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_TEST_VALIDATION_FIXEDPOINT_H__ |
| 25 | #define __ARM_COMPUTE_TEST_VALIDATION_FIXEDPOINT_H__ |
| 26 | |
| 27 | #include "Utils.h" |
Moritz Pflanzer | d0ae8b8 | 2017-06-29 14:51:57 +0100 | [diff] [blame] | 28 | #include "support/ToolchainSupport.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 29 | |
| 30 | #include <cassert> |
| 31 | #include <cstdint> |
| 32 | #include <cstdlib> |
| 33 | #include <limits> |
| 34 | #include <string> |
| 35 | #include <type_traits> |
| 36 | |
| 37 | namespace arm_compute |
| 38 | { |
| 39 | namespace test |
| 40 | { |
| 41 | namespace fixed_point_arithmetic |
| 42 | { |
| 43 | namespace detail |
| 44 | { |
| 45 | // Forward declare structs |
| 46 | struct functions; |
| 47 | template <typename T> |
| 48 | struct constant_expr; |
| 49 | } |
| 50 | |
| 51 | /** Fixed point traits */ |
| 52 | namespace traits |
| 53 | { |
| 54 | // Promote types |
| 55 | // *INDENT-OFF* |
| 56 | // clang-format off |
| 57 | template <typename T> struct promote { }; |
| 58 | template <> struct promote<uint8_t> { using type = uint16_t; }; |
| 59 | template <> struct promote<int8_t> { using type = int16_t; }; |
| 60 | template <> struct promote<uint16_t> { using type = uint32_t; }; |
| 61 | template <> struct promote<int16_t> { using type = int32_t; }; |
| 62 | template <> struct promote<uint32_t> { using type = uint64_t; }; |
| 63 | template <> struct promote<int32_t> { using type = int64_t; }; |
| 64 | template <> struct promote<uint64_t> { using type = uint64_t; }; |
| 65 | template <> struct promote<int64_t> { using type = int64_t; }; |
| 66 | // clang-format on |
| 67 | // *INDENT-ON* |
| 68 | } |
| 69 | |
| 70 | /** Strongly typed enum class representing the overflow policy */ |
| 71 | enum class OverflowPolicy |
| 72 | { |
| 73 | WRAP, /**< Wrap policy */ |
| 74 | SATURATE /**< Saturate policy */ |
| 75 | }; |
| 76 | /** Strongly typed enum class representing the rounding policy */ |
| 77 | enum class RoundingPolicy |
| 78 | { |
| 79 | TO_ZERO, /**< Round to zero policy */ |
| 80 | TO_NEAREST_EVEN /**< Round to nearest even policy */ |
| 81 | }; |
| 82 | |
| 83 | /** Arbitrary fixed-point arithmetic class */ |
| 84 | template <typename T> |
| 85 | class fixed_point |
| 86 | { |
| 87 | public: |
| 88 | // Static Checks |
| 89 | static_assert(std::is_integral<T>::value, "Type is not an integer"); |
| 90 | |
| 91 | // Friends |
| 92 | friend struct detail::functions; |
| 93 | friend struct detail::constant_expr<T>; |
| 94 | |
| 95 | /** Constructor (from different fixed point type) |
| 96 | * |
| 97 | * @param[in] val Fixed point |
| 98 | * @param[in] p Fixed point precision |
| 99 | */ |
| 100 | template <typename U> |
| 101 | fixed_point(fixed_point<U> val, uint8_t p) |
| 102 | : _value(0), _fixed_point_position(p) |
| 103 | { |
| 104 | assert(p > 0 && p < std::numeric_limits<T>::digits); |
| 105 | T v = 0; |
| 106 | |
| 107 | if(std::numeric_limits<T>::digits < std::numeric_limits<U>::digits) |
| 108 | { |
| 109 | val.rescale(p); |
| 110 | v = detail::constant_expr<T>::saturate_cast(val.raw()); |
| 111 | } |
| 112 | else |
| 113 | { |
| 114 | auto v_cast = static_cast<fixed_point<T>>(val); |
| 115 | v_cast.rescale(p); |
| 116 | v = v_cast.raw(); |
| 117 | } |
| 118 | _value = static_cast<T>(v); |
| 119 | } |
| 120 | /** Constructor (from integer) |
| 121 | * |
| 122 | * @param[in] val Integer value to be represented as fixed point |
| 123 | * @param[in] p Fixed point precision |
| 124 | * @param[in] is_raw If true val is a raw fixed point value else an integer |
| 125 | */ |
| 126 | template <typename U, typename = typename std::enable_if<std::is_integral<U>::value>::type> |
| 127 | fixed_point(U val, uint8_t p, bool is_raw = false) |
| 128 | : _value(val << p), _fixed_point_position(p) |
| 129 | { |
| 130 | if(is_raw) |
| 131 | { |
| 132 | _value = val; |
| 133 | } |
| 134 | } |
| 135 | /** Constructor (from float) |
| 136 | * |
| 137 | * @param[in] val Float value to be represented as fixed point |
| 138 | * @param[in] p Fixed point precision |
| 139 | */ |
| 140 | fixed_point(float val, uint8_t p) |
| 141 | : _value(detail::constant_expr<T>::to_fixed(val, p)), _fixed_point_position(p) |
| 142 | { |
| 143 | assert(p > 0 && p < std::numeric_limits<T>::digits); |
| 144 | } |
| 145 | /** Constructor (from float string) |
| 146 | * |
| 147 | * @param[in] str Float string to be represented as fixed point |
| 148 | * @param[in] p Fixed point precision |
