File size: 22,002 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 |
#pragma once
#ifndef FP16_FP16_H
#define FP16_FP16_H
#if defined(__cplusplus) && (__cplusplus >= 201103L)
#include <cstdint>
#include <cmath>
#elif !defined(__OPENCL_VERSION__)
#include <stdint.h>
#include <math.h>
#endif
#ifdef _MSC_VER
#include <intrin.h>
#endif
#include <fp16/bitcasts.h>
/*
* Convert a 16-bit floating-point number in IEEE half-precision format, in bit representation, to
* a 32-bit floating-point number in IEEE single-precision format, in bit representation.
*
* @note The implementation doesn't use any floating-point operations.
*/
static inline uint32_t fp16_ieee_to_fp32_bits(uint16_t h) {
/*
* Extend the half-precision floating-point number to 32 bits and shift to the upper part of the 32-bit word:
* +---+-----+------------+-------------------+
* | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 31 26-30 16-25 0-15
*
* S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.
*/
const uint32_t w = (uint32_t) h << 16;
/*
* Extract the sign of the input number into the high bit of the 32-bit word:
*
* +---+----------------------------------+
* | S |0000000 00000000 00000000 00000000|
* +---+----------------------------------+
* Bits 31 0-31
*/
const uint32_t sign = w & UINT32_C(0x80000000);
/*
* Extract mantissa and biased exponent of the input number into the bits 0-30 of the 32-bit word:
*
* +---+-----+------------+-------------------+
* | 0 |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 30 27-31 17-26 0-16
*/
const uint32_t nonsign = w & UINT32_C(0x7FFFFFFF);
/*
* Renorm shift is the number of bits to shift mantissa left to make the half-precision number normalized.
* If the initial number is normalized, some of its high 6 bits (sign == 0 and 5-bit exponent) equals one.
* In this case renorm_shift == 0. If the number is denormalize, renorm_shift > 0. Note that if we shift
* denormalized nonsign by renorm_shift, the unit bit of mantissa will shift into exponent, turning the
* biased exponent into 1, and making mantissa normalized (i.e. without leading 1).
*/
#ifdef _MSC_VER
unsigned long nonsign_bsr;
_BitScanReverse(&nonsign_bsr, (unsigned long) nonsign);
uint32_t renorm_shift = (uint32_t) nonsign_bsr ^ 31;
#else
uint32_t renorm_shift = __builtin_clz(nonsign);
#endif
renorm_shift = renorm_shift > 5 ? renorm_shift - 5 : 0;
/*
* Iff half-precision number has exponent of 15, the addition overflows it into bit 31,
* and the subsequent shift turns the high 9 bits into 1. Thus
* inf_nan_mask ==
* 0x7F800000 if the half-precision number had exponent of 15 (i.e. was NaN or infinity)
* 0x00000000 otherwise
*/
const int32_t inf_nan_mask = ((int32_t) (nonsign + 0x04000000) >> 8) & INT32_C(0x7F800000);
/*
* Iff nonsign is 0, it overflows into 0xFFFFFFFF, turning bit 31 into 1. Otherwise, bit 31 remains 0.
* The signed shift right by 31 broadcasts bit 31 into all bits of the zero_mask. Thus
* zero_mask ==
* 0xFFFFFFFF if the half-precision number was zero (+0.0h or -0.0h)
* 0x00000000 otherwise
*/
const int32_t zero_mask = (int32_t) (nonsign - 1) >> 31;
/*
* 1. Shift nonsign left by renorm_shift to normalize it (if the input was denormal)
* 2. Shift nonsign right by 3 so the exponent (5 bits originally) becomes an 8-bit field and 10-bit mantissa
* shifts into the 10 high bits of the 23-bit mantissa of IEEE single-precision number.
* 3. Add 0x70 to the exponent (starting at bit 23) to compensate the different in exponent bias
* (0x7F for single-precision number less 0xF for half-precision number).
* 4. Subtract renorm_shift from the exponent (starting at bit 23) to account for renormalization. As renorm_shift
* is less than 0x70, this can be combined with step 3.
* 5. Binary OR with inf_nan_mask to turn the exponent into 0xFF if the input was NaN or infinity.
* 6. Binary ANDNOT with zero_mask to turn the mantissa and exponent into zero if the input was zero.
* 7. Combine with the sign of the input number.
*/
return sign | ((((nonsign << renorm_shift >> 3) + ((0x70 - renorm_shift) << 23)) | inf_nan_mask) & ~zero_mask);
}
/*
* Convert a 16-bit floating-point number in IEEE half-precision format, in bit representation, to
* a 32-bit floating-point number in IEEE single-precision format.
*
* @note The implementation relies on IEEE-like (no assumption about rounding mode and no operations on denormals)
* floating-point operations and bitcasts between integer and floating-point variables.
