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/NEChannelCombineKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/IAccessWindow.h" |
| 29 | #include "arm_compute/core/IMultiImage.h" |
| 30 | #include "arm_compute/core/ITensor.h" |
| 31 | #include "arm_compute/core/MultiImageInfo.h" |
| 32 | #include "arm_compute/core/TensorInfo.h" |
| 33 | #include "arm_compute/core/Types.h" |
| 34 | #include "arm_compute/core/Validate.h" |
| 35 | #include "arm_compute/core/Window.h" |
| 36 | |
| 37 | #include <arm_neon.h> |
| 38 | |
| 39 | using namespace arm_compute; |
| 40 | |
| 41 | namespace arm_compute |
| 42 | { |
| 43 | class Coordinates; |
| 44 | } // namespace arm_compute |
| 45 | |
| 46 | NEChannelCombineKernel::NEChannelCombineKernel() |
| 47 | : _func(nullptr), _planes{ { nullptr } }, _output(nullptr), _output_multi(nullptr), _x_subsampling{ { 1, 1, 1 } }, _y_subsampling{ { 1, 1, 1 } }, _num_elems_processed_per_iteration(8), |
| 48 | _is_parallelizable(true) |
| 49 | { |
| 50 | } |
| 51 | |
| 52 | void NEChannelCombineKernel::configure(const ITensor *plane0, const ITensor *plane1, const ITensor *plane2, const ITensor *plane3, ITensor *output) |
| 53 | { |
| 54 | ARM_COMPUTE_ERROR_ON_NULLPTR(plane0, plane1, plane2, output); |
| 55 | ARM_COMPUTE_ERROR_ON(plane0 == output); |
| 56 | ARM_COMPUTE_ERROR_ON(plane1 == output); |
| 57 | ARM_COMPUTE_ERROR_ON(plane2 == output); |
| 58 | |
| 59 | set_format_if_unknown(*plane0->info(), Format::U8); |
| 60 | set_format_if_unknown(*plane1->info(), Format::U8); |
| 61 | set_format_if_unknown(*plane2->info(), Format::U8); |
| 62 | |
| 63 | if(plane3 != nullptr) |
| 64 | { |
| 65 | set_format_if_unknown(*plane3->info(), Format::U8); |
| 66 | } |
| 67 | |
| 68 | set_shape_if_empty(*output->info(), plane0->info()->tensor_shape()); |
| 69 | |
| 70 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(plane0, 1, DataType::U8); |
| 71 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(plane1, 1, DataType::U8); |
| 72 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(plane2, 1, DataType::U8); |
| 73 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(output, Format::RGB888, Format::RGBA8888, Format::UYVY422, Format::YUYV422); |
| 74 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(plane0, plane1, plane2); |
| 75 | |
| 76 | if(plane3 != nullptr) |
| 77 | { |
| 78 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(plane0, plane3); |
| 79 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(plane0, plane3); |
| 80 | } |
| 81 | |
| 82 | const Format &output_format = output->info()->format(); |
| 83 | |
| 84 | if(output_format == Format::RGBA8888) |
| 85 | { |
| 86 | ARM_COMPUTE_ERROR_ON(plane3 == output); |
| 87 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(plane3, 1, DataType::U8); |
| 88 | } |
| 89 | |
| 90 | _planes[0] = plane0; |
| 91 | _planes[1] = plane1; |
| 92 | _planes[2] = plane2; |
| 93 | _planes[3] = plane3; |
| 94 | _output = output; |
| 95 | _output_multi = nullptr; |
| 96 | |
| 97 | _num_elems_processed_per_iteration = 8; |
| 98 | _is_parallelizable = true; |
| 99 | |
| 100 | switch(output_format) |
| 101 | { |
| 102 | case Format::RGB888: |
