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/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 | |
Ioan-Cristian Szabo | 33fd07b | 2017-10-26 15:42:24 +0100 | [diff] [blame] | 54 | #ifdef ARM_COMPUTE_AARCH64_V8_2 |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 55 | namespace fp16 |
| 56 | { |
| 57 | inline float16x8_t inv(float16x8_t x) |
| 58 | { |
| 59 | const float16x8_t estimate = vrecpeq_f16(x); |
| 60 | return vmulq_f16(estimate, vrecpsq_f16(x, estimate)); |
| 61 | } |
| 62 | |
| 63 | inline float16x8_t atan2_fast(float16x8_t gx, float16x8_t gy, float16x8_t scale) |
| 64 | { |
| 65 | static const float16x8_t one = vdupq_n_f16(1.0f); |
| 66 | static const float16x8_t ninety = vdupq_n_f16(90.f * SCALE_FACTOR); |
| 67 | static const float16x8_t epsilon = vdupq_n_f16(1e-9f); |
| 68 | static const float16x8_t piover4 = vdupq_n_f16(PI_4); |
| 69 | static const float16x8_t coeff1 = vdupq_n_f16(COEFF1); |
| 70 | static const float16x8_t coeff2 = vdupq_n_f16(COEFF2); |
| 71 | |
| 72 | const float16x8_t abs_gx = vabsq_f16(gx); |
| 73 | const float16x8_t abs_gy = vabsq_f16(gy); |
| 74 | const float16x8_t tmin = vminq_f16(abs_gx, abs_gy); |
| 75 | const float16x8_t tmax = vmaxq_f16(abs_gx, abs_gy); |
| 76 | |
| 77 | // z = min(x, y) / max(x, y) |
| 78 | const float16x8_t z = vmulq_f16(tmin, inv(vaddq_f16(tmax, epsilon))); |
| 79 | const float16x8_t absz = vabsq_f16(z); |
| 80 | |
| 81 | // = x * [pi/4 + (1 - |x|) * (0.2447 + 0.0663 * |x|)] |
| 82 | float16x8_t arctan = vmulq_f16(z, vfmaq_f16(piover4, |
| 83 | vsubq_f16(one, absz), |
| 84 | vfmaq_f16(coeff2, coeff1, absz))); |
| 85 | |
| 86 | // Radians to degrees conversion with applied a scale factor in order to have the result [0, 255] |
| 87 | arctan = vmulq_f16(arctan, scale); |
| 88 | |
| 89 | /* If z > 1, result = 90 - result */ |
| 90 | return vbslq_f16(vcgeq_f16(abs_gx, abs_gy), arctan, vsubq_f16(ninety, arctan)); |
| 91 | } |
| 92 | |
| 93 | inline float16x8_t atan2_0_360(float16x8_t gx, float16x8_t gy) |
| 94 | { |
| 95 | static const float16x8_t scale = vdupq_n_f16(SCALE_360); |
| 96 | static const float16x8_t threesixty = vdupq_n_f16(360.0f * SCALE_FACTOR); |
| 97 | static const float16x8_t zero = vdupq_n_f16(0.0f); |
| 98 | static const float16x8_t oneeighty = vdupq_n_f16(180.0f * SCALE_FACTOR); |
| 99 | |
| 100 | float16x8_t arctan = atan2_fast(gx, gy, scale); |
| 101 | |
| 102 | // Choose correct quadrant |
| 103 | arctan = vbslq_f16(vcltq_f16(gx, zero), vsubq_f16(oneeighty, arctan), arctan); |
| 104 | arctan = vbslq_f16(vcltq_f16(gy, zero), vsubq_f16(threesixty, arctan), arctan); |
| 105 | |
| 106 | return arctan; |
| 107 | } |
| 108 | |
| 109 | inline float16x8_t atan2_0_180(float16x8_t gx, float16x8_t gy) |
| 110 | { |
| 111 | static const float16x8_t scale = vdupq_n_f16(SCALE_180); |
| 112 | static const float16x8_t threesixty = vdupq_n_f16(360.0f * SCALE_FACTOR); |
| 113 | static const float16x8_t oneeighty = vdupq_n_f16(180.0f * SCALE_FACTOR); |
| 114 | static const float16x8_t zero = vdupq_n_f16(0.0f); |
| 115 | |
| 116 | float16x8_t arctan = atan2_fast(gx, gy, scale); |
| 117 | |
| 118 | // Choose correct quadrant |
| 119 | arctan = vbslq_f16(vcltq_f16(gx, zero), vsubq_f16(oneeighty, arctan), arctan); |
| 120 | arctan = vbslq_f16(vcltq_f16(gy, zero), vsubq_f16(threesixty, arctan), arctan); |
| 121 | arctan = vbslq_f16(vcgtq_f16(arctan, oneeighty), vsubq_f16(arctan, oneeighty), arctan); |
| 122 | |
| 123 | return arctan; |
| 124 | } |
| 125 | |
| 126 | inline float32x4_t invsqrtv(float32x4_t x) |
| 127 | { |
| 128 | float32x4_t sqrt_reciprocal = vrsqrteq_f32(x); |
| 129 | |
| 130 | sqrt_reciprocal = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), |
| 131 | sqrt_reciprocal); |
| 132 | sqrt_reciprocal = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), |
| 133 | sqrt_reciprocal); |
| 134 | |
| 135 | return sqrt_reciprocal; |
| 136 | } |
| 137 | |
| 138 | inline float32x4_t sqrtv(float32x4_t x) |
| 139 | { |
| 140 | float32x4_t res = vdupq_n_f32(0.5f); |
| 141 | return vmlaq_f32(res, x, invsqrtv(x)); |
| 142 | } |
| 143 | |
| 144 | inline int16x8_t magnitude_l1(int16x8_t input1, int16x8_t input2) |
