Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 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 | #ifndef __ARM_COMPUTE_NEFIXEDPOINT_H__ |
| 25 | #define __ARM_COMPUTE_NEFIXEDPOINT_H__ |
| 26 | |
| 27 | #include "arm_compute/core/FixedPoint.h" |
| 28 | |
| 29 | #include <arm_neon.h> |
| 30 | |
| 31 | namespace arm_compute |
| 32 | { |
| 33 | using qint8x8_t = int8x8_t; /**< 8 bit fixed point vector with 8 elements */ |
| 34 | using qint8x8x2_t = int8x8x2_t; /**< 8 bit fixed point vector with 16 elements */ |
| 35 | using qint8x8x3_t = int8x8x3_t; /**< 8 bit fixed point vector with 24 elements */ |
| 36 | using qint8x8x4_t = int8x8x4_t; /**< 8 bit fixed point vector with 32 elements */ |
| 37 | using qint8x16_t = int8x16_t; /**< 8 bit fixed point vector with 16 elements */ |
| 38 | using qint8x16x2_t = int8x16x2_t; /**< 8 bit fixed point vector with 32 elements */ |
| 39 | using qint8x16x3_t = int8x16x3_t; /**< 8 bit fixed point vector with 48 elements */ |
| 40 | using qint8x16x4_t = int8x16x4_t; /**< 8 bit fixed point vector with 64 elements */ |
| 41 | using qint16x4_t = int16x4_t; /**< 16 bit fixed point vector with 4 elements */ |
| 42 | using qint16x4x2_t = int16x4x2_t; /**< 16 bit fixed point vector with 8 elements */ |
| 43 | using qint16x4x3_t = int16x4x3_t; /**< 16 bit fixed point vector with 12 elements */ |
| 44 | using qint16x4x4_t = int16x4x4_t; /**< 16 bit fixed point vector with 16 elements */ |
| 45 | using qint16x8_t = int16x8_t; /**< 16 bit fixed point vector with 8 elements */ |
| 46 | using qint16x8x2_t = int16x8x2_t; /**< 16 bit fixed point vector with 16 elements */ |
| 47 | using qint16x8x3_t = int16x8x3_t; /**< 16 bit fixed point vector with 24 elements */ |
| 48 | using qint16x8x4_t = int16x8x4_t; /**< 16 bit fixed point vector with 32 elements */ |
| 49 | |
| 50 | /** Get the lower half of a 16 elements vector |
| 51 | * |
| 52 | * @param[in] a vector of 16 elements |
| 53 | * |
| 54 | * @return 8 bit fixed point vector (8 elements) |
| 55 | */ |
| 56 | qint8x8_t vget_low_qs8(qint8x16_t a); |
| 57 | |
| 58 | /** Get the higher half of a 16 elements vector |
| 59 | * |
| 60 | * @param[in] a vector of 16 elements |
| 61 | * |
| 62 | * @return 8 bit fixed point vector (8 elements) |
| 63 | */ |
| 64 | qint8x8_t vget_high_qs8(qint8x16_t a); |
| 65 | |
| 66 | /** Load a single 8 bit fixed point vector from memory (8 elements) |
| 67 | * |
| 68 | * @param[in] addr Memory address of the 8 bit fixed point vector to load |
| 69 | * |
| 70 | * @return 8 bit fixed point vector (8 elements) |
| 71 | */ |
| 72 | qint8x8_t vld1_qs8(const qint8_t *addr); |
| 73 | |
| 74 | /** Load a single 8 bit fixed point vector from memory (16 elements) |
| 75 | * |
| 76 | * @param[in] addr Memory address of the 8 bit fixed point vector to load |
| 77 | * |
| 78 | * @return 8 bit fixed point vector (16 elements) |
| 79 | */ |
| 80 | qint8x16_t vld1q_qs8(const qint8_t *addr); |
| 81 | |
| 82 | /** Load a single 16 bit fixed point vector from memory (4 elements) |
| 83 | * |
| 84 | * @param[in] addr Memory address of the 16 bit fixed point vector to load |
| 85 | * |
| 86 | * @return 16 bit fixed point vector (4 elements) |
| 87 | */ |
| 88 | qint16x4_t vld1_qs16(const qint16_t *addr); |
| 89 | |
| 90 | /** Load a single 16 bit fixed point vector from memory (8 elements) |
| 91 | * |
| 92 | * @param[in] addr Memory address of the 16 bit fixed point vector to load |
| 93 | * |
| 94 | * @return 16 bit fixed point vector (8 elements) |
| 95 | */ |
| 96 | qint16x8_t vld1q_qs16(const qint16_t *addr); |
| 97 | |
| 98 | /** Load all lanes of 8 bit fixed point vector with same value from memory (8 elements) |
| 99 | * |
| 100 | * @param[in] addr Memory address of the 8 bit fixed point scalar value to load |
| 101 | * |
