Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2016-2019 Arm Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +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/core/NEON/kernels/NECannyEdgeKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/Error.h" |
| 28 | #include "arm_compute/core/Helpers.h" |
| 29 | #include "arm_compute/core/ITensor.h" |
| 30 | #include "arm_compute/core/TensorInfo.h" |
| 31 | #include "arm_compute/core/Types.h" |
| 32 | #include "arm_compute/core/Utils.h" |
| 33 | #include "arm_compute/core/Validate.h" |
| 34 | |
| 35 | #include <arm_neon.h> |
| 36 | #include <cstddef> |
| 37 | #include <cstdint> |
| 38 | #include <tuple> |
| 39 | |
| 40 | using namespace arm_compute; |
| 41 | |
| 42 | namespace arm_compute |
| 43 | { |
| 44 | class Coordinates; |
| 45 | } // namespace arm_compute |
| 46 | |
| 47 | namespace |
| 48 | { |
| 49 | constexpr int NO_EDGE = 0; |
| 50 | constexpr int EDGE = 255; |
| 51 | constexpr int MAYBE = 127; |
| 52 | } // namespace |
| 53 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 54 | namespace |
| 55 | { |
| 56 | inline uint8x8_t phase_quantization(const float32x4x2_t &gx, const float32x4x2_t &gy) |
| 57 | { |
| 58 | // Constant use for evaluating score1 and score3 |
| 59 | static const float32x4_t const45 = vdupq_n_f32(0.70710678118655f); |
| 60 | static const float32x4_t zero = vdupq_n_f32(0.0f); |
| 61 | static const float32x4_t one = vdupq_n_f32(1.0f); |
| 62 | static const float32x4_t two = vdupq_n_f32(2.0f); |
| 63 | static const float32x4_t three = vdupq_n_f32(3.0f); |
| 64 | |
| 65 | // Score0: (1, 0) |
| 66 | const float32x4x2_t score0 = |
| 67 | { |
| 68 | { |
| 69 | vabsq_f32(gx.val[0]), |
| 70 | vabsq_f32(gx.val[1]) |
| 71 | } |
| 72 | }; |
| 73 | |
| 74 | // Score2: ( 0, 1 ) |
| 75 | const float32x4x2_t score2 = |
| 76 | { |
| 77 | { |
| 78 | vabsq_f32(gy.val[0]), |
| 79 | vabsq_f32(gy.val[1]) |
| 80 | } |
| 81 | }; |
| 82 | |
| 83 | // Score1 and Score3: ( sqrt(2) / 2, sqrt(2) / 2 ) - ( -sqrt(2) / 2, sqrt(2) / 2 ) |
| 84 | float32x4x2_t score1 = |
| 85 | { |
| 86 | { |
| 87 | vmulq_f32(gy.val[0], const45), |
| 88 | vmulq_f32(gy.val[1], const45) |
| 89 | } |
| 90 | }; |
| 91 | |
| 92 | float32x4x2_t score3 = score1; |
| 93 | |
| 94 | score1.val[0] = vmlaq_f32(score1.val[0], gx.val[0], const45); |
| 95 | score1.val[1] = vmlaq_f32(score1.val[1], gx.val[1], const45); |
| 96 | score3.val[0] = vmlsq_f32(score3.val[0], gx.val[0], const45); |
| 97 | score3.val[1] = vmlsq_f32(score3.val[1], gx.val[1], const45); |
| 98 | |
| 99 | score1.val[0] = vabsq_f32(score1.val[0]); |
| 100 | score1.val[1] = vabsq_f32(score1.val[1]); |
| 101 | score3.val[0] = vabsq_f32(score3.val[0]); |
| 102 | score3.val[1] = vabsq_f32(score3.val[1]); |
| 103 | |
| 104 | float32x4x2_t phase = |
| 105 | { |
| 106 | { |
| 107 | zero, |
| 108 | zero |
| 109 | } |
| 110 | }; |
| 111 | |
| 112 | float32x4x2_t old_score = score0; |
| 113 | |
| 114 | // score1 > old_score? |
| 115 | uint32x4x2_t mask = |
| 116 | { |
| 117 | { |
| 118 | vcgtq_f32(score1.val[0], old_score.val[0]), |
| 119 | vcgtq_f32(score1.val[1], old_score.val[1]) |
| 120 | } |
| 121 | }; |
| 122 | |
| 123 | phase.val[0] = vbslq_f32(mask.val[0], one, phase.val[0]); |
| 124 | phase.val[1] = vbslq_f32(mask.val[1], one, phase.val[1]); |
| 125 | old_score.val[0] = vbslq_f32(mask.val[0], score1.val[0], old_score.val[0]); |
| 126 | old_score.val[1] = vbslq_f32(mask.val[1], score1.val[1], old_score.val[1]); |
| 127 | |
| 128 | // score2 > old_score? |
| 129 | mask.val[0] = vcgtq_f32(score2.val[0], old_score.val[0]); |
| 130 | mask.val[1] = vcgtq_f32(score2.val[1], old_score.val[1]); |
| 131 | |
| 132 | phase.val[0] = vbslq_f32(mask.val[0], two, phase.val[0]); |
| 133 | phase.val[1] = vbslq_f32(mask.val[1], two, phase.val[1]); |
| 134 | old_score.val[0] = vbslq_f32(mask.val[0], score2.val[0], old_score.val[0]); |
| 135 | old_score.val[1] = vbslq_f32(mask.val[1], score2.val[1], old_score.val[1]); |
| 136 | |
| 137 | // score3 > old_score? |
| 138 | mask.val[0] = vcgtq_f32(score3.val[0], old_score.val[0]); |
| 139 | mask.val[1] = vcgtq_f32(score3.val[1], old_score.val[1]); |
| 140 | |
| 141 | phase.val[0] = vbslq_f32(mask.val[0], three, phase.val[0]); |
| 142 | phase.val[1] = vbslq_f32(mask.val[1], three, phase.val[1]); |
| 143 | old_score.val[0] = vbslq_f32(mask.val[0], score3.val[0], old_score.val[0]); |
| 144 | old_score.val[1] = vbslq_f32(mask.val[1], score3.val[1], old_score.val[1]); |
| 145 | |
| 146 | // Convert from float32x4_t to uint8x8_t |
| 147 | return vmovn_u16(vcombine_u16(vmovn_u32(vcvtq_u32_f32(phase.val[0])), |
| 148 | vmovn_u32(vcvtq_u32_f32(phase.val[1])))); |
| 149 | } |
| 150 | |
| 151 | /* Computes the gradient phase if gradient_size = 3 or 5. The output is quantized. |
| 152 | * 0 = 0°, 1 = 45°, 2 = 90°, 3 = 135° |
| 153 | * |
| 154 | * @param[in] gx Gx component |
| 155 | * @param[in] gy Gy component |
| 156 | * |
| 157 | * @return quantized phase for 8 pixels |
| 158 | */ |
| 159 | inline uint8x8_t phase_quantization_S16_S16(int16x8_t gx, int16x8_t gy) |
| 160 | { |
| 161 | // Convert to float |
| 162 | const float32x4x2_t gx_f32 = |
| 163 | { |
| 164 | { |
| 165 | vcvtq_f32_s32(vmovl_s16(vget_low_s16(gx))), |
| 166 | vcvtq_f32_s32(vmovl_s16(vget_high_s16(gx))) |
| 167 | } |
| 168 | }; |
| 169 | |
| 170 | const float32x4x2_t gy_f32 = |
| 171 | { |
| 172 | { |
| 173 | vcvtq_f32_s32(vmovl_s16(vget_low_s16(gy))), |
| 174 | vcvtq_f32_s32(vmovl_s16(vget_high_s16(gy))) |
| 175 | } |
| 176 | }; |
| 177 | |
| 178 | return phase_quantization(gx_f32, gy_f32); |