| 145 | { |
| 146 | return vqaddq_s16(vabsq_s16(input1), vabsq_s16(input2)); |
| 147 | } |
| 148 | |
| 149 | inline int16x8_t magnitude_l2(int16x8_t input1, int16x8_t input2) |
| 150 | { |
| 151 | const int32x4x2_t square_x = |
| 152 | { |
| 153 | vmull_s16(vget_low_s16(input1), vget_low_s16(input1)), |
| 154 | vmull_s16(vget_high_s16(input1), vget_high_s16(input1)) |
| 155 | }; |
| 156 | |
| 157 | const int32x4x2_t square_y = |
| 158 | { |
| 159 | vmull_s16(vget_low_s16(input2), vget_low_s16(input2)), |
| 160 | vmull_s16(vget_high_s16(input2), vget_high_s16(input2)) |
| 161 | }; |
| 162 | |
| 163 | const uint32x4x2_t sum = |
| 164 | { |
| 165 | vaddq_u32(vreinterpretq_u32_s32(square_x.val[0]), |
| 166 | vreinterpretq_u32_s32(square_y.val[0])), |
| 167 | vaddq_u32(vreinterpretq_u32_s32(square_x.val[1]), |
| 168 | vreinterpretq_u32_s32(square_y.val[1])) |
| 169 | }; |
| 170 | |
| 171 | const float32x4x2_t res = |
| 172 | { |
| 173 | sqrtv(vcvtq_f32_u32(sum.val[0])), |
| 174 | sqrtv(vcvtq_f32_u32(sum.val[1])) |
| 175 | }; |
| 176 | |
| 177 | return vcombine_s16(vqmovn_s32(vcvtq_s32_f32(res.val[0])), |
| 178 | vqmovn_s32(vcvtq_s32_f32(res.val[1]))); |
| 179 | } |
| 180 | |
| 181 | inline uint8x8_t phase_signed(int16x8_t input1, int16x8_t input2) |
| 182 | { |
| 183 | static const float16x8_t zeropointfive = vdupq_n_f16(0.5f); |
| 184 | |
| 185 | const float16x8_t inputx_f16 = vcvtq_f16_s16(input1); |
| 186 | const float16x8_t inputy_f16 = vcvtq_f16_s16(input2); |
| 187 | |
| 188 | // Compute fast atan2 |
| 189 | const float16x8_t angle = atan2_0_360(inputx_f16, inputy_f16); |
| 190 | |
| 191 | return vqmovun_s16(vcvtq_s16_f16(vaddq_f16(angle, zeropointfive))); |
| 192 | } |
| 193 | |
| 194 | inline uint8x8_t phase_unsigned(int16x8_t input1, int16x8_t input2) |
| 195 | { |
| 196 | static const float16x8_t zeropointfive = vdupq_n_f16(0.5f); |
| 197 | |
| 198 | const float16x8_t inputx_f16 = vcvtq_f16_s16(input1); |
| 199 | const float16x8_t inputy_f16 = vcvtq_f16_s16(input2); |
| 200 | |
| 201 | // Compute fast atan2 |
| 202 | const float16x8_t angle = atan2_0_180(inputx_f16, inputy_f16); |
| 203 | |
| 204 | return vqmovun_s16(vcvtq_s16_f16(vaddq_f16(angle, zeropointfive))); |
| 205 | } |
| 206 | |
| 207 | template <MagnitudeType mag_type> |
| 208 | inline int16x8x2_t compute_magnitude(const int16x8x2_t &in0, const int16x8x2_t &gx); |
| 209 | |
| 210 | template <> |
| 211 | inline int16x8x2_t compute_magnitude<MagnitudeType::L2NORM>(const int16x8x2_t &in0, const int16x8x2_t &gx) |
| 212 | { |
| 213 | const int16x8x2_t mag = |
| 214 | { |
| 215 | magnitude_l2(in0.val[0], gx.val[0]), |
| 216 | magnitude_l2(in0.val[1], gx.val[1]) |
| 217 | }; |
| 218 | |
| 219 | return mag; |
| 220 | } |
| 221 | |
| 222 | template <> |
| 223 | inline int16x8x2_t compute_magnitude<MagnitudeType::L1NORM>(const int16x8x2_t &in0, const int16x8x2_t &gx) |
| 224 | { |
| 225 | const int16x8x2_t mag = |
| 226 | { |
| 227 | magnitude_l1(in0.val[0], gx.val[0]), |
| 228 | magnitude_l1(in0.val[1], gx.val[1]) |
| 229 | }; |
| 230 | |
| 231 | return mag; |
| 232 | } |
| 233 | |
| 234 | template <PhaseType phase_type> |
| 235 | inline uint8x16_t compute_phase(const int16x8x2_t &in0, const int16x8x2_t &gx); |
| 236 | |
| 237 | template <> |
| 238 | inline uint8x16_t compute_phase<PhaseType::SIGNED>(const int16x8x2_t &in0, const int16x8x2_t &gx) |
| 239 | { |
| 240 | return vcombine_u8(phase_signed(in0.val[0], gx.val[0]), |
| 241 | phase_signed(in0.val[1], gx.val[1])); |
| 242 | } |
| 243 | |
| 244 | template <> |
| 245 | inline uint8x16_t compute_phase<PhaseType::UNSIGNED>(const int16x8x2_t &in0, const int16x8x2_t &gx) |
| 246 | { |
| 247 | return vcombine_u8(phase_unsigned(in0.val[0], gx.val[0]), |
| 248 | phase_unsigned(in0.val[1], gx.val[1])); |
| 249 | } |
| 250 | } // namespace fp16 |
| 251 | |
| 252 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 253 | NEMagnitudePhaseFP16Kernel<mag_type, phase_type>::NEMagnitudePhaseFP16Kernel() |
| 254 | : _func(nullptr), _gx(nullptr), _gy(nullptr), _magnitude(nullptr), _phase(nullptr) |
| 255 | { |
| 256 | } |
| 257 | |
