Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +0100 | [diff] [blame] | 1 | /* |
Georgios Pinitas | dea2d2d | 2018-12-19 16:23:17 +0000 | [diff] [blame] | 2 | * Copyright (c) 2018-2019 ARM Limited. |
Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +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 | */ |
| 24 | #include "arm_compute/runtime/CL/functions/CLPadLayer.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/ICLTensor.h" |
| 27 | #include "arm_compute/core/Types.h" |
Manuel Bottini | 60f3911 | 2019-03-18 15:25:15 +0000 | [diff] [blame] | 28 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +0100 | [diff] [blame] | 29 | #include "support/ToolchainSupport.h" |
| 30 | |
| 31 | namespace arm_compute |
| 32 | { |
| 33 | CLPadLayer::CLPadLayer() |
Manuel Bottini | 60f3911 | 2019-03-18 15:25:15 +0000 | [diff] [blame] | 34 | : _copy_kernel(), _mode(), _padding(), _memset_kernel(), _num_dimensions(0), _slice_functions(nullptr), _concat_functions(nullptr), _slice_results(nullptr), _concat_results(nullptr) |
Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +0100 | [diff] [blame] | 35 | { |
| 36 | } |
| 37 | |
Manuel Bottini | 60f3911 | 2019-03-18 15:25:15 +0000 | [diff] [blame] | 38 | void CLPadLayer::configure_constant_mode(ICLTensor *input, ICLTensor *output, const PaddingList &padding, const PixelValue constant_value) |
| 39 | { |
| 40 | // Set the pages of the output to the constant_value. |
| 41 | _memset_kernel.configure(output, constant_value); |
| 42 | |
| 43 | // Fill out padding list with zeroes. |
| 44 | PaddingList padding_extended = padding; |
| 45 | for(size_t i = padding.size(); i < TensorShape::num_max_dimensions; i++) |
| 46 | { |
| 47 | padding_extended.emplace_back(PaddingInfo{ 0, 0 }); |
| 48 | } |
| 49 | |
| 50 | // Create a window within the output tensor where the input will be copied. |
| 51 | Window copy_window = Window(); |
| 52 | for(uint32_t i = 0; i < output->info()->num_dimensions(); ++i) |
| 53 | { |
| 54 | copy_window.set(i, Window::Dimension(padding_extended[i].first, padding_extended[i].first + input->info()->dimension(i), 1)); |
| 55 | } |
| 56 | // Copy the input to the output, leaving the padding filled with the constant_value. |
| 57 | _copy_kernel.configure(input, output, PaddingList(), ©_window); |
| 58 | } |
| 59 | |
| 60 | void CLPadLayer::configure_reflect_symmetric_mode(ICLTensor *input, ICLTensor *output) |
| 61 | { |
| 62 | int64_t last_padding_dimension = _padding.size() - 1; |
| 63 | // Reflecting can be performed by effectively unfolding the input as follows: |
| 64 | // For each dimension starting at DimX: |
| 65 | // Create a before and after slice, which values depend on the selected padding mode |
| 66 | // Concatenate the before and after padding with the tensor to be padded |
| 67 | |
| 68 | // Two strided slice functions will be required for each dimension padded as well as a |
| 69 | // concatenate function and the tensors to hold the temporary results. |
| 70 | _slice_functions = arm_compute::support::cpp14::make_unique<CLStridedSlice[]>(2 * _num_dimensions); |
| 71 | _slice_results = arm_compute::support::cpp14::make_unique<CLTensor[]>(2 * _num_dimensions); |
| 72 | _concat_functions = arm_compute::support::cpp14::make_unique<CLConcatenateLayer[]>(_num_dimensions); |
| 73 | _concat_results = arm_compute::support::cpp14::make_unique<CLTensor[]>(_num_dimensions - 1); |
| 74 | Coordinates starts_before, ends_before, starts_after, ends_after, strides; |
