Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [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 | */ |
Gian Marco Iodice | bd9097d | 2019-07-26 15:31:02 +0100 | [diff] [blame^] | 24 | #include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h" |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/NEON/wrapper/traits.h" |
| 28 | #include "arm_compute/core/NEON/wrapper/wrapper.h" |
| 29 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 30 | |
| 31 | namespace arm_compute |
| 32 | { |
| 33 | namespace |
| 34 | { |
| 35 | template <typename T, int S, bool has_biases> |
| 36 | void depthwise_loop_multiplier1(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, |
| 37 | const Size2D &dilation, const Window &window) |
| 38 | { |
| 39 | using VectorType = typename wrapper::traits::neon_vector<T, S>::type; |
| 40 | using TagType = typename wrapper::traits::neon_vector<T, S>::tag_type; |
| 41 | |
| 42 | const size_t input_stride_y = input->info()->strides_in_bytes().y(); |
| 43 | const size_t input_stride_z = input->info()->strides_in_bytes().z(); |
| 44 | const size_t input_max_offset = input->info()->strides_in_bytes().z() * input->info()->dimension(2) - (input->info()->padding().bottom + input->info()->padding().top) * |
| 45 | input->info()->strides_in_bytes().y(); |
| 46 | const size_t weights_width = weights->info()->dimension(1); |
| 47 | const size_t weights_height = weights->info()->dimension(2); |
| 48 | const size_t weights_stride_y = weights->info()->strides_in_bytes().y(); |
| 49 | const size_t weights_stride_z = weights->info()->strides_in_bytes().z(); |
| 50 | const size_t conv_stride_x = conv_info.stride().first; |
| 51 | const size_t conv_stride_y = conv_info.stride().second; |
| 52 | const size_t conv_pad_left = conv_info.pad_left(); |
| 53 | const size_t conv_pad_top = conv_info.pad_top(); |
| 54 | |
| 55 | Window win_input = window; |
| 56 | win_input.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 57 | win_input.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 58 | |
| 59 | Window win_weights = win_input; |
| 60 | win_weights.set(3, Window::Dimension(0, 0, 0)); |
| 61 | |
| 62 | Iterator input_it(input, win_input); |
| 63 | Iterator weights_it(weights, win_weights); |
| 64 | Iterator output_it(output, window); |
| 65 | Iterator biases_it{}; |
| 66 | |
| 67 | if(has_biases) |
| 68 | { |
| 69 | biases_it = Iterator(biases, win_weights); |
| 70 | } |
| 71 | |
| 72 | execute_window_loop(window, [&](const Coordinates & id) |
| 73 | { |
| 74 | VectorType acc = wrapper::vdup_n(static_cast<T>(0), TagType{}); |
| 75 | |
| 76 | const int input_y = id.y() * conv_stride_x - conv_pad_left; |
| 77 | const int input_z = id.z() * conv_stride_y - conv_pad_top; |
| 78 | int input_offset = input_y * input_stride_y + input_z * input_stride_z; |
| 79 | |
| 80 | auto weights_ptr = weights_it.ptr(); |
| 81 | for(size_t h = 0; h < weights_height; ++h) |
| 82 | { |
| 83 | int offs = input_offset; |
| 84 | for(size_t w = 0; w < weights_width; ++w) |
| 85 | { |
| 86 | const auto input_vals = wrapper::vload(reinterpret_cast<T *>(input_it.ptr() + std::min(static_cast<size_t>(offs), input_max_offset))); |
| 87 | const auto weights_vals = wrapper::vload(reinterpret_cast<T *>(weights_ptr + w * weights_stride_y)); |
| 88 | |
| 89 | acc = wrapper::vmla(acc, weights_vals, input_vals); |
| 90 | offs += dilation.x() * input_stride_y; |
| 91 | } |
| 92 | |
| 93 | weights_ptr += weights_stride_z; |
| 94 | input_offset += dilation.y() * input_stride_z; |
| 95 | } |
| 96 | |
| 97 | if(has_biases) |
| 98 | { |
| 99 | const auto biases_vals = wrapper::vload(reinterpret_cast<T *>(biases_it.ptr())); |
| 100 | acc = wrapper::vadd(acc, biases_vals); |
| 101 | } |
| 102 | |