| 102 | * @return 8 bit fixed point vector (8 elements) |
| 103 | */ |
| 104 | qint8x8_t vld1_dup_qs8(const qint8_t *addr); |
| 105 | |
| 106 | /** Load all lanes of 8 bit fixed point vector with same value from memory (16 elements) |
| 107 | * |
| 108 | * @param[in] addr Memory address of the 8 bit fixed point scalar value to load |
| 109 | * |
| 110 | * @return 8 bit fixed point vector (16 elements) |
| 111 | */ |
| 112 | qint8x16_t vld1q_dup_qs8(const qint8_t *addr); |
| 113 | |
| 114 | /** Store a single 8 bit fixed point vector to memory (8 elements) |
| 115 | * |
| 116 | * @param[in] addr Memory address where the 8 bit fixed point vector should be stored |
| 117 | * @param[in] b 8 bit fixed point vector to store |
| 118 | * |
| 119 | */ |
| 120 | void vst1_qs8(qint8_t *addr, qint8x8_t b); |
| 121 | |
| 122 | /** Store a single 8 bit fixed point vector to memory (16 elements) |
| 123 | * |
| 124 | * @param[in] addr Memory address where the 8 bit fixed point vector should be stored |
| 125 | * @param[in] b 8 bit fixed point vector to store |
| 126 | * |
| 127 | */ |
| 128 | void vst1q_qs8(qint8_t *addr, qint8x16_t b); |
| 129 | |
| 130 | /** Store a single 16 bit fixed point vector to memory (4 elements) |
| 131 | * |
| 132 | * @param[in] addr Memory address where the 16 bit fixed point vector should be stored |
| 133 | * @param[in] b 16 bit fixed point vector to store |
| 134 | * |
| 135 | */ |
| 136 | void vst1_qs16(qint16_t *addr, qint16x4_t b); |
| 137 | |
| 138 | /** Store a single 8 bit fixed point vector to memory (16 elements) |
| 139 | * |
| 140 | * @param[in] addr Memory address where the 16 bit fixed point vector should be stored |
| 141 | * @param[in] b 16 bit fixed point vector to store |
| 142 | * |
| 143 | */ |
| 144 | void vst1q_qs16(qint16_t *addr, qint16x8_t b); |
| 145 | |
| 146 | /** 16 bit fixed point vector saturating narrow (8 elements) |
| 147 | * |
| 148 | * @param[in] a 16 bit fixed point vector to convert |
| 149 | * |
| 150 | * @return 8 bit fixed point vector |
| 151 | */ |
| 152 | qint8x8_t vqmovn_q16(qint16x8_t a); |
| 153 | |
| 154 | /** 8 bit fixed point vector duplicate (8 elements) |
| 155 | * |
| 156 | * @param[in] a 8 bit fixed point to duplicate |
| 157 | * |
| 158 | * @return The result of the vector duplication |
| 159 | */ |
| 160 | qint8x8_t vdup_n_qs8(qint8_t a); |
| 161 | |
| 162 | /** 8 bit fixed point vector duplicate (16 elements) |
| 163 | * |
| 164 | * @param[in] a 8 bit fixed point to duplicate |
| 165 | * |
| 166 | * @return The result of the vector duplication |
| 167 | */ |
| 168 | qint8x16_t vdupq_n_qs8(qint8_t a); |
| 169 | |
| 170 | /** Duplicate a float and convert it to 8 bit fixed point vector (16 elements) |
| 171 | * |
| 172 | * @param[in] a 8 bit fixed point to duplicate |
| 173 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 174 | * |
| 175 | * @return The result of the vector duplication |
| 176 | */ |
| 177 | qint8x16_t vdupq_n_qs8_f32(float a, int fixed_point_position); |
| 178 | |
| 179 | /** 16 bit fixed point vector duplicate (8 elements) |
| 180 | * |
| 181 | * @param[in] a 16 bit fixed point to duplicate |
| 182 | * |
| 183 | * @return The result of the vector duplication |
| 184 | */ |
| 185 | qint16x8_t vdupq_n_qs16(qint16x8_t a); |
| 186 | |
| 187 | /** Absolute value of 8 bit fixed point vector (8 elements) |
| 188 | * |
| 189 | * @param[in] a 8 bit fixed point input vector |
| 190 | * |
| 191 | * @return The result of the 8 bit fixed point vector absolute value |
| 192 | */ |
| 193 | qint8x8_t vabs_qs8(qint8x8_t a); |
| 194 | |
| 195 | /** Absolute value of 8 bit fixed point vector (16 elements) |
| 196 | * |
| 197 | * @param[in] a 8 bit fixed point input vector |
| 198 | * |
| 199 | * @return The result of the 8 bit fixed point vector absolute value |