| 149 | */ |
| 150 | fixed_point(std::string str, uint8_t p) |
Moritz Pflanzer | d0ae8b8 | 2017-06-29 14:51:57 +0100 | [diff] [blame] | 151 | : _value(detail::constant_expr<T>::to_fixed(support::cpp11::stof(str), p)), _fixed_point_position(p) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 152 | { |
| 153 | assert(p > 0 && p < std::numeric_limits<T>::digits); |
| 154 | } |
| 155 | /** Default copy constructor */ |
| 156 | fixed_point &operator=(const fixed_point &) = default; |
| 157 | /** Default move constructor */ |
| 158 | fixed_point &operator=(fixed_point &&) = default; |
| 159 | /** Default copy assignment operator */ |
| 160 | fixed_point(const fixed_point &) = default; |
| 161 | /** Default move assignment operator */ |
| 162 | fixed_point(fixed_point &&) = default; |
| 163 | |
| 164 | /** Float conversion operator |
| 165 | * |
| 166 | * @return Float representation of fixed point |
| 167 | */ |
| 168 | operator float() const |
| 169 | { |
| 170 | return detail::constant_expr<T>::to_float(_value, _fixed_point_position); |
| 171 | } |
| 172 | /** Integer conversion operator |
| 173 | * |
| 174 | * @return Integer representation of fixed point |
| 175 | */ |
| 176 | template <typename U, typename = typename std::enable_if<std::is_integral<T>::value>::type> |
| 177 | operator U() const |
| 178 | { |
| 179 | return detail::constant_expr<T>::to_int(_value, _fixed_point_position); |
| 180 | } |
| 181 | /** Convert to different fixed point of different type but same precision |
| 182 | * |
| 183 | * @note Down-conversion might fail. |
| 184 | */ |
| 185 | template <typename U> |
| 186 | operator fixed_point<U>() |
| 187 | { |
| 188 | U val = static_cast<U>(_value); |
| 189 | if(std::numeric_limits<U>::digits < std::numeric_limits<T>::digits) |
| 190 | { |
| 191 | val = detail::constant_expr<U>::saturate_cast(_value); |
| 192 | } |
| 193 | return fixed_point<U>(val, _fixed_point_position, true); |
| 194 | } |
| 195 | |
| 196 | /** Arithmetic += assignment operator |
| 197 | * |
| 198 | * @param[in] rhs Fixed point operand |
| 199 | * |
| 200 | * @return Reference to this fixed point |
| 201 | */ |
| 202 | template <typename U> |
| 203 | fixed_point<T> &operator+=(const fixed_point<U> &rhs) |
| 204 | { |
| 205 | fixed_point<T> val(rhs, _fixed_point_position); |
| 206 | _value += val.raw(); |
| 207 | return *this; |
| 208 | } |
| 209 | /** Arithmetic -= assignment operator |
| 210 | * |
| 211 | * @param[in] rhs Fixed point operand |
| 212 | * |
| 213 | * @return Reference to this fixed point |
| 214 | */ |
| 215 | template <typename U> |
| 216 | fixed_point<T> &operator-=(const fixed_point<U> &rhs) |
| 217 | { |
| 218 | fixed_point<T> val(rhs, _fixed_point_position); |
| 219 | _value -= val.raw(); |
| 220 | return *this; |
| 221 | } |
| 222 | |
| 223 | /** Raw value accessor |
| 224 | * |
| 225 | * @return Raw fixed point value |
| 226 | */ |
| 227 | T raw() const |
| 228 | { |
| 229 | return _value; |
| 230 | } |
| 231 | /** Precision accessor |
| 232 | * |
| 233 | * @return Precision of fixed point |
| 234 | */ |
| 235 | uint8_t precision() const |
| 236 | { |
| 237 | return _fixed_point_position; |
| 238 | } |
| 239 | /** Rescale a fixed point to a new precision |
| 240 | * |
| 241 | * @param[in] p New fixed point precision |
| 242 | */ |
| 243 | void rescale(uint8_t p) |
| 244 | { |
| 245 | assert(p > 0 && p < std::numeric_limits<T>::digits); |
| 246 | |
| 247 | if(p > _fixed_point_position) |
| 248 | { |
| 249 | _value <<= (p - _fixed_point_position); |
| 250 | } |
| 251 | else if(p < _fixed_point_position) |
| 252 | { |
| 253 | _value >>= (_fixed_point_position - p); |
| 254 | } |
| 255 | |
| 256 | _fixed_point_position = p; |
| 257 | } |
| 258 | |
| 259 | private: |
| 260 | T _value; /**< Fixed point raw value */ |
| 261 | uint8_t _fixed_point_position; /**< Fixed point precision */ |
| 262 | }; |
| 263 | |
| 264 | namespace detail |
| 265 | { |
| 266 | /** Count the number of leading zero bits in the given value. |
| 267 | * |
| 268 | * @param[in] value Input value. |
| 269 | * |
| 270 | * @return Number of leading zero bits. |
| 271 | */ |
| 272 | template <typename T> |
| 273 | constexpr int clz(T value) |
| 274 | { |
| 275 | using unsigned_T = typename std::make_unsigned<T>::type; |
| 276 | // __builtin_clz is available for int. Need to correct reported number to |
| 277 | // match the original type. |
| 278 | return __builtin_clz(value) - (32 - std::numeric_limits<unsigned_T>::digits); |
