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/NESobel7x7Kernel.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/Types.h" |
| 31 | #include "arm_compute/core/Utils.h" |
| 32 | #include "arm_compute/core/Validate.h" |
| 33 | |
| 34 | #include <arm_neon.h> |
| 35 | #include <cstdint> |
| 36 | |
| 37 | using namespace arm_compute; |
| 38 | |
| 39 | namespace arm_compute |
| 40 | { |
| 41 | class Coordinates; |
| 42 | } // namespace arm_compute |
| 43 | |
| 44 | namespace |
| 45 | { |
| 46 | const int32x4_t minusfour = vdupq_n_s32(-4); |
| 47 | const int32x4_t minusfive = vdupq_n_s32(-5); |
| 48 | const int32x4_t four = vdupq_n_s32(4); |
| 49 | const int32x4_t five = vdupq_n_s32(5); |
| 50 | const int32x4_t six = vdupq_n_s32(6); |
| 51 | const int32x4_t fifteen = vdupq_n_s32(15); |
| 52 | const int32x4_t twenty = vdupq_n_s32(20); |
| 53 | |
| 54 | inline int32x4x2_t compute_hor_sobel_x(const int32x4x4_t &data) |
| 55 | { |
| 56 | int32x4x2_t out = |
| 57 | { |
| 58 | { |
| 59 | vnegq_s32(data.val[0]), |
| 60 | vnegq_s32(data.val[1]) |
| 61 | } |
| 62 | }; |
| 63 | |
| 64 | out.val[0] = vmlaq_s32(out.val[0], |
| 65 | vextq_s32(data.val[0], data.val[1], 1), minusfour); |
| 66 | |
| 67 | out.val[0] = vmlaq_s32(out.val[0], |
| 68 | vextq_s32(data.val[0], data.val[1], 2), minusfive); |
| 69 | |
| 70 | out.val[0] = vmlaq_s32(out.val[0], data.val[1], five); |
| 71 | |
| 72 | out.val[0] = vmlaq_s32(out.val[0], |
| 73 | vextq_s32(data.val[1], data.val[2], 1), four); |
| 74 | |
| 75 | out.val[0] = vaddq_s32(out.val[0], |
| 76 | vextq_s32(data.val[1], data.val[2], 2)); |
| 77 | |
| 78 | out.val[1] = vmlaq_s32(out.val[1], |
| 79 | vextq_s32(data.val[1], data.val[2], 1), minusfour); |
| 80 | |
| 81 | out.val[1] = vmlaq_s32(out.val[1], |
| 82 | vextq_s32(data.val[1], data.val[2], 2), minusfive); |
| 83 | |
| 84 | out.val[1] = vmlaq_s32(out.val[1], data.val[2], five); |
| 85 | |
| 86 | out.val[1] = vmlaq_s32(out.val[1], |
| 87 | vextq_s32(data.val[2], data.val[3], 1), four); |
| 88 | |
| 89 | out.val[1] = vaddq_s32(out.val[1], |
| 90 | vextq_s32(data.val[2], data.val[3], 2)); |
| 91 | |
| 92 | return out; |
| 93 | } |
| 94 | |
| 95 | inline int32x4x2_t compute_hor_sobel_y(const int32x4x4_t &data) |
| 96 | { |
| 97 | int32x4x2_t out = |
| 98 | { |
| 99 | { |
| 100 | data.val[0], |
| 101 | data.val[1] |
| 102 | } |
| 103 | }; |
| 104 | |
| 105 | out.val[0] = vmlaq_s32(out.val[0], |
| 106 | vextq_s32(data.val[0], data.val[1], 1), six); |
| 107 | |
| 108 | out.val[0] = vmlaq_s32(out.val[0], |
| 109 | vextq_s32(data.val[0], data.val[1], 2), fifteen); |
| 110 | |
| 111 | out.val[0] = vmlaq_s32(out.val[0], |
| 112 | vextq_s32(data.val[0], data.val[1], 3), twenty); |
| 113 | |
| 114 | out.val[0] = vmlaq_s32(out.val[0], data.val[1], fifteen); |
| 115 | |
| 116 | out.val[0] = vmlaq_s32(out.val[0], |
| 117 | vextq_s32(data.val[1], data.val[2], 1), six); |
| 118 | |
| 119 | out.val[0] = vaddq_s32(out.val[0], |
| 120 | vextq_s32(data.val[1], data.val[2], 2)); |
| 121 | |
| 122 | out.val[1] = vmlaq_s32(out.val[1], |
| 123 | vextq_s32(data.val[1], data.val[2], 1), six); |
