glibc/sysdeps/ieee754/ldbl-96/e_j1l.c
Joseph Myers ad180676b8 Adjust thresholds in Bessel function implementations (bug 14469).
A recent discussion in bug 14469 notes that a threshold in float
Bessel function implementations, used to determine when to use a
simpler implementation approach, results in substantially inaccurate
results.

As I discussed in
<https://sourceware.org/ml/libc-alpha/2013-03/msg00345.html>, a
heuristic argument suggests 2^(S+P) as the right order of magnitude
for a suitable threshold, where S is the number of significand bits in
the floating-point type and P is the number of significant bits in the
representation of the floating-point type, and the float and ldbl-96
implementations use thresholds that are too small.  Some threshold
does need using, there or elsewhere in the implementation, to avoid
spurious underflow and overflow for large arguments.

This patch sets the thresholds in the affected implementations to more
heuristically justifiable values.  Results will still be inaccurate
close to zeroes of the functions (thus this patch does *not* fix any
of the bugs for Bessel function inaccuracy); fixing that would require
a different implementation approach, likely along the lines described
in <http://www.cl.cam.ac.uk/~jrh13/papers/bessel.ps.gz>.

So the justification for a change such as this would be statistical
rather than based on particular tests that had excessive errors and no
longer do so (no doubt such tests could be found, but would probably
be too fragile to add to the testsuite, as liable to give large errors
again from very small implementation changes or even from compiler
changes).  See
<https://sourceware.org/ml/libc-alpha/2020-02/msg00638.html> for such
statistics of the resulting improvements for float functions.

Tested (glibc testsuite) for x86_64.
2020-02-14 14:16:25 +00:00

553 lines
17 KiB
C

/*
* ====================================================
* Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
*
* Developed at SunPro, a Sun Microsystems, Inc. business.
* Permission to use, copy, modify, and distribute this
* software is freely granted, provided that this notice
* is preserved.
* ====================================================
*/
/* Long double expansions are
Copyright (C) 2001 Stephen L. Moshier <moshier@na-net.ornl.gov>
and are incorporated herein by permission of the author. The author
reserves the right to distribute this material elsewhere under different
copying permissions. These modifications are distributed here under
the following terms:
This library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
This library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with this library; if not, see
<https://www.gnu.org/licenses/>. */
/* __ieee754_j1(x), __ieee754_y1(x)
* Bessel function of the first and second kinds of order zero.
* Method -- j1(x):
* 1. For tiny x, we use j1(x) = x/2 - x^3/16 + x^5/384 - ...
* 2. Reduce x to |x| since j1(x)=-j1(-x), and
* for x in (0,2)
* j1(x) = x/2 + x*z*R0/S0, where z = x*x;
* for x in (2,inf)
* j1(x) = sqrt(2/(pi*x))*(p1(x)*cos(x1)-q1(x)*sin(x1))
* y1(x) = sqrt(2/(pi*x))*(p1(x)*sin(x1)+q1(x)*cos(x1))
* where x1 = x-3*pi/4. It is better to compute sin(x1),cos(x1)
* as follow:
* cos(x1) = cos(x)cos(3pi/4)+sin(x)sin(3pi/4)
* = 1/sqrt(2) * (sin(x) - cos(x))
* sin(x1) = sin(x)cos(3pi/4)-cos(x)sin(3pi/4)
* = -1/sqrt(2) * (sin(x) + cos(x))
* (To avoid cancellation, use
* sin(x) +- cos(x) = -cos(2x)/(sin(x) -+ cos(x))
* to compute the worse one.)
*
* 3 Special cases
* j1(nan)= nan
* j1(0) = 0
* j1(inf) = 0
*
* Method -- y1(x):
* 1. screen out x<=0 cases: y1(0)=-inf, y1(x<0)=NaN
* 2. For x<2.
* Since
* y1(x) = 2/pi*(j1(x)*(ln(x/2)+Euler)-1/x-x/2+5/64*x^3-...)
* therefore y1(x)-2/pi*j1(x)*ln(x)-1/x is an odd function.
* We use the following function to approximate y1,
* y1(x) = x*U(z)/V(z) + (2/pi)*(j1(x)*ln(x)-1/x), z= x^2
* Note: For tiny x, 1/x dominate y1 and hence
* y1(tiny) = -2/pi/tiny
* 3. For x>=2.
* y1(x) = sqrt(2/(pi*x))*(p1(x)*sin(x1)+q1(x)*cos(x1))
* where x1 = x-3*pi/4. It is better to compute sin(x1),cos(x1)
* by method mentioned above.