| 258 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 259 | void NEMagnitudePhaseFP16Kernel<mag_type, phase_type>::configure(const ITensor *gx, const ITensor *gy, ITensor *magnitude, ITensor *phase) |
| 260 | { |
| 261 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(gx, Format::S16); |
| 262 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(gy, Format::S16); |
| 263 | ARM_COMPUTE_ERROR_ON((nullptr == magnitude) && (nullptr == phase)); |
| 264 | |
| 265 | const bool run_mag = magnitude != nullptr; |
| 266 | const bool run_phase = phase != nullptr; |
| 267 | |
| 268 | if(run_mag) |
| 269 | { |
| 270 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(magnitude, Format::S16); |
| 271 | } |
| 272 | |
| 273 | if(run_phase) |
| 274 | { |
| 275 | ARM_COMPUTE_ERROR_ON_FORMAT_NOT_IN(phase, Format::U8); |
| 276 | } |
| 277 | |
| 278 | _gx = gx; |
| 279 | _gy = gy; |
| 280 | _magnitude = magnitude; |
| 281 | _phase = phase; |
| 282 | |
| 283 | if(run_mag && run_phase) |
| 284 | { |
| 285 | /* Run magnitude and phase */ |
| 286 | _func = &NEMagnitudePhaseFP16Kernel<mag_type, phase_type>::magnitude_phase; |
| 287 | } |
| 288 | else if(run_mag) |
| 289 | { |
| 290 | /* Run magnitude */ |
| 291 | _func = &NEMagnitudePhaseFP16Kernel<mag_type, phase_type>::magnitude; |
| 292 | } |
| 293 | else if(run_phase) |
| 294 | { |
| 295 | /* Run phase */ |
| 296 | _func = &NEMagnitudePhaseFP16Kernel<mag_type, phase_type>::phase; |
| 297 | } |
| 298 | else |
| 299 | { |
| 300 | ARM_COMPUTE_ERROR("At least one output must be NOT NULL"); |
| 301 | } |
| 302 | |
| 303 | const unsigned int num_elems_processed_per_iteration = 16; |
| 304 | |
| 305 | // Configure kernel window |
| 306 | Window win = calculate_max_window(*gx->info(), Steps(num_elems_processed_per_iteration)); |
| 307 | AccessWindowHorizontal magnitude_access(magnitude == nullptr ? nullptr : magnitude->info(), 0, num_elems_processed_per_iteration); |
| 308 | AccessWindowHorizontal phase_access(phase == nullptr ? nullptr : phase->info(), 0, num_elems_processed_per_iteration); |
| 309 | |
| 310 | update_window_and_padding(win, |
| 311 | AccessWindowHorizontal(gx->info(), 0, num_elems_processed_per_iteration), |
| 312 | AccessWindowHorizontal(gy->info(), 0, num_elems_processed_per_iteration), |
| 313 | magnitude_access, |
| 314 | phase_access); |
| 315 | |
| 316 | ValidRegion valid_region = intersect_valid_regions(gx->info()->valid_region(), |
| 317 | gy->info()->valid_region()); |
| 318 | |
| 319 | magnitude_access.set_valid_region(win, valid_region); |
| 320 | phase_access.set_valid_region(win, valid_region); |
| 321 | |
| 322 | INEKernel::configure(win); |
| 323 | } |
| 324 | |
| 325 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 326 | void NEMagnitudePhaseFP16Kernel<mag_type, phase_type>::magnitude(const Window &window) |
| 327 | { |
| 328 | Iterator gx(_gx, window); |
| 329 | Iterator gy(_gy, window); |
| 330 | Iterator magnitude(_magnitude, window); |
| 331 | |
| 332 | execute_window_loop(window, [&](const Coordinates & id) |
| 333 | { |
| 334 | const int16x8x2_t input1 = |
| 335 | { |
| 336 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr())), |
| 337 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr()) + 8) |
| 338 | }; |
| 339 | |
| 340 | const int16x8x2_t input2 = |
| 341 | { |
| 342 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr())), |
| 343 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr()) + 8) |
| 344 | }; |
| 345 | |
| 346 | // Compute and store magnitude |
| 347 | const int16x8x2_t mag = fp16::compute_magnitude<mag_type>(input1, input2); |
| 348 | |
| 349 | /* Store magnitude */ |
| 350 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()), mag.val[0]); |
| 351 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()) + 8, mag.val[1]); |
| 352 | }, |
| 353 | gx, gy, magnitude); |
| 354 | } |
| 355 | |
| 356 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 357 | void NEMagnitudePhaseFP16Kernel<mag_type, phase_type>::phase(const Window &window) |
| 358 | { |
| 359 | Iterator gx(_gx, window); |
| 360 | Iterator gy(_gy, window); |
| 361 | Iterator phase(_phase, window); |
| 362 | |