| 124 | |
| 125 | out.val[1] = vmlaq_s32(out.val[1], |
| 126 | vextq_s32(data.val[1], data.val[2], 2), fifteen); |
| 127 | |
| 128 | out.val[1] = vmlaq_s32(out.val[1], |
| 129 | vextq_s32(data.val[1], data.val[2], 3), twenty); |
| 130 | |
| 131 | out.val[1] = vmlaq_s32(out.val[1], data.val[2], fifteen); |
| 132 | |
| 133 | out.val[1] = vmlaq_s32(out.val[1], |
| 134 | vextq_s32(data.val[2], data.val[3], 1), six); |
| 135 | |
| 136 | out.val[1] = vaddq_s32(out.val[1], |
| 137 | vextq_s32(data.val[2], data.val[3], 2)); |
| 138 | |
| 139 | return out; |
| 140 | } |
| 141 | } // namespace |
| 142 | |
| 143 | NESobel7x7HorKernel::NESobel7x7HorKernel() |
| 144 | : _input(nullptr), _output_x(nullptr), _output_y(nullptr), _run_sobel_x(false), _run_sobel_y(false), _border_size(0) |
| 145 | { |
| 146 | } |
| 147 | |
| 148 | BorderSize NESobel7x7HorKernel::border_size() const |
| 149 | { |
| 150 | return _border_size; |
| 151 | } |
| 152 | |
| 153 | void NESobel7x7HorKernel::configure(const ITensor *input, ITensor *output_x, ITensor *output_y, bool border_undefined) |
| 154 | { |
| 155 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(input, Format::U8); |
| 156 | ARM_COMPUTE_ERROR_ON((output_x == nullptr) && (output_y == nullptr)); |
| 157 | |
| 158 | _run_sobel_x = output_x != nullptr; |
| 159 | _run_sobel_y = output_y != nullptr; |
| 160 | |
| 161 | if(_run_sobel_x) |
| 162 | { |
| 163 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(output_x, Format::S32); |
| 164 | } |
| 165 | |
| 166 | if(_run_sobel_y) |
| 167 | { |
| 168 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(output_y, Format::S32); |
| 169 | } |
| 170 | |
| 171 | _input = input; |
| 172 | _output_x = output_x; |
| 173 | _output_y = output_y; |
| 174 | _border_size = BorderSize(border_undefined ? 0 : 3, 3); |
| 175 | |
| 176 | // Configure kernel window |
| 177 | constexpr unsigned int num_elems_processed_per_iteration = 8; |
| 178 | constexpr unsigned int num_elems_read_per_iteration = 16; |
| 179 | constexpr unsigned int num_elems_written_per_iteration = 8; |
| 180 | |
| 181 | Window win = calculate_max_window_horizontal(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); |
| 182 | AccessWindowHorizontal output_x_access(output_x == nullptr ? nullptr : output_x->info(), 0, num_elems_written_per_iteration); |
| 183 | AccessWindowHorizontal output_y_access(output_y == nullptr ? nullptr : output_y->info(), 0, num_elems_written_per_iteration); |
| 184 | |
| 185 | update_window_and_padding(win, |
| 186 | AccessWindowHorizontal(input->info(), -border_size().left, num_elems_read_per_iteration), |
| 187 | output_x_access, |
| 188 | output_y_access); |
| 189 | |
| 190 | output_x_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size()); |
| 191 | output_y_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size()); |
| 192 | |
| 193 | INEKernel::configure(win); |
| 194 | } |
| 195 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame^] | 196 | void NESobel7x7HorKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 197 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame^] | 198 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 199 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 200 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 201 | |