| 279 | } |
| 280 | |
| 281 | template <typename T> |
| 282 | struct constant_expr |
| 283 | { |
| 284 | /** Calculate representation of 1 in fixed point given a fixed point precision |
| 285 | * |
| 286 | * @param[in] p Fixed point precision |
| 287 | * |
| 288 | * @return Representation of value 1 in fixed point. |
| 289 | */ |
| 290 | static constexpr T fixed_one(uint8_t p) |
| 291 | { |
| 292 | return (1 << p); |
| 293 | } |
| 294 | /** Calculate fixed point precision step given a fixed point precision |
| 295 | * |
| 296 | * @param[in] p Fixed point precision |
| 297 | * |
| 298 | * @return Fixed point precision step |
| 299 | */ |
| 300 | static constexpr float fixed_step(uint8_t p) |
| 301 | { |
| 302 | return (1.0f / static_cast<float>(1 << p)); |
| 303 | } |
| 304 | |
| 305 | /** Convert a fixed point value to float given its precision. |
| 306 | * |
| 307 | * @param[in] val Fixed point value |
| 308 | * @param[in] p Fixed point precision |
| 309 | * |
| 310 | * @return Float representation of the fixed point number |
| 311 | */ |
| 312 | static constexpr float to_float(T val, uint8_t p) |
| 313 | { |
| 314 | return static_cast<float>(val * fixed_step(p)); |
| 315 | } |
| 316 | /** Convert a fixed point value to integer given its precision. |
| 317 | * |
| 318 | * @param[in] val Fixed point value |
| 319 | * @param[in] p Fixed point precision |
| 320 | * |
| 321 | * @return Integer of the fixed point number |
| 322 | */ |
| 323 | static constexpr T to_int(T val, uint8_t p) |
| 324 | { |
| 325 | return val >> p; |
| 326 | } |
| 327 | /** Convert a single precision floating point value to a fixed point representation given its precision. |
| 328 | * |
| 329 | * @param[in] val Floating point value |
| 330 | * @param[in] p Fixed point precision |
| 331 | * |
| 332 | * @return The raw fixed point representation |
| 333 | */ |
| 334 | static constexpr T to_fixed(float val, uint8_t p) |
| 335 | { |
Georgios Pinitas | 21efeb4 | 2017-07-04 12:47:17 +0100 | [diff] [blame] | 336 | return static_cast<T>(saturate_cast<float>(val * fixed_one(p) + ((val >= 0) ? 0.5 : -0.5))); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 337 | } |
| 338 | /** Clamp value between two ranges |
| 339 | * |
| 340 | * @param[in] val Value to clamp |
| 341 | * @param[in] min Minimum value to clamp to |
| 342 | * @param[in] max Maximum value to clamp to |
| 343 | * |
| 344 | * @return clamped value |
| 345 | */ |
| 346 | static constexpr T clamp(T val, T min, T max) |
| 347 | { |
| 348 | return std::min(std::max(val, min), max); |
| 349 | } |
| 350 | /** Saturate given number |
| 351 | * |
| 352 | * @param[in] val Value to saturate |
| 353 | * |
| 354 | * @return Saturated value |
| 355 | */ |
| 356 | template <typename U> |
| 357 | static constexpr T saturate_cast(U val) |
| 358 | { |
| 359 | return static_cast<T>(std::min<U>(std::max<U>(val, static_cast<U>(std::numeric_limits<T>::min())), static_cast<U>(std::numeric_limits<T>::max()))); |
| 360 | } |
| 361 | }; |
| 362 | struct functions |
| 363 | { |
| 364 | /** Output stream operator |
| 365 | * |
| 366 | * @param[in] s Output stream |
| 367 | * @param[in] x Fixed point value |
| 368 | * |
| 369 | * @return Reference output to updated stream |
| 370 | */ |
| 371 | template <typename T, typename U, typename traits> |
| 372 | static std::basic_ostream<T, traits> &write(std::basic_ostream<T, traits> &s, fixed_point<U> &x) |
| 373 | { |
| 374 | return s << static_cast<float>(x); |
| 375 | } |
| 376 | /** Signbit of a fixed point number. |
| 377 | * |
| 378 | * @param[in] x Fixed point number |
| 379 | * |
| 380 | * @return True if negative else false. |
| 381 | */ |
| 382 | template <typename T> |
| 383 | static bool signbit(fixed_point<T> x) |
| 384 | { |
| 385 | return ((x._value >> std::numeric_limits<T>::digits) != 0); |
| 386 | } |
| 387 | /** Checks if two fixed point numbers are equal |
| 388 | * |
| 389 | * @param[in] x First fixed point operand |
| 390 | * @param[in] y Second fixed point operand |
| 391 | * |
| 392 | * @return True if fixed points are equal else false |
| 393 | */ |
| 394 | template <typename T> |
| 395 | static bool isequal(fixed_point<T> x, fixed_point<T> y) |
| 396 | { |
| 397 | uint8_t p = std::min(x._fixed_point_position, y._fixed_point_position); |
| 398 | x.rescale(p); |
| 399 | y.rescale(p); |
| 400 | return (x._value == y._value); |
| 401 | } |
| 402 | /** Checks if two fixed point number are not equal |
| 403 | * |
| 404 | * @param[in] x First fixed point operand |
