (dilation.x() - 1) > input->dimension(idx_w) + conv_info.pad_left() + conv_info.pad_right()); |
| 49 | ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_h) + (weights->dimension(idx_h) - 1) * (dilation.y() - 1) > input->dimension(idx_h) + conv_info.pad_top() + conv_info.pad_bottom()); |
| 50 | |
| 51 | if(biases != nullptr) |
| 52 | { |
| 53 | const unsigned int channel_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL); |
| 54 | ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); |
| 55 | ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(channel_idx)); |
| 56 | } |
| 57 | |
Michele Di Giorgio | 601ba3f | 2019-08-22 16:20:04 +0100 | [diff] [blame] | 58 | const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type()); |
| 59 | |
| 60 | if(is_quantized) |
| 61 | { |
| 62 | const UniformQuantizationInfo iq_info = input->quantization_info().uniform(); |
| 63 | const UniformQuantizationInfo wq_info = weights->quantization_info().uniform(); |
| 64 | const UniformQuantizationInfo oq_info = output->quantization_info().uniform(); |
| 65 | |
| 66 | float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale; |
| 67 | ARM_COMPUTE_UNUSED(multiplier); |
| 68 | ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f); |
| 69 | } |
| 70 | |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 71 | if(!NEDepthwiseConvolutionAssemblyDispatch::is_optimized_supported(input, weights, conv_info, depth_multiplier, dilation)) |
| 72 | { |
Michele Di Giorgio | 601ba3f | 2019-08-22 16:20:04 +0100 | [diff] [blame] | 73 | TensorInfo accumulator = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32)); |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 74 | ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionLayer3x3Kernel::validate(input, weights, is_quantized ? &accumulator : output, conv_info, depth_multiplier, dilation)); |
| 75 | |
| 76 | if(is_quantized) |
| 77 | { |
| 78 | ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerOutputStageKernel::validate(&accumulator, biases, output)); |
| 79 | } |
| 80 | } |
| 81 | else |
| 82 | { |
| 83 | ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionAssemblyDispatch::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation)); |
| 84 | } |
| 85 | |
| 86 | //Validate Activation Layer |
| 87 | if(act_info.enabled()) |
| 88 | { |
| 89 | ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(output, nullptr, act_info)); |
| 90 | } |
| 91 | |
| 92 | return Status{}; |
| 93 | } |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 94 | } // namespace |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 95 | |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 96 | NEDepthwiseConvolutionLayerOptimized::NEDepthwiseConvolutionLayerOptimized(std::shared_ptr<IMemoryManager> memory_manager) |
| 97 | : _func(std::move(memory_manager)) |
| 98 | { |
| 99 | } |
| 100 | |
| 101 | void NEDepthwiseConvolutionLayerOptimized::configure(ITensor *input, |
| 102 | const ITensor *weights, |
| 103 | const ITensor *biases, |
| 104 | ITensor *output, const PadStrideInfo &conv_info, |
| 105 | unsigned int depth_multiplier, |
| 106 | const ActivationLayerInfo &act_info, |
| 107 | const Size2D &dilation) |
| 108 | { |
| 109 | _func.configure(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation); |
| 110 | } |
| 111 | |
| 112 | Status NEDepthwiseConvolutionLayerOptimized::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, |
| 113 | unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation) |
| 114 | { |
| 115 | return validate_arguments_optimized(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation); |
| 116 | } |
| 117 | |
| 118 | void NEDepthwiseConvolutionLayerOptimized::run() |
