| 1 |
/* Half-float conversion routines. |
| 2 |
|
| 3 |
Copyright (C) 2008-2025 Free Software Foundation, Inc. |
| 4 |
Contributed by CodeSourcery. |
| 5 |
|
| 6 |
This file is free software; you can redistribute it and/or modify it |
| 7 |
under the terms of the GNU General Public License as published by the |
| 8 |
Free Software Foundation; either version 3, or (at your option) any |
| 9 |
later version. |
| 10 |
|
| 11 |
This file is distributed in the hope that it will be useful, but |
| 12 |
WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 14 |
General Public License for more details. |
| 15 |
|
| 16 |
Under Section 7 of GPL version 3, you are granted additional |
| 17 |
permissions described in the GCC Runtime Library Exception, version |
| 18 |
3.1, as published by the Free Software Foundation. |
| 19 |
|
| 20 |
You should have received a copy of the GNU General Public License and |
| 21 |
a copy of the GCC Runtime Library Exception along with this program; |
| 22 |
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see |
| 23 |
<http://www.gnu.org/licenses/>. */ |
| 24 |
|
| 25 |
struct format |
| 26 |
{ |
| 27 |
/* Number of bits. */ |
| 28 |
unsigned long long size; |
| 29 |
/* Exponent bias. */ |
| 30 |
unsigned long long bias; |
| 31 |
/* Exponent width in bits. */ |
| 32 |
unsigned long long exponent; |
| 33 |
/* Significand precision in explicitly stored bits. */ |
| 34 |
unsigned long long significand; |
| 35 |
}; |
| 36 |
|
| 37 |
static const struct format |
| 38 |
binary32 = |
| 39 |
{ |
| 40 |
32, /* size. */ |
| 41 |
127, /* bias. */ |
| 42 |
8, /* exponent. */ |
| 43 |
23 /* significand. */ |
| 44 |
}; |
| 45 |
|
| 46 |
static const struct format |
| 47 |
binary64 = |
| 48 |
{ |
| 49 |
64, /* size. */ |
| 50 |
1023, /* bias. */ |
| 51 |
11, /* exponent. */ |
| 52 |
52 /* significand. */ |
| 53 |
}; |
| 54 |
|
| 55 |
/* Function prototypes. */ |
| 56 |
unsigned short __gnu_f2h_ieee (unsigned int a); |
| 57 |
unsigned int __gnu_h2f_ieee (unsigned short a); |
| 58 |
unsigned short __gnu_f2h_alternative (unsigned int x); |
| 59 |
unsigned int __gnu_h2f_alternative (unsigned short a); |
| 60 |
unsigned short __gnu_d2h_ieee (unsigned long long a); |
| 61 |
unsigned short __gnu_d2h_alternative (unsigned long long x); |
| 62 |
|
| 63 |
static inline unsigned short |
| 64 |
__gnu_float2h_internal (const struct format* fmt, |
| 65 |
unsigned long long a, int ieee) |
| 66 |
{ |
| 67 |
unsigned long long point = 1ULL << fmt->significand; |
| 68 |
unsigned short sign = (a >> (fmt->size - 16)) & 0x8000; |
| 69 |
int aexp; |
| 70 |
unsigned long long mantissa; |
| 71 |
unsigned long long mask; |
| 72 |
unsigned long long increment; |
| 73 |
|
| 74 |
/* Get the exponent and mantissa encodings. */ |
| 75 |
mantissa = a & (point - 1); |
| 76 |
|
| 77 |
mask = (1 << fmt->exponent) - 1; |
| 78 |
aexp = (a >> fmt->significand) & mask; |
| 79 |
|
| 80 |
/* Infinity, NaN and alternative format special case. */ |
| 81 |
if (((unsigned int) aexp) == mask) |
| 82 |
{ |
| 83 |
if (!ieee) |
| 84 |
return sign; |
| 85 |
if (mantissa == 0) |
| 86 |
return sign | 0x7c00; /* Infinity. */ |
| 87 |
/* Remaining cases are NaNs. Convert SNaN to QNaN. */ |
| 88 |
return sign | 0x7e00 | (mantissa >> (fmt->significand - 10)); |
| 89 |
} |
| 90 |
|
| 91 |
/* Zero. */ |
| 92 |
if (aexp == 0 && mantissa == 0) |
| 93 |
return sign; |
| 94 |
|
| 95 |
/* Construct the exponent and mantissa. */ |
| 96 |
aexp -= fmt->bias; |
| 97 |
|
| 98 |
/* Decimal point is immediately after the significand. */ |
| 99 |
mantissa |= point; |
| 100 |
|
| 101 |
if (aexp < -14) |
| 102 |
{ |
| 103 |
mask = point | (point - 1); |
| 104 |
/* Minimum exponent for half-precision is 2^-24. */ |
| 105 |
if (aexp >= -25) |
| 106 |
mask >>= 25 + aexp; |
| 107 |
} |
| 108 |