| 405 | * @param[in] y Second fixed point operand |
| 406 | * |
| 407 | * @return True if fixed points are not equal else false |
| 408 | */ |
| 409 | template <typename T> |
| 410 | static bool isnotequal(fixed_point<T> x, fixed_point<T> y) |
| 411 | { |
| 412 | return !isequal(x, y); |
| 413 | } |
| 414 | /** Checks if one fixed point is greater than the other |
| 415 | * |
| 416 | * @param[in] x First fixed point operand |
| 417 | * @param[in] y Second fixed point operand |
| 418 | * |
| 419 | * @return True if fixed point is greater than other |
| 420 | */ |
| 421 | template <typename T> |
| 422 | static bool isgreater(fixed_point<T> x, fixed_point<T> y) |
| 423 | { |
| 424 | uint8_t p = std::min(x._fixed_point_position, y._fixed_point_position); |
| 425 | x.rescale(p); |
| 426 | y.rescale(p); |
| 427 | return (x._value > y._value); |
| 428 | } |
| 429 | /** Checks if one fixed point is greater or equal than the other |
| 430 | * |
| 431 | * @param[in] x First fixed point operand |
| 432 | * @param[in] y Second fixed point operand |
| 433 | * |
| 434 | * @return True if fixed point is greater or equal than other |
| 435 | */ |
| 436 | template <typename T> |
| 437 | static bool isgreaterequal(fixed_point<T> x, fixed_point<T> y) |
| 438 | { |
| 439 | uint8_t p = std::min(x._fixed_point_position, y._fixed_point_position); |
| 440 | x.rescale(p); |
| 441 | y.rescale(p); |
| 442 | return (x._value >= y._value); |
| 443 | } |
| 444 | /** Checks if one fixed point is less than the other |
| 445 | * |
| 446 | * @param[in] x First fixed point operand |
| 447 | * @param[in] y Second fixed point operand |
| 448 | * |
| 449 | * @return True if fixed point is less than other |
| 450 | */ |
| 451 | template <typename T> |
| 452 | static bool isless(fixed_point<T> x, fixed_point<T> y) |
| 453 | { |
| 454 | uint8_t p = std::min(x._fixed_point_position, y._fixed_point_position); |
| 455 | x.rescale(p); |
| 456 | y.rescale(p); |
| 457 | return (x._value < y._value); |
| 458 | } |
| 459 | /** Checks if one fixed point is less or equal than the other |
| 460 | * |
| 461 | * @param[in] x First fixed point operand |
| 462 | * @param[in] y Second fixed point operand |
| 463 | * |
| 464 | * @return True if fixed point is less or equal than other |
| 465 | */ |
| 466 | template <typename T> |
| 467 | static bool islessequal(fixed_point<T> x, fixed_point<T> y) |
| 468 | { |
| 469 | uint8_t p = std::min(x._fixed_point_position, y._fixed_point_position); |
| 470 | x.rescale(p); |
| 471 | y.rescale(p); |
| 472 | return (x._value <= y._value); |
| 473 | } |
| 474 | /** Checks if one fixed point is less or greater than the other |
| 475 | * |
| 476 | * @param[in] x First fixed point operand |
| 477 | * @param[in] y Second fixed point operand |
| 478 | * |
| 479 | * @return True if fixed point is less or greater than other |
| 480 | */ |
| 481 | template <typename T> |
| 482 | static bool islessgreater(fixed_point<T> x, fixed_point<T> y) |
| 483 | { |
| 484 | return isnotequal(x, y); |
| 485 | } |
| 486 | /** Clamp fixed point to specific range. |
| 487 | * |
| 488 | * @param[in] x Fixed point operand |
| 489 | * @param[in] min Minimum value to clamp to |
| 490 | * @param[in] max Maximum value to clamp to |
| 491 | * |
| 492 | * @return Clamped result |
| 493 | */ |
| 494 | template <typename T> |
| 495 | static fixed_point<T> clamp(fixed_point<T> x, T min, T max) |
| 496 | { |
| 497 | return fixed_point<T>(constant_expr<T>::clamp(x._value, min, max), x._fixed_point_position, true); |
| 498 | } |
| 499 | /** Negate number |
| 500 | * |
| 501 | * @param[in] x Fixed point operand |
| 502 | * |
| 503 | * @return Negated fixed point result |
| 504 | */ |
| 505 | template <OverflowPolicy OP = OverflowPolicy::SATURATE, typename T> |
| 506 | static fixed_point<T> negate(fixed_point<T> x) |
| 507 | { |
| 508 | using promoted_T = typename traits::promote<T>::type; |
| 509 | promoted_T val = -x._value; |
| 510 | if(OP == OverflowPolicy::SATURATE) |
| 511 | { |
| 512 | val = constant_expr<T>::saturate_cast(val); |
| 513 | } |
| 514 | return fixed_point<T>(static_cast<T>(val), x._fixed_point_position, true); |
| 515 | } |
| 516 | /** Perform addition among two fixed point numbers |
| 517 | * |
| 518 | * @param[in] x First fixed point operand |
| 519 | * @param[in] y Second fixed point operand |
| 520 | * |
| 521 | * @return Result fixed point with precision equal to minimum precision of both operands |
| 522 | */ |