*/
#include <errno.h>
#include <float.h>
#include <math.h>
#include <math_private.h>
#include <math-underflow.h>
#include <libm-alias-finite.h>
static long double pone (long double), qone (long double);
static const long double
huge = 1e4930L,
one = 1.0L,
invsqrtpi = 5.6418958354775628694807945156077258584405e-1L,
tpi = 6.3661977236758134307553505349005744813784e-1L,
/* J1(x) = .5 x + x x^2 R(x^2) / S(x^2)
0 <= x <= 2
Peak relative error 4.5e-21 */
R[5] = {
-9.647406112428107954753770469290757756814E7L,
2.686288565865230690166454005558203955564E6L,
-3.689682683905671185891885948692283776081E4L,
2.195031194229176602851429567792676658146E2L,
-5.124499848728030297902028238597308971319E-1L,
},
S[4] =
{
1.543584977988497274437410333029029035089E9L,
2.133542369567701244002565983150952549520E7L,
1.394077011298227346483732156167414670520E5L,
5.252401789085732428842871556112108446506E2L,
/* 1.000000000000000000000000000000000000000E0L, */
};
static const long double zero = 0.0;
long double
__ieee754_j1l (long double x)
{
long double z, c, r, s, ss, cc, u, v, y;
int32_t ix;
uint32_t se;
GET_LDOUBLE_EXP (se, x);
ix = se & 0x7fff;
if (__glibc_unlikely (ix >= 0x7fff))
return one / x;
y = fabsl (x);
if (ix >= 0x4000)
{ /* |x| >= 2.0 */
__sincosl (y, &s, &c);
ss = -s - c;
cc = s - c;
if (ix < 0x7ffe)
{ /* make sure y+y not overflow */
z = __cosl (y + y);
if ((s * c) > zero)
cc = z / ss;
else
ss = z / cc;
}
/*
* j1(x) = 1/sqrt(pi) * (P(1,x)*cc - Q(1,x)*ss) / sqrt(x)
* y1(x) = 1/sqrt(pi) * (P(1,x)*ss + Q(1,x)*cc) / sqrt(x)
*/
if (__glibc_unlikely (ix > 0x408e))
z = (invsqrtpi * cc) / sqrtl (y);
else
{
u = pone (y);
v = qone (y);
z = invsqrtpi * (u * cc - v * ss) / sqrtl (y);
}
if (se & 0x8000)
return -z;
else
return z;
}
if (__glibc_unlikely (ix < 0x3fde)) /* |x| < 2^-33 */
{
if (huge + x > one) /* inexact if x!=0 necessary */
{
long double ret = 0.5 * x;
math_check_force_underflow (ret);
if (ret == 0 && x != 0)
__set_errno (ERANGE);
return ret;
}
}
z = x * x;
r = z * (R[0] + z * (R[1]+ z * (R[2] + z * (R[3] + z * R[4]))));
s = S[0] + z * (S[1] + z * (S[2] + z * (S[3] + z)));
r *= x;
return (x * 0.5 + r / s);
}
libm_alias_finite (__ieee754_j1l, __j1l)
/* Y1(x) = 2/pi * (log(x) * j1(x) - 1/x) + x R(x^2)
0 <= x <= 2
Peak relative error 2.3e-23 */
static const long double U0[6] = {
-5.908077186259914699178903164682444848615E10L,
1.546219327181478013495975514375773435962E10L,
-6.438303331169223128870035584107053228235E8L,
9.708540045657182600665968063824819371216E6L,
-6.138043997084355564619377183564196265471E4L,
1.418503228220927321096904291501161800215E2L,
};
static const long double V0[5] = {
3.013447341682896694781964795373783679861E11L,
4.669546565705981649470005402243136124523E9L,
3.595056091631351184676890179233695857260E7L,
1.761554028569108722903944659933744317994E5L,
5.668480419646516568875555062047234534863E2L,
/* 1.000000000000000000000000000000000000000E0L, */
};
long double
__ieee754_y1l (long double x)
{
long double z, s, c, ss, cc, u, v;
int32_t ix;
uint32_t se, i0, i1;
GET_LDOUBLE_WORDS (se, i0, i1, x);
ix = se & 0x7fff;
/* if Y1(NaN) is NaN, Y1(-inf) is NaN, Y1(inf) is 0 */
if (__glibc_unlikely (se & 0x8000))
return zero / (zero * x);
if (__glibc_unlikely (ix >= 0x7fff))
return one / (x + x * x);
if (__glibc_unlikely ((i0 | i1) == 0))
return -HUGE_VALL + x; /* -inf and overflow exception. */
if (ix >= 0x4000)
{ /* |x| >= 2.0 */
__sincosl (x, &s, &c);
ss = -s - c;
cc = s - c;
if (ix < 0x7ffe)
{ /* make sure x+x not overflow */
z = __cosl (x + x);
if ((s * c) > zero)
cc = z / ss;
else
ss = z / cc;
}
/* y1(x) = sqrt(2/(pi*x))*(p1(x)*sin(x0)+q1(x)*cos(x0))
* where x0 = x-3pi/4
* Better formula:
* cos(x0) = cos(x)cos(3pi/4)+sin(x)sin(3pi/4)
* = 1/sqrt(2) * (sin(x) - cos(x))
* sin(x0) = sin(x)cos(3pi/4)-cos(x)sin(3pi/4)
* = -1/sqrt(2) * (cos(x) + sin(x))
* To avoid cancellation, use
* sin(x) +- cos(x) = -cos(2x)/(sin(x) -+ cos(x))
* to compute the worse one.
