Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Gian Marco Iodice | fbf3ecc | 2018-08-23 17:26:21 +0100 | [diff] [blame] | 2 | * Copyright (c) 2016-2018 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/NEMagnitudePhaseKernel.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/ITensor.h" |
| 30 | #include "arm_compute/core/Validate.h" |
| 31 | |
| 32 | #include <arm_neon.h> |
| 33 | #include <cstdint> |
| 34 | |
| 35 | using namespace arm_compute; |
| 36 | |
| 37 | namespace arm_compute |
| 38 | { |
| 39 | class Coordinates; |
| 40 | } // namespace arm_compute |
| 41 | |
| 42 | namespace |
| 43 | { |
| 44 | // Defines for computing atan2 |
| 45 | constexpr float SCALE_FACTOR = 0.7111111111111111f; |
| 46 | constexpr float PI = 3.141592653589793f; |
| 47 | constexpr float SCALE_180 = 180.0f / PI; |
| 48 | constexpr float SCALE_360 = SCALE_180 * SCALE_FACTOR; |
| 49 | constexpr float PI_4 = 0.7853981633974483f; |
| 50 | constexpr float COEFF1 = 0.0663f; |
| 51 | constexpr float COEFF2 = 0.2447f; |
| 52 | } // namespace |
| 53 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 54 | namespace |
| 55 | { |
| 56 | inline float32x4_t inv(float32x4_t x) |
| 57 | { |
| 58 | float32x4_t result = vrecpeq_f32(x); |
| 59 | result = vmulq_f32(vrecpsq_f32(x, result), result); |
| 60 | return result; |
| 61 | } |
| 62 | |
| 63 | inline float32x4_t atan2_0_360(float32x4_t gx, float32x4_t gy) |
| 64 | { |
| 65 | const float32x4_t zero = vdupq_n_f32(0.0f); |
| 66 | const float32x4_t epsilon = vdupq_n_f32(1e-9f); |
| 67 | const float32x4_t piover4 = vdupq_n_f32(PI_4); |
| 68 | const float32x4_t coeff1 = vdupq_n_f32(COEFF1); |
| 69 | const float32x4_t coeff2 = vdupq_n_f32(COEFF2); |
| 70 | const float32x4_t ninety = vdupq_n_f32(90.0f * SCALE_FACTOR); |
| 71 | const float32x4_t oneeighty = vdupq_n_f32(180.0f * SCALE_FACTOR); |
| 72 | const float32x4_t threesixty = vdupq_n_f32(360.0f * SCALE_FACTOR); |
| 73 | const float32x4_t scale = vdupq_n_f32(SCALE_360); |
| 74 | |
| 75 | float32x4_t abs_gx = vabsq_f32(gx); |
| 76 | float32x4_t abs_gy = vabsq_f32(gy); |
| 77 | float32x4_t tmin = vminq_f32(abs_gx, abs_gy); |
| 78 | float32x4_t tmax = vmaxq_f32(abs_gx, abs_gy); |
| 79 | float32x4_t z = vmulq_f32(tmin, inv(vaddq_f32(tmax, epsilon))); |
| 80 | float32x4_t absz = vabsq_f32(z); |
| 81 | float32x4_t term = vmulq_f32(z, vsubq_f32(vdupq_n_f32(1.0f), absz)); |
| 82 | |
| 83 | /* Compute y = pi/4 * x - x*(abs(x)-1)*(0.2447+0.0663 * abs(x) */ |
| 84 | float32x4_t result = vaddq_f32(coeff2, vmulq_f32(absz, coeff1)); |
| 85 | result = vmulq_f32(result, term); |
| 86 | result = vmlaq_f32(result, piover4, z); |
| 87 | |
| 88 | /* Radians to degrees conversion with applied a scale factor in order to have the result [0, 255] */ |
| 89 | result = vmulq_f32(result, scale); |
| 90 | |
| 91 | /* If z > 1, result = 90 - result */ |
| 92 | result = vbslq_f32(vcgeq_f32(abs_gx, abs_gy), result, vsubq_f32(ninety, result)); |
| 93 | |
| 94 | /* Choose correct quadrant */ |
| 95 | result = vbslq_f32(vcltq_f32(gx, zero), vsubq_f32(oneeighty, result), result); |