| 103 | wrapper::vstore(reinterpret_cast<T *>(output_it.ptr()), acc); |
| 104 | }, |
| 105 | input_it, weights_it, biases_it, output_it); |
| 106 | } |
| 107 | |
| 108 | template <typename T, bool has_biases> |
| 109 | void depthwise_loop_generic(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, |
| 110 | const Size2D &dilation, unsigned int depth_multiplier, const Window &window) |
| 111 | { |
| 112 | const size_t input_stride_y = input->info()->strides_in_bytes().y(); |
| 113 | const size_t input_stride_z = input->info()->strides_in_bytes().z(); |
| 114 | const size_t input_max_offset = input->info()->strides_in_bytes().z() * input->info()->dimension(2) - (input->info()->padding().bottom + input->info()->padding().top) * |
| 115 | input->info()->strides_in_bytes().y(); |
| 116 | const size_t weights_width = weights->info()->dimension(1); |
| 117 | const size_t weights_height = weights->info()->dimension(2); |
| 118 | const size_t weights_stride_y = weights->info()->strides_in_bytes().y(); |
| 119 | const size_t weights_stride_z = weights->info()->strides_in_bytes().z(); |
| 120 | const size_t conv_stride_x = conv_info.stride().first; |
| 121 | const size_t conv_stride_y = conv_info.stride().second; |
| 122 | const size_t conv_pad_left = conv_info.pad_left(); |
| 123 | const size_t conv_pad_top = conv_info.pad_top(); |
| 124 | |
| 125 | Window win_input = window; |
| 126 | win_input.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 127 | win_input.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 128 | |
| 129 | Window win_weights = win_input; |
| 130 | win_weights.set(3, Window::Dimension(0, 0, 0)); |
| 131 | |
| 132 | win_input.set_dimension_step(Window::DimX, 1); |
| 133 | |
| 134 | Iterator input_it(input, win_input); |
| 135 | Iterator weights_it(weights, win_weights); |
| 136 | Iterator output_it(output, window); |
| 137 | Iterator biases_it{}; |
| 138 | |
| 139 | if(has_biases) |
| 140 | { |
| 141 | biases_it = Iterator(biases, win_weights); |
| 142 | } |
| 143 | |
| 144 | execute_window_loop(window, [&](const Coordinates & id) |
| 145 | { |
| 146 | std::vector<T> acc(depth_multiplier, static_cast<T>(0)); |
| 147 | |
| 148 | const int input_y = id.y() * conv_stride_x - conv_pad_left; |
| 149 | const int input_z = id.z() * conv_stride_y - conv_pad_top; |
| 150 | int input_offset = input_y * input_stride_y + input_z * input_stride_z; |
| 151 | |
| 152 | auto weights_ptr = weights_it.ptr(); |
| 153 | for(size_t h = 0; h < weights_height; ++h) |
| 154 | { |
| 155 | int offs = input_offset; |
| 156 | for(size_t w = 0; w < weights_width; ++w) |
| 157 | { |
| 158 | const auto input_val = *(reinterpret_cast<T *>(input_it.ptr() + std::min(static_cast<size_t>(offs), input_max_offset))); |
| 159 | |
| 160 | for(size_t m = 0; m < depth_multiplier; ++m) |
| 161 | { |
| 162 | const auto weights_val = *(reinterpret_cast<T *>(weights_ptr + m * sizeof(T) + w * weights_stride_y)); |
| 163 | acc.at(m) = std::fma(weights_val, input_val, acc.at(m)); |
| 164 | } |
| 165 | |
| 166 | offs += dilation.x() * input_stride_y; |
| 167 | } |
| 168 | |
| 169 | weights_ptr += weights_stride_z; |
| 170 | input_offset += dilation.y() * input_stride_z; |
| 171 | } |
| 172 | |
| 173 | if(has_biases) |
| 174 | { |
| 175 | for(size_t m = 0; m < depth_multiplier; ++m) |
| 176 | { |
| 177 | const auto biases_val = *(reinterpret_cast<T *>(biases_it.ptr() + m * sizeof(T))); |
| 178 | *(reinterpret_cast<T *>(output_it.ptr() + m * sizeof(T))) = acc.at(m) + biases_val; |
| 179 | } |
| 180 | } |
| 181 | else |
| 182 | { |
| 183 | for(size_t m = 0; m < depth_multiplier; ++m) |
| 184 | { |
| 185 | *(reinterpret_cast<T *>(output_it.ptr() + m * sizeof(T))) = acc.at(m); |
| 186 | } |
| 187 | } |