| 200 | */ |
| 201 | qint8x16_t vabsq_qs8(qint8x16_t a); |
| 202 | |
| 203 | /** Saturating absolute value of 8 bit fixed point vector (8 elements) |
| 204 | * |
| 205 | * @param[in] a 8 bit fixed point input vector |
| 206 | * |
| 207 | * @return The result of the 8 bit fixed point vector absolute value |
| 208 | */ |
| 209 | qint8x8_t vqabs_qs8(qint8x8_t a); |
| 210 | |
| 211 | /** Saturating absolute value of 8 bit fixed point vector (16 elements) |
| 212 | * |
| 213 | * @param[in] a 8 bit fixed point input vector |
| 214 | * |
| 215 | * @return The result of the 8 bit fixed point vector absolute value |
| 216 | */ |
| 217 | qint8x16_t vqabsq_qs8(qint8x16_t a); |
| 218 | |
| 219 | /** 8 bit fixed point vector max (8 elements) |
| 220 | * |
| 221 | * @param[in] a First 8 bit fixed point input vector |
| 222 | * @param[in] b Second 8 bit fixed point input vector |
| 223 | * |
| 224 | * @return The result of the 8 bit fixed point vector max operation |
| 225 | */ |
| 226 | qint8x8_t vmax_qs8(qint8x8_t a, qint8x8_t b); |
| 227 | |
| 228 | /** 8 bit fixed point vector max (16 elements) |
| 229 | * |
| 230 | * @param[in] a First 8 bit fixed point input vector |
| 231 | * @param[in] b Second 8 bit fixed point input vector |
| 232 | * |
| 233 | * @return The result of the 8 bit fixed point vector max operation |
| 234 | */ |
| 235 | qint8x16_t vmaxq_qs8(qint8x16_t a, qint8x16_t b); |
| 236 | |
| 237 | /** 8 bit fixed point vector pairwise max (8 elements) |
| 238 | * |
| 239 | * @param[in] a First 8 bit fixed point input vector |
| 240 | * @param[in] b Second 8 bit fixed point input vector |
| 241 | * |
| 242 | * @return The result of the 8 bit fixed point vector pairwise max operation |
| 243 | */ |
| 244 | qint8x8_t vpmax_qs8(qint8x8_t a, qint8x8_t b); |
| 245 | |
| 246 | /** 8 bit fixed point vector min (8 elements) |
| 247 | * |
| 248 | * @param[in] a First 8 bit fixed point input vector |
| 249 | * @param[in] b Second 8 bit fixed point input vector |
| 250 | * |
| 251 | * @return The result of the 8 bit fixed point vector max operation |
| 252 | */ |
| 253 | qint8x8_t vmin_qs8(qint8x8_t a, qint8x8_t b); |
| 254 | |
| 255 | /** 8 bit fixed point vector min (16 elements) |
| 256 | * |
| 257 | * @param[in] a First 8 bit fixed point input vector |
| 258 | * @param[in] b Second 8 bit fixed point input vector |
| 259 | * |
| 260 | * @return The result of the 8 bit fixed point vector min operation |
| 261 | */ |
| 262 | qint8x16_t vminq_qs8(qint8x16_t a, qint8x16_t b); |
| 263 | |
| 264 | /** 8 bit fixed point vector pairwise min (8 elements) |
| 265 | * |
| 266 | * @param[in] a First 8 bit fixed point input vector |
| 267 | * @param[in] b Second 8 bit fixed point input vector |
| 268 | * |
| 269 | * @return The result of the 8 bit fixed point vector pairwise min operation |
| 270 | */ |
| 271 | qint8x8_t vpmin_qs8(qint8x8_t a, qint8x8_t b); |
| 272 | |
| 273 | /** 8 bit fixed point vector add (8 elements) |
| 274 | * |
| 275 | * @param[in] a First 8 bit fixed point input vector |
| 276 | * @param[in] b Second 8 bit fixed point input vector |
| 277 | * |
| 278 | * @return The result of the 8 bit fixed point vector addition |
| 279 | */ |
| 280 | qint8x8_t vadd_qs8(qint8x8_t a, qint8x8_t b); |
| 281 | |
| 282 | /** 8 bit fixed point vector add (16 elements) |
| 283 | * |
| 284 | * @param[in] a First 8 bit fixed point input vector |
| 285 | * @param[in] b Second 8 bit fixed point input vector |
| 286 | * |
| 287 | * @return The result of the 8 bit fixed point vector addition |
| 288 | */ |
| 289 | qint8x16_t vaddq_qs8(qint8x16_t a, qint8x16_t b); |
| 290 | |
| 291 | /** 8 bit fixed point vector saturating add (8 elements) |
| 292 | * |
| 293 | * @param[in] a First 8 bit fixed point input vector |