| 523 | template <OverflowPolicy OP = OverflowPolicy::SATURATE, typename T> |
| 524 | static fixed_point<T> add(fixed_point<T> x, fixed_point<T> y) |
| 525 | { |
| 526 | uint8_t p = std::min(x._fixed_point_position, y._fixed_point_position); |
| 527 | x.rescale(p); |
| 528 | y.rescale(p); |
| 529 | if(OP == OverflowPolicy::SATURATE) |
| 530 | { |
| 531 | using type = typename traits::promote<T>::type; |
| 532 | type val = static_cast<type>(x._value) + static_cast<type>(y._value); |
| 533 | val = constant_expr<T>::saturate_cast(val); |
| 534 | return fixed_point<T>(static_cast<T>(val), p, true); |
| 535 | } |
| 536 | else |
| 537 | { |
| 538 | return fixed_point<T>(x._value + y._value, p, true); |
| 539 | } |
| 540 | } |
| 541 | /** Perform subtraction among two fixed point numbers |
| 542 | * |
| 543 | * @param[in] x First fixed point operand |
| 544 | * @param[in] y Second fixed point operand |
| 545 | * |
| 546 | * @return Result fixed point with precision equal to minimum precision of both operands |
| 547 | */ |
| 548 | template <OverflowPolicy OP = OverflowPolicy::SATURATE, typename T> |
| 549 | static fixed_point<T> sub(fixed_point<T> x, fixed_point<T> y) |
| 550 | { |
| 551 | uint8_t p = std::min(x._fixed_point_position, y._fixed_point_position); |
| 552 | x.rescale(p); |
| 553 | y.rescale(p); |
| 554 | if(OP == OverflowPolicy::SATURATE) |
| 555 | { |
| 556 | using type = typename traits::promote<T>::type; |
| 557 | type val = static_cast<type>(x._value) - static_cast<type>(y._value); |
| 558 | val = constant_expr<T>::saturate_cast(val); |
| 559 | return fixed_point<T>(static_cast<T>(val), p, true); |
| 560 | } |
| 561 | else |
| 562 | { |
| 563 | return fixed_point<T>(x._value - y._value, p, true); |
| 564 | } |
| 565 | } |
| 566 | /** Perform multiplication among two fixed point numbers |
| 567 | * |
| 568 | * @param[in] x First fixed point operand |
| 569 | * @param[in] y Second fixed point operand |
| 570 | * |
| 571 | * @return Result fixed point with precision equal to minimum precision of both operands |
| 572 | */ |
| 573 | template <OverflowPolicy OP = OverflowPolicy::SATURATE, typename T> |
| 574 | static fixed_point<T> mul(fixed_point<T> x, fixed_point<T> y) |
| 575 | { |
| 576 | using promoted_T = typename traits::promote<T>::type; |
| 577 | uint8_t p_min = std::min(x._fixed_point_position, y._fixed_point_position); |
| 578 | uint8_t p_max = std::max(x._fixed_point_position, y._fixed_point_position); |
| 579 | promoted_T round_factor = (1 << (p_max - 1)); |
| 580 | promoted_T val = ((static_cast<promoted_T>(x._value) * static_cast<promoted_T>(y._value)) + round_factor) >> p_max; |
| 581 | if(OP == OverflowPolicy::SATURATE) |
| 582 | { |
| 583 | val = constant_expr<T>::saturate_cast(val); |
| 584 | } |
| 585 | return fixed_point<T>(static_cast<T>(val), p_min, true); |
| 586 | } |
| 587 | /** Perform division among two fixed point numbers |
| 588 | * |
| 589 | * @param[in] x First fixed point operand |
| 590 | * @param[in] y Second fixed point operand |
| 591 | * |
| 592 | * @return Result fixed point with precision equal to minimum precision of both operands |
| 593 | */ |
| 594 | template <OverflowPolicy OP = OverflowPolicy::SATURATE, typename T> |
| 595 | static fixed_point<T> div(fixed_point<T> x, fixed_point<T> y) |
| 596 | { |
| 597 | using promoted_T = typename traits::promote<T>::type; |
| 598 | uint8_t p = std::min(x._fixed_point_position, y._fixed_point_position); |
| 599 | promoted_T denom = static_cast<promoted_T>(y._value); |
| 600 | if(denom != 0) |
| 601 | { |
| 602 | promoted_T val = (static_cast<promoted_T>(x._value) << std::max(x._fixed_point_position, y._fixed_point_position)) / denom; |
| 603 | if(OP == OverflowPolicy::SATURATE) |
| 604 | { |
| 605 | val = constant_expr<T>::saturate_cast(val); |
| 606 | } |
| 607 | return fixed_point<T>(static_cast<T>(val), p, true); |
| 608 | } |
| 609 | else |
| 610 | { |
| 611 | T val = (x._value < 0) ? std::numeric_limits<T>::min() : std::numeric_limits<T>::max(); |
| 612 | return fixed_point<T>(val, p, true); |
| 613 | } |
| 614 | } |
| 615 | /** Shift left |
| 616 | * |
| 617 | * @param[in] x Fixed point operand |
| 618 | * @param[in] shift Shift value |
| 619 | * |
| 620 | * @return Shifted value |
| 621 | */ |
| 622 | template <OverflowPolicy OP = OverflowPolicy::SATURATE, typename T> |
| 623 | static fixed_point<T> shift_left(fixed_point<T> x, size_t shift) |
| 624 | { |
| 625 | using promoted_T = typename traits::promote<T>::type; |