| 188 | }, |
| 189 | input_it, weights_it, biases_it, output_it); |
| 190 | } |
| 191 | |
| 192 | Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, |
| 193 | const Size2D &dilation) |
| 194 | { |
| 195 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); |
| 196 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); |
| 197 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); |
| 198 | ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier == 0); |
| 199 | ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(0) * depth_multiplier) != weights->dimension(0)); |
| 200 | ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1)); |
| 201 | ARM_COMPUTE_RETURN_ERROR_ON((conv_info.stride().first < 1) || (conv_info.stride().second < 1)); |
| 202 | |
| 203 | if(biases != nullptr) |
| 204 | { |
| 205 | ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); |
| 206 | ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0)); |
| 207 | } |
| 208 | |
| 209 | if(output->total_size() != 0) |
| 210 | { |
| 211 | const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation); |
| 212 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); |
| 213 | } |
| 214 | |
| 215 | return Status{}; |
| 216 | } |
| 217 | } // namespace |
| 218 | |
| 219 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *biases, |
| 220 | ITensorInfo *output, const PadStrideInfo &conv_info, |
| 221 | unsigned int depth_multiplier, const Size2D &dilation) |
| 222 | { |
| 223 | // Get convolved dimensions |
| 224 | const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation); |
| 225 | |
| 226 | // Output auto inizialitation if not yet initialized |
| 227 | auto_init_if_empty(*output, input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape)); |
| 228 | |
| 229 | // Configure kernel window (generic) |
| 230 | const unsigned int num_elems_read_per_iteration = (depth_multiplier == 1) ? 8 / element_size_from_data_type(input->data_type()) : 1; |
| 231 | const unsigned int num_elems_written_per_iteration = num_elems_read_per_iteration * depth_multiplier; |
| 232 | |
| 233 | // Configure kernel window |
| 234 | Window win = calculate_max_window(*output, Steps(num_elems_written_per_iteration)); |
| 235 | |
| 236 | AccessWindowStatic input_access(input, 0, -conv_info.pad_left(), ceil_to_multiple(num_elems_read_per_iteration, input->dimension(0)), |
| 237 | input->dimension(1) + std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top())); |
| 238 | AccessWindowHorizontal weights_access(weights, 0, num_elems_written_per_iteration); |
| 239 | AccessWindowHorizontal output_access(output, 0, num_elems_written_per_iteration); |
| 240 | |
| 241 | bool window_changed = update_window_and_padding(win, input_access, weights_access, output_access); |
| 242 | |
| 243 | if(biases != nullptr) |
| 244 | { |
| 245 | AccessWindowHorizontal biases_access(biases, 0, num_elems_written_per_iteration); |
| 246 | window_changed |= update_window_and_padding(win, biases_access); |
| 247 | } |
| 248 | |
| 249 | output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); |
| 250 | |
| 251 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 252 | return std::make_pair(err, win); |
| 253 | } |
| 254 | |
Gian Marco Iodice | bd9097d | 2019-07-26 15:31:02 +0100 | [diff] [blame^] | 255 | NEDepthwiseConvolutionLayerNativeKernel::NEDepthwiseConvolutionLayerNativeKernel() |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 256 | : _func(), _border_size(0), _input(), _weights(), _biases(), _output(), _conv_info(), _depth_multiplier(1), _dilation() |
| 257 | { |
| 258 | } |
| 259 | |
Gian Marco Iodice | bd9097d | 2019-07-26 15:31:02 +0100 | [diff] [blame^] | 260 | BorderSize NEDepthwiseConvolutionLayerNativeKernel::border_size() const |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 261 | { |