| 294 | * @param[in] b Second 8 bit fixed point input vector |
| 295 | * |
| 296 | * @return The result of the 8 bit fixed point vector addition. The result is saturated in case of overflow |
| 297 | */ |
| 298 | qint8x8_t vqadd_qs8(qint8x8_t a, qint8x8_t b); |
| 299 | |
| 300 | /** 8 bit fixed point vector saturating add (16 elements) |
| 301 | * |
| 302 | * @param[in] a First 8 bit fixed point input vector |
| 303 | * @param[in] b Second 8 bit fixed point input vector |
| 304 | * |
| 305 | * @return The result of the 8 bit fixed point vector addition. The result is saturated in case of overflow |
| 306 | */ |
| 307 | qint8x16_t vqaddq_qs8(qint8x16_t a, qint8x16_t b); |
| 308 | |
| 309 | /** 16 bit fixed point vector saturating add (4 elements) |
| 310 | * |
| 311 | * @param[in] a First 16 bit fixed point input vector |
| 312 | * @param[in] b Second 16 bit fixed point input vector |
| 313 | * |
| 314 | * @return The result of the 16 bit fixed point vector addition. The result is saturated in case of overflow |
| 315 | */ |
| 316 | qint16x4_t vqadd_qs16(qint16x4_t a, qint16x4_t b); |
| 317 | |
| 318 | /** 16 bit fixed point vector saturating add (8 elements) |
| 319 | * |
| 320 | * @param[in] a First 16 bit fixed point input vector |
| 321 | * @param[in] b Second 16 bit fixed point input vector |
| 322 | * |
| 323 | * @return The result of the 16 bit fixed point vector addition. The result is saturated in case of overflow |
| 324 | */ |
| 325 | qint16x8_t vqaddq_qs16(qint16x8_t a, qint16x8_t b); |
| 326 | |
| 327 | /** 8 bit fixed point vector saturating pairwise add (8 elements) |
| 328 | * |
| 329 | * @param[in] a 8 bit fixed point input vector |
| 330 | * |
| 331 | * @return The result of the 16 bit fixed point vector addition. The result is saturated in case of overflow |
| 332 | */ |
| 333 | int16x4_t vpaddl_qs8(qint8x8_t a); |
| 334 | |
| 335 | /** 8 bit fixed point vector subtraction (8 elements) |
| 336 | * |
| 337 | * @param[in] a First 8 bit fixed point input vector |
| 338 | * @param[in] b Second 8 bit fixed point input vector |
| 339 | * |
| 340 | * @return The result of the 8 bit fixed point vector subtraction |
| 341 | */ |
| 342 | qint8x8_t vsub_qs8(qint8x8_t a, qint8x8_t b); |
| 343 | |
| 344 | /** 8 bit fixed point vector subtraction (16 elements) |
| 345 | * |
| 346 | * @param[in] a First 8 bit fixed point input vector |
| 347 | * @param[in] b Second 8 bit fixed point input vector |
| 348 | * |
| 349 | * @return The result of the 8 bit fixed point vector subtraction |
| 350 | */ |
| 351 | qint8x16_t vsubq_qs8(qint8x16_t a, qint8x16_t b); |
| 352 | |
| 353 | /** 8 bit fixed point vector saturating subtraction (8 elements) |
| 354 | * |
| 355 | * @param[in] a First 8 bit fixed point input vector |
| 356 | * @param[in] b Second 8 bit fixed point input vector |
| 357 | * |
| 358 | * @return The result of the 8 bit fixed point vector subtraction. The result is saturated in case of overflow |
| 359 | */ |
| 360 | qint8x8_t vqsub_qs8(qint8x8_t a, qint8x8_t b); |
| 361 | |
| 362 | /** 8 bit fixed point vector saturating subtraction (16 elements) |
| 363 | * |
| 364 | * @param[in] a First 8 bit fixed point input vector |
| 365 | * @param[in] b Second 8 bit fixed point input vector |
| 366 | * |
| 367 | * @return The result of the 8 bit fixed point vector subtraction. The result is saturated in case of overflow |
| 368 | */ |
| 369 | qint8x16_t vqsubq_qs8(qint8x16_t a, qint8x16_t b); |
| 370 | |
| 371 | /** 8 bit fixed point vector multiply (8 elements) |
| 372 | * |
| 373 | * @param[in] a First 8 bit fixed point input vector |
| 374 | * @param[in] b Second 8 bit fixed point input vector |
| 375 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 376 | * |
| 377 | * @return The result of the 8 bit fixed point vector multiplication. |
| 378 | */ |
