| 1 |
/*========================== begin_copyright_notice ============================ |
| 2 |
|
| 3 |
Copyright (C) 2017-2021 Intel Corporation |
| 4 |
|
| 5 |
SPDX-License-Identifier: MIT |
| 6 |
|
| 7 |
============================= end_copyright_notice ===========================*/ |
| 8 |
|
| 9 |
#include "Floats.hpp" |
| 10 |
#include "../strings.hpp" |
| 11 |
|
| 12 |
#if defined(_MSC_VER) |
| 13 |
// This only affects defined(_DEBUG) and we could scope this to debug builds |
| 14 |
// only, but we prefer to keep Debug and Release configs as close as possible, |
| 15 |
// and formatting/disassembly isn't considered a performance critical path. |
| 16 |
// |
| 17 |
// There's a defect in Windows Debug runtime libraries in emitting denorms |
| 18 |
// when denorm mode is FTZ by the calling library and with the Debug runtime |
| 19 |
// https://developercommunity.visualstudio.com/content/problem/1187587/printf-assert-failure-unexpected-input-value-log10.html |
| 20 |
// |
| 21 |
// To test this place _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON); |
| 22 |
// in before calling into this module (and compiled for Debug on MSVC++). |
| 23 |
#define IGA_NEEDS_DENORM_WORKAROUND |
| 24 |
#endif |
| 25 |
|
| 26 |
// For non-NaN cases this flag enables use of a native AVX instruction |
| 27 |
// for half float conversion _cvtss_sh. |
| 28 |
// |
| 29 |
// #define IGA_USE_FP16C |
| 30 |
|
| 31 |
#include <cmath> |
| 32 |
#include <cstdint> |
| 33 |
#include <cstdlib> |
| 34 |
#include <iomanip> |
| 35 |
#include <iostream> |
| 36 |
#include <limits> |
| 37 |
#include <sstream> |
| 38 |
#if defined(IGA_NEEDS_DENORM_WORKAROUND) || defined(IGA_USE_FP16C) |
| 39 |
#include <immintrin.h> |
| 40 |
#endif |
| 41 |
|
| 42 |
using namespace iga; |
| 43 |
|
| 44 |
// Big Theory Statement (BTS): on NaN's in IGA |
| 45 |
// |
| 46 |
// NaN is an equivalence class of floating point values. Different |
| 47 |
// libraries, runtimes, and hardware use different elements of the |
| 48 |
// class as the characteristic value when they want a NaN value. |
| 49 |
// For instance, the sign bit can be anything and the mantissa payload |
| 50 |
// bits can be anything except for the leading mantissa bit (which |
| 51 |
// distinguishes a qnan from an snan). |
| 52 |
// |
| 53 |
// NOTE: we distinguish between 'qnan' and 'snan' instead of just using 'nan' |
| 54 |
// with the whole payload since we parse as 64b and only narrow once |
| 55 |
// we see the type; we considered 'nan' '(' HEXLIT ')' and just letting |
| 56 |
// the user sort the meaning of the mantissa out, but this gets messy |
| 57 |
// when widening literals during parse or narrowing them during |
| 58 |
// formatting. |
| 59 |
// |
| 60 |
// Without loss of generality (this generalizes to all IEEE sizes), |
| 61 |