| 119 | { |
| 120 | _func.run(); |
| 121 | } |
| 122 | |
| 123 | void NEDepthwiseConvolutionLayerOptimized::prepare() |
| 124 | { |
| 125 | _func.prepare(); |
| 126 | } |
| 127 | |
| 128 | NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::NEDepthwiseConvolutionLayerOptimizedInternal(std::shared_ptr<IMemoryManager> memory_manager) |
| 129 | : _memory_group(memory_manager), _dwc_kernel(), _dwc_optimized_func(memory_manager), _output_stage_kernel(), _border_handler(), _permute_input(), _permute_weights(), _permute_output(), |
| 130 | _activationlayer_function(), _accumulator(), _permuted_input(), _permuted_weights(), _permuted_output(), _original_weights(nullptr), _has_bias(false), _is_quantized(false), _is_optimized(false), |
| 131 | _is_nchw(true), _permute(false), _is_activationlayer_enabled(false), _is_prepared(false) |
| 132 | { |
| 133 | } |
| 134 | |
| 135 | void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::configure_generic(ITensor *input, |
| 136 | const ITensor *weights, |
| 137 | const ITensor *biases, |
| 138 | ITensor *output, |
| 139 | const PadStrideInfo &conv_info, |
| 140 | unsigned int depth_multiplier, |
| 141 | const ActivationLayerInfo &act_info, |
| 142 | const Size2D &dilation) |
| 143 | { |
| 144 | ARM_COMPUTE_UNUSED(act_info); |
| 145 | |
| 146 | PixelValue zero_value(0.f); |
| 147 | |
| 148 | // Initialize the intermediate accumulator tensor in case of quantized input |
| 149 | if(_is_quantized) |
| 150 | { |
| 151 | TensorShape accum_shape = output->info()->tensor_shape(); |
| 152 | DataLayout accum_layout = output->info()->data_layout(); |
| 153 | if(!_is_nchw) |
| 154 | { |
| 155 | permute(accum_shape, PermutationVector(1U, 2U, 0U)); |
| 156 | accum_layout = DataLayout::NCHW; |
| 157 | } |
| 158 | |
| 159 | _memory_group.manage(&_accumulator); |
| 160 | _accumulator.allocator()->init(TensorInfo(accum_shape, 1, DataType::S32, output->info()->quantization_info())); |
| 161 | _accumulator.info()->set_data_layout(accum_layout); |
| 162 | zero_value = PixelValue(static_cast<uint32_t>(input->info()->quantization_info().uniform().offset)); |
| 163 | } |
| 164 | |
| 165 | if(!_is_nchw) |
| 166 | { |
| 167 | _memory_group.manage(&_permuted_input); |
| 168 | _memory_group.manage(&_permuted_output); |
| 169 | |
| 170 | // Configure the function to transform the input tensor from NHWC -> NCHW |
| 171 | _permute_input.configure(input, &_permuted_input, PermutationVector(1U, 2U, 0U)); |
| 172 | _permuted_input.info()->set_data_layout(DataLayout::NCHW); |
| 173 | |
| 174 | // Configure the function to transform the weights tensor from HWI -> IHW |
| 175 | _permute_weights.configure(weights, &_permuted_weights, PermutationVector(1U, 2U, 0U)); |
| 176 | _permuted_weights.info()->set_data_layout(DataLayout::NCHW); |
| 177 | _permuted_output.info()->set_quantization_info(output->info()->quantization_info()); |
| 178 | |
| 179 | // Configure depthwise |
| 180 | _dwc_kernel.configure(&_permuted_input, &_permuted_weights, (_is_quantized) ? &_accumulator : &_permuted_output, conv_info, depth_multiplier, dilation); |
| 181 | |
| 182 | // Configure border handler |
| 183 | _border_handler.configure(&_permuted_input, _dwc_kernel.border_size(), BorderMode::CONSTANT, zero_value); |
| 184 | |
| 185 | // Allocate tensors |
| 186 | _permuted_input.allocator()->allocate(); |
| 187 | } |
| 188 | else |
| 189 | { |
| 190 | // Configure depthwise convolution kernel |
| 191 | _dwc_kernel.configure(input, weights, (_is_quantized) ? &_accumulator : output, conv_info, depth_multiplier, dilation); |
| 192 | |
| 193 | // Configure border handler |
| 194 | _border_handler.configure(input, _dwc_kernel.border_size(), BorderMode::CONSTANT, zero_value); |