| 379 | qint8x8_t vmul_qs8(qint8x8_t a, qint8x8_t b, int fixed_point_position); |
| 380 | |
| 381 | /** 8 bit fixed point vector multiply (16 elements) |
| 382 | * |
| 383 | * @param[in] a First 8 bit fixed point input vector |
| 384 | * @param[in] b Second 8 bit fixed point input vector |
| 385 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 386 | * |
| 387 | * @return The result of the 8 bit fixed point vector multiplication. |
| 388 | */ |
| 389 | qint8x16_t vmulq_qs8(qint8x16_t a, qint8x16_t b, int fixed_point_position); |
| 390 | |
| 391 | /** 8 bit fixed point vector saturating multiply (8 elements) |
| 392 | * |
| 393 | * @param[in] a First 8 bit fixed point input vector |
| 394 | * @param[in] b Second 8 bit fixed point input vector |
| 395 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 396 | * |
| 397 | * @return The result of the 8 bit fixed point vector multiplication. The result is saturated in case of overflow |
| 398 | */ |
| 399 | qint8x8_t vqmul_qs8(qint8x8_t a, qint8x8_t b, int fixed_point_position); |
| 400 | |
| 401 | /** 8 bit fixed point vector saturating multiply (16 elements) |
| 402 | * |
| 403 | * @param[in] a First 8 bit fixed point input vector |
| 404 | * @param[in] b Second 8 bit fixed point input vector |
| 405 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 406 | * |
| 407 | * @return The result of the 8 bit fixed point vector multiplication. The result is saturated in case of overflow |
| 408 | */ |
| 409 | qint8x16_t vqmulq_qs8(qint8x16_t a, qint8x16_t b, int fixed_point_position); |
| 410 | |
| 411 | /** 8 bit fixed point vector long multiply (8 elements) |
| 412 | * |
| 413 | * @param[in] a First 8 bit fixed point input vector |
| 414 | * @param[in] b Second 8 bit fixed point input vector |
| 415 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 416 | * |
| 417 | * @return The result of the 8 bit fixed point long vector multiplication. |
| 418 | */ |
| 419 | qint16x8_t vmull_qs8(qint8x8_t a, qint8x8_t b, int fixed_point_position); |
| 420 | |
| 421 | /** 8 bit fixed point vector multiply-accumulate (8 elements). This operation performs the product between @p b and @p c and add the result to @p a (a + b * c). |
| 422 | * |
| 423 | * @param[in] a First 8 bit fixed point input vector where the result of multiplication must be added to |
| 424 | * @param[in] b Second 8 bit fixed point input vector |
| 425 | * @param[in] c Third 8 bit fixed point input vector |
| 426 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 427 | * |
| 428 | * @return The result of the 8 bit fixed point vector multiply-accumulate |
| 429 | */ |
| 430 | qint8x8_t vmla_qs8(qint8x8_t a, qint8x8_t b, qint8x8_t c, int fixed_point_position); |
| 431 | |
| 432 | /** 8 bit fixed point vector multiply-accumulate (16 elements). This operation performs the product between @p b and @p c and add the result to @p a (a + b * c). |
| 433 | * |
| 434 | * @param[in] a First 8 bit fixed point input vector where the result of multiplication must be added to |
| 435 | * @param[in] b Second 8 bit fixed point input vector |
| 436 | * @param[in] c Third 8 bit fixed point input vector |
| 437 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 438 | * |
| 439 | * @return The result of the 8 bit fixed point vector multiply-accumulate |
| 440 | */ |
| 441 | qint8x16_t vmlaq_qs8(qint8x16_t a, qint8x16_t b, qint8x16_t c, int fixed_point_position); |
| 442 | |
| 443 | /** 8 bit fixed point vector saturating multiply-accumulate (8 elements). This operation performs the product between @p b and @p c and add the result to @p a (a + b * c). |
| 444 | * |
| 445 | * @param[in] a First 8 bit fixed point input vector where the result of multiplication must be added to |
| 446 | * @param[in] b Second 8 bit fixed point input vector |