| 626 | promoted_T val = static_cast<promoted_T>(x._value) << shift; |
| 627 | if(OP == OverflowPolicy::SATURATE) |
| 628 | { |
| 629 | val = constant_expr<T>::saturate_cast(val); |
| 630 | } |
| 631 | return fixed_point<T>(static_cast<T>(val), x._fixed_point_position, true); |
| 632 | } |
| 633 | /** Shift right |
| 634 | * |
| 635 | * @param[in] x Fixed point operand |
| 636 | * @param[in] shift Shift value |
| 637 | * |
| 638 | * @return Shifted value |
| 639 | */ |
| 640 | template <typename T> |
| 641 | static fixed_point<T> shift_right(fixed_point<T> x, size_t shift) |
| 642 | { |
| 643 | return fixed_point<T>(x._value >> shift, x._fixed_point_position, true); |
| 644 | } |
| 645 | /** Calculate absolute value |
| 646 | * |
| 647 | * @param[in] x Fixed point operand |
| 648 | * |
| 649 | * @return Absolute value of operand |
| 650 | */ |
| 651 | template <typename T> |
| 652 | static fixed_point<T> abs(fixed_point<T> x) |
| 653 | { |
| 654 | using promoted_T = typename traits::promote<T>::type; |
| 655 | T val = (x._value < 0) ? constant_expr<T>::saturate_cast(-static_cast<promoted_T>(x._value)) : x._value; |
| 656 | return fixed_point<T>(val, x._fixed_point_position, true); |
| 657 | } |
| 658 | /** Calculate the logarithm of a fixed point number |
| 659 | * |
| 660 | * @param[in] x Fixed point operand |
| 661 | * |
| 662 | * @return Logarithm value of operand |
| 663 | */ |
| 664 | template <typename T> |
| 665 | static fixed_point<T> log(fixed_point<T> x) |
| 666 | { |
| 667 | uint8_t p = x._fixed_point_position; |
| 668 | auto const_one = fixed_point<T>(static_cast<T>(1), p); |
| 669 | |
| 670 | // Logarithm of 1 is zero and logarithm of negative values is not defined in R, so return 0. |
| 671 | // Also, log(x) == -log(1/x) for 0 < x < 1. |
| 672 | if(isequal(x, const_one) || islessequal(x, fixed_point<T>(static_cast<T>(0), p))) |
| 673 | { |
| 674 | return fixed_point<T>(static_cast<T>(0), p, true); |
| 675 | } |
| 676 | else if(isless(x, const_one)) |
| 677 | { |
| 678 | return mul(log(div(const_one, x)), fixed_point<T>(-1, p)); |
| 679 | } |
| 680 | |
| 681 | // Remove even powers of 2 |
| 682 | T shift_val = 31 - __builtin_clz(x._value >> p); |
| 683 | x = shift_right(x, shift_val); |
| 684 | x = sub(x, const_one); |
| 685 | |
| 686 | // Constants |
| 687 | auto ln2 = fixed_point<T>(0.6931471, p); |
| 688 | auto A = fixed_point<T>(1.4384189, p); |
| 689 | auto B = fixed_point<T>(-0.67719, p); |
| 690 | auto C = fixed_point<T>(0.3218538, p); |
| 691 | auto D = fixed_point<T>(-0.0832229, p); |
| 692 | |
| 693 | // Polynomial expansion |
| 694 | auto sum = add(mul(x, D), C); |
| 695 | sum = add(mul(x, sum), B); |
| 696 | sum = add(mul(x, sum), A); |
| 697 | sum = mul(x, sum); |
| 698 | |
| 699 | return mul(add(sum, fixed_point<T>(static_cast<T>(shift_val), p)), ln2); |
| 700 | } |
| 701 | /** Calculate the exponential of a fixed point number. |
| 702 | * |
| 703 | * exp(x) = exp(floor(x)) * exp(x - floor(x)) |
| 704 | * = pow(2, floor(x) / ln(2)) * exp(x - floor(x)) |
| 705 | * = exp(x - floor(x)) << (floor(x) / ln(2)) |
| 706 | * |
| 707 | * @param[in] x Fixed point operand |
| 708 | * |
| 709 | * @return Exponential value of operand |
| 710 | */ |
| 711 | template <typename T> |
| 712 | static fixed_point<T> exp(fixed_point<T> x) |
| 713 | { |
| 714 | uint8_t p = x._fixed_point_position; |
| 715 | // Constants |
| 716 | auto const_one = fixed_point<T>(1, p); |
| 717 | auto ln2 = fixed_point<T>(0.6931471, p); |
| 718 | auto inv_ln2 = fixed_point<T>(1.442695, p); |
| 719 | auto A = fixed_point<T>(0.9978546, p); |
| 720 | auto B = fixed_point<T>(0.4994721, p); |
| 721 | auto C = fixed_point<T>(0.1763723, p); |
| 722 | auto D = fixed_point<T>(0.0435108, p); |
| 723 | |
| 724 | T scaled_int_part = detail::constant_expr<T>::to_int(mul(x, inv_ln2)._value, p); |
| 725 | |
| 726 | // Polynomial expansion |
| 727 | auto frac_part = sub(x, mul(ln2, fixed_point<T>(scaled_int_part, p))); |
| 728 | auto taylor = add(mul(frac_part, D), C); |
| 729 | taylor = add(mul(frac_part, taylor), B); |
| 730 | taylor = add(mul(frac_part, taylor), A); |
| 731 | taylor = mul(frac_part, taylor); |
| 732 | taylor = add(taylor, const_one); |
| 733 | |
| 734 | // Saturate value |
| 735 | if(static_cast<T>(clz(taylor.raw())) <= scaled_int_part) |
| 736 | { |
| 737 | return fixed_point<T>(std::numeric_limits<T>::max(), p, true); |
| 738 | } |
| 739 | |
| 740 | return (scaled_int_part < 0) ? shift_right(taylor, -scaled_int_part) : shift_left(taylor, scaled_int_part); |
