|
Packit |
6c4009 |
/* @(#)s_log1p.c 5.1 93/09/24 */
|
|
Packit |
6c4009 |
/*
|
|
Packit |
6c4009 |
* ====================================================
|
|
Packit |
6c4009 |
* Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
|
|
Packit |
6c4009 |
*
|
|
Packit |
6c4009 |
* Developed at SunPro, a Sun Microsystems, Inc. business.
|
|
Packit |
6c4009 |
* Permission to use, copy, modify, and distribute this
|
|
Packit |
6c4009 |
* software is freely granted, provided that this notice
|
|
Packit |
6c4009 |
* is preserved.
|
|
Packit |
6c4009 |
* ====================================================
|
|
Packit |
6c4009 |
*/
|
|
Packit |
6c4009 |
/* Modified by Naohiko Shimizu/Tokai University, Japan 1997/08/25,
|
|
Packit |
6c4009 |
for performance improvement on pipelined processors.
|
|
Packit |
6c4009 |
*/
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* double log1p(double x)
|
|
Packit |
6c4009 |
*
|
|
Packit |
6c4009 |
* Method :
|
|
Packit |
6c4009 |
* 1. Argument Reduction: find k and f such that
|
|
Packit |
6c4009 |
* 1+x = 2^k * (1+f),
|
|
Packit |
6c4009 |
* where sqrt(2)/2 < 1+f < sqrt(2) .
|
|
Packit |
6c4009 |
*
|
|
Packit |
6c4009 |
* Note. If k=0, then f=x is exact. However, if k!=0, then f
|
|
Packit |
6c4009 |
* may not be representable exactly. In that case, a correction
|
|
Packit |
6c4009 |
* term is need. Let u=1+x rounded. Let c = (1+x)-u, then
|
|
Packit |
6c4009 |
* log(1+x) - log(u) ~ c/u. Thus, we proceed to compute log(u),
|
|
Packit |
6c4009 |
* and add back the correction term c/u.
|
|
Packit |
6c4009 |
* (Note: when x > 2**53, one can simply return log(x))
|
|
Packit |
6c4009 |
*
|
|
Packit |
6c4009 |
* 2. Approximation of log1p(f).
|
|
Packit |
6c4009 |
* Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s)
|
|
Packit |
6c4009 |
* = 2s + 2/3 s**3 + 2/5 s**5 + .....,
|
|
Packit |
6c4009 |
* = 2s + s*R
|
|
Packit |
6c4009 |
* We use a special Reme algorithm on [0,0.1716] to generate
|
|
Packit |
6c4009 |
* a polynomial of degree 14 to approximate R The maximum error
|
|
Packit |
6c4009 |
* of this polynomial approximation is bounded by 2**-58.45. In
|
|
Packit |
6c4009 |
* other words,
|
|
Packit |
6c4009 |
* 2 4 6 8 10 12 14
|
|
Packit |
6c4009 |
* R(z) ~ Lp1*s +Lp2*s +Lp3*s +Lp4*s +Lp5*s +Lp6*s +Lp7*s
|
|
Packit |
6c4009 |
* (the values of Lp1 to Lp7 are listed in the program)
|
|
Packit |
6c4009 |
* and
|
|
Packit |
6c4009 |
* | 2 14 | -58.45
|
|
Packit |
6c4009 |
* | Lp1*s +...+Lp7*s - R(z) | <= 2
|
|
Packit |
6c4009 |
* | |
|
|
Packit |
6c4009 |
* Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2.
|
|
Packit |
6c4009 |
* In order to guarantee error in log below 1ulp, we compute log
|
|
Packit |
6c4009 |
* by
|
|
Packit |
6c4009 |
* log1p(f) = f - (hfsq - s*(hfsq+R)).
|
|
Packit |
6c4009 |
*
|
|
Packit |
6c4009 |
* 3. Finally, log1p(x) = k*ln2 + log1p(f).
|
|
Packit |
6c4009 |
* = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo)))
|
|
Packit |
6c4009 |
* Here ln2 is split into two floating point number:
|
|
Packit |
6c4009 |
* ln2_hi + ln2_lo,
|
|
Packit |
6c4009 |
* where n*ln2_hi is always exact for |n| < 2000.