| 447 | * @param[in] c Third 8 bit fixed point input vector |
| 448 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 449 | * |
| 450 | * @return The result of the 8 bit fixed point vector multiply-accumulate. The result is saturated in case of overflow |
| 451 | */ |
| 452 | qint8x8_t vqmla_qs8(qint8x8_t a, qint8x8_t b, qint8x8_t c, int fixed_point_position); |
| 453 | |
| 454 | /** 8 bit fixed point vector saturating multiply-accumulate (16 elements). This operation performs the product between @p b and @p c and add the result to @p a (a + b * c). |
| 455 | * |
| 456 | * @param[in] a First 8 bit fixed point input vector where the result of multiplication must be added to |
| 457 | * @param[in] b Second 8 bit fixed point input vector |
| 458 | * @param[in] c Third 8 bit fixed point input vector |
| 459 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 460 | * |
| 461 | * @return The result of the 8 bit fixed point vector multiply-accumulate.The result is saturated in case of overflow |
| 462 | */ |
| 463 | qint8x16_t vqmlaq_qs8(qint8x16_t a, qint8x16_t b, qint8x16_t c, int fixed_point_position); |
| 464 | |
| 465 | /** 8 bit fixed point vector multiply-accumulate long (8 elements). |
| 466 | * This operation performs the product between @p b and @p c and add the result to the 16 bit fixed point vector @p a (a + b * c). 8 elements |
| 467 | * |
| 468 | * @param[in] a First 16 bit fixed point input vector where the result of multiplication must be added to |
| 469 | * @param[in] b Second 8 bit fixed point input vector |
| 470 | * @param[in] c Third 8 bit fixed point input vector |
| 471 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 472 | * |
| 473 | * @return The result of the 8 bit fixed point vector multiply-accumulate long |
| 474 | */ |
| 475 | qint16x8_t vmlal_qs8(qint16x8_t a, qint8x8_t b, qint8x8_t c, int fixed_point_position); |
| 476 | |
| 477 | /** 8 bit fixed point vector saturating multiply-accumulate long (8 elements). The saturation is performed on the 16 bit fixed point output vector. |
| 478 | * This operation performs the product between @p b and @p c and add the result to the 16 bit fixed point vector @p a (a + b * c). 8 elements |
| 479 | * |
| 480 | * @param[in] a First 16 bit fixed point input vector where the result of multiplication must be added to |
| 481 | * @param[in] b Second 8 bit fixed point input vector |
| 482 | * @param[in] c Third 8 bit fixed point input vector |
| 483 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 484 | * |
| 485 | * @return The result of the 8 bit fixed point vector multiply-accumulate long |
| 486 | */ |
| 487 | qint16x8_t vqmlal_qs8(qint16x8_t a, qint8x8_t b, qint8x8_t c, int fixed_point_position); |
| 488 | |
| 489 | /** Convert a float vector with 4x2 elements to 8 bit fixed point vector with 8 elements |
| 490 | * |
| 491 | * @param[in] a Float input vector |
| 492 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 493 | * |
| 494 | * @return The result of the conversion float -> 8 bit fixed point |
| 495 | */ |
| 496 | qint8x8_t vcvt_qs8_f32(const float32x4x2_t &a, int fixed_point_position); |
| 497 | |
| 498 | /** Convert a float vector with 4x4 elements to 8 bit fixed point vector with 16 elements |
| 499 | * |
| 500 | * @param[in] a Float input vector |
| 501 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 502 | * |
| 503 | * @return The result of the conversion float -> 8 bit fixed point |
| 504 | */ |
| 505 | qint8x16_t vcvtq_qs8_f32(const float32x4x4_t &a, int fixed_point_position); |
| 506 | |
| 507 | /** Convert a 8 bit fixed point vector with 8 elements to a float vector with 4x2 elements |
| 508 | * |
| 509 | * @param[in] a 8 bit fixed point input vector |
| 510 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 511 | * |