// the 32-bit IEEE 754 (C.f. section 6.2.1 of the spec) we have |
| 62 |
// s eeeeeeee mmm`mmmm`mmmm`mmmm`mmmm`mmmm |
| 63 |
// (where the s and the m's can be anything such that at least one m is |
| 64 |
// non-zero; all 0's would imply infinity) |
| 65 |
// |
| 66 |
// SNAN: "snan" (signaling NaN) is: |
| 67 |
// s 11111111 0xx`xxxx`xxxx`xxxx`xxxx`xxxx |
| 68 |
// ^ leading bit of mantissa is 0; at least one bit |
| 69 |
// of the rest of the payload (x's must be non-zero) |
| 70 |
// |
| 71 |
// QNAN: "qnan" (quiet NaN) is: |
| 72 |
// s 11111111 1xx`xxxx`xxxx`xxxx`xxxx`xxxx |
| 73 |
// ^ leading bit of mantiss is 1 and the rest can be anything |
| 74 |
// including all zeros since the top mantissa bit is set |
| 75 |
// (NaN only needs one mantissa bit set and the quiet bit |
| 76 |
// counts) |
| 77 |
// |
| 78 |
// IGA supports the following syntax (from examples) |
| 79 |
// NaNLit ::= '-'? ('snan'|'qnan') '(' HEXLIT ')' |
| 80 |
// |
| 81 |
// Examples: |
| 82 |
// * "-snan(0xA):f" parses as |
| 83 |
// s eeeeeeee mmm`mmmm`mmmm`mmmm`mmmm`mmmm |
| 84 |
// 1 11111111 000`0000`0000`0000`0000`1010 |
| 85 |
// ^ negative ^ snan ^^^^ 0xA payload |
| 86 |
// sign bit is respected |
| 87 |
// |
| 88 |
// * "qnan(0x1B):f" parses as |
| 89 |
// s eeeeeeee mmm`mmmm`mmmm`mmmm`mmmm`mmmm |
| 90 |
// 1 11111111 100`0000`0000`0000`0001`1011 |
| 91 |
// ^ positive ^ qnan ^^^^^^ 0x1B |
| 92 |
// |
| 93 |
// * "snan(0x0):f" would be a parse error since the mantissa must not be 0 |
| 94 |
// s eeeeeeee mmm`mmmm`mmmm`mmmm`mmmm`mmmm |
| 95 |
// 1 11111111 000`0000`0000`0000`0000`0000 |
| 96 |
// ^ positive ^ snan^^^^^^^^^^^^^^^^^^^^^^^ at least one bit must be 0 |
| 97 |
// |
| 98 |
// * "snan(0x00800000):f" illegal: payload value too large for bitfield |
| 99 |
// s eeeeeeee mmm`mmmm`mmmm`mmmm`mmmm`mmmm |
| 100 |
// 1 11111111 000`0000`0000`0000`0000`0000 |
| 101 |
// ^ high bit of payload overflows to here |
| 102 |
// |
| 103 |
// * "snan(0x00800000):df" payload fits since df has 53 bits |
| 104 |
// s eeeeeeeeeee |
| 105 |
// m`mmmm`mmmm`mmmm`mmmm`mmmm`mmmm`mmmm`mmmm`mmmm`mmmm`mmmm`mmmm 1 |
| 106 |
// 11111111111 |
| 107 |
// 0`0000`0000`0000`0000`0000`0000`1000`0000`0000`0000`0000`0000 |
| 108 |
// ^ 0x800000 |
| 109 |
// |
| 110 |
// * "snan(0x400000):f" is legal, but just means qnan(0). This is because |
| 111 |
// the parser parses the value as 64b and then converts it to 32b when it |
| 112 |
// sees the :f. The payload fits while parsing, but conversion cannot |
| 113 |
// tell that we were dealing with qnan |
| 114 |
|
| 115 |
template <typename F> int FloatMantissaBits(); |
| 116 |
template <typename F> int FloatExponentBits(); |
| 117 |
template <> int FloatMantissaBits<double>() { return 52; } |
| 118 |
template <> int FloatExponentBits<double>() { return 11; } |
