int fractal_mpfr_calculate_line(image_info* img, int line)
{
int ret = 1;
int ix = 0;
int mx = 0;
int chk_px = ((rthdata*)img->rth_ptr)->check_stop_px;
int img_width = img->real_width;
int* raw_data = &img->raw_data[line * img_width];
depth_t depth = img->depth;
mpfr_t x, y;
mpfr_t x2, y2;
mpfr_t c_re, c_im;
/* working variables: */
mpfr_t wre, wim;
mpfr_t wre2, wim2;
mpfr_t frs_bail;
mpfr_t width, img_rw, img_xmin;
mpfr_t t1;
mpfr_init2(x, img->precision);
mpfr_init2(y, img->precision);
mpfr_init2(x2, img->precision);
mpfr_init2(y2, img->precision);
mpfr_init2(c_re, img->precision);
mpfr_init2(c_im, img->precision);
mpfr_init2(wre, img->precision);
mpfr_init2(wim, img->precision);
mpfr_init2(wre2, img->precision);
mpfr_init2(wim2, img->precision);
mpfr_init2(frs_bail,img->precision);
mpfr_init2(width, img->precision);
mpfr_init2(img_rw, img->precision);
mpfr_init2(img_xmin,img->precision);
mpfr_init2(t1, img->precision);
mpfr_set_si(frs_bail, 4, GMP_RNDN);
mpfr_set_si(img_rw, img_width, GMP_RNDN);
mpfr_set( img_xmin, img->xmin, GMP_RNDN);
mpfr_set( width, img->width, GMP_RNDN);
/* y = img->ymax - ((img->xmax - img->xmin)
/ (long double)img->real_width)
* (long double)img->lines_done; */
mpfr_div( t1, width, img_rw, GMP_RNDN);
mpfr_mul_si( t1, t1, line, GMP_RNDN);
mpfr_sub( y, img->ymax, t1, GMP_RNDN);
mpfr_mul( y2, y, y, GMP_RNDN);
while (ix < img_width)
{
mx += chk_px;
if (mx > img_width)
mx = img_width;
for (; ix < mx; ++ix, ++raw_data)
{
/* x = ((long double)ix / (long double)img->real_width)
* (img->xmax - img->xmin) + img->xmin; */
mpfr_si_div(t1, ix, img_rw, GMP_RNDN);
mpfr_mul(x, t1, width, GMP_RNDN);
mpfr_add(x, x, img_xmin, GMP_RNDN);
mpfr_mul( x2, x, x, GMP_RNDN);
mpfr_set( wre, x, GMP_RNDN);
mpfr_set( wim, y, GMP_RNDN);
mpfr_set( wre2, x2, GMP_RNDN);
mpfr_set( wim2, y2, GMP_RNDN);
switch (img->family)
{
case FAMILY_MANDEL:
mpfr_set(c_re, x, GMP_RNDN);
mpfr_set(c_im, y, GMP_RNDN);
break;
case FAMILY_JULIA:
mpfr_set(c_re, img->u.julia.c_re, GMP_RNDN);
mpfr_set(c_im, img->u.julia.c_im, GMP_RNDN);
break;
}
switch(img->fractal)
{
case BURNING_SHIP:
*raw_data = frac_burning_ship_mpfr(
depth, frs_bail,
wim, wre,
c_im, c_re,
wim2, wre2, t1);
break;
case GENERALIZED_CELTIC:
*raw_data = frac_generalized_celtic_mpfr(
depth, frs_bail,
wim, wre,
c_im, c_re,
wim2, wre2, t1);
break;
case VARIANT:
*raw_data = frac_variant_mpfr(
depth, frs_bail,
wim, wre,
c_im, c_re,
wim2, wre2, t1);
break;
case MANDELBROT:
default:
*raw_data = frac_mandel_mpfr(depth, frs_bail,
wim, wre,
c_im, c_re,
wim2, wre2, t1);
}
}
if (rth_render_should_stop((rthdata*)img->rth_ptr))
{
ret = 0;
break;
}
}
mpfr_clear(x);
mpfr_clear(y);
mpfr_clear(x2);
mpfr_clear(y2);
mpfr_clear(c_re);
mpfr_clear(c_im);
mpfr_clear(wre);
mpfr_clear(wim);
mpfr_clear(wre2);
mpfr_clear(wim2);
mpfr_clear(frs_bail);
mpfr_clear(width);
mpfr_clear(img_rw);
mpfr_clear(t1);
return ret;
}
int
mpfr_sinh (mpfr_ptr y, mpfr_srcptr xt, mp_rnd_t rnd_mode)
{
/****** Declarations ******/
mpfr_t x;
mp_prec_t Nxt = MPFR_PREC(xt);
int flag_neg=0, inexact =0;
if (MPFR_IS_NAN(xt))
{
MPFR_SET_NAN(y);
MPFR_RET_NAN;
}
MPFR_CLEAR_NAN(y);
if (MPFR_IS_INF(xt))
{
MPFR_SET_INF(y);
MPFR_SET_SAME_SIGN(y, xt);
MPFR_RET(0);
}
MPFR_CLEAR_INF(y);
if (MPFR_IS_ZERO(xt))
{
MPFR_SET_ZERO(y); /* sinh(0) = 0 */
MPFR_SET_SAME_SIGN(y, xt);
MPFR_RET(0);
}
mpfr_init2 (x, Nxt);
mpfr_set (x, xt, GMP_RNDN);
if(MPFR_SIGN(x)<0)
{
MPFR_CHANGE_SIGN(x);