*/
static inline float fp16_ieee_to_fp32_value(uint16_t h) {
/*
* Extend the half-precision floating-point number to 32 bits and shift to the upper part of the 32-bit word:
* +---+-----+------------+-------------------+
* | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 31 26-30 16-25 0-15
*
* S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.
*/
const uint32_t w = (uint32_t) h << 16;
/*
* Extract the sign of the input number into the high bit of the 32-bit word:
*
* +---+----------------------------------+
* | S |0000000 00000000 00000000 00000000|
* +---+----------------------------------+
* Bits 31 0-31
*/
const uint32_t sign = w & UINT32_C(0x80000000);
/*
* Extract mantissa and biased exponent of the input number into the high bits of the 32-bit word:
*
* +-----+------------+---------------------+
* |EEEEE|MM MMMM MMMM|0 0000 0000 0000 0000|
* +-----+------------+---------------------+
* Bits 27-31 17-26 0-16
*/
const uint32_t two_w = w + w;
/*
* Shift mantissa and exponent into bits 23-28 and bits 13-22 so they become mantissa and exponent
* of a single-precision floating-point number:
*
* S|Exponent | Mantissa
* +-+---+-----+------------+----------------+
* |0|000|EEEEE|MM MMMM MMMM|0 0000 0000 0000|
* +-+---+-----+------------+----------------+
* Bits | 23-31 | 0-22
*
* Next, there are some adjustments to the exponent:
* - The exponent needs to be corrected by the difference in exponent bias between single-precision and half-precision
* formats (0x7F - 0xF = 0x70)
* - Inf and NaN values in the inputs should become Inf and NaN values after conversion to the single-precision number.
* Therefore, if the biased exponent of the half-precision input was 0x1F (max possible value), the biased exponent
* of the single-precision output must be 0xFF (max possible value). We do this correction in two steps:
* - First, we adjust the exponent by (0xFF - 0x1F) = 0xE0 (see exp_offset below) rather than by 0x70 suggested
* by the difference in the exponent bias (see above).
* - Then we multiply the single-precision result of exponent adjustment by 2**(-112) to reverse the effect of
* exponent adjustment by 0xE0 less the necessary exponent adjustment by 0x70 due to difference in exponent bias.
* The floating-point multiplication hardware would ensure than Inf and NaN would retain their value on at least
* partially IEEE754-compliant implementations.
*
* Note that the above operations do not handle denormal inputs (where biased exponent == 0). However, they also do not
* operate on denormal inputs, and do not produce denormal results.
*/
const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float exp_scale = 0x1.0p-112f;
#else
const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
/*
* Convert denormalized half-precision inputs into single-precision results (always normalized).
* Zero inputs are also handled here.
*
* In a denormalized number the biased exponent is zero, and mantissa has on-zero bits.
* First, we shift mantissa into bits 0-9 of the 32-bit word.
*
* zeros | mantissa
* +---------------------------+------------+
* |0000 0000 0000 0000 0000 00|MM MMMM MMMM|
* +---------------------------+------------+
* Bits 10-31 0-9
*
* Now, remember that denormalized half-precision numbers are represented as:
* FP16 = mantissa * 2**(-24).
* The trick is to construct a normalized single-precision number with the same mantissa and thehalf-precision input
* and with an exponent which would scale the corresponding mantissa bits to 2**(-24).
* A normalized single-precision floating-point number is represented as:
* FP32 = (1 + mantissa * 2**(-23)) * 2**(exponent - 127)
* Therefore, when the biased exponent is 126, a unit change in the mantissa of the input denormalized half-precision
* number causes a change of the constructud single-precision number by 2**(-24), i.e. the same ammount.
*
* The last step is to adjust the bias of the constructed single-precision number. When the input half-precision number
* is zero, the constructed single-precision number has the value of
* FP32 = 1 * 2**(126 - 127) = 2**(-1) = 0.5
* Therefore, we need to subtract 0.5 from the constructed single-precision number to get the numerical equivalent of
* the input half-precision number.
*/
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
/*
* - Choose either results of conversion of input as a normalized number, or as a denormalized number, depending on the
* input exponent. The variable two_w contains input exponent in bits 27-31, therefore if its smaller than 2**27, the
* input is either a denormal number, or zero.
* - Combine the result of conversion of exponent and mantissa with the sign of the input number.
*/
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
/*
* Convert a 32-bit floating-point number in IEEE single-precision format to a 16-bit floating-point number in
* IEEE half-precision format, in bit representation.
*
* @note The implementation relies on IEEE-like (no assumption about rounding mode and no operations on denormals)
* floating-point operations and bitcasts between integer and floating-point variables.