| 262 | return _border_size; |
| 263 | } |
| 264 | |
Gian Marco Iodice | bd9097d | 2019-07-26 15:31:02 +0100 | [diff] [blame^] | 265 | void NEDepthwiseConvolutionLayerNativeKernel::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, |
| 266 | const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation) |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 267 | { |
| 268 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); |
| 269 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, dilation)); |
| 270 | |
| 271 | _input = input; |
| 272 | _weights = weights; |
| 273 | _biases = biases; |
| 274 | _output = output; |
| 275 | _conv_info = conv_info; |
| 276 | _depth_multiplier = depth_multiplier; |
| 277 | _border_size = BorderSize(_conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0); |
| 278 | _dilation = dilation; |
| 279 | |
| 280 | switch(_input->info()->data_type()) |
| 281 | { |
| 282 | case DataType::F32: |
Gian Marco Iodice | bd9097d | 2019-07-26 15:31:02 +0100 | [diff] [blame^] | 283 | _func = (biases != nullptr) ? &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<float, 2, true> : &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise<float, 2, false>; |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 284 | break; |
| 285 | default: |
| 286 | ARM_COMPUTE_ERROR("Data type not supported"); |
| 287 | break; |
| 288 | } |
| 289 | |
| 290 | auto win_config = validate_and_configure_window(_input->info(), _weights->info(), (biases != nullptr) ? biases->info() : nullptr, _output->info(), _conv_info, _depth_multiplier, dilation); |
| 291 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 292 | INEKernel::configure(win_config.second); |
| 293 | } |
| 294 | |
Gian Marco Iodice | bd9097d | 2019-07-26 15:31:02 +0100 | [diff] [blame^] | 295 | Status NEDepthwiseConvolutionLayerNativeKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, |
| 296 | unsigned int depth_multiplier, |
| 297 | const Size2D &dilation) |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 298 | { |
| 299 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, dilation)); |
| 300 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), (biases != nullptr) ? biases->clone().get() : nullptr, output->clone().get(), conv_info, |
| 301 | depth_multiplier, dilation) |
| 302 | .first); |
| 303 | return Status{}; |
| 304 | } |
| 305 | |
Gian Marco Iodice | bd9097d | 2019-07-26 15:31:02 +0100 | [diff] [blame^] | 306 | void NEDepthwiseConvolutionLayerNativeKernel::run(const Window &window, const ThreadInfo &info) |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 307 | { |
| 308 | ARM_COMPUTE_UNUSED(info); |
| 309 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 310 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 311 | |
| 312 | (this->*_func)(window); |
| 313 | } |
| 314 | |
| 315 | template <typename T, int S, bool has_biases> |
Gian Marco Iodice | bd9097d | 2019-07-26 15:31:02 +0100 | [diff] [blame^] | 316 | void NEDepthwiseConvolutionLayerNativeKernel::run_depthwise(const Window &window) |
Giorgio Arena | 44f5572 | 2019-07-12 14:49:49 +0100 | [diff] [blame] | 317 | { |
| 318 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 319 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 320 | |
| 321 | if(_depth_multiplier == 1) |
| 322 | { |
| 323 | depthwise_loop_multiplier1<T, S, has_biases>(_input, _weights, _biases, _output, _conv_info, _dilation, window); |
| 324 | } |
| 325 | else |
| 326 | { |
| 327 | depthwise_loop_generic<T, has_biases>(_input, _weights, _biases, _output, _conv_info, _dilation, _depth_multiplier, window); |
| 328 | } |
| 329 | } |
| 330 | } // namespace arm_compute |