| 179 | } |
| 180 | |
| 181 | /* Computes the gradient phase if gradient_size = 7. The output is quantized. |
| 182 | * 0 = 0°, 1 = 45°, 2 = 90°, 3 = 135° |
| 183 | * |
| 184 | * @param[in] gx Gx component |
| 185 | * @param[in] gy Gy component |
| 186 | * |
| 187 | * @return quantized phase for 8 pixels |
| 188 | */ |
| 189 | inline uint8x8_t phase_quantization_S32_S32(const int32x4x2_t &gx, const int32x4x2_t &gy) |
| 190 | { |
| 191 | // Convert to float |
| 192 | const float32x4x2_t gx_f32 = |
| 193 | { |
| 194 | { |
| 195 | vcvtq_f32_s32(gx.val[0]), |
| 196 | vcvtq_f32_s32(gx.val[1]) |
| 197 | } |
| 198 | }; |
| 199 | |
| 200 | const float32x4x2_t gy_f32 = |
| 201 | { |
| 202 | { |
| 203 | vcvtq_f32_s32(gy.val[0]), |
| 204 | vcvtq_f32_s32(gy.val[1]) |
| 205 | } |
| 206 | }; |
| 207 | |
| 208 | return phase_quantization(gx_f32, gy_f32); |
| 209 | } |
| 210 | |
| 211 | /* Computes the magnitude using the L1-norm type if gradient_size = 3 or 5 |
| 212 | * |
| 213 | * @param[in] gx Gx component |
| 214 | * @param[in] gy Gy component |
| 215 | * |
| 216 | * @return magnitude for 8 pixels |
| 217 | */ |
| 218 | inline uint16x8_t mag_l1_S16_S16(int16x8_t gx, int16x8_t gy) |
| 219 | { |
| 220 | return vaddq_u16(vreinterpretq_u16_s16(vabsq_s16(gx)), |
| 221 | vreinterpretq_u16_s16(vabsq_s16(gy))); |
| 222 | } |
| 223 | |
| 224 | /* Computes the magnitude using the L1-norm type if gradient_size = 7 |
| 225 | * |
| 226 | * @param[in] gx Gx component |
| 227 | * @param[in] gy Gy component |
| 228 | * |
| 229 | * @return magnitude for 8 pixels |
| 230 | */ |
| 231 | inline uint32x4x2_t mag_l1_S32_S32(const int32x4x2_t &gx, const int32x4x2_t &gy) |
| 232 | { |
| 233 | const uint32x4x2_t gx_abs = |
| 234 | { |
| 235 | { |
| 236 | vreinterpretq_u32_s32(vabsq_s32(gx.val[0])), |
| 237 | vreinterpretq_u32_s32(vabsq_s32(gx.val[1])) |
| 238 | } |
| 239 | }; |
| 240 | |
| 241 | const uint32x4x2_t gy_abs = |
| 242 | { |
| 243 | { |
| 244 | vreinterpretq_u32_s32(vabsq_s32(gy.val[0])), |
| 245 | vreinterpretq_u32_s32(vabsq_s32(gy.val[1])) |
| 246 | } |
| 247 | }; |
| 248 | |
| 249 | const uint32x4x2_t output = |
| 250 | { |
| 251 | { |
| 252 | vaddq_u32(gx_abs.val[0], gy_abs.val[0]), |
| 253 | vaddq_u32(gx_abs.val[1], gy_abs.val[1]) |
| 254 | } |
| 255 | }; |
| 256 | |
| 257 | return output; |
| 258 | } |
| 259 | |
| 260 | inline float32x4x2_t mag_l2(const float32x4x2_t &gx, const float32x4x2_t &gy) |
| 261 | { |
| 262 | // x^2 ... |
| 263 | float32x4x2_t magnitude = |
| 264 | { |
| 265 | { |
| 266 | vmulq_f32(gx.val[0], gx.val[0]), |
| 267 | vmulq_f32(gx.val[1], gx.val[1]) |
| 268 | } |
| 269 | }; |
| 270 | |
| 271 | // ... + y^2 |
| 272 | magnitude.val[0] = vmlaq_f32(magnitude.val[0], gy.val[0], gy.val[0]); |
| 273 | magnitude.val[1] = vmlaq_f32(magnitude.val[1], gy.val[1], gy.val[1]); |
| 274 | |
| 275 | // sqrt(...) |
| 276 | magnitude.val[0] = vmulq_f32(vrsqrteq_f32(magnitude.val[0]), magnitude.val[0]); |
| 277 | magnitude.val[1] = vmulq_f32(vrsqrteq_f32(magnitude.val[1]), magnitude.val[1]); |
| 278 | |
| 279 | return magnitude; |
| 280 | } |
| 281 | |
| 282 | /* Computes the magnitude using L2-norm if gradient_size = 3 or 5 |
| 283 | * |
| 284 | * @param[in] gx Gx component |
| 285 | * @param[in] gy Gy component |
| 286 | * |
| 287 | * @return magnitude for 8 pixels |
| 288 | */ |
| 289 | inline uint16x8_t mag_l2_S16_S16(int16x8_t gx, int16x8_t gy) |
| 290 | { |
| 291 | // Compute magnitude using L2 normalization |
| 292 | const float32x4x2_t gx2 = |
| 293 | { |
| 294 | { |
| 295 | vcvtq_f32_s32(vmovl_s16(vget_low_s16(gx))), |
| 296 | vcvtq_f32_s32(vmovl_s16(vget_high_s16(gx))) |
| 297 | } |
| 298 | }; |
| 299 | |
| 300 | const float32x4x2_t gy2 = |
| 301 | { |
| 302 | { |
| 303 | vcvtq_f32_s32(vmovl_s16(vget_low_s16(gy))), |
| 304 | vcvtq_f32_s32(vmovl_s16(vget_high_s16(gy))) |
| 305 | } |
| 306 | }; |
| 307 | |
| 308 | const float32x4x2_t magnitude = mag_l2(gx2, gy2); |
| 309 | |
| 310 | // Store magnitude - Convert to uint16x8 |
| 311 | return vcombine_u16(vmovn_u32(vcvtq_u32_f32(magnitude.val[0])), |
| 312 | vmovn_u32(vcvtq_u32_f32(magnitude.val[1]))); |
| 313 | } |
| 314 | |
| 315 | /* Computes the magnitude using L2-norm if gradient_size = 7 |
| 316 | * |
| 317 | * @param[in] gx Gx component |
| 318 | * @param[in] gy Gy component |
| 319 | * |
| 320 | * @return magnitude for 8 pixels |
| 321 | */ |
| 322 | inline uint32x4x2_t mag_l2_S32_S32(const int32x4x2_t &gx, const int32x4x2_t &gy) |
| 323 | { |
| 324 | // Compute magnitude using L2 normalization |
| 325 | float32x4x2_t gx2 = |
| 326 | { |
| 327 | { |
| 328 | vcvtq_f32_s32(gx.val[0]), |
| 329 | vcvtq_f32_s32(gx.val[1]) |
| 330 | } |
| 331 | }; |
| 332 | |
| 333 | float32x4x2_t gy2 = |
| 334 | { |
| 335 | { |
| 336 | vcvtq_f32_s32(gy.val[0]), |
| 337 | vcvtq_f32_s32(gy.val[1]) |
| 338 | } |
| 339 | }; |
| 340 | |
| 341 | const float32x4x2_t magnitude = mag_l2(gx2, gy2); |
| 342 | const uint32x4x2_t mag32 = |
| 343 | { |
| 344 | { |
| 345 | vcvtq_u32_f32(magnitude.val[0]), |
| 346 | vcvtq_u32_f32(magnitude.val[1]) |
| 347 | } |
| 348 | }; |
| 349 | |
| 350 | return mag32; |
| 351 | } |
| 352 | |
| 353 | /* Gradient function used when the gradient size = 3 or 5 and when the norm_type = L1-norm |
| 354 | * |
| 355 | * @param[in] gx_ptr Pointer to source image. Gx image. Data type supported S16 |
| 356 | * @param[in] gy_ptr Pointer to source image. Gy image. Data type supported S16 |
| 357 | * @param[out] magnitude_ptr Pointer to destination image. Magnitude. Data type supported U16 |