| 363 | execute_window_loop(window, [&](const Coordinates & id) |
| 364 | { |
| 365 | const int16x8x2_t input1 = |
| 366 | { |
| 367 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr())), |
| 368 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr()) + 8) |
| 369 | }; |
| 370 | |
| 371 | const int16x8x2_t input2 = |
| 372 | { |
| 373 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr())), |
| 374 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr()) + 8) |
| 375 | }; |
| 376 | |
| 377 | // Compute and store phase |
| 378 | vst1q_u8(phase.ptr(), fp16::compute_phase<phase_type>(input1, input2)); |
| 379 | }, |
| 380 | gx, gy, phase); |
| 381 | } |
| 382 | |
| 383 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 384 | void NEMagnitudePhaseFP16Kernel<mag_type, phase_type>::magnitude_phase(const Window &window) |
| 385 | { |
| 386 | Iterator gx(_gx, window); |
| 387 | Iterator gy(_gy, window); |
| 388 | Iterator magnitude(_magnitude, window); |
| 389 | Iterator phase(_phase, window); |
| 390 | |
| 391 | execute_window_loop(window, [&](const Coordinates & id) |
| 392 | { |
| 393 | const int16x8x2_t input1 = |
| 394 | { |
| 395 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr())), |
| 396 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr()) + 8) |
| 397 | }; |
| 398 | |
| 399 | const int16x8x2_t input2 = |
| 400 | { |
| 401 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr())), |
| 402 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr()) + 8) |
| 403 | }; |
| 404 | |
| 405 | // Compute and store magnitude |
| 406 | const int16x8x2_t mag = fp16::compute_magnitude<mag_type>(input1, input2); |
| 407 | |
| 408 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()), mag.val[0]); |
| 409 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()) + 8, mag.val[1]); |
| 410 | |
| 411 | // Compute and store phase |
| 412 | vst1q_u8(phase.ptr(), fp16::compute_phase<phase_type>(input1, input2)); |
| 413 | }, |
| 414 | gx, gy, magnitude, phase); |
| 415 | } |
| 416 | |
| 417 | template <MagnitudeType mag_type, PhaseType phase_type> |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 418 | void NEMagnitudePhaseFP16Kernel<mag_type, phase_type>::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 419 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 420 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 421 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 422 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 423 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 424 | |
| 425 | (this->*_func)(window); |
| 426 | } |
| 427 | |
| 428 | template class arm_compute::NEMagnitudePhaseFP16Kernel<MagnitudeType::L1NORM, PhaseType::SIGNED>; |
| 429 | template class arm_compute::NEMagnitudePhaseFP16Kernel<MagnitudeType::L2NORM, PhaseType::SIGNED>; |
| 430 | template class arm_compute::NEMagnitudePhaseFP16Kernel<MagnitudeType::L1NORM, PhaseType::UNSIGNED>; |
| 431 | template class arm_compute::NEMagnitudePhaseFP16Kernel<MagnitudeType::L2NORM, PhaseType::UNSIGNED>; |
Ioan-Cristian Szabo | 33fd07b | 2017-10-26 15:42:24 +0100 | [diff] [blame] | 432 | #endif /* ARM_COMPUTE_AARCH64_V8_2 */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 433 | |
| 434 | namespace |
| 435 | { |
| 436 | inline float32x4_t inv(float32x4_t x) |
| 437 | { |
| 438 | float32x4_t result = vrecpeq_f32(x); |
| 439 | result = vmulq_f32(vrecpsq_f32(x, result), result); |
| 440 | return result; |
| 441 | } |
| 442 | |
| 443 | inline float32x4_t atan2_0_360(float32x4_t gx, float32x4_t gy) |
| 444 | { |
| 445 | const float32x4_t zero = vdupq_n_f32(0.0f); |
| 446 | const float32x4_t epsilon = vdupq_n_f32(1e-9f); |
| 447 | const float32x4_t piover4 = vdupq_n_f32(PI_4); |
| 448 | const float32x4_t coeff1 = vdupq_n_f32(COEFF1); |
| 449 | const float32x4_t coeff2 = vdupq_n_f32(COEFF2); |
| 450 | const float32x4_t ninety = vdupq_n_f32(90.0f * SCALE_FACTOR); |
| 451 | const float32x4_t oneeighty = vdupq_n_f32(180.0f * SCALE_FACTOR); |
| 452 | const float32x4_t threesixty = vdupq_n_f32(360.0f * SCALE_FACTOR); |
| 453 | const float32x4_t scale = vdupq_n_f32(SCALE_360); |
| 454 | |
