| 1 |
#include <stdint.h> |
| 2 |
#include <math.h> |
| 3 |
#include "libm.h" |
| 4 |
#include "sqrt_data.h" |
| 5 |
|
| 6 |
#define FENV_SUPPORT 1 |
| 7 |
|
| 8 |
static inline uint32_t mul32(uint32_t a, uint32_t b) |
| 9 |
{ |
| 10 |
return (uint64_t)a*b >> 32; |
| 11 |
} |
| 12 |
|
| 13 |
/* see sqrt.c for more detailed comments. */ |
| 14 |
|
| 15 |
float sqrtf(float x) |
| 16 |
{ |
| 17 |
uint32_t ix, m, m1, m0, even, ey; |
| 18 |
|
| 19 |
ix = asuint(x); |
| 20 |
if (predict_false(ix - 0x00800000 >= 0x7f800000 - 0x00800000)) { |
| 21 |
/* x < 0x1p-126 or inf or nan. */ |
| 22 |
if (ix * 2 == 0) |
| 23 |
return x; |
| 24 |
if (ix == 0x7f800000) |
| 25 |
return x; |
| 26 |
if (ix > 0x7f800000) |
| 27 |
return __math_invalidf(x); |
| 28 |
/* x is subnormal, normalize it. */ |
| 29 |
ix = asuint(x * 0x1p23f); |
| 30 |
ix -= 23 << 23; |
| 31 |
} |
| 32 |
|
| 33 |
/* x = 4^e m; with int e and m in [1, 4). */ |
| 34 |
even = ix & 0x00800000; |
| 35 |
m1 = (ix << 8) | 0x80000000; |
| 36 |
m0 = (ix << 7) & 0x7fffffff; |
| 37 |
m = even ? m0 : m1; |
| 38 |
|
| 39 |
/* 2^e is the exponent part of the return value. */ |
| 40 |
ey = ix >> 1; |
| 41 |
ey += 0x3f800000 >> 1; |
| 42 |
ey &= 0x7f800000; |
| 43 |
|
| 44 |
/* compute r ~ 1/sqrt(m), s ~ sqrt(m) with 2 goldschmidt iterations. */ |
| 45 |
static const uint32_t three = 0xc0000000; |
| 46 |
uint32_t r, s, d, u, i; |
| 47 |
i = (ix >> 17) % 128; |
| 48 |
r = (uint32_t)__rsqrt_tab[i] << 16; |
| 49 |
/* |r*sqrt(m) - 1| < 0x1p-8 */ |
| 50 |
s = mul32(m, r); |
| 51 |
/* |s/sqrt(m) - 1| < 0x1p-8 */ |
| 52 |
d = mul32(s, r); |
| 53 |
u = three - d; |
| 54 |
r = mul32(r, u) << 1; |
| 55 |
/* |r*sqrt(m) - 1| < 0x1.7bp-16 */ |
| 56 |
s = mul32(s, u) << 1; |
| 57 |
/* |s/sqrt(m) - 1| < 0x1.7bp-16 */ |
| 58 |
d = mul32(s, r); |
| 59 |
u = three - d; |
| 60 |
s = mul32(s, u); |
| 61 |
/* -0x1.03p-28 < s/sqrt(m) - 1 < 0x1.fp-31 */ |
| 62 |
s = (s - 1)>>6; |
| 63 |
/* s < sqrt(m) < s + 0x1.08p-23 */ |
| 64 |
|
| 65 |
/* compute nearest rounded result. */ |
| 66 |
uint32_t d0, d1, d2; |
| 67 |
float y, t; |
| 68 |
d0 = (m << 16) - s*s; |
| 69 |
d1 = s - d0; |
| 70 |
d2 = d1 + s + 1; |
| 71 |
s += d1 >> 31; |
| 72 |
s &= 0x007fffff; |
| 73 |
s |= ey; |
| 74 |
y = asfloat(s); |
| 75 |
if (FENV_SUPPORT) { |
| 76 |
/* handle rounding and inexact exception. */ |
| 77 |
uint32_t tiny = predict_false(d2==0) ? 0 : 0x01000000; |
| 78 |
tiny |= (d1^d2) & 0x80000000; |
| 79 |
t = asfloat(tiny); |
| 80 |
y = eval_as_float(y + t); |
| 81 |
} |
| 82 |
return y; |
| 83 |
} |