| 358 | * @param[out] phase_ptr Pointer to destination image. Quantized phase. Data type supported U8 |
| 359 | */ |
| 360 | void mag_phase_l1norm_S16_S16_U16_U8(const void *__restrict gx_ptr, const void *__restrict gy_ptr, void *__restrict magnitude_ptr, void *__restrict phase_ptr) |
| 361 | { |
| 362 | const auto gx = static_cast<const int16_t *__restrict>(gx_ptr); |
| 363 | const auto gy = static_cast<const int16_t *__restrict>(gy_ptr); |
| 364 | const auto magnitude = static_cast<uint16_t *__restrict>(magnitude_ptr); |
| 365 | const auto phase = static_cast<uint8_t *__restrict>(phase_ptr); |
| 366 | |
| 367 | const int16x8x4_t gx_val = |
| 368 | { |
| 369 | { |
| 370 | vld1q_s16(gx), |
| 371 | vld1q_s16(gx + 8), |
| 372 | vld1q_s16(gx + 16), |
| 373 | vld1q_s16(gx + 24) |
| 374 | } |
| 375 | }; |
| 376 | |
| 377 | const int16x8x4_t gy_val = |
| 378 | { |
| 379 | { |
| 380 | vld1q_s16(gy), |
| 381 | vld1q_s16(gy + 8), |
| 382 | vld1q_s16(gy + 16), |
| 383 | vld1q_s16(gy + 24) |
| 384 | } |
| 385 | }; |
| 386 | |
| 387 | // Compute and store phase |
| 388 | vst1_u8(phase + 0, phase_quantization_S16_S16(gx_val.val[0], gy_val.val[0])); |
| 389 | vst1_u8(phase + 8, phase_quantization_S16_S16(gx_val.val[1], gy_val.val[1])); |
| 390 | vst1_u8(phase + 16, phase_quantization_S16_S16(gx_val.val[2], gy_val.val[2])); |
| 391 | vst1_u8(phase + 24, phase_quantization_S16_S16(gx_val.val[3], gy_val.val[3])); |
| 392 | |
| 393 | // Compute ans store magnitude using L1 normalization |
| 394 | vst1q_u16(magnitude + 0, mag_l1_S16_S16(gx_val.val[0], gy_val.val[0])); |
| 395 | vst1q_u16(magnitude + 8, mag_l1_S16_S16(gx_val.val[1], gy_val.val[1])); |
| 396 | vst1q_u16(magnitude + 16, mag_l1_S16_S16(gx_val.val[2], gy_val.val[2])); |
| 397 | vst1q_u16(magnitude + 24, mag_l1_S16_S16(gx_val.val[3], gy_val.val[3])); |
| 398 | } |
| 399 | |
| 400 | /* Gradient function used when the gradient size = 3 or 5 and when the norm_type = L2-norm |
| 401 | * |
| 402 | * @param[in] gx_ptr Pointer to source image. Gx image. Data type supported S16 |
| 403 | * @param[in] gy_ptr Pointer to source image. Gy image. Data type supported S16 |
| 404 | * @param[out] magnitude_ptr Pointer to destination image. Magnitude. Data type supported U16 |
| 405 | * @param[out] phase_ptr Pointer to destination image. Quantized phase. Data type supported U8 |
| 406 | */ |
| 407 | void mag_phase_l2norm_S16_S16_U16_U8(const void *__restrict gx_ptr, const void *__restrict gy_ptr, void *__restrict magnitude_ptr, void *__restrict phase_ptr) |
| 408 | { |
| 409 | const auto gx = static_cast<const int16_t *__restrict>(gx_ptr); |
| 410 | const auto gy = static_cast<const int16_t *__restrict>(gy_ptr); |
| 411 | const auto magnitude = static_cast<uint16_t *__restrict>(magnitude_ptr); |
| 412 | const auto phase = static_cast<uint8_t *__restrict>(phase_ptr); |
| 413 | |
| 414 | const int16x8x4_t gx_val = |
| 415 | { |
| 416 | { |
| 417 | vld1q_s16(gx), |
| 418 | vld1q_s16(gx + 8), |
| 419 | vld1q_s16(gx + 16), |
| 420 | vld1q_s16(gx + 24) |
| 421 | } |
| 422 | }; |
| 423 | |
| 424 | const int16x8x4_t gy_val = |
| 425 | { |
| 426 | { |
| 427 | vld1q_s16(gy), |
| 428 | vld1q_s16(gy + 8), |
| 429 | vld1q_s16(gy + 16), |
| 430 | vld1q_s16(gy + 24) |
| 431 | } |
| 432 | }; |
| 433 | |
| 434 | // Compute and store phase |
| 435 | vst1_u8(phase + 0, phase_quantization_S16_S16(gx_val.val[0], gy_val.val[0])); |
| 436 | vst1_u8(phase + 8, phase_quantization_S16_S16(gx_val.val[1], gy_val.val[1])); |
| 437 | vst1_u8(phase + 16, phase_quantization_S16_S16(gx_val.val[2], gy_val.val[2])); |
| 438 | vst1_u8(phase + 24, phase_quantization_S16_S16(gx_val.val[3], gy_val.val[3])); |
| 439 | |
| 440 | // Compute and store magnitude using L2 normalization |
| 441 | vst1q_u16(magnitude + 0, mag_l2_S16_S16(gx_val.val[0], gy_val.val[0])); |
| 442 | vst1q_u16(magnitude + 8, mag_l2_S16_S16(gx_val.val[1], gy_val.val[1])); |
| 443 | vst1q_u16(magnitude + 16, mag_l2_S16_S16(gx_val.val[2], gy_val.val[2])); |
| 444 | vst1q_u16(magnitude + 24, mag_l2_S16_S16(gx_val.val[3], gy_val.val[3])); |
| 445 | } |
| 446 | |
| 447 | /* Gradient function used when the gradient size = 7 and when the norm_type = L1-norm |
| 448 | * |
| 449 | * @param[in] gx_ptr Pointer to source image. Gx image. Data type supported S32 |
| 450 | * @param[in] gy_ptr Pointer to source image. Gy image. Data type supported S32 |
| 451 | * @param[out] magnitude_ptr Pointer to destination image. Magnitude. Data type supported U32 |
| 452 | * @param[out] phase_ptr Pointer to destination image. Quantized phase. Data type support U8 |
| 453 | */ |
| 454 | void mag_phase_l1norm_S32_S32_U32_U8(const void *__restrict gx_ptr, const void *__restrict gy_ptr, void *__restrict magnitude_ptr, void *__restrict phase_ptr) |
| 455 | { |
| 456 | auto gx = static_cast<const int32_t *__restrict>(gx_ptr); |
| 457 | auto gy = static_cast<const int32_t *__restrict>(gy_ptr); |
| 458 | auto magnitude = static_cast<uint32_t *__restrict>(magnitude_ptr); |
| 459 | auto phase = static_cast<uint8_t *__restrict>(phase_ptr); |
| 460 | |
| 461 | // Process low and high part |
| 462 | for(size_t i = 0; i < 2; ++i, gx += 16, gy += 16, magnitude += 16, phase += 16) |
| 463 | { |
| 464 | const int32x4x2_t gx0 = |
| 465 | { |
| 466 | { |
| 467 | vld1q_s32(gx + 0), |
| 468 | vld1q_s32(gx + 4) |
| 469 | } |
| 470 | }; |
| 471 | |
| 472 | const int32x4x2_t gx1 = |
| 473 | { |
| 474 | { |
| 475 | vld1q_s32(gx + 8), |
| 476 | vld1q_s32(gx + 12) |
| 477 | } |
| 478 | }; |
| 479 | |
| 480 | const int32x4x2_t gy0 = |
| 481 | { |
| 482 | { |
| 483 | vld1q_s32(gy + 0), |
| 484 | vld1q_s32(gy + 4) |