|
|
Packit |
6c4009 |
*
|
|
Packit |
6c4009 |
* Special cases:
|
|
Packit |
6c4009 |
* log1p(x) is NaN with signal if x < -1 (including -INF) ;
|
|
Packit |
6c4009 |
* log1p(+INF) is +INF; log1p(-1) is -INF with signal;
|
|
Packit |
6c4009 |
* log1p(NaN) is that NaN with no signal.
|
|
Packit |
6c4009 |
*
|
|
Packit |
6c4009 |
* Accuracy:
|
|
Packit |
6c4009 |
* according to an error analysis, the error is always less than
|
|
Packit |
6c4009 |
* 1 ulp (unit in the last place).
|
|
Packit |
6c4009 |
*
|
|
Packit |
6c4009 |
* Constants:
|
|
Packit |
6c4009 |
* The hexadecimal values are the intended ones for the following
|
|
Packit |
6c4009 |
* constants. The decimal values may be used, provided that the
|
|
Packit |
6c4009 |
* compiler will convert from decimal to binary accurately enough
|
|
Packit |
6c4009 |
* to produce the hexadecimal values shown.
|
|
Packit |
6c4009 |
*
|
|
Packit |
6c4009 |
* Note: Assuming log() return accurate answer, the following
|
|
Packit |
6c4009 |
* algorithm can be used to compute log1p(x) to within a few ULP:
|
|
Packit |
6c4009 |
*
|
|
Packit |
6c4009 |
* u = 1+x;
|
|
Packit |
6c4009 |
* if(u==1.0) return x ; else
|
|
Packit |
6c4009 |
* return log(u)*(x/(u-1.0));
|
|
Packit |
6c4009 |
*
|
|
Packit |
6c4009 |
* See HP-15C Advanced Functions Handbook, p.193.
|
|
Packit |
6c4009 |
*/
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#include <float.h>
|
|
Packit |
6c4009 |
#include <math.h>
|
|
Packit |
6c4009 |
#include <math-barriers.h>
|
|
Packit |
6c4009 |
#include <math_private.h>
|
|
Packit |
6c4009 |
#include <math-underflow.h>
|
|
Packit |
6c4009 |
#include <libc-diag.h>
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
static const double
|
|
Packit |
6c4009 |
ln2_hi = 6.93147180369123816490e-01, /* 3fe62e42 fee00000 */
|
|
Packit |
6c4009 |
ln2_lo = 1.90821492927058770002e-10, /* 3dea39ef 35793c76 */
|
|
Packit |
6c4009 |
two54 = 1.80143985094819840000e+16, /* 43500000 00000000 */
|
|
Packit |
6c4009 |
Lp[] = { 0.0, 6.666666666666735130e-01, /* 3FE55555 55555593 */
|
|
Packit |
6c4009 |
3.999999999940941908e-01, /* 3FD99999 9997FA04 */
|
|
Packit |
6c4009 |
2.857142874366239149e-01, /* 3FD24924 94229359 */
|
|
Packit |
6c4009 |
2.222219843214978396e-01, /* 3FCC71C5 1D8E78AF */
|
|
Packit |
6c4009 |
1.818357216161805012e-01, /* 3FC74664 96CB03DE */
|
|
Packit |
6c4009 |
1.531383769920937332e-01, /* 3FC39A09 D078C69F */
|
|
Packit |
6c4009 |
1.479819860511658591e-01 }; /* 3FC2F112 DF3E5244 */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
static const double zero = 0.0;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
double
|
|
Packit |
6c4009 |
__log1p (double x)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
double hfsq, f, c, s, z, R, u, z2, z4, z6, R1, R2, R3, R4;
|
|
Packit |
6c4009 |
int32_t k, hx, hu, ax;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
GET_HIGH_WORD (hx, x);
|
|
Packit |
6c4009 |
ax = hx & 0x7fffffff;
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
k = 1;
|
|
Packit |
6c4009 |
if (hx < 0x3FDA827A) /* x < 0.41422 */
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
if (__glibc_unlikely (ax >= 0x3ff00000)) /* x <= -1.0 */
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
if (x == -1.0)
|
|
Packit |
6c4009 |
return -two54 / zero; /* log1p(-1)=-inf */
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
return (x - x) / (x - x); /* log1p(x<-1)=NaN */
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
if (__glibc_unlikely (ax < 0x3e200000)) /* |x| < 2**-29 */