| 512 | * @return The result of the conversion 8 bit fixed point -> float32x2x4 |
| 513 | */ |
| 514 | float32x4x2_t vcvt_f32_qs8(qint8x8_t a, int fixed_point_position); |
| 515 | |
| 516 | /** Convert a 8 bit fixed point vector with 16 elements to a float vector with 4x4 elements |
| 517 | * |
| 518 | * @param[in] a 8 bit fixed point input vector |
| 519 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 520 | * |
| 521 | * @return The result of the conversion 8 bit fixed point -> float32x4x4 |
| 522 | */ |
| 523 | float32x4x4_t vcvtq_qs8_f32(qint8x16_t a, int fixed_point_position); |
| 524 | |
| 525 | /** Calculate reciprocal of a fixed point 8bit number using the Newton-Raphson method. (8 elements) |
| 526 | * |
| 527 | * @param[in] a 8bit fixed point input vector |
| 528 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 529 | * |
| 530 | * @return The result of the 8bit reciprocal (1/a). |
| 531 | */ |
| 532 | qint8x8_t vrecip_qs8(qint8x8_t a, int fixed_point_position); |
| 533 | |
| 534 | /** Calculate reciprocal of a fixed point 8bit number using the Newton-Raphson method. (16 elements) |
| 535 | * |
| 536 | * @param[in] a 8bit fixed point input vector |
| 537 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 538 | * |
| 539 | * @return The result of the 8bit reciprocal (1/a). |
| 540 | */ |
| 541 | qint8x16_t vrecipq_qs8(qint8x16_t a, int fixed_point_position); |
| 542 | |
| 543 | /** Division fixed point 8bit (8 elements) |
| 544 | * |
| 545 | * @param[in] a First 8bit fixed point input vector |
| 546 | * @param[in] b Second 8bit fixed point input vector |
| 547 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 548 | * |
| 549 | * @return The quotient and remainder number in fixed point format. |
| 550 | */ |
| 551 | qint8x8_t vdiv_qs8(qint8x8_t a, int8x8_t b, int fixed_point_position); |
| 552 | |
| 553 | /** Division fixed point 8bit (16 elements) |
| 554 | * |
| 555 | * @param[in] a First 8bit fixed point input vector |
| 556 | * @param[in] b Second 8bit fixed point input vector |
| 557 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 558 | * |
| 559 | * @return The quotient and remainder number in 8bit fixed point format. |
| 560 | */ |
| 561 | qint8x16_t vdivq_qs8(qint8x16_t a, int8x16_t b, int fixed_point_position); |
| 562 | |
| 563 | /** Perform a 4th degree polynomial approximation. (8 elements) |
| 564 | * |
| 565 | * @param[in] a 8bit fixed point input vector |
| 566 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 567 | * |
| 568 | * @return The result of the 8bit taylor approximation. |
| 569 | */ |
| 570 | template <bool islog> |
| 571 | qint8x8_t vtaylor_poly_qs8(qint8x8_t a, int fixed_point_position); |
| 572 | |
| 573 | /** Perform a 4th degree polynomial approximation. (16 elements) |
| 574 | * |
| 575 | * @param[in] a 8bit fixed point input vector |
| 576 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 577 | * |
| 578 | * @return The result of the 8bit taylor approximation. |
| 579 | */ |
| 580 | template <bool islog> |
| 581 | qint8x16_t vtaylor_polyq_qs8(qint8x16_t a, int fixed_point_position); |
| 582 | |
| 583 | /** Calculate saturating exponential fixed point 8bit (8 elements) |
| 584 | * |
| 585 | * @param[in] a 8bit fixed point input vector |
| 586 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 587 | * |
| 588 | * @return The result of the 8bit saturating exponential |
| 589 | */ |
| 590 | qint8x8_t vqexp_qs8(qint8x8_t a, int fixed_point_position); |
| 591 | |
| 592 | /** Calculate saturating exponential fixed point 8bit (16 elements) |
| 593 | * |
| 594 | * @param[in] a 8bit fixed point input vector |
| 595 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 596 | * |
| 597 | * @return The result of the 8bit saturating exponential |
| 598 | */ |