| 485 | } |
| 486 | }; |
| 487 | |
| 488 | const int32x4x2_t gy1 = |
| 489 | { |
| 490 | { |
| 491 | vld1q_s32(gy + 8), |
| 492 | vld1q_s32(gy + 12) |
| 493 | } |
| 494 | }; |
| 495 | |
| 496 | // Compute and store phase |
| 497 | vst1_u8(phase + 0, phase_quantization_S32_S32(gx0, gy0)); |
| 498 | vst1_u8(phase + 8, phase_quantization_S32_S32(gx1, gy1)); |
| 499 | |
| 500 | // Compute magnitude using L1 normalization |
| 501 | const uint32x4x2_t mag0 = mag_l1_S32_S32(gx0, gy0); |
| 502 | const uint32x4x2_t mag1 = mag_l1_S32_S32(gx1, gy1); |
| 503 | |
| 504 | // Store magnitude |
| 505 | vst1q_u32(magnitude + 0, mag0.val[0]); |
| 506 | vst1q_u32(magnitude + 4, mag0.val[1]); |
| 507 | vst1q_u32(magnitude + 8, mag1.val[0]); |
| 508 | vst1q_u32(magnitude + 12, mag1.val[1]); |
| 509 | } |
| 510 | } |
| 511 | |
| 512 | /* Gradient function used when the gradient size = 7 and when the norm_type = L2-norm |
| 513 | * |
| 514 | * @param[in] gx_ptr Pointer to source image. Gx image. Data type supported S32 |
| 515 | * @param[in] gy_ptr Pointer to source image. Gy image. Data type supported S32 |
| 516 | * @param[out] magnitude_ptr Pointer to destination image. Magnitude. Data type supported U32 |
| 517 | * @param[out] phase_ptr Pointer to destination image. Quantized phase. Data type supported U8 |
| 518 | */ |
| 519 | void mag_phase_l2norm_S32_S32_U32_U8(const void *__restrict gx_ptr, const void *__restrict gy_ptr, void *__restrict magnitude_ptr, void *__restrict phase_ptr) |
| 520 | { |
| 521 | auto gx = static_cast<const int32_t *__restrict>(gx_ptr); |
| 522 | auto gy = static_cast<const int32_t *__restrict>(gy_ptr); |
| 523 | auto magnitude = static_cast<uint32_t *__restrict>(magnitude_ptr); |
| 524 | auto phase = static_cast<uint8_t *__restrict>(phase_ptr); |
| 525 | |
| 526 | // Process low and high part |
| 527 | for(size_t i = 0; i < 2; ++i, gx += 16, gy += 16, magnitude += 16, phase += 16) |
| 528 | { |
| 529 | const int32x4x2_t gx0 = |
| 530 | { |
| 531 | { |
| 532 | vld1q_s32(gx + 0), |
| 533 | vld1q_s32(gx + 4) |
| 534 | } |
| 535 | }; |
| 536 | |
| 537 | const int32x4x2_t gx1 = |
| 538 | { |
| 539 | { |
| 540 | vld1q_s32(gx + 8), |
| 541 | vld1q_s32(gx + 12) |
| 542 | } |
| 543 | }; |
| 544 | |
| 545 | const int32x4x2_t gy0 = |
| 546 | { |
| 547 | { |
| 548 | vld1q_s32(gy + 0), |
| 549 | vld1q_s32(gy + 4) |
| 550 | } |
| 551 | }; |
| 552 | |
| 553 | const int32x4x2_t gy1 = |
| 554 | { |
| 555 | { |
| 556 | vld1q_s32(gy + 8), |
| 557 | vld1q_s32(gy + 12) |
| 558 | } |
| 559 | }; |
| 560 | |
| 561 | // Compute and store phase |
| 562 | vst1_u8(phase + 0, phase_quantization_S32_S32(gx0, gy0)); |
| 563 | vst1_u8(phase + 8, phase_quantization_S32_S32(gx1, gy1)); |
| 564 | |
| 565 | // Compute magnitude using L2 normalization |
| 566 | const uint32x4x2_t mag0 = mag_l2_S32_S32(gx0, gy0); |
| 567 | const uint32x4x2_t mag1 = mag_l2_S32_S32(gx1, gy1); |
| 568 | |
| 569 | // Store magnitude |
| 570 | vst1q_u32(magnitude + 0, mag0.val[0]); |
| 571 | vst1q_u32(magnitude + 4, mag0.val[1]); |
| 572 | vst1q_u32(magnitude + 8, mag1.val[0]); |
| 573 | vst1q_u32(magnitude + 12, mag1.val[1]); |
| 574 | } |
| 575 | } |
| 576 | |
| 577 | /* Computes non-maxima suppression and hysteresis when the gradient size = 3 or 5 |
| 578 | * |
| 579 | * @param[in] magnitude_ptr Pointer to source image. Magnitude. Data type supported U16 |
| 580 | * @param[in] phase_ptr Pointer to source image. Quantized phase. Data type supported U8 |
| 581 | * @param[out] output_ptr Pointer to output image. Data type supported U8 |
| 582 | * @param[in] stride_mag Stride of magnitude image |
| 583 | * @param[in] lower_thr Lower threshold used for the hysteresis |
| 584 | * @param[in] upper_thr Upper threshold used for the hysteresis |
| 585 | */ |
| 586 | void non_max_suppression_U16_U8_U8(const void *__restrict magnitude_ptr, const void *__restrict phase_ptr, void *__restrict output_ptr, const uint32_t stride_mag, const int32_t lower_thr, |
| 587 | const int32_t upper_thr) |
| 588 | { |
| 589 | const auto magnitude = static_cast<const uint16_t *__restrict>(magnitude_ptr); |
| 590 | const auto phase = static_cast<const uint8_t *__restrict>(phase_ptr); |
| 591 | const auto output = static_cast<uint8_t *__restrict>(output_ptr); |
| 592 | |
| 593 | // Get magnitude and phase of the centre pixels |
| 594 | uint16x8_t mc = vld1q_u16(magnitude); |
| 595 | |
| 596 | // Angle_quantized: 0 = 0°, 1 = 45°, 2 = 90°, 3 = 135° |
| 597 | const uint16x8_t pc16 = vmovl_u8(vld1_u8(phase)); |
| 598 | |
| 599 | // 0 degree |
| 600 | const uint16x8_t mk0_0 = vld1q_u16(magnitude - 1); |
| 601 | const uint16x8_t mk0_1 = vld1q_u16(magnitude + 1); |
| 602 | uint16x8_t mask0 = vceqq_u16(pc16, vdupq_n_u16(0)); |
Abe Mbise | 1b99338 | 2017-12-19 13:51:59 +0000 | [diff] [blame] | 603 | mask0 = vandq_u16(mask0, vcgtq_u16(mc, mk0_0)); |
| 604 | mask0 = vandq_u16(mask0, vcgtq_u16(mc, mk0_1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 605 | |
| 606 | // 45 degree |
| 607 | const uint16x8_t mk45_0 = vld1q_u16(magnitude - stride_mag - 1); |
| 608 | const uint16x8_t mk45_1 = vld1q_u16(magnitude + stride_mag + 1); |
| 609 | uint16x8_t mask1 = vceqq_u16(pc16, vdupq_n_u16(1)); |
Abe Mbise | 1b99338 | 2017-12-19 13:51:59 +0000 | [diff] [blame] | 610 | mask1 = vandq_u16(mask1, vcgtq_u16(mc, mk45_0)); |
| 611 | mask1 = vandq_u16(mask1, vcgtq_u16(mc, mk45_1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 612 | |
| 613 | // 90 degree |
| 614 | const uint16x8_t mk90_0 = vld1q_u16(magnitude - stride_mag); |