| 119 |
template <> int FloatMantissaBits<float>() { return 23; } |
| 120 |
template <> int FloatExponentBits<float>() { return 8; } |
| 121 |
template <> int FloatMantissaBits<uint16_t>() { return 10; } |
| 122 |
template <> int FloatExponentBits<uint16_t>() { return 5; } |
| 123 |
|
| 124 |
// E.g. FloatBias<float>() == 127 |
| 125 |
// |
| 126 |
template <typename F> static int FloatBias() { |
| 127 |
return (1 << (FloatExponentBits<F>() - 1)) - 1; |
| 128 |
} |
| 129 |
|
| 130 |
// FloatBiasDiff<double,float> => 1023 - 127 |
| 131 |
// FloatBiasDiff<float,half> => 127 - 15 (0x70) |
| 132 |
template <typename FBIG, typename FSML> static int FloatBiasDiff() { |
| 133 |
return FloatBias<FBIG>() - FloatBias<FSML>(); |
| 134 |
} |
| 135 |
template <typename FBIG, typename FSML> static int FloatMantissaBitsDiff() { |
| 136 |
return FloatMantissaBits<FBIG>() - FloatMantissaBits<FSML>(); |
| 137 |
} |
| 138 |
|
| 139 |
template <typename F, typename I> |
| 140 |
static void FormatFloatImplNaN(std::ostream &os, I bits) { |
| 141 |
const int MANT_LEN = FloatMantissaBits<F>(); |
| 142 |
const int EXPN_LEN = FloatExponentBits<F>(); |
| 143 |
const I SIGN_BIT = ((I)1 << (MANT_LEN + EXPN_LEN)); |
| 144 |
if (bits & SIGN_BIT) { |
| 145 |
os << '-'; |
| 146 |
} |
| 147 |
// split the payload into the quiet/signaling bit |
| 148 |
// (high bit of the mantissa) and the lower bits |
| 149 |
const I QNAN_BIT = (I)1 << (MANT_LEN - 1); |
| 150 |
if (bits & QNAN_BIT) { |
| 151 |
os << "qnan"; |
| 152 |
} else { |
| 153 |
os << "snan"; |
| 154 |
} |
| 155 |
os << "("; |
| 156 |
I lowerPayload = bits & (QNAN_BIT - 1); // lower bits of mantissa |
| 157 |
fmtHex(os, (uint64_t)lowerPayload); |
| 158 |
os << ")"; |
| 159 |
} |
| 160 |
|
| 161 |
// Checks if a floating point number will reparse exactly as formatted. |
| 162 |
// |
| 163 |
// We parse floats as 64b literals and cast down. So we use strtod for all |
| 164 |
// floating point types and cast down. |
| 165 |
template <typename T> static bool willReparseExactly(T x, std::string str) { |
| 166 |
// we always parse as a double since the parser will do the same |
| 167 |
double y = strtod(str.c_str(), nullptr); |
| 168 |
return ((T)y == x); |
| 169 |
} |
| 170 |
|
| 171 |
#ifdef IGA_NEEDS_DENORM_WORKAROUND |
| 172 |
// C.f. with the #define for this above for notes on this |
| 173 |
struct ScopedDenormWorkaround { |
| 174 |
unsigned int oldDenormMode; |
| 175 |
ScopedDenormWorkaround() : oldDenormMode(_MM_GET_DENORMALS_ZERO_MODE()) { |
| 176 |
_MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_OFF); |
| 177 |
} |
| 178 |
~ScopedDenormWorkaround() { _MM_SET_DENORMALS_ZERO_MODE(oldDenormMode); } |
| 179 |
}; |
| 180 |
#endif |
| 181 |
|
| 182 |
// Formats a float in decimal or scientific format if it will not lose |