| 195 | } |
| 196 | |
| 197 | // Configure biases accumulation |
| 198 | if(_is_quantized) |
| 199 | { |
| 200 | const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform(); |
| 201 | const UniformQuantizationInfo wq_info = weights->info()->quantization_info().uniform(); |
| 202 | const UniformQuantizationInfo oq_info = (output->info()->total_size() == 0) ? iq_info : output->info()->quantization_info().uniform(); |
| 203 | |
| 204 | float multiplier = (iq_info.scale * wq_info.scale) / oq_info.scale; |
| 205 | int output_multiplier; |
| 206 | int output_shift; |
| 207 | quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); |
| 208 | _output_stage_kernel.configure(&_accumulator, biases, _is_nchw ? output : &_permuted_output, output_multiplier, output_shift, oq_info.offset); |
| 209 | _accumulator.allocator()->allocate(); |
| 210 | } |
| 211 | else if(_has_bias) |
| 212 | { |
| 213 | _output_stage_kernel.configure(_is_nchw ? output : &_permuted_output, biases); |
| 214 | } |
| 215 | |
| 216 | // Permute output |
| 217 | if(!_is_nchw) |
| 218 | { |
| 219 | // Configure the function to transform the convoluted output to NHWC |
| 220 | _permute_output.configure(&_permuted_output, output, PermutationVector(2U, 0U, 1U)); |
| 221 | _permuted_output.allocator()->allocate(); |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::configure_optimized(const ITensor *input, |
| 226 | const ITensor *weights, |
| 227 | const ITensor *biases, |
| 228 | ITensor *output, |
| 229 | const PadStrideInfo &conv_info, |
| 230 | unsigned int depth_multiplier, |
| 231 | const ActivationLayerInfo &act_info, |
| 232 | const Size2D &dilation) |
| 233 | { |
| 234 | ActivationLayerInfo act_info_to_use = ActivationLayerInfo(); |
| 235 | const bool is_relu = arm_compute::utils::info_helpers::is_relu(act_info); |
| 236 | const bool is_relu6 = arm_compute::utils::info_helpers::is_relu6(act_info); |
| 237 | _is_activationlayer_enabled = act_info.enabled() && !(is_relu || is_relu6); |
| 238 | if(!_is_activationlayer_enabled) |
| 239 | { |
| 240 | act_info_to_use = act_info; |
| 241 | } |
| 242 | |
| 243 | if(_is_nchw) |
| 244 | { |
| 245 | _memory_group.manage(&_permuted_input); |
| 246 | _memory_group.manage(&_permuted_output); |
| 247 | |
| 248 | // Configure the function to transform the input tensor from NCHW -> NHWC |
| 249 | _permute_input.configure(input, &_permuted_input, PermutationVector(2U, 0U, 1U)); |
| 250 | _permuted_input.info()->set_data_layout(DataLayout::NHWC); |
| 251 | |
| 252 | // Configure the function to transform the weights tensor from IHW -> HWI |
| 253 | _permute_weights.configure(weights, &_permuted_weights, PermutationVector(2U, 0U, 1U)); |
| 254 | _permuted_weights.info()->set_data_layout(DataLayout::NHWC); |
| 255 | |
| 256 | _permuted_output.info()->set_data_layout(DataLayout::NHWC); |
| 257 | _permuted_output.info()->set_quantization_info(output->info()->quantization_info()); |
| 258 | |
| 259 | // Configure optimized depthwise |
| 260 | _dwc_optimized_func.configure(&_permuted_input, &_permuted_weights, biases, &_permuted_output, conv_info, depth_multiplier, act_info_to_use, dilation); |
| 261 | |
| 262 | // Configure the function to transform the convoluted output to ACL's native ordering format NCHW |
| 263 | _permuted_output.info()->set_data_layout(DataLayout::NHWC); |
| 264 | _permute_output.configure(&_permuted_output, output, PermutationVector(1U, 2U, 0U)); |
| 265 | |
| 266 | // Allocate tensors |
| 267 | _permuted_input.allocator()->allocate(); |
| 268 | _permuted_output.allocator()->allocate(); |
| 269 | } |
| 270 | else |
| 271 | { |
| 272 | _dwc_optimized_func.configure(input, weights, biases, output, conv_info, depth_multiplier, act_info_to_use, dilation); |