| 202 | Iterator input(_input, window); |
| 203 | Iterator output_x; |
| 204 | Iterator output_y; |
| 205 | |
| 206 | if(_run_sobel_x) |
| 207 | { |
| 208 | output_x = Iterator(_output_x, window); |
| 209 | } |
| 210 | |
| 211 | if(_run_sobel_y) |
| 212 | { |
| 213 | output_y = Iterator(_output_y, window); |
| 214 | } |
| 215 | |
| 216 | if(_run_sobel_y && _run_sobel_x) |
| 217 | { |
| 218 | execute_window_loop(window, [&](const Coordinates & id) |
| 219 | { |
| 220 | const uint8x16_t data = vld1q_u8(input.ptr() - 3); |
| 221 | |
| 222 | const uint16x8_t tmp_low_u16 = vmovl_u8(vget_low_u8(data)); |
| 223 | const uint16x8_t tmp_high_u16 = vmovl_u8(vget_high_u8(data)); |
| 224 | |
| 225 | const int32x4x4_t data_s32 = |
| 226 | { |
| 227 | { |
| 228 | vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(tmp_low_u16))), |
| 229 | vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(tmp_low_u16))), |
| 230 | vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(tmp_high_u16))), |
| 231 | vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(tmp_high_u16))) |
| 232 | } |
| 233 | }; |
| 234 | |
| 235 | const int32x4x2_t out_y = compute_hor_sobel_y(data_s32); |
| 236 | vst1q_s32(reinterpret_cast<int32_t *>(output_y.ptr()), out_y.val[0]); |
| 237 | vst1q_s32(reinterpret_cast<int32_t *>(output_y.ptr()) + 4, out_y.val[1]); |
| 238 | |
| 239 | const int32x4x2_t out_x = compute_hor_sobel_x(data_s32); |
| 240 | vst1q_s32(reinterpret_cast<int32_t *>(output_x.ptr()), out_x.val[0]); |
| 241 | vst1q_s32(reinterpret_cast<int32_t *>(output_x.ptr()) + 4, out_x.val[1]); |
| 242 | }, |
| 243 | input, output_x, output_y); |
| 244 | } |
| 245 | else if(_run_sobel_x) |
| 246 | { |
| 247 | execute_window_loop(window, [&](const Coordinates & id) |
| 248 | { |
| 249 | const uint8x16_t data = vld1q_u8(input.ptr() - 3); |
| 250 | |
| 251 | const uint16x8_t tmp_low_u16 = vmovl_u8(vget_low_u8(data)); |
| 252 | const uint16x8_t tmp_high_u16 = vmovl_u8(vget_high_u8(data)); |
| 253 | |
| 254 | const int32x4x4_t data_s32 = |
| 255 | { |
| 256 | { |
| 257 | vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(tmp_low_u16))), |
| 258 | vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(tmp_low_u16))), |
| 259 | vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(tmp_high_u16))), |
| 260 | vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(tmp_high_u16))) |
| 261 | } |
| 262 | }; |
| 263 | |
| 264 | const int32x4x2_t out = compute_hor_sobel_x(data_s32); |
| 265 | vst1q_s32(reinterpret_cast<int32_t *>(output_x.ptr()), out.val[0]); |
| 266 | vst1q_s32(reinterpret_cast<int32_t *>(output_x.ptr()) + 4, out.val[1]); |
| 267 | }, |
| 268 | input, output_x); |
| 269 | } |
| 270 | else if(_run_sobel_y) |
| 271 | { |
| 272 | execute_window_loop(window, [&](const Coordinates & id) |
| 273 | { |
| 274 | const uint8x16_t data = vld1q_u8(input.ptr() - 3); |
| 275 | |
| 276 | const uint16x8_t tmp_low_u16 = vmovl_u8(vget_low_u8(data)); |