| 183 |
// precision during reparse |
| 184 |
// |
| 185 |
// Returns false if we were able to format it as something representable. |
| 186 |
template <typename F, typename I> |
| 187 |
static bool TryFormatFloatImplNonHex(std::ostream &os, F x) { |
| 188 |
#ifdef IGA_NEEDS_DENORM_WORKAROUND |
| 189 |
ScopedDenormWorkaround sdw; |
| 190 |
#endif |
| 191 |
auto fpc = std::fpclassify(x); |
| 192 |
switch (fpc) { |
| 193 |
case FP_INFINITE: |
| 194 |
if (x < 0) { |
| 195 |
os << '-'; |
| 196 |
} |
| 197 |
os << "inf"; |
| 198 |
return true; |
| 199 |
case FP_NAN: |
| 200 |
FormatFloatImplNaN<F, I>(os, iga::FloatToBits(x)); |
| 201 |
return true; |
| 202 |
/////////////////////////////////////////////////////////////////////////// |
| 203 |
// case FP_ZERO: |
| 204 |
// case FP_NORMAL: |
| 205 |
// case FP_SUBNORMAL: |
| 206 |
default: |
| 207 |
// Fall through and go through the pretty-print algorithm |
| 208 |
break; |
| 209 |
} |
| 210 |
|
| 211 |
// we try and render it several different ways until we find something |
| 212 |
// that we can parse back bit-exact. |
| 213 |
// We try: |
| 214 |
// - decimal |
| 215 |
// - exponential |
| 216 |
// - and then fall back on hex |
| 217 |
// static_assert(sizeof(F) == sizeof(I)); |
| 218 |
|
| 219 |
// try as default, this lets STL pick the format. |
| 220 |
// it sometimes gives nice terse output |
| 221 |
// e.g. "3" for "3.0" instead of "3.0000000000..." (when possible) |
| 222 |
std::stringstream ss; |
| 223 |
ss.unsetf(std::ios_base::floatfield); |
| 224 |
ss << x; |
| 225 |
if (willReparseExactly(x, ss.str())) { |
| 226 |
auto str = ss.str(); |
| 227 |
os << str; |
| 228 |
if (str.find('.') == std::string::npos && |
| 229 |
str.find('e') == std::string::npos && |
| 230 |
str.find('E') == std::string::npos) { |
| 231 |
// floats need a ".0" suffixing them if not in scientific form |
| 232 |
// STL default float sometimes drops the . |
| 233 |
// |
| 234 |
// e.g. given "-0.0f", if we were to format that |
| 235 |
// as "-0:f" (which MSVCRT does), then it parses as |
| 236 |
// "0" since this is the negation the S64 value 0 (during parse) |
| 237 |
// |
| 238 |
// NOTE: we also have to ensure that we aren't using an |
| 239 |
// exponential (scientific) form already. |
| 240 |
// |
| 241 |
// e.g. 1e-007 should not convert to 1e-007.0 |
| 242 |
os << ".0"; |
| 243 |
} |
| 244 |
return true; |
| 245 |
} |
| 246 |
|
| 247 |
// try as scientific |
| 248 |
ss.str(std::string()); // reset |
| 249 |
ss << std::scientific << x; |
| 250 |
if (willReparseExactly(x, ss.str())) { |
| 251 |
os << ss.str(); |
| 252 |
return true; |
| 253 |
} |
| 254 |
|
| 255 |
// TODO: IFDEF this given a new enough compiler |
| 256 |
// (must parse too) |
| 257 |
// e.g. 0x1.47ae147ae147bp-7 (need to parse first) |
| 258 |
// NOTE: parsing should use >> |