| 273 | } |
| 274 | } |
| 275 | |
| 276 | void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::configure(ITensor *input, |
| 277 | const ITensor *weights, |
| 278 | const ITensor *biases, |
| 279 | ITensor *output, const PadStrideInfo &conv_info, |
| 280 | unsigned int depth_multiplier, |
| 281 | const ActivationLayerInfo &act_info, |
| 282 | const Size2D &dilation) |
| 283 | { |
| 284 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); |
| 285 | // Perform validation step |
| 286 | ARM_COMPUTE_ERROR_THROW_ON(NEDepthwiseConvolutionLayerOptimizedInternal::validate(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), |
| 287 | output->info(), conv_info, depth_multiplier, act_info, dilation)); |
| 288 | |
| 289 | _original_weights = weights; |
| 290 | _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type()); |
| 291 | _has_bias = biases != nullptr; |
| 292 | _is_optimized = NEDepthwiseConvolutionAssemblyDispatch::is_optimized_supported(input->info(), |
| 293 | weights->info(), |
| 294 | conv_info, |
| 295 | depth_multiplier, |
| 296 | dilation); |
| 297 | _is_nchw = input->info()->data_layout() == DataLayout::NCHW; |
| 298 | _permute = _is_optimized == _is_nchw; |
| 299 | _is_prepared = false; |
| 300 | _is_activationlayer_enabled = act_info.enabled(); |
| 301 | |
| 302 | // Configure appropriate pipeline |
| 303 | if(_is_optimized) |
| 304 | { |
| 305 | configure_optimized(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation); |
| 306 | } |
| 307 | else |
| 308 | { |
| 309 | configure_generic(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation); |
| 310 | } |
| 311 | |
| 312 | // Configure activation |
| 313 | if(_is_activationlayer_enabled) |
| 314 | { |
| 315 | _activationlayer_function.configure(output, nullptr, act_info); |
| 316 | } |
| 317 | } |
| 318 | |
| 319 | Status NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::validate(const ITensorInfo *input, |
| 320 | const ITensorInfo *weights, |
| 321 | const ITensorInfo *biases, |
| 322 | const ITensorInfo *output, |
| 323 | const PadStrideInfo &conv_info, |
| 324 | unsigned int depth_multiplier, |
| 325 | const ActivationLayerInfo &act_info, |
| 326 | const Size2D &dilation) |
| 327 | { |
| 328 | return validate_arguments_optimized(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation); |
| 329 | } |
| 330 | |
| 331 | void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::run_generic() |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 332 | { |
| 333 | // Fill border |
| 334 | NEScheduler::get().schedule(&_border_handler, Window::DimX); |
| 335 | |
| 336 | // Execute depthwise convolution |
| 337 | NEScheduler::get().schedule(&_dwc_kernel, Window::DimX); |
| 338 | |
| 339 | // Add biases |
| 340 | if(_has_bias || _is_quantized) |
| 341 | { |
| 342 | NEScheduler::get().schedule(&_output_stage_kernel, Window::DimX); |
| 343 | } |
| 344 | |
| 345 | // Permute output |
| 346 | if(!_is_nchw) |
| 347 | { |
| 348 | _permute_output.run(); |
| 349 | } |
| 350 | } |
| 351 | |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 352 | void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::run_optimized() |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 353 | { |
| 354 | // Run assembly function |
| 355 | _dwc_optimized_func.run(); |
| 356 | |
| 357 | // Permute output |
| 358 | if(_is_nchw) |
| 359 | { |
| 360 | _permute_output.run(); |
| 361 | } |
| 362 | } |
| 363 | |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 364 | void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::run() |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 365 | { |
| 366 | prepare(); |
| 367 | |