| 75 | ICLTensor *prev = input; |
| 76 | for(uint32_t i = 0; i < _num_dimensions; ++i) |
| 77 | { |
| 78 | // Values in strides from the previous dimensions need to be set to 1 to avoid reversing again. |
| 79 | if(i > 0) |
| 80 | { |
| 81 | strides.set(i - 1, 1); |
| 82 | } |
| 83 | |
| 84 | if(_padding[i].first > 0 || _padding[i].second > 0) |
| 85 | { |
| 86 | // Set the starts, ends, and strides values for the current dimension. |
| 87 | // Due to the bit masks passed to strided slice, the values below the current dimension in |
| 88 | // starts and ends will be ignored so do not need to be modified. |
| 89 | if(_mode == PaddingMode::REFLECT) |
| 90 | { |
| 91 | starts_before.set(i, _padding[i].first); |
| 92 | ends_before.set(i, 0); |
| 93 | starts_after.set(i, input->info()->dimension(i) - 2); |
| 94 | ends_after.set(i, input->info()->dimension(i) - _padding[i].second - 2); |
| 95 | strides.set(i, -1); |
| 96 | } |
| 97 | else |
| 98 | { |
| 99 | starts_before.set(i, _padding[i].first - 1); |
| 100 | ends_before.set(i, -1); |
| 101 | starts_after.set(i, input->info()->dimension(i) - 1); |
| 102 | ends_after.set(i, input->info()->dimension(i) - _padding[i].second - 1); |
| 103 | strides.set(i, -1); |
| 104 | } |
| 105 | |
| 106 | // Strided slice wraps negative indexes around to the end of the range, |
| 107 | // instead this should indicate use of the full range and so the bit mask will be modified. |
| 108 | const int32_t begin_mask_before = starts_before[i] < 0 ? ~0 : ~(1u << i); |
| 109 | const int32_t end_mask_before = ends_before[i] < 0 ? ~0 : ~(1u << i); |
| 110 | const int32_t begin_mask_after = starts_after[i] < 0 ? ~0 : ~(1u << i); |
| 111 | const int32_t end_mask_after = ends_after[i] < 0 ? ~0 : ~(1u << i); |
| 112 | |
| 113 | // Reflect the input values for the padding before and after the input. |
| 114 | std::vector<ICLTensor *> concat_vector; |
| 115 | if(_padding[i].first > 0) |
| 116 | { |
| 117 | if(i < prev->info()->num_dimensions()) |
| 118 | { |
| 119 | _slice_functions[2 * i].configure(prev, &_slice_results[2 * i], starts_before, ends_before, strides, begin_mask_before, end_mask_before); |
| 120 | concat_vector.push_back(&_slice_results[2 * i]); |
| 121 | } |
| 122 | else |
| 123 | { |
| 124 | // Performing the slice is unnecessary if the result would simply be a copy of the tensor. |
| 125 | concat_vector.push_back(prev); |
| 126 | } |
| 127 | } |
| 128 | concat_vector.push_back(prev); |
| 129 | if(_padding[i].second > 0) |
| 130 | { |
| 131 | if(i < prev->info()->num_dimensions()) |
| 132 | { |
| 133 | _slice_functions[2 * i + 1].configure(prev, &_slice_results[2 * i + 1], starts_after, ends_after, strides, begin_mask_after, end_mask_after); |
| 134 | concat_vector.push_back(&_slice_results[2 * i + 1]); |
| 135 | } |
| 136 | else |
| 137 | { |
| 138 | // Performing the slice is unnecessary if the result would simply be a copy of the tensor. |
| 139 | concat_vector.push_back(prev); |
| 140 | } |
| 141 | } |
| 142 | // Concatenate the padding before and after with the input. |
| 143 | ICLTensor *out = (static_cast<int32_t>(i) == last_padding_dimension) ? output : &_concat_results[i]; |
| 144 | _concat_functions[i].configure(concat_vector, out, get_index_data_layout_dimension(prev->info()->data_layout(), i)); |
| 145 | prev = out; |
| 146 | } |
| 147 | } |
| 148 | for(uint32_t i = 0; i < _num_dimensions; ++i) |
| 149 | { |
| 150 | if((static_cast<int32_t>(i) != last_padding_dimension)) |