| 259 |
// float f; |
| 260 |
// e.g. std::istringstream(""0x1P-1022") >> std::hexfloat >> f; |
| 261 |
|
| 262 |
// fallback to hex integral |
| 263 |
// FormatFloatImplHex<F,I>(os, x); |
| 264 |
return false; |
| 265 |
} |
| 266 |
|
| 267 |
template <typename F> static void FormatFloatAsHex(std::ostream &os, F f) { |
| 268 |
fmtHex(os, (uint64_t)iga::FloatToBits(f)); |
| 269 |
} |
| 270 |
|
| 271 |
void iga::FormatFloat(std::ostream &os, float x) { |
| 272 |
if (!TryFormatFloatImplNonHex<float, uint32_t>(os, x)) { |
| 273 |
FormatFloatAsHex<float>(os, x); |
| 274 |
} |
| 275 |
} |
| 276 |
|
| 277 |
void iga::FormatFloat(std::ostream &os, double x) { |
| 278 |
if (!TryFormatFloatImplNonHex<double, uint64_t>(os, x)) { |
| 279 |
FormatFloatAsHex<double>(os, x); |
| 280 |
} |
| 281 |
} |
| 282 |
|
| 283 |
void iga::FormatFloat(std::ostream &os, uint16_t w16) { |
| 284 |
#if 0 |
| 285 |
// this would turn off all non-hex floats |
| 286 |
fmtHex(os, (uint64_t)w16); |
| 287 |
#else |
| 288 |
// trys to format a half float in a friendly format |
| 289 |
// falling back to hex if all else fails |
| 290 |
float f32 = ConvertHalfToFloat(w16); |
| 291 |
if (IS_NAN(f32)) { |
| 292 |
// So we get the correct payload size for NaNs |
| 293 |
FormatFloatImplNaN<uint16_t, uint16_t>(os, w16); |
| 294 |
} else if (ConvertFloatToHalf(f32) != w16 || |
| 295 |
!TryFormatFloatImplNonHex<float, uint32_t>(os, f32)) { |
| 296 |
FormatFloatAsHex<uint16_t>(os, w16); |
| 297 |
} // else: FormatFloatImplNonHex worked |
| 298 |
#endif |
| 299 |
} |
| 300 |
|
| 301 |
void iga::FormatFloat(std::ostream &os, uint8_t x) { |
| 302 |
FormatFloat(os, ConvertQuarterToFloatGEN(x)); |
| 303 |
} |
| 304 |
|
| 305 |
uint32_t iga::ConvertDoubleToFloatBits(double f) { |
| 306 |
uint64_t f64 = FloatToBits(f); |
| 307 |
|
| 308 |
uint64_t m64 = f64 & F64_MANT_MASK; |
| 309 |
uint64_t e64 = (f64 & F64_EXP_MASK) >> FloatMantissaBits<double>(); |
| 310 |
if (e64 == (F64_EXP_MASK >> FloatMantissaBits<double>()) && m64 != 0) { |
| 311 |
// f64 NaN |
| 312 |
uint32_t m32 = (uint32_t)m64 & F32_MANT_MASK; |
| 313 |
m32 |= (uint32_t)((m64 & F64_QNAN_BIT) >> |
| 314 |
(FloatMantissaBits<double>() - |
| 315 |
FloatMantissaBits<float>())); // preserve snan |
| 316 |
if (m32 == 0) { |
| 317 |
// The payload was only in the high bits which we dropped; |
| 318 |
// make it non-zero so we retain NaN'ness |
| 319 |
m32 = 1; |
| 320 |
} |
| 321 |
uint32_t s32 = (uint32_t)(f64 >> 32) & F32_SIGN_BIT; |
| 322 |
return (s32 | F32_EXP_MASK | m32); |
| 323 |
} else { |
| 324 |
// regular conversion can deal with all the other special cases |
| 325 |
return FloatToBits((float)f); |
| 326 |
} |
| 327 |
} |
| 328 |
|
| 329 |
float iga::ConvertDoubleToFloat(double f) { |
| 330 |
return FloatFromBits(ConvertDoubleToFloatBits(f)); |