| 96 | result = vbslq_f32(vcltq_f32(gy, zero), vsubq_f32(threesixty, result), result); |
| 97 | |
| 98 | return result; |
| 99 | } |
| 100 | |
| 101 | inline float32x4_t atan2_0_180(float32x4_t gx, float32x4_t gy) |
| 102 | { |
| 103 | const float32x4_t zero = vdupq_n_f32(0.0f); |
| 104 | const float32x4_t epsilon = vdupq_n_f32(1e-9f); // epsilon used to avoiding division by 0 |
| 105 | const float32x4_t piover4 = vdupq_n_f32(PI_4); |
| 106 | const float32x4_t coeff1 = vdupq_n_f32(COEFF1); |
| 107 | const float32x4_t coeff2 = vdupq_n_f32(COEFF2); |
| 108 | const float32x4_t ninety = vdupq_n_f32(90.0f); |
| 109 | const float32x4_t oneeighty = vdupq_n_f32(180.0f); |
| 110 | const float32x4_t threesixty = vdupq_n_f32(360.0f); |
| 111 | const float32x4_t scale = vdupq_n_f32(SCALE_180); |
| 112 | |
| 113 | float32x4_t abs_gx = vabsq_f32(gx); |
| 114 | float32x4_t abs_gy = vabsq_f32(gy); |
| 115 | float32x4_t tmin = vminq_f32(abs_gx, abs_gy); |
| 116 | float32x4_t tmax = vmaxq_f32(abs_gx, abs_gy); |
| 117 | float32x4_t z = vmulq_f32(tmin, inv(vaddq_f32(tmax, epsilon))); |
| 118 | float32x4_t absz = vabsq_f32(z); |
| 119 | |
| 120 | /* Compute y = pi/4 * z - z*(abs(z)-1)*(0.2447+0.0663 * abs(z) */ |
| 121 | float32x4_t term = vmulq_f32(z, vsubq_f32(vdupq_n_f32(1.0f), absz)); |
| 122 | float32x4_t result = vaddq_f32(coeff2, vmulq_f32(absz, coeff1)); |
| 123 | result = vmulq_f32(result, term); |
| 124 | result = vmlaq_f32(result, piover4, z); |
| 125 | |
| 126 | /* Radians to degrees conversion */ |
| 127 | result = vmulq_f32(result, scale); |
| 128 | |
| 129 | /* If z > 1, result = 90 - result */ |
| 130 | result = vbslq_f32(vcgeq_f32(abs_gx, abs_gy), result, vsubq_f32(ninety, result)); |
| 131 | |
| 132 | /* Choose correct quadrant */ |
| 133 | result = vbslq_f32(vcltq_f32(gx, zero), vsubq_f32(oneeighty, result), result); |
| 134 | result = vbslq_f32(vcltq_f32(gy, zero), vsubq_f32(threesixty, result), result); |
| 135 | result = vbslq_f32(vcgtq_f32(result, oneeighty), vsubq_f32(result, oneeighty), result); |
| 136 | |
| 137 | return result; |
| 138 | } |
| 139 | |
| 140 | inline float32x4_t invsqrtv(float32x4_t x) |
| 141 | { |
| 142 | float32x4_t sqrt_reciprocal = vrsqrteq_f32(x); |
| 143 | |
| 144 | sqrt_reciprocal = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), |
| 145 | sqrt_reciprocal); |
| 146 | sqrt_reciprocal = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), |
| 147 | sqrt_reciprocal); |
| 148 | |
| 149 | return sqrt_reciprocal; |
| 150 | } |
| 151 | |
| 152 | inline float32x4_t sqrtv(float32x4_t x) |
| 153 | { |
| 154 | float32x4_t res = vdupq_n_f32(0.5f); |
| 155 | return vmlaq_f32(res, x, invsqrtv(x)); |
| 156 | } |
| 157 | |
| 158 | inline int16x8_t magnitude_l2(int16x8_t input1, int16x8_t input2) |
| 159 | { |
| 160 | const int32x4x2_t square_x = |
| 161 | { |
| 162 | { |
| 163 | vmull_s16(vget_low_s16(input1), vget_low_s16(input1)), |
| 164 | vmull_s16(vget_high_s16(input1), vget_high_s16(input1)) |
| 165 | } |
| 166 | }; |
| 167 | |
| 168 | const int32x4x2_t square_y = |
| 169 | { |
| 170 | { |
| 171 | vmull_s16(vget_low_s16(input2), vget_low_s16(input2)), |