*/
if (__glibc_unlikely (ix > 0x408e))
z = (invsqrtpi * ss) / sqrtl (x);
else
{
u = pone (x);
v = qone (x);
z = invsqrtpi * (u * ss + v * cc) / sqrtl (x);
}
return z;
}
if (__glibc_unlikely (ix <= 0x3fbe))
{ /* x < 2**-65 */
z = -tpi / x;
if (isinf (z))
__set_errno (ERANGE);
return z;
}
z = x * x;
u = U0[0] + z * (U0[1] + z * (U0[2] + z * (U0[3] + z * (U0[4] + z * U0[5]))));
v = V0[0] + z * (V0[1] + z * (V0[2] + z * (V0[3] + z * (V0[4] + z))));
return (x * (u / v) +
tpi * (__ieee754_j1l (x) * __ieee754_logl (x) - one / x));
}
libm_alias_finite (__ieee754_y1l, __y1l)
/* For x >= 8, the asymptotic expansions of pone is
* 1 + 15/128 s^2 - 4725/2^15 s^4 - ..., where s = 1/x.
* We approximate pone by
* pone(x) = 1 + (R/S)
*/
/* J1(x) cosX + Y1(x) sinX = sqrt( 2/(pi x)) P1(x)
P1(x) = 1 + z^2 R(z^2), z=1/x
8 <= x <= inf (0 <= z <= 0.125)
Peak relative error 5.2e-22 */
static const long double pr8[7] = {
8.402048819032978959298664869941375143163E-9L,
1.813743245316438056192649247507255996036E-6L,
1.260704554112906152344932388588243836276E-4L,
3.439294839869103014614229832700986965110E-3L,
3.576910849712074184504430254290179501209E-2L,
1.131111483254318243139953003461511308672E-1L,
4.480715825681029711521286449131671880953E-2L,
};
static const long double ps8[6] = {
7.169748325574809484893888315707824924354E-8L,
1.556549720596672576431813934184403614817E-5L,
1.094540125521337139209062035774174565882E-3L,
3.060978962596642798560894375281428805840E-2L,
3.374146536087205506032643098619414507024E-1L,
1.253830208588979001991901126393231302559E0L,
/* 1.000000000000000000000000000000000000000E0L, */
};
/* J1(x) cosX + Y1(x) sinX = sqrt( 2/(pi x)) P1(x)
P1(x) = 1 + z^2 R(z^2), z=1/x
4.54541015625 <= x <= 8
Peak relative error 7.7e-22 */
static const long double pr5[7] = {
4.318486887948814529950980396300969247900E-7L,
4.715341880798817230333360497524173929315E-5L,
1.642719430496086618401091544113220340094E-3L,
2.228688005300803935928733750456396149104E-2L,
1.142773760804150921573259605730018327162E-1L,
1.755576530055079253910829652698703791957E-1L,
3.218803858282095929559165965353784980613E-2L,
};
static const long double ps5[6] = {
3.685108812227721334719884358034713967557E-6L,
4.069102509511177498808856515005792027639E-4L,
1.449728676496155025507893322405597039816E-2L,
2.058869213229520086582695850441194363103E-1L,
1.164890985918737148968424972072751066553E0L,
2.274776933457009446573027260373361586841E0L,
/* 1.000000000000000000000000000000000000000E0L,*/
};
/* J1(x) cosX + Y1(x) sinX = sqrt( 2/(pi x)) P1(x)