| 615 | const uint16x8_t mk90_1 = vld1q_u16(magnitude + stride_mag); |
| 616 | uint16x8_t mask2 = vceqq_u16(pc16, vdupq_n_u16(2)); |
Abe Mbise | 1b99338 | 2017-12-19 13:51:59 +0000 | [diff] [blame] | 617 | mask2 = vandq_u16(mask2, vcgtq_u16(mc, mk90_0)); |
| 618 | mask2 = vandq_u16(mask2, vcgtq_u16(mc, mk90_1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 619 | |
| 620 | // 135 degree |
| 621 | const uint16x8_t mk135_0 = vld1q_u16(magnitude - stride_mag + 1); |
| 622 | const uint16x8_t mk135_1 = vld1q_u16(magnitude + stride_mag - 1); |
| 623 | uint16x8_t mask3 = vceqq_u16(pc16, vdupq_n_u16(3)); |
Abe Mbise | 1b99338 | 2017-12-19 13:51:59 +0000 | [diff] [blame] | 624 | mask3 = vandq_u16(mask3, vcgtq_u16(mc, mk135_0)); |
| 625 | mask3 = vandq_u16(mask3, vcgtq_u16(mc, mk135_1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 626 | |
| 627 | // Merge masks |
| 628 | mask0 = vorrq_u16(mask0, mask1); |
| 629 | mask2 = vorrq_u16(mask2, mask3); |
| 630 | mask0 = vorrq_u16(mask0, mask2); |
| 631 | |
| 632 | mc = vbslq_u16(mask0, mc, vdupq_n_u16(0)); |
| 633 | |
| 634 | // mc > upper_thr |
| 635 | mask0 = vcgtq_u16(mc, vdupq_n_u16(upper_thr)); |
| 636 | |
| 637 | // mc <= lower_thr |
| 638 | mask1 = vcleq_u16(mc, vdupq_n_u16(lower_thr)); |
| 639 | |
| 640 | // mc <= upper_thr && mc > lower_thr |
| 641 | mask2 = vcleq_u16(mc, vdupq_n_u16(upper_thr)); |
| 642 | mask2 = vandq_u16(mask2, vcgtq_u16(mc, vdupq_n_u16(lower_thr))); |
| 643 | |
| 644 | mc = vbslq_u16(mask0, vdupq_n_u16(EDGE), mc); |
| 645 | mc = vbslq_u16(mask1, vdupq_n_u16(NO_EDGE), mc); |
| 646 | mc = vbslq_u16(mask2, vdupq_n_u16(MAYBE), mc); |
| 647 | |
| 648 | vst1_u8(output, vmovn_u16(mc)); |
| 649 | } |
| 650 | |
| 651 | inline uint16x4_t non_max_U32_helper(const uint32_t *input, const uint16x4_t pc, const uint32_t stride_mag, const int32_t lower_thr, const int32_t upper_thr) |
| 652 | { |
| 653 | // Phase for 4 pixel |
| 654 | const uint32x4_t pc32 = vmovl_u16(pc); |
| 655 | |
| 656 | // Get magnitude for 4 pixel |
| 657 | uint32x4_t mc = vld1q_u32(input); |
| 658 | |
| 659 | // Angle_quantized: 0 = 0°, 1 = 45°, 2 = 90°, 3 = 135° |
| 660 | // 0 degree |
| 661 | const uint32x4_t mk0_0 = vld1q_u32(input - 1); |
| 662 | const uint32x4_t mk0_1 = vld1q_u32(input + 1); |
| 663 | uint32x4_t mask0 = vceqq_u32(pc32, vdupq_n_u32(0)); |
Abe Mbise | 1b99338 | 2017-12-19 13:51:59 +0000 | [diff] [blame] | 664 | mask0 = vandq_u32(mask0, vcgtq_u32(mc, mk0_0)); |
| 665 | mask0 = vandq_u32(mask0, vcgtq_u32(mc, mk0_1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 666 | |
| 667 | // 45 degree |
| 668 | const uint32x4_t mk45_0 = vld1q_u32(input - stride_mag - 1); |
| 669 | const uint32x4_t mk45_1 = vld1q_u32(input + stride_mag + 1); |
| 670 | uint32x4_t mask1 = vceqq_u32(pc32, vdupq_n_u32(1)); |
Abe Mbise | 1b99338 | 2017-12-19 13:51:59 +0000 | [diff] [blame] | 671 | mask1 = vandq_u32(mask1, vcgtq_u32(mc, mk45_0)); |
| 672 | mask1 = vandq_u32(mask1, vcgtq_u32(mc, mk45_1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 673 | |
| 674 | // 90 degree |
| 675 | const uint32x4_t mk90_0 = vld1q_u32(input - stride_mag); |
| 676 | const uint32x4_t mk90_1 = vld1q_u32(input + stride_mag); |
| 677 | uint32x4_t mask2 = vceqq_u32(pc32, vdupq_n_u32(2)); |
Abe Mbise | 1b99338 | 2017-12-19 13:51:59 +0000 | [diff] [blame] | 678 | mask2 = vandq_u32(mask2, vcgtq_u32(mc, mk90_0)); |
| 679 | mask2 = vandq_u32(mask2, vcgtq_u32(mc, mk90_1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 680 | |
| 681 | // 135 degree |
| 682 | const uint32x4_t mk135_0 = vld1q_u32(input - stride_mag + 1); |
| 683 | const uint32x4_t mk135_1 = vld1q_u32(input + stride_mag - 1); |
| 684 | uint32x4_t mask3 = vceqq_u32(pc32, vdupq_n_u32(3)); |
Abe Mbise | 1b99338 | 2017-12-19 13:51:59 +0000 | [diff] [blame] | 685 | mask3 = vandq_u32(mask3, vcgtq_u32(mc, mk135_0)); |
| 686 | mask3 = vandq_u32(mask3, vcgtq_u32(mc, mk135_1)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 687 | |
| 688 | // Merge masks |
| 689 | mask0 = vorrq_u32(mask0, mask1); |
| 690 | mask2 = vorrq_u32(mask2, mask3); |
| 691 | mask0 = vorrq_u32(mask0, mask2); |
| 692 | |
| 693 | mc = vbslq_u32(mask0, mc, vdupq_n_u32(0)); |
| 694 | |
| 695 | // mc > upper_thr |
| 696 | mask0 = vcgtq_u32(mc, vdupq_n_u32(upper_thr)); |
| 697 | |
| 698 | // mc <= lower_thr |
| 699 | mask1 = vcleq_u32(mc, vdupq_n_u32(lower_thr)); |
| 700 | |
| 701 | // mc <= upper_thr && mc > lower_thr |
| 702 | mask2 = vcleq_u32(mc, vdupq_n_u32(upper_thr)); |
| 703 | mask2 = vandq_u32(mask2, vcgtq_u32(mc, vdupq_n_u32(lower_thr))); |
| 704 | |
| 705 | mc = vbslq_u32(mask0, vdupq_n_u32(EDGE), mc); |
| 706 | mc = vbslq_u32(mask1, vdupq_n_u32(NO_EDGE), mc); |
| 707 | mc = vbslq_u32(mask2, vdupq_n_u32(MAYBE), mc); |
| 708 | |
| 709 | return vmovn_u32(mc); |
| 710 | } |
| 711 | |
| 712 | /* Computes non-maxima suppression and hysteresis when the gradient_size = 7 |
| 713 | * |
| 714 | * @param[in] magnitude_ptr Pointer to source image. Magnitude. Data type supported U32 |
| 715 | * @param[in] phase_ptr Pointer to source image. Quantized phase. Data type supported U8 |
| 716 | * @param[out] output_ptr Pointer to destination image. Data type supported U8 |
| 717 | * @param[in] stride_mag Stride of magnitude image |
| 718 | * @param[in] lower_thr Lower threshold used for the hysteresis |
| 719 | * @param[in] upper_thr Upper threshold used for the hysteresis |
| 720 | */ |
| 721 | void non_max_suppression_U32_U8_U8(const void *__restrict magnitude_ptr, const void *__restrict phase_ptr, void *__restrict output_ptr, const uint32_t stride_mag, const int32_t lower_thr, |