| 172 | vmull_s16(vget_high_s16(input2), vget_high_s16(input2)) |
| 173 | } |
| 174 | }; |
| 175 | |
| 176 | const uint32x4x2_t sum = |
| 177 | { |
| 178 | { |
| 179 | vaddq_u32(vreinterpretq_u32_s32(square_x.val[0]), vreinterpretq_u32_s32(square_y.val[0])), |
| 180 | vaddq_u32(vreinterpretq_u32_s32(square_x.val[1]), vreinterpretq_u32_s32(square_y.val[1])) |
| 181 | } |
| 182 | }; |
| 183 | |
| 184 | const float32x4x2_t res = |
| 185 | { |
| 186 | { |
| 187 | sqrtv(vcvtq_f32_u32(sum.val[0])), |
| 188 | sqrtv(vcvtq_f32_u32(sum.val[1])) |
| 189 | } |
| 190 | }; |
| 191 | |
| 192 | return vcombine_s16(vqmovn_s32(vcvtq_s32_f32(res.val[0])), |
| 193 | vqmovn_s32(vcvtq_s32_f32(res.val[1]))); |
| 194 | } |
| 195 | |
| 196 | inline int16x8_t magnitude_l1(int16x8_t input1, int16x8_t input2) |
| 197 | { |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 198 | /* Saturating add */ |
John Richardson | 3c5f949 | 2017-10-04 15:27:37 +0100 | [diff] [blame] | 199 | return vqaddq_s16(vqabsq_s16(input1), vqabsq_s16(input2)); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 200 | } |
| 201 | |
| 202 | inline uint8x8_t phase_signed(int16x8_t input1, int16x8_t input2) |
| 203 | { |
| 204 | const float32x4_t zeropointfive = vdupq_n_f32(0.5f); |
| 205 | |
| 206 | float32x4_t inputx_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(input1))); |
| 207 | float32x4_t inputx_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(input1))); |
| 208 | float32x4_t inputy_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(input2))); |
| 209 | float32x4_t inputy_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(input2))); |
| 210 | |
| 211 | /* Compute fast atan2 */ |
| 212 | float32x4_t angle_high = atan2_0_360(inputx_f32_high, inputy_f32_high); |
| 213 | float32x4_t angle_low = atan2_0_360(inputx_f32_low, inputy_f32_low); |
| 214 | |
| 215 | angle_high = vaddq_f32(angle_high, zeropointfive); |
| 216 | angle_low = vaddq_f32(angle_low, zeropointfive); |
| 217 | |
| 218 | return vmovn_u16(vcombine_u16(vqmovun_s32(vcvtq_s32_f32(angle_low)), |
| 219 | vqmovun_s32(vcvtq_s32_f32(angle_high)))); |
| 220 | } |
| 221 | |
| 222 | inline uint8x8_t phase_unsigned(int16x8_t input1, int16x8_t input2) |
| 223 | { |
| 224 | const float32x4_t zeropointfive = vdupq_n_f32(0.5f); |
| 225 | |
| 226 | float32x4_t inputx_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(input1))); |
| 227 | float32x4_t inputx_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(input1))); |
| 228 | float32x4_t inputy_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(input2))); |
| 229 | float32x4_t inputy_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(input2))); |
| 230 | |
| 231 | /* Compute fast atan2 */ |
| 232 | float32x4_t angle_high = atan2_0_180(inputx_f32_high, inputy_f32_high); |
| 233 | float32x4_t angle_low = atan2_0_180(inputx_f32_low, inputy_f32_low); |
| 234 | |
| 235 | angle_high = vaddq_f32(angle_high, zeropointfive); |
| 236 | angle_low = vaddq_f32(angle_low, zeropointfive); |
| 237 | |
| 238 | return vmovn_u16(vcombine_u16(vqmovun_s32(vcvtq_s32_f32(angle_low)), |