P1(x) = 1 + z^2 R(z^2), z=1/x
2.85711669921875 <= x <= 4.54541015625
Peak relative error 6.5e-21 */
static const long double pr3[7] = {
1.265251153957366716825382654273326407972E-5L,
8.031057269201324914127680782288352574567E-4L,
1.581648121115028333661412169396282881035E-2L,
1.179534658087796321928362981518645033967E-1L,
3.227936912780465219246440724502790727866E-1L,
2.559223765418386621748404398017602935764E-1L,
2.277136933287817911091370397134882441046E-2L,
};
static const long double ps3[6] = {
1.079681071833391818661952793568345057548E-4L,
6.986017817100477138417481463810841529026E-3L,
1.429403701146942509913198539100230540503E-1L,
1.148392024337075609460312658938700765074E0L,
3.643663015091248720208251490291968840882E0L,
3.990702269032018282145100741746633960737E0L,
/* 1.000000000000000000000000000000000000000E0L, */
};
/* J1(x) cosX + Y1(x) sinX = sqrt( 2/(pi x)) P1(x)
P1(x) = 1 + z^2 R(z^2), z=1/x
2 <= x <= 2.85711669921875
Peak relative error 3.5e-21 */
static const long double pr2[7] = {
2.795623248568412225239401141338714516445E-4L,
1.092578168441856711925254839815430061135E-2L,
1.278024620468953761154963591853679640560E-1L,
5.469680473691500673112904286228351988583E-1L,
8.313769490922351300461498619045639016059E-1L,
3.544176317308370086415403567097130611468E-1L,
1.604142674802373041247957048801599740644E-2L,
};
static const long double ps2[6] = {
2.385605161555183386205027000675875235980E-3L,
9.616778294482695283928617708206967248579E-2L,
1.195215570959693572089824415393951258510E0L,
5.718412857897054829999458736064922974662E0L,
1.065626298505499086386584642761602177568E1L,
6.809140730053382188468983548092322151791E0L,
/* 1.000000000000000000000000000000000000000E0L, */
};
static long double
pone (long double x)
{
const long double *p, *q;
long double z, r, s;
int32_t ix;
uint32_t se, i0, i1;
GET_LDOUBLE_WORDS (se, i0, i1, x);
ix = se & 0x7fff;
/* ix >= 0x4000 for all calls to this function. */
if (ix >= 0x4002) /* x >= 8 */
{
p = pr8;
q = ps8;
}
else
{
i1 = (ix << 16) | (i0 >> 16);
if (i1 >= 0x40019174) /* x >= 4.54541015625 */
{
p = pr5;
q = ps5;
}
else if (i1 >= 0x4000b6db) /* x >= 2.85711669921875 */
{
p = pr3;
q = ps3;
}
else /* x >= 2 */
{
p = pr2;
q = ps2;
}
}
z = one / (x * x);
r = p[0] + z * (p[1] +
z * (p[2] + z * (p[3] + z * (p[4] + z * (p[5] + z * p[6])))));
s = q[0] + z * (q[1] + z * (q[2] + z * (q[3] + z * (q[4] + z * (q[5] + z)))));
return one + z * r / s;
}
/* For x >= 8, the asymptotic expansions of qone is
* 3/8 s - 105/1024 s^3 - ..., where s = 1/x.