| 151 | { |
| 152 | _concat_results[i].allocator()->allocate(); |
| 153 | } |
| 154 | _slice_results[2 * i].allocator()->allocate(); |
| 155 | _slice_results[2 * i + 1].allocator()->allocate(); |
| 156 | } |
| 157 | } |
| 158 | |
Usama Arif | 8cf8c11 | 2019-03-14 15:36:54 +0000 | [diff] [blame] | 159 | void CLPadLayer::configure(ICLTensor *input, ICLTensor *output, const PaddingList &padding, PixelValue constant_value, PaddingMode mode) |
Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +0100 | [diff] [blame] | 160 | { |
Manuel Bottini | 60f3911 | 2019-03-18 15:25:15 +0000 | [diff] [blame] | 161 | ARM_COMPUTE_ERROR_THROW_ON(validate(input->info(), output->info(), padding, constant_value, mode)); |
Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +0100 | [diff] [blame] | 162 | |
Manuel Bottini | 60f3911 | 2019-03-18 15:25:15 +0000 | [diff] [blame] | 163 | _padding = padding; |
| 164 | _mode = mode; |
Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +0100 | [diff] [blame] | 165 | |
Manuel Bottini | 60f3911 | 2019-03-18 15:25:15 +0000 | [diff] [blame] | 166 | TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input->info()->tensor_shape(), _padding); |
| 167 | |
| 168 | auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(padded_shape)); |
| 169 | |
| 170 | // Find the last dimension requiring padding so that it is known when to write to output and whether any padding is applied. |
| 171 | int64_t last_padding_dimension = _padding.size() - 1; |
| 172 | for(; last_padding_dimension >= 0; --last_padding_dimension) |
| 173 | { |
| 174 | if(_padding[last_padding_dimension].first > 0 || _padding[last_padding_dimension].second > 0) |
| 175 | { |
| 176 | break; |
| 177 | } |
| 178 | } |
| 179 | _num_dimensions = last_padding_dimension + 1; |
| 180 | if(_num_dimensions > 0) |
| 181 | { |
| 182 | switch(_mode) |
| 183 | { |
| 184 | case PaddingMode::CONSTANT: |
| 185 | { |
| 186 | configure_constant_mode(input, output, padding, constant_value); |
| 187 | break; |
| 188 | } |
| 189 | case PaddingMode::REFLECT: |
| 190 | case PaddingMode::SYMMETRIC: |
| 191 | { |
| 192 | configure_reflect_symmetric_mode(input, output); |
| 193 | break; |
| 194 | } |
| 195 | default: |
| 196 | ARM_COMPUTE_ERROR("Padding mode not supported."); |
| 197 | } |
| 198 | } |
| 199 | else |
| 200 | { |
| 201 | // Copy the input to the whole output if no padding is applied |
| 202 | _copy_kernel.configure(input, output); |
| 203 | } |
Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +0100 | [diff] [blame] | 204 | } |
| 205 | |
Usama Arif | 8cf8c11 | 2019-03-14 15:36:54 +0000 | [diff] [blame] | 206 | Status CLPadLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding, PixelValue constant_value, PaddingMode mode) |
Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +0100 | [diff] [blame] | 207 | { |
Manuel Bottini | 60f3911 | 2019-03-18 15:25:15 +0000 | [diff] [blame] | 208 | ARM_COMPUTE_RETURN_ERROR_ON(padding.size() > input->num_dimensions()); |
Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +0100 | [diff] [blame] | 209 | |
Manuel Bottini | 60f3911 | 2019-03-18 15:25:15 +0000 | [diff] [blame] | 210 | TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding); |
| 211 | |
| 212 | // Use CLCopyKernel and CLMemsetKernel to validate all padding modes as this includes all of the shape and info validation. |
| 213 | PaddingList padding_extended = padding; |
| 214 | for(size_t i = padding.size(); i < TensorShape::num_max_dimensions; i++) |