| 741 | } |
| 742 | /** Calculate the inverse square root of a fixed point number |
| 743 | * |
| 744 | * @param[in] x Fixed point operand |
| 745 | * |
| 746 | * @return Inverse square root value of operand |
| 747 | */ |
| 748 | template <typename T> |
| 749 | static fixed_point<T> inv_sqrt(fixed_point<T> x) |
| 750 | { |
| 751 | const uint8_t p = x._fixed_point_position; |
| 752 | int8_t shift = std::numeric_limits<T>::digits - (p + detail::clz(x._value)); |
| 753 | |
| 754 | shift += std::numeric_limits<T>::is_signed ? 1 : 0; |
| 755 | |
Georgios Pinitas | 6410fb2 | 2017-07-03 14:38:50 +0100 | [diff] [blame] | 756 | // Use volatile to restrict compiler optimizations on shift as compiler reports maybe-uninitialized error on Android |
| 757 | volatile int8_t *shift_ptr = &shift; |
| 758 | |
| 759 | auto const_three = fixed_point<T>(3, p); |
| 760 | auto a = (*shift_ptr < 0) ? shift_left(x, -(shift)) : shift_right(x, shift); |
| 761 | fixed_point<T> x2 = a; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 762 | |
Michalis Spyrou | 0a8334c | 2017-06-14 18:00:05 +0100 | [diff] [blame] | 763 | // We need three iterations to find the result for QS8 and five for QS16 |
| 764 | constexpr int num_iterations = std::is_same<T, int8_t>::value ? 3 : 5; |
| 765 | for(int i = 0; i < num_iterations; ++i) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 766 | { |
Georgios Pinitas | 6410fb2 | 2017-07-03 14:38:50 +0100 | [diff] [blame] | 767 | fixed_point<T> three_minus_dx = sub(const_three, mul(a, mul(x2, x2))); |
| 768 | x2 = shift_right(mul(x2, three_minus_dx), 1); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 769 | } |
| 770 | |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 771 | return (shift < 0) ? shift_left(x2, (-shift) >> 1) : shift_right(x2, shift >> 1); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 772 | } |
| 773 | /** Calculate the hyperbolic tangent of a fixed point number |
| 774 | * |
| 775 | * @param[in] x Fixed point operand |
| 776 | * |
| 777 | * @return Hyperbolic tangent of the operand |
| 778 | */ |
| 779 | template <typename T> |
| 780 | static fixed_point<T> tanh(fixed_point<T> x) |
| 781 | { |
| 782 | uint8_t p = x._fixed_point_position; |
| 783 | // Constants |
| 784 | auto const_one = fixed_point<T>(1, p); |
| 785 | auto const_two = fixed_point<T>(2, p); |
| 786 | |
| 787 | auto exp2x = exp(const_two * x); |
| 788 | auto num = exp2x - const_one; |
| 789 | auto den = exp2x + const_one; |
| 790 | auto tanh = num / den; |
| 791 | |
| 792 | return tanh; |
| 793 | } |
| 794 | /** Calculate the a-th power of a fixed point number. |
| 795 | * |
| 796 | * The power is computed as x^a = e^(log(x) * a) |
| 797 | * |
| 798 | * @param[in] x Fixed point operand |
| 799 | * @param[in] a Fixed point exponent |
| 800 | * |
| 801 | * @return a-th power of the operand |
| 802 | */ |
| 803 | template <typename T> |
| 804 | static fixed_point<T> pow(fixed_point<T> x, fixed_point<T> a) |
| 805 | { |
| 806 | return exp(log(x) * a); |
| 807 | } |
| 808 | }; |
| 809 | |
| 810 | template <typename T> |
| 811 | bool operator==(const fixed_point<T> &lhs, const fixed_point<T> &rhs) |
| 812 | { |
| 813 | return functions::isequal(lhs, rhs); |
| 814 | } |
| 815 | template <typename T> |
| 816 | bool operator!=(const fixed_point<T> &lhs, const fixed_point<T> &rhs) |
| 817 | { |
| 818 | return !operator==(lhs, rhs); |
| 819 | } |
| 820 | template <typename T> |
| 821 | bool operator<(const fixed_point<T> &lhs, const fixed_point<T> &rhs) |
| 822 | { |
| 823 | return functions::isless(lhs, rhs); |
| 824 | } |
| 825 | template <typename T> |
| 826 | bool operator>(const fixed_point<T> &lhs, const fixed_point<T> &rhs) |
| 827 | { |
| 828 | return operator<(rhs, lhs); |
| 829 | } |
| 830 | template <typename T> |
| 831 | bool operator<=(const fixed_point<T> &lhs, const fixed_point<T> &rhs) |
| 832 | { |
| 833 | return !operator>(lhs, rhs); |
| 834 | } |
| 835 | template <typename T> |
| 836 | bool operator>=(const fixed_point<T> &lhs, const fixed_point<T> &rhs) |
| 837 | { |
| 838 | return !operator<(lhs, rhs); |
| 839 | } |
| 840 | template <typename T> |
| 841 | fixed_point<T> operator+(const fixed_point<T> &lhs, const fixed_point<T> &rhs) |
| 842 | { |
| 843 | return functions::add(lhs, rhs); |
| 844 | } |
| 845 | template <typename T> |
| 846 | fixed_point<T> operator-(const fixed_point<T> &lhs, const fixed_point<T> &rhs) |
| 847 | { |
| 848 | return functions::sub(lhs, rhs); |