| 239 | vqmovun_s32(vcvtq_s32_f32(angle_high)))); |
| 240 | } |
| 241 | } // namespace |
| 242 | |
| 243 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 244 | NEMagnitudePhaseKernel<mag_type, phase_type>::NEMagnitudePhaseKernel() |
| 245 | : _func(nullptr), _gx(nullptr), _gy(nullptr), _magnitude(nullptr), _phase(nullptr) |
| 246 | { |
| 247 | } |
| 248 | |
| 249 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 250 | void NEMagnitudePhaseKernel<mag_type, phase_type>::configure(const ITensor *gx, const ITensor *gy, ITensor *magnitude, ITensor *phase) |
| 251 | { |
| 252 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(gx, 1, DataType::S16); |
| 253 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(gy, 1, DataType::S16); |
| 254 | ARM_COMPUTE_ERROR_ON((nullptr == magnitude) && (nullptr == phase)); |
| 255 | |
| 256 | const bool run_mag = magnitude != nullptr; |
| 257 | const bool run_phase = phase != nullptr; |
| 258 | |
| 259 | if(run_mag) |
| 260 | { |
| 261 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(magnitude, 1, DataType::S16); |
| 262 | } |
| 263 | |
| 264 | if(run_phase) |
| 265 | { |
| 266 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(phase, 1, DataType::U8); |
| 267 | } |
| 268 | |
| 269 | _gx = gx; |
| 270 | _gy = gy; |
| 271 | _magnitude = magnitude; |
| 272 | _phase = phase; |
| 273 | |
| 274 | if(run_mag && run_phase) |
| 275 | { |
| 276 | /* Run magnitude and phase */ |
| 277 | _func = &NEMagnitudePhaseKernel<mag_type, phase_type>::magnitude_phase; |
| 278 | } |
| 279 | else |
| 280 | { |
| 281 | if(run_mag) |
| 282 | { |
| 283 | /* Run magnitude */ |
| 284 | _func = &NEMagnitudePhaseKernel<mag_type, phase_type>::magnitude; |
| 285 | } |
| 286 | else if(run_phase) |
| 287 | { |
| 288 | /* Run phase */ |
| 289 | _func = &NEMagnitudePhaseKernel<mag_type, phase_type>::phase; |
| 290 | } |
| 291 | else |
| 292 | { |
| 293 | ARM_COMPUTE_ERROR("At least one output must be NOT NULL"); |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 298 | |
| 299 | // Configure kernel window |
| 300 | Window win = calculate_max_window(*gx->info(), Steps(num_elems_processed_per_iteration)); |
| 301 | AccessWindowHorizontal magnitude_access(magnitude == nullptr ? nullptr : magnitude->info(), 0, num_elems_processed_per_iteration); |
| 302 | AccessWindowHorizontal phase_access(phase == nullptr ? nullptr : phase->info(), 0, num_elems_processed_per_iteration); |
| 303 | |
| 304 | update_window_and_padding(win, |
| 305 | AccessWindowHorizontal(gx->info(), 0, num_elems_processed_per_iteration), |
| 306 | AccessWindowHorizontal(gy->info(), 0, num_elems_processed_per_iteration), |
| 307 | magnitude_access, |
| 308 | phase_access); |
| 309 | |
| 310 | ValidRegion valid_region = intersect_valid_regions(gx->info()->valid_region(), |
| 311 | gy->info()->valid_region()); |
| 312 | |
| 313 | magnitude_access.set_valid_region(win, valid_region); |
| 314 | phase_access.set_valid_region(win, valid_region); |
| 315 | |
| 316 | INEKernel::configure(win); |
| 317 | } |
| 318 | |