| 103 | _func = &NEChannelCombineKernel::combine_3C; |
| 104 | break; |
| 105 | case Format::RGBA8888: |
| 106 | _func = &NEChannelCombineKernel::combine_4C; |
| 107 | break; |
| 108 | case Format::UYVY422: |
| 109 | _x_subsampling[1] = 2; |
| 110 | _x_subsampling[2] = 2; |
| 111 | _num_elems_processed_per_iteration = 16; |
| 112 | _func = &NEChannelCombineKernel::combine_YUV_1p<true>; |
| 113 | break; |
| 114 | case Format::YUYV422: |
| 115 | _x_subsampling[1] = 2; |
| 116 | _x_subsampling[2] = 2; |
| 117 | _num_elems_processed_per_iteration = 16; |
| 118 | _func = &NEChannelCombineKernel::combine_YUV_1p<false>; |
| 119 | break; |
| 120 | default: |
| 121 | ARM_COMPUTE_ERROR("Not supported format."); |
| 122 | break; |
| 123 | } |
| 124 | |
| 125 | TensorShape subsampled_shape_plane1{ plane0->info()->tensor_shape() }; |
| 126 | subsampled_shape_plane1.set(0, subsampled_shape_plane1[0] / _x_subsampling[1]); |
| 127 | TensorShape subsampled_shape_plane2{ plane0->info()->tensor_shape() }; |
| 128 | subsampled_shape_plane2.set(0, subsampled_shape_plane2[0] / _x_subsampling[2]); |
| 129 | |
| 130 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(plane1->info()->tensor_shape(), subsampled_shape_plane1); |
| 131 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(plane2->info()->tensor_shape(), subsampled_shape_plane2); |
| 132 | |
| 133 | Window win = calculate_max_window(*plane0->info(), Steps(_num_elems_processed_per_iteration)); |
| 134 | |
| 135 | AccessWindowHorizontal output_access(output->info(), 0, _num_elems_processed_per_iteration); |
| 136 | AccessWindowHorizontal plane0_access(plane0->info(), 0, _num_elems_processed_per_iteration / _x_subsampling[1], 1.f / _x_subsampling[0]); |
| 137 | AccessWindowHorizontal plane1_access(plane1->info(), 0, _num_elems_processed_per_iteration / _x_subsampling[1], 1.f / _x_subsampling[1]); |
| 138 | AccessWindowHorizontal plane2_access(plane2->info(), 0, _num_elems_processed_per_iteration / _x_subsampling[1], 1.f / _x_subsampling[2]); |
| 139 | AccessWindowHorizontal plane3_access(plane3 == nullptr ? nullptr : plane3->info(), 0, _num_elems_processed_per_iteration); |
| 140 | |
| 141 | update_window_and_padding( |
| 142 | win, |
| 143 | plane0_access, |
| 144 | plane1_access, |
| 145 | plane2_access, |
| 146 | plane3_access, |
| 147 | output_access); |
| 148 | |
| 149 | ValidRegion valid_region = intersect_valid_regions(plane0->info()->valid_region(), |
| 150 | plane1->info()->valid_region(), |
| 151 | plane2->info()->valid_region()); |
| 152 | |
| 153 | if(plane3 != nullptr) |
| 154 | { |
| 155 | valid_region = intersect_valid_regions(plane3->info()->valid_region(), valid_region); |
| 156 | } |
| 157 | |
| 158 | output_access.set_valid_region(win, ValidRegion(valid_region.anchor, output->info()->tensor_shape())); |
| 159 | |
| 160 | INEKernel::configure(win); |
| 161 | } |
| 162 | |
| 163 | void NEChannelCombineKernel::configure(const IImage *plane0, const IImage *plane1, const IImage *plane2, IMultiImage *output) |
| 164 | { |
| 165 | ARM_COMPUTE_ERROR_ON_NULLPTR(plane0, plane1, plane2, output); |
| 166 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(plane0); |