| 455 | float32x4_t abs_gx = vabsq_f32(gx); |
| 456 | float32x4_t abs_gy = vabsq_f32(gy); |
| 457 | float32x4_t tmin = vminq_f32(abs_gx, abs_gy); |
| 458 | float32x4_t tmax = vmaxq_f32(abs_gx, abs_gy); |
| 459 | float32x4_t z = vmulq_f32(tmin, inv(vaddq_f32(tmax, epsilon))); |
| 460 | float32x4_t absz = vabsq_f32(z); |
| 461 | float32x4_t term = vmulq_f32(z, vsubq_f32(vdupq_n_f32(1.0f), absz)); |
| 462 | |
| 463 | /* Compute y = pi/4 * x - x*(abs(x)-1)*(0.2447+0.0663 * abs(x) */ |
| 464 | float32x4_t result = vaddq_f32(coeff2, vmulq_f32(absz, coeff1)); |
| 465 | result = vmulq_f32(result, term); |
| 466 | result = vmlaq_f32(result, piover4, z); |
| 467 | |
| 468 | /* Radians to degrees conversion with applied a scale factor in order to have the result [0, 255] */ |
| 469 | result = vmulq_f32(result, scale); |
| 470 | |
| 471 | /* If z > 1, result = 90 - result */ |
| 472 | result = vbslq_f32(vcgeq_f32(abs_gx, abs_gy), result, vsubq_f32(ninety, result)); |
| 473 | |
| 474 | /* Choose correct quadrant */ |
| 475 | result = vbslq_f32(vcltq_f32(gx, zero), vsubq_f32(oneeighty, result), result); |
| 476 | result = vbslq_f32(vcltq_f32(gy, zero), vsubq_f32(threesixty, result), result); |
| 477 | |
| 478 | return result; |
| 479 | } |
| 480 | |
| 481 | inline float32x4_t atan2_0_180(float32x4_t gx, float32x4_t gy) |
| 482 | { |
| 483 | const float32x4_t zero = vdupq_n_f32(0.0f); |
| 484 | const float32x4_t epsilon = vdupq_n_f32(1e-9f); // epsilon used to avoiding division by 0 |
| 485 | const float32x4_t piover4 = vdupq_n_f32(PI_4); |
| 486 | const float32x4_t coeff1 = vdupq_n_f32(COEFF1); |
| 487 | const float32x4_t coeff2 = vdupq_n_f32(COEFF2); |
| 488 | const float32x4_t ninety = vdupq_n_f32(90.0f); |
| 489 | const float32x4_t oneeighty = vdupq_n_f32(180.0f); |
| 490 | const float32x4_t threesixty = vdupq_n_f32(360.0f); |
| 491 | const float32x4_t scale = vdupq_n_f32(SCALE_180); |
| 492 | |
| 493 | float32x4_t abs_gx = vabsq_f32(gx); |
| 494 | float32x4_t abs_gy = vabsq_f32(gy); |
| 495 | float32x4_t tmin = vminq_f32(abs_gx, abs_gy); |
| 496 | float32x4_t tmax = vmaxq_f32(abs_gx, abs_gy); |
| 497 | float32x4_t z = vmulq_f32(tmin, inv(vaddq_f32(tmax, epsilon))); |
| 498 | float32x4_t absz = vabsq_f32(z); |
| 499 | |
| 500 | /* Compute y = pi/4 * z - z*(abs(z)-1)*(0.2447+0.0663 * abs(z) */ |
| 501 | float32x4_t term = vmulq_f32(z, vsubq_f32(vdupq_n_f32(1.0f), absz)); |
| 502 | float32x4_t result = vaddq_f32(coeff2, vmulq_f32(absz, coeff1)); |
| 503 | result = vmulq_f32(result, term); |
| 504 | result = vmlaq_f32(result, piover4, z); |
| 505 | |
| 506 | /* Radians to degrees conversion */ |
| 507 | result = vmulq_f32(result, scale); |
| 508 | |
| 509 | /* If z > 1, result = 90 - result */ |
| 510 | result = vbslq_f32(vcgeq_f32(abs_gx, abs_gy), result, vsubq_f32(ninety, result)); |
| 511 | |
| 512 | /* Choose correct quadrant */ |
| 513 | result = vbslq_f32(vcltq_f32(gx, zero), vsubq_f32(oneeighty, result), result); |
| 514 | result = vbslq_f32(vcltq_f32(gy, zero), vsubq_f32(threesixty, result), result); |
| 515 | result = vbslq_f32(vcgtq_f32(result, oneeighty), vsubq_f32(result, oneeighty), result); |
| 516 | |
| 517 | return result; |
| 518 | } |
| 519 | |
| 520 | inline float32x4_t invsqrtv(float32x4_t x) |
| 521 | { |
| 522 | float32x4_t sqrt_reciprocal = vrsqrteq_f32(x); |
| 523 | |
| 524 | sqrt_reciprocal = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), |
| 525 | sqrt_reciprocal); |
| 526 | sqrt_reciprocal = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x, sqrt_reciprocal), sqrt_reciprocal), |
| 527 | sqrt_reciprocal); |
| 528 | |
| 529 | return sqrt_reciprocal; |
| 530 | } |
| 531 | |
| 532 | inline float32x4_t sqrtv(float32x4_t x) |
| 533 | { |
| 534 | float32x4_t res = vdupq_n_f32(0.5f); |
| 535 | return vmlaq_f32(res, x, invsqrtv(x)); |
| 536 | } |
| 537 | |
| 538 | inline int16x8_t magnitude_l2(int16x8_t input1, int16x8_t input2) |
| 539 | { |
| 540 | const int32x4x2_t square_x = |
| 541 | { |
| 542 | { |
| 543 | vmull_s16(vget_low_s16(input1), vget_low_s16(input1)), |
| 544 | vmull_s16(vget_high_s16(input1), vget_high_s16(input1)) |
| 545 | } |
| 546 | }; |