| 277 | const uint16x8_t tmp_high_u16 = vmovl_u8(vget_high_u8(data)); |
| 278 | |
| 279 | const int32x4x4_t data_s32 = |
| 280 | { |
| 281 | { |
| 282 | vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(tmp_low_u16))), |
| 283 | vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(tmp_low_u16))), |
| 284 | vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(tmp_high_u16))), |
| 285 | vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(tmp_high_u16))) |
| 286 | } |
| 287 | }; |
| 288 | |
| 289 | const int32x4x2_t out = compute_hor_sobel_x(data_s32); |
| 290 | vst1q_s32(reinterpret_cast<int32_t *>(output_y.ptr()), out.val[0]); |
| 291 | vst1q_s32(reinterpret_cast<int32_t *>(output_y.ptr()) + 4, out.val[1]); |
| 292 | }, |
| 293 | input, output_y); |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | NESobel7x7VertKernel::NESobel7x7VertKernel() |
| 298 | : _input_x(nullptr), _input_y(nullptr), _output_x(nullptr), _output_y(nullptr), _run_sobel_x(false), _run_sobel_y(false) |
| 299 | { |
| 300 | } |
| 301 | |
| 302 | BorderSize NESobel7x7VertKernel::border_size() const |
| 303 | { |
| 304 | return BorderSize(3, 0); |
| 305 | } |
| 306 | |
| 307 | void NESobel7x7VertKernel::configure(const ITensor *input_x, const ITensor *input_y, ITensor *output_x, ITensor *output_y, bool border_undefined) |
| 308 | { |
| 309 | ARM_COMPUTE_ERROR_ON((output_x == nullptr) && (output_y == nullptr)); |
| 310 | |
| 311 | _run_sobel_x = (output_x != nullptr); |
| 312 | _run_sobel_y = (output_y != nullptr); |
| 313 | |
| 314 | if(_run_sobel_x) |
| 315 | { |
| 316 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(input_x, Format::S32); |
| 317 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(output_x, Format::S32); |
| 318 | } |
| 319 | |
| 320 | if(_run_sobel_y) |
| 321 | { |
| 322 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(input_y, Format::S32); |
| 323 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(output_y, Format::S32); |
| 324 | } |
| 325 | |
| 326 | _input_x = input_x; |
| 327 | _input_y = input_y; |
| 328 | _output_x = output_x; |
| 329 | _output_y = output_y; |
| 330 | |
| 331 | const ITensor *const input = _run_sobel_x ? input_x : input_y; |
| 332 | |
| 333 | // Configure kernel window |
| 334 | constexpr unsigned int num_elems_processed_per_iteration = 8; |
| 335 | constexpr unsigned int num_elems_read_per_iteration = 8; |
| 336 | constexpr unsigned int num_elems_written_per_iteration = 8; |
| 337 | constexpr unsigned int num_rows_read_per_iteration = 7; |
| 338 | |
| 339 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); |
| 340 | AccessWindowHorizontal output_x_access(output_x == nullptr ? nullptr : output_x->info(), 0, num_elems_written_per_iteration); |
| 341 | AccessWindowHorizontal output_y_access(output_y == nullptr ? nullptr : output_y->info(), 0, num_elems_written_per_iteration); |
| 342 | |
| 343 | update_window_and_padding(win, |
| 344 | AccessWindowRectangle(input_x == nullptr ? nullptr : input_x->info(), 0, -border_size().top, num_elems_read_per_iteration, num_rows_read_per_iteration), |