| 722 | const int32_t upper_thr) |
| 723 | { |
| 724 | const auto magnitude = static_cast<const uint32_t *__restrict>(magnitude_ptr); |
| 725 | const auto phase = static_cast<const uint8_t *__restrict>(phase_ptr); |
| 726 | const auto output = static_cast<uint8_t *__restrict>(output_ptr); |
| 727 | |
| 728 | // Get phase for 8 pixel |
| 729 | const uint16x8_t pc16 = vmovl_u8(vld1_u8(phase)); |
| 730 | |
| 731 | // Compute non maxima suppression |
| 732 | const uint16x4x2_t res = |
| 733 | { |
| 734 | { |
| 735 | non_max_U32_helper(magnitude, vget_low_u16(pc16), stride_mag, lower_thr, upper_thr), |
| 736 | non_max_U32_helper(magnitude + 4, vget_high_u16(pc16), stride_mag, lower_thr, upper_thr) |
| 737 | } |
| 738 | }; |
| 739 | |
| 740 | // Store result |
| 741 | vst1_u8(output, vmovn_u16(vcombine_u16(res.val[0], res.val[1]))); |
| 742 | } |
| 743 | |
| 744 | /* Computes edge tracing when is called by edge_trace_U8_U8 recursively |
| 745 | * |
| 746 | * @param[in] input Pointer to source image. Data type supported U8 |
| 747 | * @param[out] output Pointer to destination image. Data type supported U8 |
| 748 | * @param[in] input_stride Stride of the input image |
| 749 | * @param[in] output_stride Stride of the output image |
| 750 | */ |
| 751 | void edge_trace_recursive_U8_U8(uint8_t *__restrict input, uint8_t *__restrict output, const int32_t input_stride, const int32_t output_stride) |
| 752 | { |
| 753 | // Look for MAYBE pixels in 8 directions |
| 754 | *output = EDGE; |
| 755 | |
| 756 | // (-1, 0) |
| 757 | uint8_t pixel = *(input - 1); |
| 758 | |
| 759 | if(pixel == MAYBE) |
| 760 | { |
| 761 | // Touched a MAYBE point. MAYBE becomes EDGE |
| 762 | *(input - 1) = EDGE; |
| 763 | |
| 764 | edge_trace_recursive_U8_U8(input - 1, output - 1, input_stride, output_stride); |
| 765 | } |
| 766 | |
| 767 | // (+1, 0) |
| 768 | pixel = *(input + 1); |
| 769 | |
| 770 | if(pixel == MAYBE) |
| 771 | { |
| 772 | // Touched a MAYBE point. MAYBE becomes EDGE |
| 773 | *(input + 1) = EDGE; |
| 774 | |
| 775 | edge_trace_recursive_U8_U8(input + 1, output + 1, input_stride, output_stride); |
| 776 | } |
| 777 | |
| 778 | input -= input_stride; |
| 779 | output -= output_stride; |
| 780 | |
| 781 | // (-1, -1) |
| 782 | pixel = *(input - 1); |
| 783 | |
| 784 | if(pixel == MAYBE) |
| 785 | { |
| 786 | // Touched a MAYBE point. MAYBE becomes EDGE |
| 787 | *(input - 1) = EDGE; |
| 788 | |
| 789 | edge_trace_recursive_U8_U8(input - 1, output - 1, input_stride, output_stride); |
| 790 | } |
| 791 | |
| 792 | // (0, -1) |
| 793 | pixel = *input; |
| 794 | |
| 795 | if(pixel == MAYBE) |
| 796 | { |
| 797 | // Touched a MAYBE point. MAYBE becomes EDGE |
| 798 | *input = EDGE; |
| 799 | |
| 800 | edge_trace_recursive_U8_U8(input, output, input_stride, output_stride); |
| 801 | } |
| 802 | |
| 803 | // (+1, -1) |
| 804 | pixel = *(input + 1); |
| 805 | |
| 806 | if(pixel == MAYBE) |
| 807 | { |
| 808 | // Touched a MAYBE point. MAYBE becomes EDGE |
| 809 | *(input + 1) = EDGE; |
| 810 | |
| 811 | edge_trace_recursive_U8_U8(input + 1, output + 1, input_stride, output_stride); |
| 812 | } |
| 813 | |
| 814 | input += input_stride * 2; |
| 815 | output += output_stride * 2; |
| 816 | |
| 817 | // (-1, +1) |
| 818 | pixel = *(input - 1); |
| 819 | |
| 820 | if(pixel == MAYBE) |
| 821 | { |
| 822 | // Touched a MAYBE point. MAYBE becomes EDGE |
| 823 | *(input - 1) = EDGE; |
| 824 | |
| 825 | edge_trace_recursive_U8_U8(input - 1, output - 1, input_stride, output_stride); |
| 826 | } |
| 827 | |
| 828 | // (0, +1) |
| 829 | pixel = *input; |
| 830 | |
| 831 | if(pixel == MAYBE) |
| 832 | { |
| 833 | // Touched a MAYBE point. MAYBE becomes EDGE |
| 834 | *input = EDGE; |
| 835 | |
| 836 | edge_trace_recursive_U8_U8(input, output, input_stride, output_stride); |
| 837 | } |
| 838 | |
| 839 | // (+1, +1) |
| 840 | pixel = *(input + 1); |
| 841 | |
| 842 | if(pixel == MAYBE) |
| 843 | { |
| 844 | // Touched a MAYBE point. MAYBE becomes EDGE |
| 845 | *(input + 1) = EDGE; |
| 846 | |
| 847 | edge_trace_recursive_U8_U8(input + 1, output + 1, input_stride, output_stride); |
| 848 | } |
| 849 | } |
| 850 | |
| 851 | /* Computes edge tracing |
| 852 | * |
| 853 | * @param[in] input Pointer to source image. Data type supported U8 |
| 854 | * @param[out] output Pointer to destination image. Data type supported U8 |
| 855 | * @param[in] input_stride Stride of the input image |
| 856 | * @param[in] output_stride Stride of the output image |
| 857 | */ |
| 858 | void edge_trace_U8_U8(uint8_t *__restrict input, uint8_t *__restrict output, const int32_t input_stride, const int32_t output_stride) |
| 859 | { |
| 860 | if(*input == NO_EDGE) |
| 861 | { |
| 862 | *output = NO_EDGE; |
| 863 | } |
| 864 | // Check if EDGE and not yet touched |
| 865 | else if((*input == EDGE) && (*output == NO_EDGE)) |
| 866 | { |
| 867 | edge_trace_recursive_U8_U8(input, output, input_stride, output_stride); |
| 868 | } |
| 869 | } |
| 870 | } // namespace |
| 871 | |
| 872 | NEGradientKernel::NEGradientKernel() |
| 873 | : _func(nullptr), _gx(nullptr), _gy(nullptr), _magnitude(nullptr), _phase(nullptr) |
| 874 | { |
| 875 | } |
| 876 | |
| 877 | void NEGradientKernel::configure(const ITensor *gx, const ITensor *gy, ITensor *magnitude, ITensor *phase, int32_t norm_type) |
| 878 | { |
| 879 | ARM_COMPUTE_ERROR_ON_NULLPTR(gx, gy, magnitude, phase); |
| 880 | |
| 881 | set_shape_if_empty(*magnitude->info(), gx->info()->tensor_shape()); |
| 882 | set_shape_if_empty(*phase->info(), gx->info()->tensor_shape()); |
| 883 | |