| 547 | |
| 548 | const int32x4x2_t square_y = |
| 549 | { |
| 550 | { |
| 551 | vmull_s16(vget_low_s16(input2), vget_low_s16(input2)), |
| 552 | vmull_s16(vget_high_s16(input2), vget_high_s16(input2)) |
| 553 | } |
| 554 | }; |
| 555 | |
| 556 | const uint32x4x2_t sum = |
| 557 | { |
| 558 | { |
| 559 | vaddq_u32(vreinterpretq_u32_s32(square_x.val[0]), vreinterpretq_u32_s32(square_y.val[0])), |
| 560 | vaddq_u32(vreinterpretq_u32_s32(square_x.val[1]), vreinterpretq_u32_s32(square_y.val[1])) |
| 561 | } |
| 562 | }; |
| 563 | |
| 564 | const float32x4x2_t res = |
| 565 | { |
| 566 | { |
| 567 | sqrtv(vcvtq_f32_u32(sum.val[0])), |
| 568 | sqrtv(vcvtq_f32_u32(sum.val[1])) |
| 569 | } |
| 570 | }; |
| 571 | |
| 572 | return vcombine_s16(vqmovn_s32(vcvtq_s32_f32(res.val[0])), |
| 573 | vqmovn_s32(vcvtq_s32_f32(res.val[1]))); |
| 574 | } |
| 575 | |
| 576 | inline int16x8_t magnitude_l1(int16x8_t input1, int16x8_t input2) |
| 577 | { |
| 578 | int16x8_t gx_abs = vabsq_s16(input1); |
| 579 | int16x8_t gy_abs = vabsq_s16(input2); |
| 580 | |
| 581 | /* Saturating add */ |
| 582 | return vqaddq_s16(gx_abs, gy_abs); |
| 583 | } |
| 584 | |
| 585 | inline uint8x8_t phase_signed(int16x8_t input1, int16x8_t input2) |
| 586 | { |
| 587 | const float32x4_t zeropointfive = vdupq_n_f32(0.5f); |
| 588 | |
| 589 | float32x4_t inputx_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(input1))); |
| 590 | float32x4_t inputx_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(input1))); |
| 591 | float32x4_t inputy_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(input2))); |
| 592 | float32x4_t inputy_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(input2))); |
| 593 | |
| 594 | /* Compute fast atan2 */ |
| 595 | float32x4_t angle_high = atan2_0_360(inputx_f32_high, inputy_f32_high); |
| 596 | float32x4_t angle_low = atan2_0_360(inputx_f32_low, inputy_f32_low); |
| 597 | |
| 598 | angle_high = vaddq_f32(angle_high, zeropointfive); |
| 599 | angle_low = vaddq_f32(angle_low, zeropointfive); |
| 600 | |
| 601 | return vmovn_u16(vcombine_u16(vqmovun_s32(vcvtq_s32_f32(angle_low)), |
| 602 | vqmovun_s32(vcvtq_s32_f32(angle_high)))); |
| 603 | } |
| 604 | |
| 605 | inline uint8x8_t phase_unsigned(int16x8_t input1, int16x8_t input2) |
| 606 | { |
| 607 | const float32x4_t zeropointfive = vdupq_n_f32(0.5f); |
| 608 | |
| 609 | float32x4_t inputx_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(input1))); |
| 610 | float32x4_t inputx_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(input1))); |
| 611 | float32x4_t inputy_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(input2))); |
| 612 | float32x4_t inputy_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(input2))); |
| 613 | |
| 614 | /* Compute fast atan2 */ |
| 615 | float32x4_t angle_high = atan2_0_180(inputx_f32_high, inputy_f32_high); |
| 616 | float32x4_t angle_low = atan2_0_180(inputx_f32_low, inputy_f32_low); |
| 617 | |
| 618 | angle_high = vaddq_f32(angle_high, zeropointfive); |
| 619 | angle_low = vaddq_f32(angle_low, zeropointfive); |
| 620 | |
| 621 | return vmovn_u16(vcombine_u16(vqmovun_s32(vcvtq_s32_f32(angle_low)), |
| 622 | vqmovun_s32(vcvtq_s32_f32(angle_high)))); |
| 623 | } |
| 624 | } // namespace |
| 625 | |
| 626 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 627 | NEMagnitudePhaseKernel<mag_type, phase_type>::NEMagnitudePhaseKernel() |
| 628 | : _func(nullptr), _gx(nullptr), _gy(nullptr), _magnitude(nullptr), _phase(nullptr) |
| 629 | { |
| 630 | } |
| 631 | |
| 632 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 633 | void NEMagnitudePhaseKernel<mag_type, phase_type>::configure(const ITensor *gx, const ITensor *gy, ITensor *magnitude, ITensor *phase) |
| 634 | { |
| 635 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(gx, 1, DataType::S16); |
| 636 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(gy, 1, DataType::S16); |
| 637 | ARM_COMPUTE_ERROR_ON((nullptr == magnitude) && (nullptr == phase)); |
| 638 | |
| 639 | const bool run_mag = magnitude != nullptr; |
| 640 | const bool run_phase = phase != nullptr; |
| 641 | |
| 642 | if(run_mag) |
| 643 | { |