| 345 | AccessWindowRectangle(input_y == nullptr ? nullptr : input_y->info(), 0, -border_size().top, num_elems_read_per_iteration, num_rows_read_per_iteration), |
| 346 | output_x_access, |
| 347 | output_y_access); |
| 348 | |
| 349 | output_x_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size()); |
| 350 | output_y_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size()); |
| 351 | |
| 352 | INEKernel::configure(win); |
| 353 | } |
| 354 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame^] | 355 | void NESobel7x7VertKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 356 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame^] | 357 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 358 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 359 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 360 | |
| 361 | Iterator input_x; |
| 362 | Iterator input_y; |
| 363 | Iterator output_x; |
| 364 | Iterator output_y; |
| 365 | |
| 366 | int32_t in_x_stride = 0; |
| 367 | int32_t in_y_stride = 0; |
| 368 | |
| 369 | if(_run_sobel_x) |
| 370 | { |
| 371 | input_x = Iterator(_input_x, window); |
| 372 | output_x = Iterator(_output_x, window); |
| 373 | in_x_stride = _input_x->info()->strides_in_bytes()[1] / pixel_size_from_format(_input_x->info()->format()); |
| 374 | } |
| 375 | |
| 376 | if(_run_sobel_y) |
| 377 | { |
| 378 | input_y = Iterator(_input_y, window); |
| 379 | output_y = Iterator(_output_y, window); |
| 380 | in_y_stride = _input_y->info()->strides_in_bytes()[1] / pixel_size_from_format(_input_y->info()->format()); |
| 381 | } |
| 382 | |
| 383 | if(_run_sobel_x) |
| 384 | { |
| 385 | execute_window_loop(window, [&](const Coordinates & id) |
| 386 | { |
| 387 | auto in_ptr = reinterpret_cast<int32_t *>(input_x.ptr()) - 3 * in_x_stride; |
| 388 | |
| 389 | //top3 |
| 390 | int32x4x2_t data = |
| 391 | { |
| 392 | { |
| 393 | vld1q_s32(in_ptr), |
| 394 | vld1q_s32(in_ptr + 4) |
| 395 | } |
| 396 | }; |
| 397 | |
| 398 | int32x4x2_t out = data; |
| 399 | |
| 400 | //top2 |
| 401 | in_ptr += in_x_stride; |
| 402 | data.val[0] = vld1q_s32(in_ptr); |
| 403 | out.val[0] = vmlaq_s32(out.val[0], data.val[0], six); |
| 404 | |
| 405 | data.val[1] = vld1q_s32(in_ptr + 4); |
| 406 | out.val[1] = vmlaq_s32(out.val[1], data.val[1], six); |
| 407 | |
| 408 | //top |
| 409 | in_ptr += in_x_stride; |
| 410 | data.val[0] = vld1q_s32(in_ptr); |
| 411 | out.val[0] = vmlaq_s32(out.val[0], data.val[0], fifteen); |
| 412 | |
| 413 | data.val[1] = vld1q_s32(in_ptr + 4); |
| 414 | out.val[1] = vmlaq_s32(out.val[1], data.val[1], fifteen); |
| 415 | |
| 416 | //mid |
| 417 | in_ptr += in_x_stride; |
| 418 | data.val[0] = vld1q_s32(in_ptr); |
| 419 | out.val[0] = vmlaq_s32(out.val[0], data.val[0], twenty); |
| 420 | |
| 421 | data.val[1] = vld1q_s32(in_ptr + 4); |
| 422 | out.val[1] = vmlaq_s32(out.val[1], data.val[1], twenty); |
| 423 | |
| 424 | //low |
| 425 | in_ptr += in_x_stride; |
| 426 | data.val[0] = vld1q_s32(in_ptr); |
| 427 | out.val[0] = vmlaq_s32(out.val[0], data.val[0], fifteen); |
| 428 | |
| 429 | data.val[1] = vld1q_s32(in_ptr + 4); |