| 167 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(plane1); |
| 168 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(plane2); |
| 169 | |
| 170 | set_format_if_unknown(*plane0->info(), Format::U8); |
| 171 | set_format_if_unknown(*plane1->info(), Format::U8); |
| 172 | set_format_if_unknown(*plane2->info(), Format::U8); |
| 173 | |
| 174 | set_shape_if_empty(*output->plane(0)->info(), plane0->info()->tensor_shape()); |
| 175 | |
| 176 | switch(output->info()->format()) |
| 177 | { |
| 178 | case Format::NV12: |
| 179 | case Format::NV21: |
| 180 | case Format::IYUV: |
| 181 | { |
| 182 | TensorShape subsampled_shape = plane0->info()->tensor_shape(); |
| 183 | subsampled_shape.set(0, subsampled_shape[0] / 2); |
| 184 | subsampled_shape.set(1, subsampled_shape[1] / 2); |
| 185 | |
| 186 | set_shape_if_empty(*output->plane(1)->info(), subsampled_shape); |
| 187 | |
| 188 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->plane(1)->info()->tensor_shape(), subsampled_shape); |
| 189 | |
| 190 | if(output->info()->format() == Format::IYUV) |
| 191 | { |
| 192 | set_shape_if_empty(*output->plane(2)->info(), subsampled_shape); |
| 193 | |
| 194 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->plane(2)->info()->tensor_shape(), subsampled_shape); |
| 195 | } |
| 196 | break; |
| 197 | } |
| 198 | case Format::YUV444: |
| 199 | set_shape_if_empty(*output->plane(1)->info(), plane0->info()->tensor_shape()); |
| 200 | set_shape_if_empty(*output->plane(2)->info(), plane0->info()->tensor_shape()); |
| 201 | |
| 202 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(plane1, plane2, output->plane(1), output->plane(2)); |
| 203 | break; |
| 204 | default: |
| 205 | ARM_COMPUTE_ERROR("Unsupported format"); |
| 206 | } |
| 207 | |
| 208 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(plane0, output->plane(0)); |
| 209 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(plane0, 1, DataType::U8); |
| 210 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(plane1, 1, DataType::U8); |
| 211 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(plane2, 1, DataType::U8); |
| 212 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(output, Format::NV12, Format::NV21, Format::IYUV, Format::YUV444); |
| 213 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(plane0, plane1, plane2); |
| 214 | |
| 215 | _planes[0] = plane0; |
| 216 | _planes[1] = plane1; |
| 217 | _planes[2] = plane2; |
| 218 | _planes[3] = nullptr; |
| 219 | _output = nullptr; |
| 220 | _output_multi = output; |
| 221 | bool has_two_planes = false; |
| 222 | unsigned int num_elems_written_plane1 = 8; |
| 223 | |
| 224 | _num_elems_processed_per_iteration = 8; |
| 225 | _is_parallelizable = true; |
| 226 | |
| 227 | const Format &output_format = output->info()->format(); |
| 228 | |
| 229 | switch(output_format) |
| 230 | { |
| 231 | case Format::NV12: |
| 232 | case Format::NV21: |
| 233 | _x_subsampling = { { 1, 2, 2 } }; |
| 234 | _y_subsampling = { { 1, 2, 2 } }; |
| 235 | _func = &NEChannelCombineKernel::combine_YUV_2p; |
| 236 | has_two_planes = true; |
| 237 | num_elems_written_plane1 = 16; |
| 238 | break; |
| 239 | case Format::IYUV: |
| 240 | _is_parallelizable = false; |