| 644 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(magnitude, 1, DataType::S16); |
| 645 | } |
| 646 | |
| 647 | if(run_phase) |
| 648 | { |
| 649 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(phase, 1, DataType::U8); |
| 650 | } |
| 651 | |
| 652 | _gx = gx; |
| 653 | _gy = gy; |
| 654 | _magnitude = magnitude; |
| 655 | _phase = phase; |
| 656 | |
| 657 | if(run_mag && run_phase) |
| 658 | { |
| 659 | /* Run magnitude and phase */ |
| 660 | _func = &NEMagnitudePhaseKernel<mag_type, phase_type>::magnitude_phase; |
| 661 | } |
| 662 | else |
| 663 | { |
| 664 | if(run_mag) |
| 665 | { |
| 666 | /* Run magnitude */ |
| 667 | _func = &NEMagnitudePhaseKernel<mag_type, phase_type>::magnitude; |
| 668 | } |
| 669 | else if(run_phase) |
| 670 | { |
| 671 | /* Run phase */ |
| 672 | _func = &NEMagnitudePhaseKernel<mag_type, phase_type>::phase; |
| 673 | } |
| 674 | else |
| 675 | { |
| 676 | ARM_COMPUTE_ERROR("At least one output must be NOT NULL"); |
| 677 | } |
| 678 | } |
| 679 | |
| 680 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 681 | |
| 682 | // Configure kernel window |
| 683 | Window win = calculate_max_window(*gx->info(), Steps(num_elems_processed_per_iteration)); |
| 684 | AccessWindowHorizontal magnitude_access(magnitude == nullptr ? nullptr : magnitude->info(), 0, num_elems_processed_per_iteration); |
| 685 | AccessWindowHorizontal phase_access(phase == nullptr ? nullptr : phase->info(), 0, num_elems_processed_per_iteration); |
| 686 | |
| 687 | update_window_and_padding(win, |
| 688 | AccessWindowHorizontal(gx->info(), 0, num_elems_processed_per_iteration), |
| 689 | AccessWindowHorizontal(gy->info(), 0, num_elems_processed_per_iteration), |
| 690 | magnitude_access, |
| 691 | phase_access); |
| 692 | |
| 693 | ValidRegion valid_region = intersect_valid_regions(gx->info()->valid_region(), |
| 694 | gy->info()->valid_region()); |
| 695 | |
| 696 | magnitude_access.set_valid_region(win, valid_region); |
| 697 | phase_access.set_valid_region(win, valid_region); |
| 698 | |
| 699 | INEKernel::configure(win); |
| 700 | } |
| 701 | |
| 702 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 703 | void NEMagnitudePhaseKernel<mag_type, phase_type>::magnitude(const Window &window) |
| 704 | { |
| 705 | Iterator gx(_gx, window); |
| 706 | Iterator gy(_gy, window); |
| 707 | Iterator magnitude(_magnitude, window); |
| 708 | |
| 709 | execute_window_loop(window, [&](const Coordinates & id) |
| 710 | { |
| 711 | const int16x8x2_t input1 = |
| 712 | { |
| 713 | { |
| 714 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr())), |
| 715 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr()) + 8) |
| 716 | } |
| 717 | }; |
| 718 | |
| 719 | const int16x8x2_t input2 = |
| 720 | { |
| 721 | { |
| 722 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr())), |
| 723 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr()) + 8) |
| 724 | } |
| 725 | }; |
| 726 | |
| 727 | /* Compute magnitude */ |
| 728 | int16x8x2_t mag{ {} }; |
| 729 | |
| 730 | if(MagnitudeType::L2NORM == mag_type) |
| 731 | { |
| 732 | mag.val[0] = magnitude_l2(input1.val[0], input2.val[0]); |
| 733 | mag.val[1] = magnitude_l2(input1.val[1], input2.val[1]); |
| 734 | } |
| 735 | else |
| 736 | { |
| 737 | mag.val[0] = magnitude_l1(input1.val[0], input2.val[0]); |
| 738 | mag.val[1] = magnitude_l1(input1.val[1], input2.val[1]); |
| 739 | } |
| 740 | |
| 741 | /* Store magnitude */ |
| 742 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()), mag.val[0]); |
| 743 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()) + 8, mag.val[1]); |
| 744 | }, |
| 745 | gx, gy, magnitude); |
| 746 | } |
| 747 | |
| 748 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 749 | void NEMagnitudePhaseKernel<mag_type, phase_type>::phase(const Window &window) |
| 750 | { |
| 751 | Iterator gx(_gx, window); |
| 752 | Iterator gy(_gy, window); |
| 753 | Iterator phase(_phase, window); |
| 754 | |
| 755 | execute_window_loop(window, [&](const Coordinates & id) |
| 756 | { |
| 757 | const int16x8x2_t input1 = |
| 758 | { |
| 759 | { |
| 760 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr())), |