else |
| 109 |
mask = (point - 1) >> 10; |
| 110 |
|
| 111 |
/* Round. */ |
| 112 |
if (mantissa & mask) |
| 113 |
{ |
| 114 |
increment = (mask + 1) >> 1; |
| 115 |
if ((mantissa & mask) == increment) |
| 116 |
increment = mantissa & (increment << 1); |
| 117 |
mantissa += increment; |
| 118 |
if (mantissa >= (point << 1)) |
| 119 |
{ |
| 120 |
mantissa >>= 1; |
| 121 |
aexp++; |
| 122 |
} |
| 123 |
} |
| 124 |
|
| 125 |
if (ieee) |
| 126 |
{ |
| 127 |
if (aexp > 15) |
| 128 |
return sign | 0x7c00; |
| 129 |
} |
| 130 |
else |
| 131 |
{ |
| 132 |
if (aexp > 16) |
| 133 |
return sign | 0x7fff; |
| 134 |
} |
| 135 |
|
| 136 |
if (aexp < -24) |
| 137 |
return sign; |
| 138 |
|
| 139 |
if (aexp < -14) |
| 140 |
{ |
| 141 |
mantissa >>= -14 - aexp; |
| 142 |
aexp = -14; |
| 143 |
} |
| 144 |
|
| 145 |
/* Encode the final 16-bit floating-point value. |
| 146 |
|
| 147 |
This is formed of the sign bit, the bias-adjusted exponent, and the |
| 148 |
calculated mantissa, with the following caveats: |
| 149 |
|
| 150 |
1. The mantissa calculated after rounding could have a leading 1. |
| 151 |
To compensate for this, subtract one from the exponent bias (15) |
| 152 |
before adding it to the calculated exponent. |
| 153 |
2. When we were calculating rounding, we left the mantissa with the |
| 154 |
number of bits of the source operand, it needs reduced to ten |
| 155 |
bits (+1 for the afforementioned leading 1) by shifting right by |
| 156 |
the number of bits in the source mantissa - 10. |
| 157 |
3. To ensure the leading 1 in the mantissa is applied to the exponent |
| 158 |
we need to add the mantissa rather than apply an arithmetic "or" |
| 159 |
to it. */ |
| 160 |
|
| 161 |
return sign | (((aexp + 14) << 10) + (mantissa >> (fmt->significand - 10))); |
| 162 |
} |
| 163 |
|
| 164 |
static inline unsigned short |
| 165 |
__gnu_f2h_internal (unsigned int a, int ieee) |
| 166 |
{ |
| 167 |
return __gnu_float2h_internal (&binary32, (unsigned long long) a, ieee); |
| 168 |
} |
| 169 |
|
| 170 |
static inline unsigned short |
| 171 |
__gnu_d2h_internal (unsigned long long a, int ieee) |
| 172 |
{ |
| 173 |
return __gnu_float2h_internal (&binary64, a, ieee); |
| 174 |
} |
| 175 |
|
| 176 |
static inline unsigned int |
| 177 |
__gnu_h2f_internal(unsigned short a, int ieee) |
| 178 |
{ |
| 179 |
unsigned int sign = (unsigned int)(a & 0x8000) << 16; |
| 180 |
int aexp = (a >> 10) & 0x1f; |
| 181 |
unsigned int mantissa = a & 0x3ff; |
| 182 |
|
| 183 |
if (aexp == 0x1f && ieee) |
| 184 |
return sign | 0x7f800000 | (mantissa << 13); |
| 185 |
|
| 186 |
if (aexp == 0) |
| 187 |
{ |
| 188 |
int shift; |
| 189 |
|
| 190 |
if (mantissa == 0) |
| 191 |
return sign; |
| 192 |
|
| 193 |
shift = __builtin_clz(mantissa) - 21; |
| 194 |
mantissa <<= shift; |
| 195 |
aexp = -shift; |
| 196 |
} |
| 197 |
|
| 198 |
return sign | (((aexp + 0x70) << 23) + (mantissa << 13)); |
| 199 |
} |
| 200 |
|
| 201 |
unsigned short |
| 202 |
__gnu_f2h_ieee(unsigned int a) |
| 203 |
{ |
| 204 |
return __gnu_f2h_internal(a, 1); |
| 205 |
} |
| 206 |
|
| 207 |
unsigned int |
| 208 |
__gnu_h2f_ieee(unsigned short a) |
| 209 |
{ |
| 210 |
return __gnu_h2f_internal(a, 1); |
| 211 |
} |
| 212 |
|
| 213 |
unsigned short |
| 214 |
__gnu_f2h_alternative(unsigned int x) |
| 215 |
{ |
| 216 |
return __gnu_f2h_internal(x, 0); |
| 217 |
} |
| 218 |
|
| 219 |
unsigned int |
| 220 |
__gnu_h2f_alternative(unsigned short a) |
| 221 |
{ |
| 222 |
return __gnu_h2f_internal(a, 0); |
| 223 |
} |
| 224 |
|
| 225 |
unsigned short |
| 226 |
__gnu_d2h_ieee (unsigned long long a) |
| 227 |
{ |
| 228 |
return __gnu_d2h_internal (a, 1); |
| 229 |
} |
| 230 |
|
| 231 |
unsigned short |
| 232 |
__gnu_d2h_alternative (unsigned long long x) |
| 233 |
{ |
| 234 |
return __gnu_d2h_internal (x, 0); |
| 235 |
} |