| 368 | MemoryGroupResourceScope scope_mg(_memory_group); |
| 369 | |
| 370 | // Permute input |
| 371 | if(_permute) |
| 372 | { |
| 373 | _permute_input.run(); |
| 374 | } |
| 375 | |
| 376 | _is_optimized ? run_optimized() : run_generic(); |
| 377 | |
| 378 | // Run activation |
| 379 | if(_is_activationlayer_enabled) |
| 380 | { |
| 381 | _activationlayer_function.run(); |
| 382 | } |
| 383 | } |
| 384 | |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 385 | void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerOptimizedInternal::prepare() |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 386 | { |
| 387 | if(!_is_prepared) |
| 388 | { |
| 389 | // Permute weights |
| 390 | if(_permute) |
| 391 | { |
| 392 | _permuted_weights.allocator()->allocate(); |
| 393 | _permute_weights.run(); |
| 394 | _original_weights->mark_as_unused(); |
| 395 | } |
| 396 | |
| 397 | // Prepare optimized function |
| 398 | if(_is_optimized) |
| 399 | { |
| 400 | _dwc_optimized_func.prepare(); |
| 401 | if(!_permuted_weights.is_used()) |
| 402 | { |
| 403 | _permuted_weights.allocator()->free(); |
| 404 | } |
| 405 | } |
| 406 | |
| 407 | _is_prepared = true; |
| 408 | } |
| 409 | } |
| 410 | |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 411 | NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerGeneric::NEDepthwiseConvolutionLayerGeneric() |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 412 | : _depthwise_conv_kernel(), _fill_border(), _permute_input(), _permute_weights(), _permute_output(), _activationlayer_function(), _permuted_input(), _permuted_weights(), _permuted_output(), |
| 413 | _is_prepared(false), _is_nchw(false), _is_activationlayer_enabled(false), _original_weights(nullptr) |
Michalis Spyrou | b7b3153 | 2017-11-23 12:10:21 +0000 | [diff] [blame] | 414 | { |
| 415 | } |
| 416 | |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 417 | void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerGeneric::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, |
| 418 | unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation) |
Michalis Spyrou | b7b3153 | 2017-11-23 12:10:21 +0000 | [diff] [blame] | 419 | { |
Michele Di Giorgio | ff27192 | 2019-07-17 15:59:32 +0100 | [diff] [blame] | 420 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); |
Michele Di Giorgio | ff27192 | 2019-07-17 15:59:32 +0100 | [diff] [blame] | 421 | ARM_COMPUTE_ERROR_THROW_ON(NEDepthwiseConvolutionLayer::validate(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), |
| 422 | output->info(), conv_info, depth_multiplier, act_info, dilation)); |
Michalis Spyrou | b7b3153 | 2017-11-23 12:10:21 +0000 | [diff] [blame] | 423 | |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 424 | _is_nchw = input->info()->data_layout() == DataLayout::NCHW; |
| 425 | _is_prepared = !_is_nchw; |
Giorgio Arena | 26b2216 | 2018-08-13 15:49:49 +0100 | [diff] [blame] | 426 | |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 427 | ITensor *input_to_use = input; |
| 428 | const ITensor *weights_to_use = weights; |
| 429 | ITensor *output_to_use = output; |
| 430 | if(_is_nchw) |
Giorgio Arena | 26b2216 | 2018-08-13 15:49:49 +0100 | [diff] [blame] | 431 | { |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 432 | _permute_input.configure(input, &_permuted_input, PermutationVector(2U, 0U, 1U)); |
| 433 | _permuted_input.info()->set_data_layout(DataLayout::NHWC); |
| 434 | input_to_use = &_permuted_input; |
Giorgio Arena | 26b2216 | 2018-08-13 15:49:49 +0100 | [diff] [blame] | 435 | |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 436 | _permute_weights.configure(weights, &_permuted_weights, PermutationVector(2U, 0U, 1U)); |
