| 1 |
#include <stdint.h> |
| 2 |
#include <float.h> |
| 3 |
#include <math.h> |
| 4 |
#include "atomic.h" |
| 5 |
|
| 6 |
#define ASUINT64(x) ((union {double f; uint64_t i;}){x}).i |
| 7 |
#define ZEROINFNAN (0x7ff-0x3ff-52-1) |
| 8 |
|
| 9 |
struct num { uint64_t m; int e; int sign; }; |
| 10 |
|
| 11 |
static struct num normalize(double x) |
| 12 |
{ |
| 13 |
uint64_t ix = ASUINT64(x); |
| 14 |
int e = ix>>52; |
| 15 |
int sign = e & 0x800; |
| 16 |
e &= 0x7ff; |
| 17 |
if (!e) { |
| 18 |
ix = ASUINT64(x*0x1p63); |
| 19 |
e = ix>>52 & 0x7ff; |
| 20 |
e = e ? e-63 : 0x800; |
| 21 |
} |
| 22 |
ix &= (1ull<<52)-1; |
| 23 |
ix |= 1ull<<52; |
| 24 |
ix <<= 1; |
| 25 |
e -= 0x3ff + 52 + 1; |
| 26 |
return (struct num){ix,e,sign}; |
| 27 |
} |
| 28 |
|
| 29 |
static void mul(uint64_t *hi, uint64_t *lo, uint64_t x, uint64_t y) |
| 30 |
{ |
| 31 |
uint64_t t1,t2,t3; |
| 32 |
uint64_t xlo = (uint32_t)x, xhi = x>>32; |
| 33 |
uint64_t ylo = (uint32_t)y, yhi = y>>32; |
| 34 |
|
| 35 |
t1 = xlo*ylo; |
| 36 |
t2 = xlo*yhi + xhi*ylo; |
| 37 |
t3 = xhi*yhi; |
| 38 |
*lo = t1 + (t2<<32); |
| 39 |
*hi = t3 + (t2>>32) + (t1 > *lo); |
| 40 |
} |
| 41 |
|
| 42 |
double fma(double x, double y, double z) |
| 43 |
{ |
| 44 |
#pragma STDC FENV_ACCESS ON |
| 45 |
|
| 46 |
/* normalize so top 10bits and last bit are 0 */ |
| 47 |
struct num nx, ny, nz; |
| 48 |
nx = normalize(x); |
| 49 |
ny = normalize(y); |
| 50 |
nz = normalize(z); |
| 51 |
|
| 52 |
if (nx.e >= ZEROINFNAN || ny.e >= ZEROINFNAN) |
| 53 |
return x*y + z; |
| 54 |
if (nz.e >= ZEROINFNAN) { |
| 55 |
if (nz.e > ZEROINFNAN) /* z==0 */ |
| 56 |
return x*y + z; |
| 57 |
return z; |
| 58 |
} |
| 59 |
|
| 60 |
/* mul: r = x*y */ |
| 61 |
uint64_t rhi, rlo, zhi, zlo; |
| 62 |
mul(&rhi, &rlo, nx.m, ny.m); |
| 63 |
/* either top 20 or 21 bits of rhi and last 2 bits of rlo are 0 */ |
| 64 |
|
| 65 |
/* align exponents */ |
| 66 |
int e = nx.e + ny.e; |
| 67 |
int d = nz.e - e; |
| 68 |
/* shift bits z<<=kz, r>>=kr, so kz+kr == d, set e = e+kr (== ez-kz) */ |
| 69 |
if (d > 0) { |
| 70 |
if (d < 64) { |
| 71 |
zlo = nz.m<<d; |
| 72 |
zhi = nz.m>>64-d; |
| 73 |
} else { |
| 74 |
zlo = 0; |
| 75 |
zhi = nz.m; |
| 76 |
e = nz.e - 64; |
| 77 |
d -= 64; |
| 78 |
if (d == 0) { |
| 79 |
} else if (d < 64) { |
| 80 |
rlo = rhi<<64-d | rlo>>d | !!(rlo<<64-d); |
| 81 |
rhi = rhi>>d; |
| 82 |
} else { |
| 83 |
rlo = 1; |
| 84 |
rhi = 0; |
| 85 |
} |
| 86 |
} |
| 87 |
} else { |
| 88 |
zhi = 0; |
| 89 |
d = -d; |
| 90 |
if (d == 0) { |
| 91 |
zlo = nz.m; |
| 92 |
} else if (d < 64) { |