* We approximate pone by
* qone(x) = s*(0.375 + (R/S))
*/
/* Y1(x)cosX - J1(x)sinX = sqrt( 2/(pi x)) Q1(x),
Q1(x) = z(.375 + z^2 R(z^2)), z=1/x
8 <= x <= inf
Peak relative error 8.3e-22 */
static const long double qr8[7] = {
-5.691925079044209246015366919809404457380E-10L,
-1.632587664706999307871963065396218379137E-7L,
-1.577424682764651970003637263552027114600E-5L,
-6.377627959241053914770158336842725291713E-4L,
-1.087408516779972735197277149494929568768E-2L,
-6.854943629378084419631926076882330494217E-2L,
-1.055448290469180032312893377152490183203E-1L,
};
static const long double qs8[7] = {
5.550982172325019811119223916998393907513E-9L,
1.607188366646736068460131091130644192244E-6L,
1.580792530091386496626494138334505893599E-4L,
6.617859900815747303032860443855006056595E-3L,
1.212840547336984859952597488863037659161E-1L,
9.017885953937234900458186716154005541075E-1L,
2.201114489712243262000939120146436167178E0L,
/* 1.000000000000000000000000000000000000000E0L, */
};
/* Y1(x)cosX - J1(x)sinX = sqrt( 2/(pi x)) Q1(x),
Q1(x) = z(.375 + z^2 R(z^2)), z=1/x
4.54541015625 <= x <= 8
Peak relative error 4.1e-22 */
static const long double qr5[7] = {
-6.719134139179190546324213696633564965983E-8L,
-9.467871458774950479909851595678622044140E-6L,
-4.429341875348286176950914275723051452838E-4L,
-8.539898021757342531563866270278505014487E-3L,
-6.818691805848737010422337101409276287170E-2L,
-1.964432669771684034858848142418228214855E-1L,
-1.333896496989238600119596538299938520726E-1L,
};
static const long double qs5[7] = {
6.552755584474634766937589285426911075101E-7L,
9.410814032118155978663509073200494000589E-5L,
4.561677087286518359461609153655021253238E-3L,
9.397742096177905170800336715661091535805E-2L,
8.518538116671013902180962914473967738771E-1L,
3.177729183645800174212539541058292579009E0L,
4.006745668510308096259753538973038902990E0L,
/* 1.000000000000000000000000000000000000000E0L, */
};
/* Y1(x)cosX - J1(x)sinX = sqrt( 2/(pi x)) Q1(x),
Q1(x) = z(.375 + z^2 R(z^2)), z=1/x
2.85711669921875 <= x <= 4.54541015625
Peak relative error 2.2e-21 */
static const long double qr3[7] = {
-3.618746299358445926506719188614570588404E-6L,
-2.951146018465419674063882650970344502798E-4L,
-7.728518171262562194043409753656506795258E-3L,
-8.058010968753999435006488158237984014883E-2L,
-3.356232856677966691703904770937143483472E-1L,
-4.858192581793118040782557808823460276452E-1L,
-1.592399251246473643510898335746432479373E-1L,
};
static const long double qs3[7] = {
3.529139957987837084554591421329876744262E-5L,
2.973602667215766676998703687065066180115E-3L,
8.273534546240864308494062287908662592100E-2L,
9.613359842126507198241321110649974032726E-1L,
4.853923697093974370118387947065402707519E0L,
1.002671608961669247462020977417828796933E1L,
7.028927383922483728931327850683151410267E0L,
/* 1.000000000000000000000000000000000000000E0L, */
};
/* Y1(x)cosX - J1(x)sinX = sqrt( 2/(pi x)) Q1(x),
Q1(x) = z(.375 + z^2 R(z^2)), z=1/x
2 <= x <= 2.85711669921875
Peak relative error 6.9e-22 */
static const long double qr2[7] = {
-1.372751603025230017220666013816502528318E-4L,
-6.879190253347766576229143006767218972834E-3L,
-1.061253572090925414598304855316280077828E-1L,
-6.262164224345471241219408329354943337214E-1L,
-1.423149636514768476376254324731437473915E0L,
-1.087955310491078933531734062917489870754E0L,
-1.826821119773182847861406108689273719137E-1L,
};
static const long double qs2[7] = {
1.338768933634451601814048220627185324007E-3L,
7.071099998918497559736318523932241901810E-2L,
1.200511429784048632105295629933382142221E0L,
8.327301713640367079030141077172031825276E0L,
2.468479301872299311658145549931764426840E1L,
2.961179686096262083509383820557051621644E1L,
1.201402313144305153005639494661767354977E1L,
/* 1.000000000000000000000000000000000000000E0L, */
};
static long double
qone (long double x)
{
const long double *p, *q;
long double s, r, z;
int32_t ix;
uint32_t se, i0, i1;
GET_LDOUBLE_WORDS (se, i0, i1, x);
ix = se & 0x7fff;
/* ix >= 0x4000 for all calls to this function. */
if (ix >= 0x4002) /* x >= 8 */
{
p = qr8;
q = qs8;
}
else
{
i1 = (ix << 16) | (i0 >> 16);
if (i1 >= 0x40019174) /* x >= 4.54541015625 */
{
p = qr5;
q = qs5;
}
else if (i1 >= 0x4000b6db) /* x >= 2.85711669921875 */
{
p = qr3;
q = qs3;
}
else /* x >= 2 */
{
p = qr2;
q = qs2;
}
}
z = one / (x * x);
r =
p[0] + z * (p[1] +
z * (p[2] + z * (p[3] + z * (p[4] + z * (p[5] + z * p[6])))));
s =
q[0] + z * (q[1] +
z * (q[2] +
z * (q[3] + z * (q[4] + z * (q[5] + z * (q[6] + z))))));
return (.375 + z * r / s) / x;
}