| 331 |
} |
| 332 |
|
| 333 |
uint16_t iga::ConvertFloatToHalf(float f) { |
| 334 |
static const uint32_t F32_EXP_MASK = ((1 << FloatExponentBits<float>()) - 1) |
| 335 |
<< FloatMantissaBits<float>(); |
| 336 |
|
| 337 |
const uint32_t w32 = FloatToBits(f); |
| 338 |
const uint32_t w32_u = w32 & 0x7FFFFFFF; |
| 339 |
const uint32_t sign = w32 & 0x80000000; |
| 340 |
const uint16_t sign16 = (uint16_t)(sign >> 16); |
| 341 |
|
| 342 |
if (w32_u > F32_EXP_MASK) { // NaN |
| 343 |
uint16_t m16 = 0; |
| 344 |
m16 |= (F32_QNAN_BIT & w32_u) >> // preserve qnan bit |
| 345 |
(FloatMantissaBits<float>() - FloatMantissaBits<uint16_t>()); |
| 346 |
m16 |= (F16_MANT_MASK >> 1) & w32_u; // and bottom 9b |
| 347 |
// |
| 348 |
// s eeeeeeee qmmmmmmmmmmmmmmmmmmmmm |
| 349 |
// | ||||||||| |
| 350 |
// | vvvvvvvvv |
| 351 |
// +---------->qmmmmmmmmm |
| 352 |
if (m16 == 0x0) { |
| 353 |
// if the nonzero payload is in the high bits and and gets |
| 354 |
// dropped and the signal bit is non-zero, then m16 is 0; |
| 355 |
// to maintain it as a qnan we must set at least one bit |
| 356 |
m16 = 0x1; |
| 357 |
} |
| 358 |
return sign16 | F16_EXP_MASK | m16; |
| 359 |
} else if (w32_u == F32_EXP_MASK) { // +/-Infinity |
| 360 |
return sign16 | F16_EXP_MASK; |
| 361 |
} else { |
| 362 |
// norm/denorm |
| 363 |
#ifdef IGA_USE_FP16C |
| 364 |
// should be on all 3rd generation or newer CPUs (HSW/SNB) |
| 365 |
// _MM_FROUND_NO_EXC |
| 366 |
const auto oldCsr = _mm_getcsr(); |
| 367 |
_mm_setcsr(oldCsr | _MM_FROUND_NO_EXC); |
| 368 |
uint16_t w16 = (uint16_t)_mm_cvtsi128_si32( |
| 369 |
_mm_cvtps_ph(_mm_set_ss(f), _MM_FROUND_TO_NEAREST_INT)); |
| 370 |
_mm_setcsr(oldCsr); |
| 371 |
return w16; |
| 372 |
#else // !IGA_USE_FP16C |
| 373 |
static const uint32_t F32_BIAS = FloatBias<float>(); |
| 374 |
static const uint32_t F16_BIAS = FloatBias<uint16_t>(); |
| 375 |
static const int F32_MNT_BITS = FloatMantissaBits<float>(); |
| 376 |
static const int F16_MNT_BITS = FloatMantissaBits<uint16_t>(); |
| 377 |
|
| 378 |
static const uint32_t LOWEST_OVERFLOW = // 0x47800000 |
| 379 |
(F32_BIAS + F16_BIAS + 1) << FloatMantissaBits<float>(); |
| 380 |
static const uint32_t LOWEST_NORM = // 0x38800000 |
| 381 |
(F32_BIAS - F16_BIAS + 1) << FloatMantissaBits<float>(); |
| 382 |
static const uint32_t LOWEST_DENORM = // 0x33000000 |
| 383 |
(F32_BIAS - F16_BIAS - (uint16_t)F16_MNT_BITS) |
| 384 |
<< FloatMantissaBits<float>(); |
| 385 |
auto round = [](uint32_t v, uint32_t g, uint32_t s) { |
| 386 |
return v + (g & (s | v)); |
| 387 |
}; |
| 388 |
if (w32_u >= LOWEST_OVERFLOW) { // overflows to infinity |
| 389 |
return sign16 | F16_EXP_MASK; |
| 390 |
} else if (w32_u >= LOWEST_NORM) { // fits as normalized half |
| 391 |
uint32_t v = |
| 392 |
(((w32_u >> F32_MNT_BITS) - (F32_BIAS - F16_BIAS)) << F16_MNT_BITS) | |