| 93 |
zlo = nz.m>>d | !!(nz.m<<64-d); |
| 94 |
} else { |
| 95 |
zlo = 1; |
| 96 |
} |
| 97 |
} |
| 98 |
|
| 99 |
/* add */ |
| 100 |
int sign = nx.sign^ny.sign; |
| 101 |
int samesign = !(sign^nz.sign); |
| 102 |
int nonzero = 1; |
| 103 |
if (samesign) { |
| 104 |
/* r += z */ |
| 105 |
rlo += zlo; |
| 106 |
rhi += zhi + (rlo < zlo); |
| 107 |
} else { |
| 108 |
/* r -= z */ |
| 109 |
uint64_t t = rlo; |
| 110 |
rlo -= zlo; |
| 111 |
rhi = rhi - zhi - (t < rlo); |
| 112 |
if (rhi>>63) { |
| 113 |
rlo = -rlo; |
| 114 |
rhi = -rhi-!!rlo; |
| 115 |
sign = !sign; |
| 116 |
} |
| 117 |
nonzero = !!rhi; |
| 118 |
} |
| 119 |
|
| 120 |
/* set rhi to top 63bit of the result (last bit is sticky) */ |
| 121 |
if (nonzero) { |
| 122 |
e += 64; |
| 123 |
d = a_clz_64(rhi)-1; |
| 124 |
/* note: d > 0 */ |
| 125 |
rhi = rhi<<d | rlo>>64-d | !!(rlo<<d); |
| 126 |
} else if (rlo) { |
| 127 |
d = a_clz_64(rlo)-1; |
| 128 |
if (d < 0) |
| 129 |
rhi = rlo>>1 | (rlo&1); |
| 130 |
else |
| 131 |
rhi = rlo<<d; |
| 132 |
} else { |
| 133 |
/* exact +-0 */ |
| 134 |
return x*y + z; |
| 135 |
} |
| 136 |
e -= d; |
| 137 |
|
| 138 |
/* convert to double */ |
| 139 |
int64_t i = rhi; /* i is in [1<<62,(1<<63)-1] */ |
| 140 |
if (sign) |
| 141 |
i = -i; |
| 142 |
double r = i; /* |r| is in [0x1p62,0x1p63] */ |
| 143 |
|
| 144 |
if (e < -1022-62) { |
| 145 |
/* result is subnormal before rounding */ |
| 146 |
if (e == -1022-63) { |
| 147 |
double c = 0x1p63; |
| 148 |
if (sign) |
| 149 |
c = -c; |
| 150 |
if (r == c) { |
| 151 |
/* min normal after rounding, underflow depends |
| 152 |
on arch behaviour which can be imitated by |
| 153 |
a double to float conversion */ |
| 154 |
float fltmin = 0x0.ffffff8p-63*FLT_MIN * r; |
| 155 |
return DBL_MIN/FLT_MIN * fltmin; |
| 156 |
} |
| 157 |
/* one bit is lost when scaled, add another top bit to |
| 158 |
only round once at conversion if it is inexact */ |
| 159 |
if (rhi << 53) { |
| 160 |
i = rhi>>1 | (rhi&1) | 1ull<<62; |
| 161 |
if (sign) |
| 162 |
i = -i; |
| 163 |
r = i; |
| 164 |
r = 2*r - c; /* remove top bit */ |
| 165 |
|
| 166 |
/* raise underflow portably, such that it |
| 167 |
cannot be optimized away */ |
| 168 |
{ |
| 169 |
double_t tiny = DBL_MIN/FLT_MIN * r; |
| 170 |
r += (double)(tiny*tiny) * (r-r); |
| 171 |
} |
| 172 |
} |
| 173 |
} else { |
| 174 |
/* only round once when scaled */ |
| 175 |
d = 10; |
| 176 |
i = ( rhi>>d | !!(rhi<<64-d) ) << d; |
| 177 |
if (sign) |
| 178 |
i = -i; |
| 179 |
r = i; |
| 180 |
} |
| 181 |
} |
| 182 |
return scalbn(r, e); |
| 183 |
} |