| 849 | } |
| 850 | template <typename T> |
| 851 | fixed_point<T> operator-(const fixed_point<T> &rhs) |
| 852 | { |
| 853 | return functions::negate(rhs); |
| 854 | } |
| 855 | template <typename T> |
| 856 | fixed_point<T> operator*(fixed_point<T> x, fixed_point<T> y) |
| 857 | { |
| 858 | return functions::mul(x, y); |
| 859 | } |
| 860 | template <typename T> |
| 861 | fixed_point<T> operator/(fixed_point<T> x, fixed_point<T> y) |
| 862 | { |
| 863 | return functions::div(x, y); |
| 864 | } |
| 865 | template <typename T> |
| 866 | fixed_point<T> operator>>(fixed_point<T> x, size_t shift) |
| 867 | { |
| 868 | return functions::shift_right(x, shift); |
| 869 | } |
| 870 | template <typename T> |
| 871 | fixed_point<T> operator<<(fixed_point<T> x, size_t shift) |
| 872 | { |
| 873 | return functions::shift_left(x, shift); |
| 874 | } |
| 875 | template <typename T, typename U, typename traits> |
| 876 | std::basic_ostream<T, traits> &operator<<(std::basic_ostream<T, traits> &s, fixed_point<U> x) |
| 877 | { |
| 878 | return functions::write(s, x); |
| 879 | } |
| 880 | template <typename T> |
| 881 | inline fixed_point<T> min(fixed_point<T> x, fixed_point<T> y) |
| 882 | { |
| 883 | return x > y ? y : x; |
| 884 | } |
| 885 | template <typename T> |
| 886 | inline fixed_point<T> max(fixed_point<T> x, fixed_point<T> y) |
| 887 | { |
| 888 | return x > y ? x : y; |
| 889 | } |
| 890 | template <OverflowPolicy OP = OverflowPolicy::SATURATE, typename T> |
| 891 | inline fixed_point<T> add(fixed_point<T> x, fixed_point<T> y) |
| 892 | { |
| 893 | return functions::add<OP>(x, y); |
| 894 | } |
| 895 | template <OverflowPolicy OP = OverflowPolicy::SATURATE, typename T> |
| 896 | inline fixed_point<T> sub(fixed_point<T> x, fixed_point<T> y) |
| 897 | { |
| 898 | return functions::sub<OP>(x, y); |
| 899 | } |
| 900 | template <OverflowPolicy OP = OverflowPolicy::SATURATE, typename T> |
| 901 | inline fixed_point<T> mul(fixed_point<T> x, fixed_point<T> y) |
| 902 | { |
| 903 | return functions::mul<OP>(x, y); |
| 904 | } |
| 905 | template <typename T> |
| 906 | inline fixed_point<T> div(fixed_point<T> x, fixed_point<T> y) |
| 907 | { |
| 908 | return functions::div(x, y); |
| 909 | } |
| 910 | template <typename T> |
| 911 | inline fixed_point<T> abs(fixed_point<T> x) |
| 912 | { |
| 913 | return functions::abs(x); |
| 914 | } |
| 915 | template <typename T> |
| 916 | inline fixed_point<T> clamp(fixed_point<T> x, T min, T max) |
| 917 | { |
| 918 | return functions::clamp(x, min, max); |
| 919 | } |
| 920 | template <typename T> |
| 921 | inline fixed_point<T> exp(fixed_point<T> x) |
| 922 | { |
| 923 | return functions::exp(x); |
| 924 | } |
| 925 | template <typename T> |
| 926 | inline fixed_point<T> log(fixed_point<T> x) |
| 927 | { |
| 928 | return functions::log(x); |
| 929 | } |
| 930 | template <typename T> |
| 931 | inline fixed_point<T> inv_sqrt(fixed_point<T> x) |
| 932 | { |
| 933 | return functions::inv_sqrt(x); |
| 934 | } |
| 935 | template <typename T> |
| 936 | inline fixed_point<T> tanh(fixed_point<T> x) |
| 937 | { |
| 938 | return functions::tanh(x); |
| 939 | } |
| 940 | template <typename T> |
| 941 | inline fixed_point<T> pow(fixed_point<T> x, fixed_point<T> a) |
| 942 | { |
| 943 | return functions::pow(x, a); |
| 944 | } |
| 945 | } // namespace detail |
| 946 | |
| 947 | // Expose operators |
| 948 | using detail::operator==; |
| 949 | using detail::operator!=; |
| 950 | using detail::operator<; |
| 951 | using detail::operator>; |
| 952 | using detail::operator<=; |
| 953 | using detail::operator>=; |
| 954 | using detail::operator+; |
| 955 | using detail::operator-; |
| 956 | using detail::operator*; |
| 957 | using detail::operator/; |
| 958 | using detail::operator>>; |
| 959 | using detail::operator<<; |
| 960 | |
| 961 | // Expose additional functions |
| 962 | using detail::min; |
| 963 | using detail::max; |
| 964 | using detail::add; |
| 965 | using detail::sub; |
| 966 | using detail::mul; |
| 967 | using detail::div; |
| 968 | using detail::abs; |
| 969 | using detail::clamp; |
| 970 | using detail::exp; |
| 971 | using detail::log; |
| 972 | using detail::inv_sqrt; |
| 973 | using detail::tanh; |
| 974 | using detail::pow; |
| 975 | // TODO: floor |
| 976 | // TODO: ceil |
| 977 | // TODO: sqrt |
| 978 | } // namespace fixed_point_arithmetic |
| 979 | } // namespace test |
| 980 | } // namespace arm_compute |
| 981 | #endif /*__ARM_COMPUTE_TEST_VALIDATION_FIXEDPOINT_H__ */ |