| 215 | { |
| 216 | padding_extended.emplace_back(PaddingInfo{ 0, 0 }); |
| 217 | } |
| 218 | |
| 219 | Window copy_window = Window(); |
| 220 | for(uint32_t i = 0; i < padded_shape.num_dimensions(); ++i) |
| 221 | { |
| 222 | copy_window.set(i, Window::Dimension(padding_extended[i].first, padding_extended[i].first + input->dimension(i), 1)); |
| 223 | } |
| 224 | if(output->total_size() > 0) |
| 225 | { |
| 226 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), padded_shape); |
| 227 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(output, input); |
| 228 | ARM_COMPUTE_RETURN_ON_ERROR(CLCopyKernel::validate(input, output, PaddingList(), ©_window)); |
| 229 | ARM_COMPUTE_RETURN_ON_ERROR(CLMemsetKernel::validate(output, constant_value)); |
| 230 | } |
| 231 | else |
| 232 | { |
| 233 | ARM_COMPUTE_RETURN_ON_ERROR(CLCopyKernel::validate(input, &input->clone()->set_tensor_shape(padded_shape), PaddingList(), ©_window)); |
| 234 | ARM_COMPUTE_RETURN_ON_ERROR(CLMemsetKernel::validate(&input->clone()->set_tensor_shape(padded_shape), constant_value)); |
| 235 | } |
| 236 | |
| 237 | switch(mode) |
| 238 | { |
| 239 | case PaddingMode::CONSTANT: |
| 240 | { |
| 241 | break; |
| 242 | } |
| 243 | case PaddingMode::REFLECT: |
| 244 | case PaddingMode::SYMMETRIC: |
| 245 | { |
| 246 | for(uint32_t i = 0; i < padding.size(); ++i) |
| 247 | { |
| 248 | if(mode == PaddingMode::REFLECT) |
| 249 | { |
| 250 | ARM_COMPUTE_RETURN_ERROR_ON(padding[i].first >= input->dimension(i)); |
| 251 | ARM_COMPUTE_RETURN_ERROR_ON(padding[i].second >= input->dimension(i)); |
| 252 | } |
| 253 | else |
| 254 | { |
| 255 | ARM_COMPUTE_RETURN_ERROR_ON(padding[i].first > input->dimension(i)); |
| 256 | ARM_COMPUTE_RETURN_ERROR_ON(padding[i].second > input->dimension(i)); |
| 257 | } |
| 258 | } |
| 259 | break; |
| 260 | } |
| 261 | default: |
| 262 | { |
| 263 | ARM_COMPUTE_ERROR("Invalid mode"); |
| 264 | } |
| 265 | } |
Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +0100 | [diff] [blame] | 266 | return Status{}; |
| 267 | } |
| 268 | |
| 269 | void CLPadLayer::run() |
| 270 | { |
Manuel Bottini | 60f3911 | 2019-03-18 15:25:15 +0000 | [diff] [blame] | 271 | if(_num_dimensions > 0) |
| 272 | { |
| 273 | switch(_mode) |
| 274 | { |
| 275 | case PaddingMode::CONSTANT: |
| 276 | { |
| 277 | CLScheduler::get().enqueue(_memset_kernel, false); |
| 278 | CLScheduler::get().enqueue(_copy_kernel, true); |
| 279 | break; |
| 280 | } |
| 281 | case PaddingMode::REFLECT: |
| 282 | case PaddingMode::SYMMETRIC: |
| 283 | { |
| 284 | for(uint32_t i = 0; i < _num_dimensions; ++i) |
| 285 | { |
| 286 | if(_padding[i].first > 0 || _padding[i].second > 0) |
| 287 | { |
| 288 | if(_padding[i].first > 0 && _slice_results[2 * i].info()->total_size() > 0) |
| 289 | { |
| 290 | _slice_functions[2 * i].run(); |
| 291 | } |
| 292 | if(_padding[i].second > 0 && _slice_results[2 * i + 1].info()->total_size() > 0) |
| 293 | { |
| 294 | _slice_functions[2 * i + 1].run(); |
| 295 | } |
| 296 | CLScheduler::get().sync(); |
| 297 | _concat_functions[i].run(); |
| 298 | CLScheduler::get().sync(); |
| 299 | } |
| 300 | } |
| 301 | break; |
| 302 | } |
| 303 | default: |
| 304 | ARM_COMPUTE_ERROR("Padding mode not supported."); |
| 305 | } |
| 306 | } |
| 307 | else |
| 308 | { |
| 309 | CLScheduler::get().enqueue(_copy_kernel, true); |
| 310 | } |
Giuseppe Rossini | d7647d4 | 2018-07-17 18:13:13 +0100 | [diff] [blame] | 311 | } |
| 312 | } // namespace arm_compute |