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
math_force_eval (two54 + x); /* raise inexact */
|
|
Packit |
6c4009 |
if (ax < 0x3c900000) /* |x| < 2**-54 */
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
math_check_force_underflow (x);
|
|
Packit |
6c4009 |
return x;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
return x - x * x * 0.5;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
if (hx > 0 || hx <= ((int32_t) 0xbfd2bec3))
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
k = 0; f = x; hu = 1;
|
|
Packit |
6c4009 |
} /* -0.2929
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
else if (__glibc_unlikely (hx >= 0x7ff00000))
|
|
Packit |
6c4009 |
return x + x;
|
|
Packit |
6c4009 |
if (k != 0)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
if (hx < 0x43400000)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
u = 1.0 + x;
|
|
Packit |
6c4009 |
GET_HIGH_WORD (hu, u);
|
|
Packit |
6c4009 |
k = (hu >> 20) - 1023;
|
|
Packit |
6c4009 |
c = (k > 0) ? 1.0 - (u - x) : x - (u - 1.0); /* correction term */
|
|
Packit |
6c4009 |
c /= u;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
u = x;
|
|
Packit |
6c4009 |
GET_HIGH_WORD (hu, u);
|
|
Packit |
6c4009 |
k = (hu >> 20) - 1023;
|
|
Packit |
6c4009 |
c = 0;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
hu &= 0x000fffff;
|
|
Packit |
6c4009 |
if (hu < 0x6a09e)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
SET_HIGH_WORD (u, hu | 0x3ff00000); /* normalize u */
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
k += 1;
|
|
Packit |
6c4009 |
SET_HIGH_WORD (u, hu | 0x3fe00000); /* normalize u/2 */
|
|
Packit |
6c4009 |
hu = (0x00100000 - hu) >> 2;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
f = u - 1.0;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
hfsq = 0.5 * f * f;
|
|
Packit |
6c4009 |
if (hu == 0) /* |f| < 2**-20 */
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
if (f == zero)
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
if (k == 0)
|
|
Packit |
6c4009 |
return zero;
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
c += k * ln2_lo; return k * ln2_hi + c;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
R = hfsq * (1.0 - 0.66666666666666666 * f);
|
|
Packit |
6c4009 |
if (k == 0)
|
|
Packit |
6c4009 |
return f - R;
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
return k * ln2_hi - ((R - (k * ln2_lo + c)) - f);
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
s = f / (2.0 + f);
|
|
Packit |
6c4009 |
z = s * s;
|
|
Packit |
6c4009 |
R1 = z * Lp[1]; z2 = z * z;
|
|
Packit |
6c4009 |
R2 = Lp[2] + z * Lp[3]; z4 = z2 * z2;
|
|
Packit |
6c4009 |
R3 = Lp[4] + z * Lp[5]; z6 = z4 * z2;
|
|
Packit |
6c4009 |
R4 = Lp[6] + z * Lp[7];
|
|
Packit |
6c4009 |
R = R1 + z2 * R2 + z4 * R3 + z6 * R4;
|
|
Packit |
6c4009 |
if (k == 0)
|
|
Packit |
6c4009 |
return f - (hfsq - s * (hfsq + R));
|
|
Packit |
6c4009 |
else
|
|
Packit |
6c4009 |
{
|
|
Packit |
6c4009 |
/* With GCC 7 when compiling with -Os the compiler warns that c
|
|
Packit |
6c4009 |
might be used uninitialized. This can't be true because k
|
|
Packit |
6c4009 |
must be 0 for c to be uninitialized and we handled that
|
|
Packit |
6c4009 |
computation earlier without using c. */
|
|
Packit |
6c4009 |
DIAG_PUSH_NEEDS_COMMENT;
|
|
Packit |
6c4009 |
DIAG_IGNORE_Os_NEEDS_COMMENT (7, "-Wmaybe-uninitialized");
|
|
Packit |
6c4009 |
return k * ln2_hi - ((hfsq - (s * (hfsq + R) + (k * ln2_lo + c))) - f);
|
|
Packit |
6c4009 |
DIAG_POP_NEEDS_COMMENT;
|
|
Packit |
6c4009 |
}
|
|
Packit |
6c4009 |
}
|