| 241 | _x_subsampling = { { 1, 2, 2 } }; |
| 242 | _y_subsampling = { { 1, 2, 2 } }; |
| 243 | _func = &NEChannelCombineKernel::combine_YUV_3p; |
| 244 | break; |
| 245 | case Format::YUV444: |
| 246 | _is_parallelizable = false; |
| 247 | _x_subsampling = { { 1, 1, 1 } }; |
| 248 | _y_subsampling = { { 1, 1, 1 } }; |
| 249 | _func = &NEChannelCombineKernel::combine_YUV_3p; |
| 250 | break; |
| 251 | default: |
| 252 | ARM_COMPUTE_ERROR("Not supported format."); |
| 253 | break; |
| 254 | } |
| 255 | |
| 256 | const unsigned int y_step = *std::max_element(_y_subsampling.begin(), _y_subsampling.end()); |
| 257 | |
| 258 | Window win = calculate_max_window(*plane0->info(), Steps(_num_elems_processed_per_iteration, y_step)); |
| 259 | AccessWindowRectangle output_plane0_access(output->plane(0)->info(), 0, 0, _num_elems_processed_per_iteration, 1, 1.f, 1.f / _y_subsampling[0]); |
| 260 | AccessWindowRectangle output_plane1_access(output->plane(1)->info(), 0, 0, num_elems_written_plane1, 1, 1.f / _x_subsampling[1], 1.f / _y_subsampling[1]); |
| 261 | AccessWindowRectangle output_plane2_access(has_two_planes ? nullptr : output->plane(2)->info(), 0, 0, _num_elems_processed_per_iteration, 1, 1.f / _x_subsampling[2], 1.f / _y_subsampling[2]); |
| 262 | |
| 263 | update_window_and_padding(win, |
| 264 | AccessWindowHorizontal(plane0->info(), 0, _num_elems_processed_per_iteration), |
| 265 | AccessWindowRectangle(plane1->info(), 0, 0, _num_elems_processed_per_iteration, 1, 1.f / _x_subsampling[1], 1.f / _y_subsampling[1]), |
| 266 | AccessWindowRectangle(plane2->info(), 0, 0, _num_elems_processed_per_iteration, 1, 1.f / _x_subsampling[2], 1.f / _y_subsampling[2]), |
| 267 | output_plane0_access, |
| 268 | output_plane1_access, |
| 269 | output_plane2_access); |
| 270 | |
| 271 | ValidRegion plane0_valid_region = plane0->info()->valid_region(); |
| 272 | |
| 273 | ValidRegion output_plane1_region = has_two_planes ? intersect_valid_regions(plane1->info()->valid_region(), plane2->info()->valid_region()) : plane2->info()->valid_region(); |
| 274 | |
| 275 | output_plane0_access.set_valid_region(win, ValidRegion(plane0_valid_region.anchor, output->plane(0)->info()->tensor_shape())); |
| 276 | output_plane1_access.set_valid_region(win, ValidRegion(output_plane1_region.anchor, output->plane(1)->info()->tensor_shape())); |
| 277 | output_plane2_access.set_valid_region(win, ValidRegion(plane2->info()->valid_region().anchor, output->plane(2)->info()->tensor_shape())); |
| 278 | |
| 279 | INEKernel::configure(win); |
| 280 | } |
| 281 | |
| 282 | bool NEChannelCombineKernel::is_parallelisable() const |
| 283 | { |
| 284 | return _is_parallelizable; |
| 285 | } |
| 286 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 287 | void NEChannelCombineKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 288 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 289 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 290 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 291 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 292 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 293 | |
| 294 | (this->*_func)(window); |
| 295 | } |