*/
static inline uint16_t fp16_ieee_from_fp32_value(float f) {
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float scale_to_inf = 0x1.0p+112f;
const float scale_to_zero = 0x1.0p-110f;
#else
const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
const uint32_t w = fp32_to_bits(f);
const uint32_t shl1_w = w + w;
const uint32_t sign = w & UINT32_C(0x80000000);
uint32_t bias = shl1_w & UINT32_C(0xFF000000);
if (bias < UINT32_C(0x71000000)) {
bias = UINT32_C(0x71000000);
}
base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
const uint32_t bits = fp32_to_bits(base);
const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
const uint32_t nonsign = exp_bits + mantissa_bits;
return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}
/*
* Convert a 16-bit floating-point number in ARM alternative half-precision format, in bit representation, to
* a 32-bit floating-point number in IEEE single-precision format, in bit representation.
*
* @note The implementation doesn't use any floating-point operations.
*/
static inline uint32_t fp16_alt_to_fp32_bits(uint16_t h) {
/*
* Extend the half-precision floating-point number to 32 bits and shift to the upper part of the 32-bit word:
* +---+-----+------------+-------------------+
* | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 31 26-30 16-25 0-15
*
* S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.
*/
const uint32_t w = (uint32_t) h << 16;
/*
* Extract the sign of the input number into the high bit of the 32-bit word:
*
* +---+----------------------------------+
* | S |0000000 00000000 00000000 00000000|
* +---+----------------------------------+
* Bits 31 0-31
*/
const uint32_t sign = w & UINT32_C(0x80000000);
/*
* Extract mantissa and biased exponent of the input number into the bits 0-30 of the 32-bit word:
*
* +---+-----+------------+-------------------+
* | 0 |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 30 27-31 17-26 0-16
*/
const uint32_t nonsign = w & UINT32_C(0x7FFFFFFF);
/*
* Renorm shift is the number of bits to shift mantissa left to make the half-precision number normalized.
* If the initial number is normalized, some of its high 6 bits (sign == 0 and 5-bit exponent) equals one.
* In this case renorm_shift == 0. If the number is denormalize, renorm_shift > 0. Note that if we shift
* denormalized nonsign by renorm_shift, the unit bit of mantissa will shift into exponent, turning the
* biased exponent into 1, and making mantissa normalized (i.e. without leading 1).
*/
#ifdef _MSC_VER
unsigned long nonsign_bsr;
_BitScanReverse(&nonsign_bsr, (unsigned long) nonsign);
uint32_t renorm_shift = (uint32_t) nonsign_bsr ^ 31;
#else
uint32_t renorm_shift = __builtin_clz(nonsign);
#endif
renorm_shift = renorm_shift > 5 ? renorm_shift - 5 : 0;
/*
* Iff nonsign is 0, it overflows into 0xFFFFFFFF, turning bit 31 into 1. Otherwise, bit 31 remains 0.
* The signed shift right by 31 broadcasts bit 31 into all bits of the zero_mask. Thus
* zero_mask ==
* 0xFFFFFFFF if the half-precision number was zero (+0.0h or -0.0h)
* 0x00000000 otherwise
*/
const int32_t zero_mask = (int32_t) (nonsign - 1) >> 31;
/*
* 1. Shift nonsign left by renorm_shift to normalize it (if the input was denormal)
* 2. Shift nonsign right by 3 so the exponent (5 bits originally) becomes an 8-bit field and 10-bit mantissa
* shifts into the 10 high bits of the 23-bit mantissa of IEEE single-precision number.
* 3. Add 0x70 to the exponent (starting at bit 23) to compensate the different in exponent bias
* (0x7F for single-precision number less 0xF for half-precision number).
* 4. Subtract renorm_shift from the exponent (starting at bit 23) to account for renormalization. As renorm_shift
* is less than 0x70, this can be combined with step 3.
* 5. Binary ANDNOT with zero_mask to turn the mantissa and exponent into zero if the input was zero.
* 6. Combine with the sign of the input number.
*/
return sign | (((nonsign << renorm_shift >> 3) + ((0x70 - renorm_shift) << 23)) & ~zero_mask);
}
/*
* Convert a 16-bit floating-point number in ARM alternative half-precision format, in bit representation, to
* a 32-bit floating-point number in IEEE single-precision format.
*
* @note The implementation relies on IEEE-like (no assumption about rounding mode and no operations on denormals)
* floating-point operations and bitcasts between integer and floating-point variables.
*/
static inline float fp16_alt_to_fp32_value(uint16_t h) {
/*
* Extend the half-precision floating-point number to 32 bits and shift to the upper part of the 32-bit word:
* +---+-----+------------+-------------------+
* | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000|
* +---+-----+------------+-------------------+
* Bits 31 26-30 16-25 0-15
*
* S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.