| 393 |
((w32_u >> (F32_MNT_BITS - F16_MNT_BITS)) & F16_MANT_MASK); |
| 394 |
uint32_t g = (w32_u >> (F32_MNT_BITS - F16_MNT_BITS - 1)) & 0x1; |
| 395 |
uint32_t s = (w32_u & 0x0FFF) ? 1 : 0; |
| 396 |
return (uint16_t)round(sign16 | v, g, s); |
| 397 |
} else if (w32_u >= LOWEST_DENORM) { // fits as normalized half |
| 398 |
uint32_t i = (F32_BIAS - 1 - 1) - (w32_u >> F32_MNT_BITS); |
| 399 |
uint32_t w32_u2 = (w32_u & F32_MANT_MASK) | (F32_MANT_MASK + 1); |
| 400 |
uint32_t v = sign16 | (w32_u2 >> (i + 1)); |
| 401 |
uint32_t g = (w32_u2 >> i) & 1; |
| 402 |
uint32_t s = (w32_u2 & ((1 << i) - 1)) ? 1 : 0; |
| 403 |
return (uint16_t)round(v, g, s); |
| 404 |
} else { |
| 405 |
// underflow to +-0 |
| 406 |
return sign16; |
| 407 |
} |
| 408 |
#endif // !IGA_USE_FP16C |
| 409 |
} |
| 410 |
} |
| 411 |
|
| 412 |
// GEN's 8-bit restricted float ("quarter float") |
| 413 |
// s eee mmmm (bias 2^(3-1) - 1 == 3) |
| 414 |
// s gfe dcba |
| 415 |
// mantissa shifted to the top of the float |
| 416 |
// dcb`a000`0000`0000`0000`0000 |
| 417 |
// exponenent is unpacked by expanding |
| 418 |
// gfe to the following: |
| 419 |
// |
| 420 |
// dgGGGGfe where G = ~g |
| 421 |
// negative and positive 0 bypass this logic |
| 422 |
// |
| 423 |
// => NaN and infinities are illegal in this format |
| 424 |
// => Denorms are not supported |
| 425 |
// QUOTE: |
| 426 |
// Specifically, when the exponent field is zero and the fraction |
| 427 |
// field is not zero, an implied one is still present instead of |
| 428 |
// taking a denormalized form (without an implied one). This results |
| 429 |
// in a simple implementation but with a smaller dynamic range - |
| 430 |
// the magnitude of the smallest non-zero number is 0.1328125. |
| 431 |
float iga::ConvertQuarterToFloatGEN(uint8_t u8) { |
| 432 |
if (u8 == 0x00) { |
| 433 |
return 0.0f; |
| 434 |
} else if (u8 == 0x80) { |
| 435 |
return -0.0f; |
| 436 |
} else { |
| 437 |
uint32_t f32; |
| 438 |
f32 = ((uint32_t)u8 & 0x80) << (32 - 8); // d: sign |
| 439 |
f32 |= ((0x30 & (uint32_t)u8) << (23 - 4)); // fe: low bits of exp |
| 440 |
if ((0x40 & u8) == 0) { // g = 0 |
| 441 |
// exp=011111fe |
| 442 |
f32 |= 0x1F << (23 + 2); |
| 443 |
} else { |
| 444 |
// exp=100000fe |
| 445 |
f32 |= 1ul << (23 + 7); // f32 high exp bit |
| 446 |
} |
| 447 |
f32 |= ((uint32_t)u8 & 0xF) << (23 - 4); // dcba: mantissa |
| 448 |
return FloatFromBits(f32); |
| 449 |
} |
| 450 |
} |
| 451 |
|
| 452 |
bool iga::IsNaN(uint16_t u16) { |
| 453 |
return (F16_EXP_MASK & u16) == F16_EXP_MASK && (F16_MANT_MASK & u16) != 0; |
| 454 |
} |
| 455 |
|
| 456 |
bool iga::IsInf(uint16_t u16) { |
| 457 |
return (F16_EXP_MASK & u16) == F16_EXP_MASK && (F16_MANT_MASK & u16) == 0; |
| 458 |
} |
| 459 |
|
| 460 |