| 296 | |
| 297 | void NEChannelCombineKernel::combine_3C(const Window &win) |
| 298 | { |
| 299 | Iterator p0(_planes[0], win); |
| 300 | Iterator p1(_planes[1], win); |
| 301 | Iterator p2(_planes[2], win); |
| 302 | Iterator out(_output, win); |
| 303 | |
| 304 | execute_window_loop(win, [&](const Coordinates & id) |
| 305 | { |
| 306 | const auto p0_ptr = static_cast<uint8_t *>(p0.ptr()); |
| 307 | const auto p1_ptr = static_cast<uint8_t *>(p1.ptr()); |
| 308 | const auto p2_ptr = static_cast<uint8_t *>(p2.ptr()); |
| 309 | const auto out_ptr = static_cast<uint8_t *>(out.ptr()); |
| 310 | |
| 311 | const uint8x8x3_t pixels = |
| 312 | { |
| 313 | { |
| 314 | vld1_u8(p0_ptr), |
| 315 | vld1_u8(p1_ptr), |
| 316 | vld1_u8(p2_ptr) |
| 317 | } |
| 318 | }; |
| 319 | |
| 320 | vst3_u8(out_ptr, pixels); |
| 321 | }, |
| 322 | p0, p1, p2, out); |
| 323 | } |
| 324 | |
| 325 | void NEChannelCombineKernel::combine_4C(const Window &win) |
| 326 | { |
| 327 | Iterator p0(_planes[0], win); |
| 328 | Iterator p1(_planes[1], win); |
| 329 | Iterator p2(_planes[2], win); |
| 330 | Iterator p3(_planes[3], win); |
| 331 | Iterator out(_output, win); |
| 332 | |
| 333 | execute_window_loop(win, [&](const Coordinates & id) |
| 334 | { |
| 335 | const auto p0_ptr = static_cast<uint8_t *>(p0.ptr()); |
| 336 | const auto p1_ptr = static_cast<uint8_t *>(p1.ptr()); |
| 337 | const auto p2_ptr = static_cast<uint8_t *>(p2.ptr()); |
| 338 | const auto p3_ptr = static_cast<uint8_t *>(p3.ptr()); |
| 339 | const auto out_ptr = static_cast<uint8_t *>(out.ptr()); |
| 340 | |
| 341 | const uint8x8x4_t pixels = |
| 342 | { |
| 343 | { |
| 344 | vld1_u8(p0_ptr), |
| 345 | vld1_u8(p1_ptr), |
| 346 | vld1_u8(p2_ptr), |
| 347 | vld1_u8(p3_ptr) |
| 348 | } |
| 349 | }; |
| 350 | |
| 351 | vst4_u8(out_ptr, pixels); |
| 352 | }, |
| 353 | p0, p1, p2, p3, out); |
| 354 | } |
| 355 | |
| 356 | template <bool is_uyvy> |
| 357 | void NEChannelCombineKernel::combine_YUV_1p(const Window &win) |
| 358 | { |
| 359 | // Create sub-sampled uv window and init uv planes |
| 360 | Window win_uv(win); |
| 361 | win_uv.set_dimension_step(0, win.x().step() / _x_subsampling[1]); |
| 362 | win_uv.validate(); |
| 363 | |
| 364 | Iterator p0(_planes[0], win); |
| 365 | Iterator p1(_planes[1], win_uv); |
| 366 | Iterator p2(_planes[2], win_uv); |
| 367 | Iterator out(_output, win); |
| 368 | |
| 369 | constexpr auto shift = is_uyvy ? 1 : 0; |
| 370 | |
| 371 | execute_window_loop(win, [&](const Coordinates & id) |
| 372 | { |
| 373 | const auto p0_ptr = static_cast<uint8_t *>(p0.ptr()); |
| 374 | const auto p1_ptr = static_cast<uint8_t *>(p1.ptr()); |
| 375 | const auto p2_ptr = static_cast<uint8_t *>(p2.ptr()); |
| 376 | const auto out_ptr = static_cast<uint8_t *>(out.ptr()); |
| 377 | |
| 378 | const uint8x8x2_t pixels_y = vld2_u8(p0_ptr); |
| 379 | const uint8x8x2_t pixels_uv = |
| 380 | { |
| 381 | { |
| 382 | vld1_u8(p1_ptr), |
| 383 | vld1_u8(p2_ptr) |
| 384 | } |
| 385 | }; |
| 386 | |
| 387 | uint8x8x4_t pixels{ {} }; |
| 388 | pixels.val[0 + shift] = pixels_y.val[0]; |
| 389 | pixels.val[1 - shift] = pixels_uv.val[0]; |
| 390 | pixels.val[2 + shift] = pixels_y.val[1]; |