*/
const uint32_t w = (uint32_t) h << 16;
/*
* Extract the sign of the input number into the high bit of the 32-bit word:
*
* +---+----------------------------------+
* | S |0000000 00000000 00000000 00000000|
* +---+----------------------------------+
* Bits 31 0-31
*/
const uint32_t sign = w & UINT32_C(0x80000000);
/*
* Extract mantissa and biased exponent of the input number into the high bits of the 32-bit word:
*
* +-----+------------+---------------------+
* |EEEEE|MM MMMM MMMM|0 0000 0000 0000 0000|
* +-----+------------+---------------------+
* Bits 27-31 17-26 0-16
*/
const uint32_t two_w = w + w;
/*
* Shift mantissa and exponent into bits 23-28 and bits 13-22 so they become mantissa and exponent
* of a single-precision floating-point number:
*
* S|Exponent | Mantissa
* +-+---+-----+------------+----------------+
* |0|000|EEEEE|MM MMMM MMMM|0 0000 0000 0000|
* +-+---+-----+------------+----------------+
* Bits | 23-31 | 0-22
*
* Next, the exponent is adjusted for the difference in exponent bias between single-precision and half-precision
* formats (0x7F - 0xF = 0x70). This operation never overflows or generates non-finite values, as the largest
* half-precision exponent is 0x1F and after the adjustment is can not exceed 0x8F < 0xFE (largest single-precision
* exponent for non-finite values).
*
* Note that this operation does not handle denormal inputs (where biased exponent == 0). However, they also do not
* operate on denormal inputs, and do not produce denormal results.
*/
const uint32_t exp_offset = UINT32_C(0x70) << 23;
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset);
/*
* Convert denormalized half-precision inputs into single-precision results (always normalized).
* Zero inputs are also handled here.
*
* In a denormalized number the biased exponent is zero, and mantissa has on-zero bits.
* First, we shift mantissa into bits 0-9 of the 32-bit word.
*
* zeros | mantissa
* +---------------------------+------------+
* |0000 0000 0000 0000 0000 00|MM MMMM MMMM|
* +---------------------------+------------+
* Bits 10-31 0-9
*
* Now, remember that denormalized half-precision numbers are represented as:
* FP16 = mantissa * 2**(-24).
* The trick is to construct a normalized single-precision number with the same mantissa and thehalf-precision input
* and with an exponent which would scale the corresponding mantissa bits to 2**(-24).
* A normalized single-precision floating-point number is represented as:
* FP32 = (1 + mantissa * 2**(-23)) * 2**(exponent - 127)
* Therefore, when the biased exponent is 126, a unit change in the mantissa of the input denormalized half-precision
* number causes a change of the constructud single-precision number by 2**(-24), i.e. the same ammount.
*
* The last step is to adjust the bias of the constructed single-precision number. When the input half-precision number
* is zero, the constructed single-precision number has the value of
* FP32 = 1 * 2**(126 - 127) = 2**(-1) = 0.5
* Therefore, we need to subtract 0.5 from the constructed single-precision number to get the numerical equivalent of
* the input half-precision number.
*/
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
/*
* - Choose either results of conversion of input as a normalized number, or as a denormalized number, depending on the
* input exponent. The variable two_w contains input exponent in bits 27-31, therefore if its smaller than 2**27, the
* input is either a denormal number, or zero.
* - Combine the result of conversion of exponent and mantissa with the sign of the input number.
*/
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
/*
* Convert a 32-bit floating-point number in IEEE single-precision format to a 16-bit floating-point number in
* ARM alternative half-precision format, in bit representation.
*
* @note The implementation relies on IEEE-like (no assumption about rounding mode and no operations on denormals)
* floating-point operations and bitcasts between integer and floating-point variables.
*/
static inline uint16_t fp16_alt_from_fp32_value(float f) {
const uint32_t w = fp32_to_bits(f);
const uint32_t sign = w & UINT32_C(0x80000000);
const uint32_t shl1_w = w + w;
const uint32_t shl1_max_fp16_fp32 = UINT32_C(0x8FFFC000);
const uint32_t shl1_base = shl1_w > shl1_max_fp16_fp32 ? shl1_max_fp16_fp32 : shl1_w;
uint32_t shl1_bias = shl1_base & UINT32_C(0xFF000000);
const uint32_t exp_difference = 23 - 10;
const uint32_t shl1_bias_min = (127 - 1 - exp_difference) << 24;
if (shl1_bias < shl1_bias_min) {
shl1_bias = shl1_bias_min;
}
const float bias = fp32_from_bits((shl1_bias >> 1) + ((exp_difference + 2) << 23));
const float base = fp32_from_bits((shl1_base >> 1) + (2 << 23)) + bias;
const uint32_t exp_f = fp32_to_bits(base) >> 13;
return (sign >> 16) | ((exp_f & UINT32_C(0x00007C00)) + (fp32_to_bits(base) & UINT32_C(0x00000FFF)));
}
#endif /* FP16_FP16_H */
|