| 437 | _permuted_weights.info()->set_data_layout(DataLayout::NHWC); |
| 438 | weights_to_use = &_permuted_weights; |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 439 | |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 440 | _permuted_output.allocator()->init(output->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(TensorShape())); |
| 441 | output_to_use = &_permuted_output; |
Giorgio Arena | 26b2216 | 2018-08-13 15:49:49 +0100 | [diff] [blame] | 442 | } |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 443 | _original_weights = weights_to_use; |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 444 | |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 445 | _depthwise_conv_kernel.configure(input_to_use, weights_to_use, biases, output_to_use, conv_info, depth_multiplier, dilation); |
| 446 | _fill_border.configure(input_to_use, _depthwise_conv_kernel.border_size(), BorderMode::CONSTANT, PixelValue(static_cast<uint64_t>(0), input->info()->data_type(), input->info()->quantization_info())); |
| 447 | |
| 448 | if(_is_nchw) |
| 449 | { |
| 450 | _permute_output.configure(&_permuted_output, output, PermutationVector(1U, 2U, 0U)); |
| 451 | _permuted_output.info()->set_data_layout(DataLayout::NHWC); |
| 452 | |
| 453 | _permuted_input.allocator()->allocate(); |
| 454 | _permuted_weights.allocator()->allocate(); |
| 455 | _permuted_output.allocator()->allocate(); |
Giorgio Arena | 26b2216 | 2018-08-13 15:49:49 +0100 | [diff] [blame] | 456 | } |
| 457 | |
Georgios Pinitas | 60e9825 | 2018-10-22 16:17:20 +0100 | [diff] [blame] | 458 | //Configure Activation Layer |
| 459 | _is_activationlayer_enabled = act_info.enabled(); |
Georgios Pinitas | 60e9825 | 2018-10-22 16:17:20 +0100 | [diff] [blame] | 460 | if(_is_activationlayer_enabled) |
| 461 | { |
| 462 | _activationlayer_function.configure(output, nullptr, act_info); |
| 463 | } |
Michalis Spyrou | b7b3153 | 2017-11-23 12:10:21 +0000 | [diff] [blame] | 464 | } |
| 465 | |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 466 | Status NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerGeneric::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, |
| 467 | const PadStrideInfo &conv_info, |
| 468 | unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation) |
Abe Mbise | 7784c83 | 2018-05-31 16:48:41 +0100 | [diff] [blame] | 469 | { |
| 470 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 471 | if(input->data_layout() == DataLayout::NCHW) |
Giorgio Arena | 26b2216 | 2018-08-13 15:49:49 +0100 | [diff] [blame] | 472 | { |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 473 | TensorShape permuted_input_shape = input->tensor_shape(); |
| 474 | TensorShape permuted_weights_shape = weights->tensor_shape(); |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 475 | TensorShape permuted_output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation); |
| 476 | permute(permuted_input_shape, PermutationVector(2U, 0U, 1U)); |
| 477 | permute(permuted_weights_shape, PermutationVector(2U, 0U, 1U)); |
| 478 | permute(permuted_output_shape, PermutationVector(2U, 0U, 1U)); |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 479 | |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 480 | const TensorInfo permuted_input = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_input_shape).set_data_layout(DataLayout::NHWC)); |
| 481 | const TensorInfo permuted_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_weights_shape).set_data_layout(DataLayout::NHWC)); |
| 482 | const TensorInfo permuted_output = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(permuted_output_shape).set_data_layout(DataLayout::NCHW)); |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 483 | |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 484 | ARM_COMPUTE_RETURN_ON_ERROR(NEPermute::validate(input, &permuted_input, PermutationVector(2U, 0U, 1U))); |
