Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2016, 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/core/NEON/kernels/NEFillBorderKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/ITensor.h" |
| 29 | #include "arm_compute/core/TensorInfo.h" |
| 30 | #include "arm_compute/core/Validate.h" |
| 31 | #include "arm_compute/core/Window.h" |
| 32 | |
| 33 | #include <algorithm> |
| 34 | #include <cstdint> |
| 35 | |
| 36 | using namespace arm_compute; |
| 37 | |
| 38 | namespace arm_compute |
| 39 | { |
| 40 | class Coordinates; |
| 41 | } // namespace arm_compute |
| 42 | |
| 43 | NEFillBorderKernel::NEFillBorderKernel() |
| 44 | : _tensor(nullptr), _border_size(0), _mode(BorderMode::UNDEFINED), _constant_border_value(0) |
| 45 | { |
| 46 | } |
| 47 | |
| 48 | void NEFillBorderKernel::configure(ITensor *tensor, BorderSize border_size, BorderMode border_mode, const PixelValue &constant_border_value) |
| 49 | { |
| 50 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(tensor, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F32); |
| 51 | |
| 52 | _tensor = tensor; |
| 53 | _border_size = border_size; |
| 54 | _mode = border_mode; |
| 55 | _constant_border_value = constant_border_value; |
| 56 | |
| 57 | _border_size.limit(tensor->info()->padding()); |
| 58 | |
| 59 | Window win; |
| 60 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 61 | win.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 62 | win.use_tensor_dimensions(_tensor->info(), Window::DimZ); |
| 63 | INEKernel::configure(win); |
| 64 | } |
| 65 | |
| 66 | void NEFillBorderKernel::run(const Window &window) |
| 67 | { |
| 68 | // If there is no border: early exit |
| 69 | if(_border_size.empty()) |
| 70 | { |
| 71 | return; |
| 72 | } |
| 73 | |
| 74 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 75 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 76 | |
| 77 | switch(_mode) |
| 78 | { |
| 79 | case BorderMode::CONSTANT: |
| 80 | { |
| 81 | switch(_tensor->info()->data_type()) |
| 82 | { |
| 83 | case DataType::U8: |
| 84 | fill_constant_value_single_channel<uint8_t>(window); |
| 85 | break; |
| 86 | case DataType::QS8: |
| 87 | case DataType::S8: |
| 88 | fill_constant_value_single_channel<int8_t>(window); |
| 89 | break; |
| 90 | case DataType::U16: |
| 91 | fill_constant_value_single_channel<uint16_t>(window); |
| 92 | break; |
| 93 | case DataType::S16: |
| 94 | case DataType::QS16: |
| 95 | fill_constant_value_single_channel<int16_t>(window); |
| 96 | break; |
| 97 | case DataType::U32: |
| 98 | fill_constant_value_single_channel<uint32_t>(window); |
| 99 | break; |
| 100 | case DataType::S32: |
| 101 | fill_constant_value_single_channel<int32_t>(window); |
| 102 | break; |
| 103 | case DataType::F32: |
| 104 | static_assert(sizeof(float) == 4, "Float must be 32 bit"); |
| 105 | fill_constant_value_single_channel<float>(window); |
| 106 | break; |
| 107 | default: |
| 108 | ARM_COMPUTE_ERROR("Not handled"); |
| 109 | } |
| 110 | break; |
| 111 | } |
| 112 | case BorderMode::REPLICATE: |
| 113 | { |
| 114 | switch(_tensor->info()->data_type()) |
| 115 | { |
| 116 | case DataType::U8: |
| 117 | fill_replicate_single_channel<uint8_t>(window); |
| 118 | break; |
| 119 | case DataType::QS8: |
| 120 | case DataType::S8: |
| 121 | fill_replicate_single_channel<int8_t>(window); |
| 122 | break; |
| 123 | case DataType::U16: |
| 124 | fill_replicate_single_channel<uint16_t>(window); |
| 125 | break; |
| 126 | case DataType::S16: |
| 127 | case DataType::QS16: |
| 128 | fill_replicate_single_channel<int16_t>(window); |
| 129 | break; |
| 130 | case DataType::U32: |
| 131 | fill_replicate_single_channel<uint32_t>(window); |
| 132 | break; |
| 133 | case DataType::S32: |
| 134 | fill_replicate_single_channel<int32_t>(window); |
| 135 | break; |
| 136 | case DataType::F32: |
| 137 | static_assert(sizeof(float) == 4, "Float must be 32 bit"); |
| 138 | fill_replicate_single_channel<float>(window); |
| 139 | break; |
| 140 | default: |
| 141 | ARM_COMPUTE_ERROR("Not handled"); |
| 142 | } |
| 143 | break; |
| 144 | } |
| 145 | case BorderMode::UNDEFINED: |
| 146 | break; // Nothing to do here |
| 147 | default: |
| 148 | ARM_COMPUTE_ERROR("Unknown border mode"); |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | template <typename T> |
| 153 | void NEFillBorderKernel::fill_replicate_single_channel(const Window &window) |
| 154 | { |
| 155 | uint8_t *const start_valid_region = _tensor->ptr_to_element(_tensor->info()->valid_region().anchor); |
| 156 | const size_t &width = _tensor->info()->valid_region().shape[0]; |
| 157 | const size_t &height = _tensor->info()->valid_region().shape[1]; |