| 319 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 320 | void NEMagnitudePhaseKernel<mag_type, phase_type>::magnitude(const Window &window) |
| 321 | { |
| 322 | Iterator gx(_gx, window); |
| 323 | Iterator gy(_gy, window); |
| 324 | Iterator magnitude(_magnitude, window); |
| 325 | |
| 326 | execute_window_loop(window, [&](const Coordinates & id) |
| 327 | { |
| 328 | const int16x8x2_t input1 = |
| 329 | { |
| 330 | { |
| 331 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr())), |
| 332 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr()) + 8) |
| 333 | } |
| 334 | }; |
| 335 | |
| 336 | const int16x8x2_t input2 = |
| 337 | { |
| 338 | { |
| 339 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr())), |
| 340 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr()) + 8) |
| 341 | } |
| 342 | }; |
| 343 | |
| 344 | /* Compute magnitude */ |
| 345 | int16x8x2_t mag{ {} }; |
| 346 | |
| 347 | if(MagnitudeType::L2NORM == mag_type) |
| 348 | { |
| 349 | mag.val[0] = magnitude_l2(input1.val[0], input2.val[0]); |
| 350 | mag.val[1] = magnitude_l2(input1.val[1], input2.val[1]); |
| 351 | } |
| 352 | else |
| 353 | { |
| 354 | mag.val[0] = magnitude_l1(input1.val[0], input2.val[0]); |
| 355 | mag.val[1] = magnitude_l1(input1.val[1], input2.val[1]); |
| 356 | } |
| 357 | |
| 358 | /* Store magnitude */ |
| 359 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()), mag.val[0]); |
| 360 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()) + 8, mag.val[1]); |
| 361 | }, |
| 362 | gx, gy, magnitude); |
| 363 | } |
| 364 | |
| 365 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 366 | void NEMagnitudePhaseKernel<mag_type, phase_type>::phase(const Window &window) |
| 367 | { |
| 368 | Iterator gx(_gx, window); |
| 369 | Iterator gy(_gy, window); |
| 370 | Iterator phase(_phase, window); |
| 371 | |
| 372 | execute_window_loop(window, [&](const Coordinates & id) |
| 373 | { |
| 374 | const int16x8x2_t input1 = |
| 375 | { |
| 376 | { |
| 377 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr())), |
| 378 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr()) + 8) |
| 379 | } |
| 380 | }; |
| 381 | |
| 382 | const int16x8x2_t input2 = |
| 383 | { |
| 384 | { |
| 385 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr())), |
| 386 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr()) + 8) |
| 387 | } |
| 388 | }; |
| 389 | |
| 390 | /* Compute phase */ |
| 391 | uint8x8x2_t vphase{ {} }; |
| 392 | |
| 393 | if(PhaseType::SIGNED == phase_type) |
| 394 | { |
| 395 | vphase.val[0] = phase_signed(input1.val[0], input2.val[0]); |
| 396 | vphase.val[1] = phase_signed(input1.val[1], input2.val[1]); |
| 397 | } |
| 398 | else |
| 399 | { |
| 400 | vphase.val[0] = phase_unsigned(input1.val[0], input2.val[0]); |
| 401 | vphase.val[1] = phase_unsigned(input1.val[1], input2.val[1]); |
| 402 | } |
| 403 | |
| 404 | /* Store phase */ |
| 405 | vst1q_u8(phase.ptr(), vcombine_u8(vphase.val[0], vphase.val[1])); |
| 406 | }, |
| 407 | gx, gy, phase); |
| 408 | } |
| 409 | |
| 410 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 411 | void NEMagnitudePhaseKernel<mag_type, phase_type>::magnitude_phase(const Window &window) |