float iga::ConvertHalfToFloat(uint16_t u16) { |
| 461 |
uint16_t u16_u = u16 & 0x7FFF; |
| 462 |
uint32_t s32 = ((uint32_t)u16 & F16_SIGN_BIT) << 16; |
| 463 |
uint32_t m16 = u16 & F16_MANT_MASK; |
| 464 |
if (u16_u > F16_EXP_MASK) { |
| 465 |
// NaN |
| 466 |
uint32_t m32 = |
| 467 |
(u16 & F16_QNAN_BIT) |
| 468 |
<< FloatMantissaBitsDiff<float, uint16_t>(); // preserve sNaN bit |
| 469 |
m32 |= (F16_MANT_MASK >> 1) & m16; |
| 470 |
if (m32 == 0) { |
| 471 |
m32 = 1; // ensure still NaN |
| 472 |
} |
| 473 |
return FloatFromBits(s32 | F32_EXP_MASK | m32); |
| 474 |
} |
| 475 |
#ifdef IGA_USE_FP16C |
| 476 |
float f = _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(u16))); |
| 477 |
return f; |
| 478 |
#else // !IGA_USE_FP16C |
| 479 |
uint32_t e16 = (u16 & F16_EXP_MASK) >> FloatMantissaBits<uint16_t>(); |
| 480 |
uint32_t e32, m32; |
| 481 |
if (u16_u == F16_EXP_MASK) { |
| 482 |
// +-infinity |
| 483 |
e32 = F32_EXP_MASK >> FloatMantissaBits<float>(); |
| 484 |
m32 = 0; |
| 485 |
} else if (e16 != 0 && e16 < 0x1F) { |
| 486 |
// normal number |
| 487 |
e32 = e16 + FloatBiasDiff<float, uint16_t>(); // (127 - 15); // 0x70 |
| 488 |
m32 = m16 << FloatMantissaBitsDiff<float, uint16_t>(); // (23 - 10); |
| 489 |
} else if (e16 == 0 && m16 != 0) { |
| 490 |
// denorm/subnorm number (e16 == 0) => renormalize it |
| 491 |
// shift the mantissa left until the hidden one gets set |
| 492 |
for (e32 = FloatBiasDiff<float, uint16_t>() + 1; |
| 493 |
(m16 & (F16_MANT_MASK + 1)) == 0; m16 <<= 1, e32--) |
| 494 |
; |
| 495 |
m32 = (m16 << FloatMantissaBitsDiff<float, uint16_t>()) & F32_MANT_MASK; |
| 496 |
} else { // if (e16 == 0) // +/- 0.0 |
| 497 |
e32 = 0; |
| 498 |
m32 = 0; |
| 499 |
} |
| 500 |
return FloatFromBits(s32 | (e32 << FloatMantissaBits<float>()) | m32); |
| 501 |
#endif // !IGA_USE_FP16C |
| 502 |
} |
| 503 |
|
| 504 |
double iga::ConvertFloatToDouble(float f) { |
| 505 |
if (IS_NAN(f)) { |
| 506 |
uint32_t f32 = FloatToBits(f); |
| 507 |
|
| 508 |
uint64_t m64; |
| 509 |
m64 = (uint64_t)(f32 & F32_QNAN_BIT) |
| 510 |
<< (FloatMantissaBits<double>() - |
| 511 |
FloatMantissaBits<float>()); // keep the sNaN bit |
| 512 |
m64 |= (F32_MANT_MASK >> 1) & f32; // keep the non sNaN part |
| 513 |
// lower part of the payload |
| 514 |
uint64_t bits = (((uint64_t)f32 & F32_SIGN_BIT) << 32) | // sign |
| 515 |
F64_EXP_MASK | // exp |
| 516 |
m64; // new mantissa |
| 517 |
return FloatFromBits(bits); |
| 518 |
} else { |
| 519 |
// not NaN: use default value |
| 520 |
return (double)f; |
| 521 |
} |
| 522 |
} |
| 523 |
|
| 524 |
|
| 525 |
bool iga::ParseFLTLIT(const std::string &syntax, double &value) { |
| 526 |
char *end = nullptr; |
| 527 |
value = std::strtod(syntax.c_str(), &end); |
| 528 |
if (*end) { |
| 529 |
return false; |
| 530 |
} |
| 531 |
return true; |
| 532 |
} |