| 391 | pixels.val[3 - shift] = pixels_uv.val[1]; |
| 392 | |
| 393 | vst4_u8(out_ptr, pixels); |
| 394 | }, |
| 395 | p0, p1, p2, out); |
| 396 | } |
| 397 | |
| 398 | void NEChannelCombineKernel::combine_YUV_2p(const Window &win) |
| 399 | { |
| 400 | ARM_COMPUTE_ERROR_ON(win.x().start() % _x_subsampling[1]); |
| 401 | ARM_COMPUTE_ERROR_ON(win.y().start() % _y_subsampling[1]); |
| 402 | |
| 403 | // Copy first plane |
| 404 | copy_plane(win, 0); |
| 405 | |
| 406 | // Update UV window |
| 407 | Window uv_win(win); |
| 408 | uv_win.set(Window::DimX, Window::Dimension(uv_win.x().start() / _x_subsampling[1], uv_win.x().end() / _x_subsampling[1], _num_elems_processed_per_iteration)); |
| 409 | uv_win.set(Window::DimY, Window::Dimension(uv_win.y().start() / _y_subsampling[1], uv_win.y().end() / _y_subsampling[1], 1)); |
| 410 | uv_win.validate(); |
| 411 | |
| 412 | // Update output win |
| 413 | Window out_win(win); |
| 414 | out_win.set(Window::DimX, Window::Dimension(out_win.x().start(), out_win.x().end(), out_win.x().step() * 2)); |
| 415 | out_win.set(Window::DimY, Window::Dimension(out_win.y().start() / _y_subsampling[1], out_win.y().end() / _y_subsampling[1], 1)); |
| 416 | out_win.validate(); |
| 417 | |
| 418 | // Construct second plane |
| 419 | const int shift = (Format::NV12 == _output_multi->info()->format()) ? 0 : 1; |
| 420 | Iterator p1(_planes[1 + shift], uv_win); |
| 421 | Iterator p2(_planes[2 - shift], uv_win); |
| 422 | Iterator out(_output_multi->plane(1), out_win); |
| 423 | |
| 424 | execute_window_loop(out_win, [&](const Coordinates & id) |
| 425 | { |
| 426 | const uint8x8x2_t pixels = |
| 427 | { |
| 428 | { |
| 429 | vld1_u8(p1.ptr()), |
| 430 | vld1_u8(p2.ptr()) |
| 431 | } |
| 432 | }; |
| 433 | |
| 434 | vst2_u8(out.ptr(), pixels); |
| 435 | }, |
| 436 | p1, p2, out); |
| 437 | } |
| 438 | |
| 439 | void NEChannelCombineKernel::combine_YUV_3p(const Window &win) |
| 440 | { |
| 441 | copy_plane(win, 0); |
| 442 | copy_plane(win, 1); |
| 443 | copy_plane(win, 2); |
| 444 | } |
| 445 | |
| 446 | void NEChannelCombineKernel::copy_plane(const Window &win, uint32_t plane_id) |
| 447 | { |
| 448 | ARM_COMPUTE_ERROR_ON(win.x().start() % _x_subsampling[plane_id]); |
| 449 | ARM_COMPUTE_ERROR_ON(win.y().start() % _y_subsampling[plane_id]); |
| 450 | |
| 451 | // Update window |
| 452 | Window tmp_win(win); |
| 453 | tmp_win.set(Window::DimX, Window::Dimension(tmp_win.x().start() / _x_subsampling[plane_id], tmp_win.x().end() / _x_subsampling[plane_id], _num_elems_processed_per_iteration)); |
| 454 | tmp_win.set(Window::DimY, Window::Dimension(tmp_win.y().start() / _y_subsampling[plane_id], tmp_win.y().end() / _y_subsampling[plane_id], 1)); |
| 455 | tmp_win.validate(); |
| 456 | |
| 457 | Iterator in(_planes[plane_id], tmp_win); |
| 458 | Iterator out(_output_multi->plane(plane_id), tmp_win); |
| 459 | |
| 460 | execute_window_loop(tmp_win, [&](const Coordinates & id) |
| 461 | { |
| 462 | const auto in_ptr = static_cast<uint8_t *>(in.ptr()); |
| 463 | const auto out_ptr = static_cast<uint8_t *>(out.ptr()); |
| 464 | |
| 465 | vst1_u8(out_ptr, vld1_u8(in_ptr)); |
| 466 | }, |
| 467 | in, out); |
| 468 | } |