| 485 | ARM_COMPUTE_RETURN_ON_ERROR(NEPermute::validate(weights, &permuted_weights, PermutationVector(2U, 0U, 1U))); |
| 486 | ARM_COMPUTE_RETURN_ON_ERROR(NEPermute::validate(&permuted_output, output, PermutationVector(1U, 2U, 0U))); |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 487 | |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 488 | ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionLayerNativeKernel::validate(&permuted_input, &permuted_weights, biases, &permuted_output, conv_info, depth_multiplier, dilation)); |
Giorgio Arena | 26b2216 | 2018-08-13 15:49:49 +0100 | [diff] [blame] | 489 | } |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 490 | else |
Giorgio Arena | 26b2216 | 2018-08-13 15:49:49 +0100 | [diff] [blame] | 491 | { |
Gian Marco Iodice | bd9097d | 2019-07-26 15:31:02 +0100 | [diff] [blame] | 492 | ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionLayerNativeKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, dilation)); |
Abe Mbise | 7784c83 | 2018-05-31 16:48:41 +0100 | [diff] [blame] | 493 | } |
| 494 | |
Georgios Pinitas | 60e9825 | 2018-10-22 16:17:20 +0100 | [diff] [blame] | 495 | // Validate Activation Layer |
| 496 | if(act_info.enabled()) |
| 497 | { |
| 498 | ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(output, nullptr, act_info)); |
| 499 | } |
| 500 | |
Abe Mbise | 7784c83 | 2018-05-31 16:48:41 +0100 | [diff] [blame] | 501 | return Status{}; |
| 502 | } |
| 503 | |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 504 | void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerGeneric::run() |
Michalis Spyrou | b7b3153 | 2017-11-23 12:10:21 +0000 | [diff] [blame] | 505 | { |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 506 | if(_is_nchw) |
Giorgio Arena | 26b2216 | 2018-08-13 15:49:49 +0100 | [diff] [blame] | 507 | { |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 508 | prepare(); |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 509 | _permute_input.run(); |
Giorgio Arena | 26b2216 | 2018-08-13 15:49:49 +0100 | [diff] [blame] | 510 | } |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 511 | |
| 512 | NEScheduler::get().schedule(&_fill_border, Window::DimX); |
| 513 | NEScheduler::get().schedule(&_depthwise_conv_kernel, Window::DimY); |
| 514 | |
| 515 | if(_is_nchw) |
Georgios Pinitas | d05dce4 | 2018-01-22 16:29:17 +0000 | [diff] [blame] | 516 | { |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 517 | _permute_output.run(); |
Giorgio Arena | 26b2216 | 2018-08-13 15:49:49 +0100 | [diff] [blame] | 518 | } |
Georgios Pinitas | 60e9825 | 2018-10-22 16:17:20 +0100 | [diff] [blame] | 519 | |
| 520 | if(_is_activationlayer_enabled) |
| 521 | { |
| 522 | _activationlayer_function.run(); |
| 523 | } |
Anthony Barbier | fb8dda2 | 2018-01-30 09:27:05 +0000 | [diff] [blame] | 524 | } |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 525 | |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 526 | void NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayerGeneric::prepare() |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 527 | { |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 528 | if(!_is_prepared) |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 529 | { |
| 530 | ARM_COMPUTE_ERROR_ON(!_original_weights->is_used()); |
| 531 | |
Giorgio Arena | d93e263 | 2019-10-15 11:09:33 +0100 | [diff] [blame] | 532 | _permute_weights.run(); |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 533 | _original_weights->mark_as_unused(); |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 534 | _is_prepared = true; |