| 884 | Format magnitude_format = gx->info()->data_type() == DataType::S16 ? Format::U16 : Format::U32; |
| 885 | set_format_if_unknown(*magnitude->info(), magnitude_format); |
| 886 | set_format_if_unknown(*phase->info(), Format::U8); |
| 887 | |
| 888 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(gx, gy, magnitude, phase); |
| 889 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(gx, 1, DataType::S16, DataType::S32); |
| 890 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(gy, 1, DataType::S16, DataType::S32); |
| 891 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(magnitude, 1, DataType::U16, DataType::U32); |
| 892 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(phase, 1, DataType::U8); |
| 893 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(gx, gy); |
| 894 | ARM_COMPUTE_ERROR_ON_MSG(element_size_from_data_type(gx->info()->data_type()) != element_size_from_data_type(magnitude->info()->data_type()), "Magnitude must have the same element size as Gx and Gy"); |
| 895 | |
| 896 | _gx = gx; |
| 897 | _gy = gy; |
| 898 | _magnitude = magnitude; |
| 899 | _phase = phase; |
| 900 | |
| 901 | if(_gx->info()->data_type() == DataType::S16) |
| 902 | { |
| 903 | if(norm_type == 1) |
| 904 | { |
| 905 | _func = &mag_phase_l1norm_S16_S16_U16_U8; |
| 906 | } |
| 907 | else |
| 908 | { |
| 909 | _func = &mag_phase_l2norm_S16_S16_U16_U8; |
| 910 | } |
| 911 | } |
| 912 | else |
| 913 | { |
| 914 | if(norm_type == 1) |
| 915 | { |
| 916 | _func = &mag_phase_l1norm_S32_S32_U32_U8; |
| 917 | } |
| 918 | else |
| 919 | { |
| 920 | _func = &mag_phase_l2norm_S32_S32_U32_U8; |
| 921 | } |
| 922 | } |
| 923 | |
| 924 | constexpr unsigned int num_elems_processed_per_iteration = 32; |
| 925 | |
| 926 | // Configure kernel window |
| 927 | Window win = calculate_max_window(*_gx->info(), Steps(num_elems_processed_per_iteration)); |
| 928 | |
| 929 | AccessWindowHorizontal gx_access(_gx->info(), 0, num_elems_processed_per_iteration); |
| 930 | AccessWindowHorizontal gy_access(_gy->info(), 0, num_elems_processed_per_iteration); |
| 931 | AccessWindowHorizontal mag_access(_magnitude->info(), 0, num_elems_processed_per_iteration); |
| 932 | AccessWindowHorizontal phase_access(_phase->info(), 0, num_elems_processed_per_iteration); |
| 933 | |
Michalis Spyrou | ebdde65 | 2019-07-08 11:52:46 +0100 | [diff] [blame] | 934 | ARM_COMPUTE_UNUSED(update_window_and_padding(win, gx_access, gy_access, mag_access, phase_access)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 935 | |
| 936 | mag_access.set_valid_region(win, _gx->info()->valid_region()); |
| 937 | phase_access.set_valid_region(win, _gx->info()->valid_region()); |
| 938 | |
| 939 | INEKernel::configure(win); |
| 940 | } |
| 941 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 942 | void NEGradientKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 943 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 944 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 945 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 946 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 947 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 948 | Iterator gx(_gx, window); |
| 949 | Iterator gy(_gy, window); |
| 950 | Iterator magnitude(_magnitude, window); |
| 951 | Iterator phase(_phase, window); |
| 952 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 953 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 954 | { |
| 955 | (*_func)(gx.ptr(), gy.ptr(), magnitude.ptr(), phase.ptr()); |
| 956 | }, |
| 957 | gx, gy, magnitude, phase); |
| 958 | } |
| 959 | |
| 960 | NEEdgeNonMaxSuppressionKernel::NEEdgeNonMaxSuppressionKernel() |
| 961 | : _func(nullptr), _magnitude(nullptr), _phase(nullptr), _output(nullptr), _lower_thr(0), _upper_thr(0) |
| 962 | { |
| 963 | } |
| 964 | |
| 965 | BorderSize NEEdgeNonMaxSuppressionKernel::border_size() const |
| 966 | { |
| 967 | return BorderSize(1); |
| 968 | } |
| 969 | |
| 970 | void NEEdgeNonMaxSuppressionKernel::configure(const ITensor *magnitude, const ITensor *phase, ITensor *output, |
| 971 | int32_t upper_thr, int32_t lower_thr, bool border_undefined) |
| 972 | { |
| 973 | ARM_COMPUTE_ERROR_ON_NULLPTR(magnitude, phase, output); |
| 974 | |
| 975 | set_shape_if_empty(*output->info(), magnitude->info()->tensor_shape()); |
| 976 | |
| 977 | set_format_if_unknown(*phase->info(), Format::U8); |
| 978 | set_format_if_unknown(*output->info(), Format::U8); |
| 979 | |
| 980 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(magnitude, phase, output); |
| 981 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(magnitude, 1, DataType::U16, DataType::U32); |
| 982 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(phase, 1, DataType::U8); |
| 983 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8); |
| 984 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(phase, output); |
| 985 | |
| 986 | _magnitude = magnitude; |
| 987 | _phase = phase; |
| 988 | _output = output; |
| 989 | |
| 990 | switch(_magnitude->info()->data_type()) |
| 991 | { |
| 992 | case DataType::U16: |
| 993 | _func = &non_max_suppression_U16_U8_U8; |
| 994 | break; |
| 995 | case DataType::U32: |
| 996 | _func = &non_max_suppression_U32_U8_U8; |
| 997 | break; |
| 998 | default: |
| 999 | ARM_COMPUTE_ERROR("Unsupported data type!"); |
| 1000 | } |
| 1001 | |
| 1002 | // Set thresholds |
| 1003 | _lower_thr = lower_thr; |
| 1004 | _upper_thr = upper_thr; |
| 1005 | |
| 1006 | constexpr unsigned int num_elems_processed_per_iteration = 8; |
| 1007 | constexpr unsigned int num_elems_read_per_iteration = 10; |