| 599 | qint8x16_t vqexpq_qs8(qint8x16_t a, int fixed_point_position); |
| 600 | |
| 601 | /** Calculate logarithm fixed point 16bit (8 elements) |
| 602 | * |
| 603 | * @param[in] a 8bit fixed point input vector |
| 604 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 605 | * |
| 606 | * @return The result of the 8bit logarithm. |
| 607 | */ |
| 608 | qint8x8_t vlog_qs8(qint8x8_t a, int fixed_point_position); |
| 609 | |
| 610 | /** Calculate logarithm fixed point 16bit (16 elements) |
| 611 | * |
| 612 | * @param[in] a 8bit fixed point input vector |
| 613 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 614 | * |
| 615 | * @return The result of the 8bit logarithm. |
| 616 | */ |
| 617 | qint8x16_t vlogq_qs8(qint8x16_t a, int fixed_point_position); |
| 618 | |
| 619 | /** Calculate inverse square root for fixed point 8bit using Newton-Raphosn method (8 elements) |
| 620 | * |
| 621 | * @param[in] a 8bit fixed point input vector |
| 622 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 623 | * |
| 624 | * @return The result of the 8bit inverse sqrt. |
| 625 | */ |
| 626 | qint8x8_t vinvsqrt_qs8(qint8x8_t a, int fixed_point_position); |
| 627 | |
| 628 | /** Calculate saturating inverse square root for fixed point 8bit using Newton-Raphosn method (8 elements) |
| 629 | * |
| 630 | * @param[in] a 8bit fixed point input vector |
| 631 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 632 | * |
| 633 | * @return The result of the 8bit inverse sqrt. |
| 634 | */ |
| 635 | qint8x8_t vqinvsqrt_qs8(qint8x8_t a, int fixed_point_position); |
| 636 | |
| 637 | /** Calculate inverse square root for fixed point 8bit using Newton-Raphosn method (16 elements) |
| 638 | * |
| 639 | * @param[in] a 8bit fixed point input vector |
| 640 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 641 | * |
| 642 | * @return The result of the 8bit inverse sqrt. |
| 643 | */ |
| 644 | qint8x16_t vinvsqrtq_qs8(qint8x16_t a, int fixed_point_position); |
| 645 | |
| 646 | /** Calculate saturating inverse square root for fixed point 8bit using Newton-Raphosn method (16 elements) |
| 647 | * |
| 648 | * @param[in] a 8bit fixed point input vector |
| 649 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 650 | * |
| 651 | * @return The result of the 8bit inverse sqrt. |
| 652 | */ |
| 653 | qint8x16_t vqinvsqrtq_qs8(qint8x16_t a, int fixed_point_position); |
| 654 | |
| 655 | /** Calculate hyperbolic tangent for fixed point 8bit (8 elements) |
| 656 | * |
| 657 | * @param[in] a 8bit fixed point input vector |
| 658 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 659 | * |
| 660 | * @return The calculated Hyperbolic Tangent. |
| 661 | */ |
| 662 | qint8x8_t vtanh_qs8(qint8x8_t a, int fixed_point_position); |
| 663 | |
| 664 | /** Calculate hyperbolic tangent for fixed point 8bit (16 elements) |
| 665 | * |
| 666 | * @param[in] a 8bit fixed point input vector |
| 667 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 668 | * |
| 669 | * @return The calculated Hyperbolic Tangent. |
| 670 | */ |
| 671 | qint8x16_t vtanhq_qs8(qint8x16_t a, int fixed_point_position); |
| 672 | |
| 673 | /** Calculate saturating n power for fixed point 8bit (16 elements). |
| 674 | * |
| 675 | * pow(a,b) = e^(b*log(a)) |
| 676 | * |
| 677 | * @param[in] a 8bit fixed point input vector |
| 678 | * @param[in] b 8bit fixed point power vector |
| 679 | * @param[in] fixed_point_position Fixed point position that expresses the number of bits for the fractional part of the number |
| 680 | * |
| 681 | * @return The result of the 8bit power. |
| 682 | */ |
| 683 | qint8x8_t vqpowq_qs8(qint8x8_t a, qint8x16_t b, int fixed_point_position); |
| 684 | } |
| 685 | #include "arm_compute/core/NEON/NEFixedPoint.inl" |
| 686 | #endif /* __ARM_COMPUTE_NEFIXEDPOINT_H__ */ |