| 430 | out.val[1] = vmlaq_s32(out.val[1], data.val[1], fifteen); |
| 431 | |
| 432 | //low2 |
| 433 | in_ptr += in_x_stride; |
| 434 | data.val[0] = vld1q_s32(in_ptr); |
| 435 | out.val[0] = vmlaq_s32(out.val[0], data.val[0], six); |
| 436 | |
| 437 | data.val[1] = vld1q_s32(in_ptr + 4); |
| 438 | out.val[1] = vmlaq_s32(out.val[1], data.val[1], six); |
| 439 | |
| 440 | //low3 |
| 441 | in_ptr += in_x_stride; |
| 442 | data.val[0] = vld1q_s32(in_ptr); |
| 443 | out.val[0] = vaddq_s32(out.val[0], data.val[0]); |
| 444 | |
| 445 | data.val[1] = vld1q_s32(in_ptr + 4); |
| 446 | out.val[1] = vaddq_s32(out.val[1], data.val[1]); |
| 447 | |
| 448 | vst1q_s32(reinterpret_cast<int32_t *>(output_x.ptr()) + 0, out.val[0]); |
| 449 | vst1q_s32(reinterpret_cast<int32_t *>(output_x.ptr()) + 4, out.val[1]); |
| 450 | }, |
| 451 | input_x, output_x); |
| 452 | } |
| 453 | |
| 454 | if(_run_sobel_y) |
| 455 | { |
| 456 | execute_window_loop(window, [&](const Coordinates & id) |
| 457 | { |
| 458 | auto in_ptr = reinterpret_cast<int32_t *>(input_y.ptr()) - 3 * in_y_stride; |
| 459 | |
| 460 | //top3 |
| 461 | int32x4x2_t data = |
| 462 | { |
| 463 | { |
| 464 | vld1q_s32(in_ptr), |
| 465 | vld1q_s32(in_ptr + 4) |
| 466 | } |
| 467 | }; |
| 468 | |
| 469 | int32x4x2_t out = |
| 470 | { |
| 471 | { |
| 472 | vnegq_s32(data.val[0]), |
| 473 | vnegq_s32(data.val[1]) |
| 474 | } |
| 475 | }; |
| 476 | |
| 477 | //top2 |
| 478 | in_ptr += in_y_stride; |
| 479 | data.val[0] = vld1q_s32(in_ptr); |
| 480 | out.val[0] = vmlaq_s32(out.val[0], data.val[0], minusfour); |
| 481 | |
| 482 | data.val[1] = vld1q_s32(in_ptr + 4); |
| 483 | out.val[1] = vmlaq_s32(out.val[1], data.val[1], minusfour); |
| 484 | |
| 485 | //top |
| 486 | in_ptr += in_y_stride; |
| 487 | data.val[0] = vld1q_s32(in_ptr); |
| 488 | out.val[0] = vmlaq_s32(out.val[0], data.val[0], minusfive); |
| 489 | |
| 490 | data.val[1] = vld1q_s32(in_ptr + 4); |
| 491 | out.val[1] = vmlaq_s32(out.val[1], data.val[1], minusfive); |
| 492 | |
| 493 | //low |
| 494 | in_ptr += (2 * in_y_stride); |
| 495 | data.val[0] = vld1q_s32(in_ptr); |
| 496 | out.val[0] = vmlaq_s32(out.val[0], data.val[0], five); |
| 497 | |
| 498 | data.val[1] = vld1q_s32(in_ptr + 4); |
| 499 | out.val[1] = vmlaq_s32(out.val[1], data.val[1], five); |
| 500 | |
| 501 | //low2 |
| 502 | in_ptr += in_y_stride; |
| 503 | data.val[0] = vld1q_s32(in_ptr); |
| 504 | out.val[0] = vmlaq_s32(out.val[0], data.val[0], four); |
| 505 | |
| 506 | data.val[1] = vld1q_s32(in_ptr + 4); |
| 507 | out.val[1] = vmlaq_s32(out.val[1], data.val[1], four); |
| 508 | |
| 509 | //low3 |
| 510 | in_ptr += in_y_stride; |
| 511 | data.val[0] = vld1q_s32(in_ptr); |
| 512 | out.val[0] = vaddq_s32(out.val[0], data.val[0]); |
| 513 | |
| 514 | data.val[1] = vld1q_s32(in_ptr + 4); |
| 515 | out.val[1] = vaddq_s32(out.val[1], data.val[1]); |
| 516 | |
| 517 | vst1q_s32(reinterpret_cast<int32_t *>(output_y.ptr()) + 0, out.val[0]); |
| 518 | vst1q_s32(reinterpret_cast<int32_t *>(output_y.ptr()) + 4, out.val[1]); |
| 519 | }, |
| 520 | input_y, output_y); |
| 521 | } |
| 522 | } |