| 1008 | constexpr unsigned int num_rows_read_per_iteration = 3; |
| 1009 | |
| 1010 | // Configure kernel window |
| 1011 | Window win = calculate_max_window(*_magnitude->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); |
| 1012 | |
| 1013 | AccessWindowRectangle mag_access(_magnitude->info(), -border_size().left, -border_size().top, num_elems_read_per_iteration, num_rows_read_per_iteration); |
| 1014 | AccessWindowHorizontal phase_access(_phase->info(), 0, num_elems_processed_per_iteration); |
| 1015 | AccessWindowHorizontal output_access(_output->info(), 0, num_elems_processed_per_iteration); |
| 1016 | |
| 1017 | update_window_and_padding(win, mag_access, phase_access, output_access); |
| 1018 | |
| 1019 | output_access.set_valid_region(win, _magnitude->info()->valid_region(), border_undefined, border_size()); |
| 1020 | |
| 1021 | INEKernel::configure(win); |
| 1022 | } |
| 1023 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1024 | void NEEdgeNonMaxSuppressionKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1025 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1026 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1027 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 1028 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 1029 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 1030 | Iterator magnitude(_magnitude, window); |
| 1031 | Iterator phase(_phase, window); |
| 1032 | Iterator output(_output, window); |
| 1033 | |
| 1034 | const size_t input1_stride = _magnitude->info()->strides_in_bytes()[1]; |
| 1035 | const size_t input1_stride_ushort = input1_stride / data_size_from_type(_magnitude->info()->data_type()); |
| 1036 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 1037 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1038 | { |
| 1039 | (*_func)(magnitude.ptr(), phase.ptr(), output.ptr(), input1_stride_ushort, _lower_thr, _upper_thr); |
| 1040 | }, |
| 1041 | magnitude, phase, output); |
| 1042 | } |
| 1043 | |
| 1044 | NEEdgeTraceKernel::NEEdgeTraceKernel() |
| 1045 | : _input(nullptr), _output(nullptr) |
| 1046 | { |
| 1047 | } |
| 1048 | |
| 1049 | BorderSize NEEdgeTraceKernel::border_size() const |
| 1050 | { |
| 1051 | return BorderSize(1); |
| 1052 | } |
| 1053 | |
| 1054 | bool NEEdgeTraceKernel::is_parallelisable() const |
| 1055 | { |
| 1056 | return false; |
| 1057 | } |
| 1058 | |
| 1059 | void NEEdgeTraceKernel::configure(ITensor *input, ITensor *output) |
| 1060 | { |
| 1061 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 1062 | |
| 1063 | set_shape_if_empty(*output->info(), input->info()->tensor_shape()); |
| 1064 | |
| 1065 | set_format_if_unknown(*input->info(), Format::U8); |
| 1066 | set_format_if_unknown(*output->info(), Format::U8); |
| 1067 | |
| 1068 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); |
| 1069 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); |
| 1070 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8); |
| 1071 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 1072 | |
| 1073 | _input = input; |
| 1074 | _output = output; |
| 1075 | |
| 1076 | constexpr unsigned int num_elems_processed_per_iteration = 1; |
| 1077 | |
| 1078 | // Configure kernel window |
| 1079 | Window win = calculate_max_window(*_input->info(), Steps(num_elems_processed_per_iteration)); |
| 1080 | |
| 1081 | const ValidRegion &input_valid_region = input->info()->valid_region(); |
| 1082 | const ValidRegion &output_valid_region = output->info()->valid_region(); |
| 1083 | |
| 1084 | // Reads can occur within the valid region of the input + border |
| 1085 | AccessWindowStatic input_access(input->info(), |
| 1086 | input_valid_region.anchor[0] - border_size().left, |
| 1087 | input_valid_region.anchor[1] - border_size().top, |
| 1088 | input_valid_region.anchor[0] + input_valid_region.shape[0] + border_size().right, |
| 1089 | input_valid_region.anchor[1] + input_valid_region.shape[1] + border_size().bottom); |
| 1090 | |
| 1091 | // Writes can occur within the valid region of the output + border |
| 1092 | AccessWindowStatic output_access(output->info(), |
| 1093 | output_valid_region.anchor[0] - border_size().left, |
| 1094 | output_valid_region.anchor[1] - border_size().top, |
| 1095 | output_valid_region.anchor[0] + output_valid_region.shape[0] + border_size().right, |
| 1096 | output_valid_region.anchor[1] + output_valid_region.shape[1] + border_size().bottom); |
| 1097 | |
| 1098 | update_window_and_padding(win, input_access, output_access); |
| 1099 | |
| 1100 | output_access.set_valid_region(win, _input->info()->valid_region()); |
| 1101 | |
| 1102 | INEKernel::configure(win); |
| 1103 | } |
| 1104 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1105 | void NEEdgeTraceKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1106 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1107 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1108 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 1109 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 1110 | Iterator input(_input, window); |
| 1111 | Iterator output(_output, window); |
| 1112 | |
| 1113 | const size_t input_stride = _input->info()->strides_in_bytes()[1]; |
| 1114 | const size_t output_stride = _output->info()->strides_in_bytes()[1]; |
| 1115 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 1116 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1117 | { |
| 1118 | edge_trace_U8_U8(input.ptr(), output.ptr(), input_stride, output_stride); |
| 1119 | }, |
| 1120 | input, output); |
| 1121 | } |