| 412 | { |
| 413 | Iterator gx(_gx, window); |
| 414 | Iterator gy(_gy, window); |
| 415 | Iterator magnitude(_magnitude, window); |
| 416 | Iterator phase(_phase, window); |
| 417 | |
| 418 | execute_window_loop(window, [&](const Coordinates & id) |
| 419 | { |
| 420 | const int16x8x2_t input1 = |
| 421 | { |
| 422 | { |
| 423 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr())), |
| 424 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr()) + 8) |
| 425 | } |
| 426 | }; |
| 427 | |
| 428 | const int16x8x2_t input2 = |
| 429 | { |
| 430 | { |
| 431 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr())), |
| 432 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr()) + 8) |
| 433 | } |
| 434 | }; |
| 435 | |
| 436 | /* Compute magnitude */ |
| 437 | int16x8x2_t mag{ {} }; |
| 438 | |
| 439 | if(MagnitudeType::L2NORM == mag_type) |
| 440 | { |
| 441 | mag.val[0] = magnitude_l2(input1.val[0], input2.val[0]); |
| 442 | mag.val[1] = magnitude_l2(input1.val[1], input2.val[1]); |
| 443 | } |
| 444 | else |
| 445 | { |
| 446 | mag.val[0] = magnitude_l1(input1.val[0], input2.val[0]); |
| 447 | mag.val[1] = magnitude_l1(input1.val[1], input2.val[1]); |
| 448 | } |
| 449 | |
| 450 | /* Store magnitude */ |
| 451 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()), mag.val[0]); |
| 452 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()) + 8, mag.val[1]); |
| 453 | |
| 454 | /* Compute phase */ |
| 455 | uint8x8x2_t vphase{ {} }; |
| 456 | |
| 457 | if(PhaseType::SIGNED == phase_type) |
| 458 | { |
| 459 | vphase.val[0] = phase_signed(input1.val[0], input2.val[0]); |
| 460 | vphase.val[1] = phase_signed(input1.val[1], input2.val[1]); |
| 461 | } |
| 462 | else |
| 463 | { |
| 464 | vphase.val[0] = phase_unsigned(input1.val[0], input2.val[0]); |
| 465 | vphase.val[1] = phase_unsigned(input1.val[1], input2.val[1]); |
| 466 | } |
| 467 | |
| 468 | /* Store phase */ |
| 469 | vst1q_u8(phase.ptr(), vcombine_u8(vphase.val[0], vphase.val[1])); |
| 470 | }, |
| 471 | gx, gy, magnitude, phase); |
| 472 | } |
| 473 | |
| 474 | template <MagnitudeType mag_type, PhaseType phase_type> |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 475 | void NEMagnitudePhaseKernel<mag_type, phase_type>::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 476 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 477 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 478 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 479 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 480 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 481 | |
| 482 | (this->*_func)(window); |
| 483 | } |
| 484 | |
| 485 | template class arm_compute::NEMagnitudePhaseKernel<MagnitudeType::L1NORM, PhaseType::SIGNED>; |
| 486 | template class arm_compute::NEMagnitudePhaseKernel<MagnitudeType::L2NORM, PhaseType::SIGNED>; |
| 487 | template class arm_compute::NEMagnitudePhaseKernel<MagnitudeType::L1NORM, PhaseType::UNSIGNED>; |
| 488 | template class arm_compute::NEMagnitudePhaseKernel<MagnitudeType::L2NORM, PhaseType::UNSIGNED>; |