| 535 | } |
| 536 | } |
Manuel Bottini | 05069f0 | 2019-09-26 17:18:26 +0100 | [diff] [blame] | 537 | |
| 538 | NEDepthwiseConvolutionLayer::NEDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) |
| 539 | : _depth_conv_func(DepthwiseConvolutionFunction::GENERIC), _func_optimized(std::move(memory_manager)), _func_generic() |
| 540 | { |
| 541 | } |
| 542 | |
| 543 | void NEDepthwiseConvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, |
| 544 | const ActivationLayerInfo &act_info, const Size2D &dilation) |
| 545 | { |
| 546 | _depth_conv_func = get_depthwiseconvolution_function(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info, dilation); |
| 547 | switch(_depth_conv_func) |
| 548 | { |
| 549 | case DepthwiseConvolutionFunction::OPTIMIZED: |
| 550 | _func_optimized.configure(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation); |
| 551 | break; |
| 552 | case DepthwiseConvolutionFunction::GENERIC: |
| 553 | _func_generic.configure(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation); |
| 554 | break; |
| 555 | default: |
| 556 | ARM_COMPUTE_ERROR("Unsupported DepthwiseConvolutionFunction"); |
| 557 | } |
| 558 | } |
| 559 | |
| 560 | Status NEDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, |
| 561 | unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation) |
| 562 | { |
| 563 | DepthwiseConvolutionFunction depth_conv_func = get_depthwiseconvolution_function(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation); |
| 564 | switch(depth_conv_func) |
| 565 | { |
| 566 | case DepthwiseConvolutionFunction::OPTIMIZED: |
| 567 | return NEDepthwiseConvolutionLayerOptimized::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation); |
| 568 | break; |
| 569 | case DepthwiseConvolutionFunction::GENERIC: |
| 570 | return NEDepthwiseConvolutionLayerGeneric::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation); |
| 571 | break; |
| 572 | default: |
| 573 | ARM_COMPUTE_ERROR("Unsupported DepthwiseConvolutionFunction"); |
| 574 | } |
| 575 | } |
| 576 | |
| 577 | DepthwiseConvolutionFunction NEDepthwiseConvolutionLayer::get_depthwiseconvolution_function(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, |
| 578 | const PadStrideInfo &conv_info, |
| 579 | unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation) |
| 580 | { |
| 581 | if(bool(NEDepthwiseConvolutionLayerOptimized::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation))) |
| 582 | { |
| 583 | return DepthwiseConvolutionFunction::OPTIMIZED; |
| 584 | } |
| 585 | else |
| 586 | { |
| 587 | return DepthwiseConvolutionFunction::GENERIC; |
| 588 | } |
| 589 | } |
| 590 | |
| 591 | void NEDepthwiseConvolutionLayer::run() |
| 592 | { |
| 593 | switch(_depth_conv_func) |
| 594 | { |
| 595 | case DepthwiseConvolutionFunction::OPTIMIZED: |
| 596 | _func_optimized.run(); |
| 597 | break; |
| 598 | case DepthwiseConvolutionFunction::GENERIC: |
| 599 | _func_generic.run(); |
| 600 | break; |
| 601 | default: |
| 602 | ARM_COMPUTE_ERROR("DepthwiseConvolutionFunction not properly configured"); |
| 603 | } |
| 604 | } |
| 605 | |
| 606 | void NEDepthwiseConvolutionLayer::prepare() |
| 607 | { |
| 608 | switch(_depth_conv_func) |
| 609 | { |
| 610 | case DepthwiseConvolutionFunction::OPTIMIZED: |
| 611 | _func_optimized.prepare(); |
| 612 | break; |
| 613 | case DepthwiseConvolutionFunction::GENERIC: |
| 614 | _func_generic.prepare(); |
| 615 | break; |
| 616 | default: |
| 617 | ARM_COMPUTE_ERROR("DepthwiseConvolutionFunction not properly configured"); |
| 618 | } |
| 619 | } |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 620 | } // namespace arm_compute |