| 158 | |
| 159 | // Left and right border |
| 160 | Window vertical(window); |
| 161 | vertical.set(Window::DimY, Window::Dimension(0, height, 1)); |
| 162 | |
| 163 | Iterator vertical_it(_tensor, vertical); |
| 164 | |
| 165 | execute_window_loop(vertical, [&](const Coordinates & id) |
| 166 | { |
| 167 | const auto row_start = reinterpret_cast<T *>(start_valid_region + vertical_it.offset()); |
| 168 | const auto left_val = *reinterpret_cast<T *>(vertical_it.ptr()); |
| 169 | const auto right_val = *(reinterpret_cast<T *>(vertical_it.ptr()) + width - 1); |
| 170 | |
| 171 | // Fill left and right borders |
| 172 | std::fill_n(row_start - _border_size.left, _border_size.left, left_val); |
| 173 | std::fill_n(row_start + width, _border_size.right, right_val); |
| 174 | }, |
| 175 | vertical_it); |
| 176 | |
| 177 | // Top and bottom border |
| 178 | Iterator plane_it(_tensor, window); |
| 179 | |
| 180 | // Iterate over all XY planes |
| 181 | execute_window_loop(window, [&](const Coordinates & id) |
| 182 | { |
| 183 | const auto first_row = reinterpret_cast<T *>(start_valid_region + plane_it.offset()); |
| 184 | |
| 185 | // Top border |
| 186 | for(int i = -_border_size.top; i < 0; ++i) |
| 187 | { |
| 188 | const auto row_start = reinterpret_cast<T *>(start_valid_region + plane_it.offset() + i * _tensor->info()->strides_in_bytes()[1]); |
| 189 | |
| 190 | // Copy top rows including left/right borders |
| 191 | std::copy_n(first_row - _border_size.left, _border_size.left + width + _border_size.right, row_start - _border_size.left); |
| 192 | } |
| 193 | |
| 194 | const auto last_row = reinterpret_cast<T *>(start_valid_region + plane_it.offset() + (height - 1) * _tensor->info()->strides_in_bytes()[1]); |
| 195 | |
| 196 | // Bottom border |
| 197 | for(unsigned int i = height; i < height + _border_size.bottom; ++i) |
| 198 | { |
| 199 | const auto row_start = reinterpret_cast<T *>(start_valid_region + plane_it.offset() + i * _tensor->info()->strides_in_bytes()[1]); |
| 200 | |
| 201 | // Copy bottom rows including left/right borders |
| 202 | std::copy_n(last_row - _border_size.left, _border_size.left + width + _border_size.right, row_start - _border_size.left); |
| 203 | } |
| 204 | }, |
| 205 | plane_it); |
| 206 | } |
| 207 | |
| 208 | template <typename T> |
| 209 | void NEFillBorderKernel::fill_constant_value_single_channel(const Window &window) |
| 210 | { |
| 211 | T constant_border_value; |
| 212 | _constant_border_value.get(constant_border_value); |
| 213 | |
| 214 | uint8_t *const start_valid_region = _tensor->ptr_to_element(_tensor->info()->valid_region().anchor); |
| 215 | const size_t &width = _tensor->info()->valid_region().shape[0]; |
| 216 | const size_t &height = _tensor->info()->valid_region().shape[1]; |
| 217 | |
| 218 | // Left and right border |
| 219 | Window vertical(window); |
| 220 | vertical.set(Window::DimY, Window::Dimension(0, height, 1)); |
| 221 | |
| 222 | Iterator vertical_it(_tensor, vertical); |
| 223 | |
| 224 | execute_window_loop(vertical, [&](const Coordinates & id) |
| 225 | { |
| 226 | const auto row_start = reinterpret_cast<T *>(start_valid_region + vertical_it.offset()); |
| 227 | |
| 228 | // Fill left and right borders |
| 229 | std::fill_n(row_start - _border_size.left, _border_size.left, constant_border_value); |
| 230 | std::fill_n(row_start + width, _border_size.right, constant_border_value); |
| 231 | }, |
| 232 | vertical_it); |
| 233 | |
| 234 | // Top and bottom border |
| 235 | Iterator plane_it(_tensor, window); |
| 236 | |
| 237 | // Iterate over all XY planes |
| 238 | execute_window_loop(window, [&](const Coordinates & id) |
| 239 | { |
| 240 | // Top border |
| 241 | for(int i = -_border_size.top; i < 0; ++i) |
| 242 | { |
| 243 | const auto row_start = reinterpret_cast<T *>(start_valid_region + plane_it.offset() + i * _tensor->info()->strides_in_bytes()[1]); |
| 244 | |
| 245 | // Fill top rows including left/right borders |
| 246 | std::fill_n(row_start - _border_size.left, _border_size.left + width + _border_size.right, constant_border_value); |
| 247 | } |
| 248 | |
| 249 | // Bottom border |
| 250 | for(unsigned int i = height; i < height + _border_size.bottom; ++i) |
| 251 | { |
| 252 | const auto row_start = reinterpret_cast<T *>(start_valid_region + plane_it.offset() + i * _tensor->info()->strides_in_bytes()[1]); |
| 253 | |
| 254 | // Fill bottom rows including left/right borders |
| 255 | std::fill_n(row_start - _border_size.left, _border_size.left + width + _border_size.right, constant_border_value); |
| 256 | } |
| 257 | }, |
| 258 | plane_it); |
| 259 | } |