| 761 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr()) + 8) |
| 762 | } |
| 763 | }; |
| 764 | |
| 765 | const int16x8x2_t input2 = |
| 766 | { |
| 767 | { |
| 768 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr())), |
| 769 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr()) + 8) |
| 770 | } |
| 771 | }; |
| 772 | |
| 773 | /* Compute phase */ |
| 774 | uint8x8x2_t vphase{ {} }; |
| 775 | |
| 776 | if(PhaseType::SIGNED == phase_type) |
| 777 | { |
| 778 | vphase.val[0] = phase_signed(input1.val[0], input2.val[0]); |
| 779 | vphase.val[1] = phase_signed(input1.val[1], input2.val[1]); |
| 780 | } |
| 781 | else |
| 782 | { |
| 783 | vphase.val[0] = phase_unsigned(input1.val[0], input2.val[0]); |
| 784 | vphase.val[1] = phase_unsigned(input1.val[1], input2.val[1]); |
| 785 | } |
| 786 | |
| 787 | /* Store phase */ |
| 788 | vst1q_u8(phase.ptr(), vcombine_u8(vphase.val[0], vphase.val[1])); |
| 789 | }, |
| 790 | gx, gy, phase); |
| 791 | } |
| 792 | |
| 793 | template <MagnitudeType mag_type, PhaseType phase_type> |
| 794 | void NEMagnitudePhaseKernel<mag_type, phase_type>::magnitude_phase(const Window &window) |
| 795 | { |
| 796 | Iterator gx(_gx, window); |
| 797 | Iterator gy(_gy, window); |
| 798 | Iterator magnitude(_magnitude, window); |
| 799 | Iterator phase(_phase, window); |
| 800 | |
| 801 | execute_window_loop(window, [&](const Coordinates & id) |
| 802 | { |
| 803 | const int16x8x2_t input1 = |
| 804 | { |
| 805 | { |
| 806 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr())), |
| 807 | vld1q_s16(reinterpret_cast<int16_t *>(gx.ptr()) + 8) |
| 808 | } |
| 809 | }; |
| 810 | |
| 811 | const int16x8x2_t input2 = |
| 812 | { |
| 813 | { |
| 814 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr())), |
| 815 | vld1q_s16(reinterpret_cast<int16_t *>(gy.ptr()) + 8) |
| 816 | } |
| 817 | }; |
| 818 | |
| 819 | /* Compute magnitude */ |
| 820 | int16x8x2_t mag{ {} }; |
| 821 | |
| 822 | if(MagnitudeType::L2NORM == mag_type) |
| 823 | { |
| 824 | mag.val[0] = magnitude_l2(input1.val[0], input2.val[0]); |
| 825 | mag.val[1] = magnitude_l2(input1.val[1], input2.val[1]); |
| 826 | } |
| 827 | else |
| 828 | { |
| 829 | mag.val[0] = magnitude_l1(input1.val[0], input2.val[0]); |
| 830 | mag.val[1] = magnitude_l1(input1.val[1], input2.val[1]); |
| 831 | } |
| 832 | |
| 833 | /* Store magnitude */ |
| 834 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()), mag.val[0]); |
| 835 | vst1q_s16(reinterpret_cast<int16_t *>(magnitude.ptr()) + 8, mag.val[1]); |
| 836 | |
| 837 | /* Compute phase */ |
| 838 | uint8x8x2_t vphase{ {} }; |
| 839 | |
| 840 | if(PhaseType::SIGNED == phase_type) |
| 841 | { |
| 842 | vphase.val[0] = phase_signed(input1.val[0], input2.val[0]); |
| 843 | vphase.val[1] = phase_signed(input1.val[1], input2.val[1]); |
| 844 | } |
| 845 | else |
| 846 | { |
| 847 | vphase.val[0] = phase_unsigned(input1.val[0], input2.val[0]); |
| 848 | vphase.val[1] = phase_unsigned(input1.val[1], input2.val[1]); |
| 849 | } |
| 850 | |
| 851 | /* Store phase */ |
| 852 | vst1q_u8(phase.ptr(), vcombine_u8(vphase.val[0], vphase.val[1])); |
| 853 | }, |
| 854 | gx, gy, magnitude, phase); |
| 855 | } |
| 856 | |
| 857 | template <MagnitudeType mag_type, PhaseType phase_type> |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 858 | 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] | 859 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 860 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 861 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 862 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 863 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 864 | |
| 865 | (this->*_func)(window); |
| 866 | } |
| 867 | |
| 868 | template class arm_compute::NEMagnitudePhaseKernel<MagnitudeType::L1NORM, PhaseType::SIGNED>; |
| 869 | template class arm_compute::NEMagnitudePhaseKernel<MagnitudeType::L2NORM, PhaseType::SIGNED>; |
| 870 | template class arm_compute::NEMagnitudePhaseKernel<MagnitudeType::L1NORM, PhaseType::UNSIGNED>; |
| 871 | template class arm_compute::NEMagnitudePhaseKernel<MagnitudeType::L2NORM, PhaseType::UNSIGNED>; |