flag_neg=1;
}
/* General case */
{
/* Declaration of the intermediary variable */
mpfr_t t, te, ti;
int d;
/* Declaration of the size variable */
mp_prec_t Nx = Nxt; /* Precision of input variable */
mp_prec_t Ny = MPFR_PREC(y); /* Precision of input variable */
mp_prec_t Nt; /* Precision of the intermediary variable */
long int err; /* Precision of error */
/* compute the precision of intermediary variable */
Nt = MAX(Nx, Ny);
/* the optimal number of bits : see algorithms.ps */
Nt = Nt + _mpfr_ceil_log2 (5) + _mpfr_ceil_log2 (Nt);
/* initialise of intermediary variable */
mpfr_init (t);
mpfr_init (te);
mpfr_init (ti);
/* First computation of sinh */
do {
/* reactualisation of the precision */
mpfr_set_prec (t, Nt);
mpfr_set_prec (te, Nt);
mpfr_set_prec (ti, Nt);
/* compute sinh */
mpfr_exp (te, x, GMP_RNDD); /* exp(x) */
mpfr_ui_div (ti, 1, te, GMP_RNDU); /* 1/exp(x) */
mpfr_sub (t, te, ti, GMP_RNDN); /* exp(x) - 1/exp(x) */
mpfr_div_2ui (t, t, 1, GMP_RNDN); /* 1/2(exp(x) - 1/exp(x)) */
/* it may be that t is zero (in fact, it can only occur when te=1,
and thus ti=1 too) */
if (MPFR_IS_ZERO(t))
err = -1;
else
{
/* calculation of the error */
d = MPFR_EXP(te) - MPFR_EXP(t) + 2;
/* estimation of the error */
/* err = Nt-(_mpfr_ceil_log2(1+pow(2,d)));*/
err = Nt - (MAX(d,0) + 1);
}
/* actualisation of the precision */
Nt += 10;
} while ((err < 0) || !mpfr_can_round(t, err, GMP_RNDN, rnd_mode, Ny));
if (flag_neg == 1)
MPFR_CHANGE_SIGN(t);
inexact = mpfr_set (y, t, rnd_mode);
mpfr_clear (t);
mpfr_clear (ti);
mpfr_clear (te);
}
mpfr_clear (x);
return inexact;
}
static void
check_underflow (void)
{
mpfr_t sum1, sum2, t[NUNFL];
mpfr_ptr p[NUNFL];
mpfr_prec_t precmax = 444;
mpfr_exp_t emin, emax;
unsigned int ex_flags, flags;
int c, i;
emin = mpfr_get_emin ();
emax = mpfr_get_emax ();
set_emin (MPFR_EMIN_MIN);
set_emax (MPFR_EMAX_MAX);
ex_flags = MPFR_FLAGS_UNDERFLOW | MPFR_FLAGS_INEXACT;
mpfr_init2 (sum1, MPFR_PREC_MIN);
mpfr_init2 (sum2, precmax);
for (i = 0; i < NUNFL; i++)
{
mpfr_init2 (t[i], precmax);
p[i] = t[i];
}
for (c = 0; c < 8; c++)
{
mpfr_prec_t fprec;
int n, neg, r;
fprec = MPFR_PREC_MIN + (randlimb () % (precmax - MPFR_PREC_MIN + 1));
n = 3 + (randlimb () % (NUNFL - 2));
MPFR_ASSERTN (n <= NUNFL);
mpfr_set_prec (sum2, (randlimb () & 1) ? MPFR_PREC_MIN : precmax);
mpfr_set_prec (t[0], fprec + 64);
mpfr_set_zero (t[0], 1);
for (i = 1; i < n; i++)
{
int inex;
mpfr_set_prec (t[i], MPFR_PREC_MIN +
(randlimb () % (fprec - MPFR_PREC_MIN + 1)));
do
mpfr_urandomb (t[i], RANDS);
while (MPFR_IS_ZERO (t[i]));
mpfr_set_exp (t[i], MPFR_EMIN_MIN);
inex = mpfr_sub (t[0], t[0], t[i], MPFR_RNDN);
MPFR_ASSERTN (inex == 0);
}
neg = randlimb () & 1;
if (neg)
mpfr_nextbelow (t[0]);
else
mpfr_nextabove (t[0]);
RND_LOOP(r)
{
int inex1, inex2;
mpfr_set_zero (sum1, 1);
if (neg)
mpfr_nextbelow (sum1);
else
mpfr_nextabove (sum1);
inex1 = mpfr_div_2ui (sum1, sum1, 2, (mpfr_rnd_t) r);
mpfr_clear_flags ();
inex2 = mpfr_sum (sum2, p, n, (mpfr_rnd_t) r);
flags = __gmpfr_flags;
MPFR_ASSERTN (mpfr_check (sum1));
MPFR_ASSERTN (mpfr_check (sum2));
if (flags != ex_flags)
{
printf ("Bad flags in check_underflow on %s, c = %d\n",
mpfr_print_rnd_mode ((mpfr_rnd_t) r), c);
printf ("Expected flags:");
flags_out (ex_flags);
printf ("Got flags: ");
flags_out (flags);
printf ("sum = ");
mpfr_dump (sum2);
exit (1);
}
if (!(mpfr_equal_p (sum1, sum2) && SAME_SIGN (inex1, inex2)))
{
printf ("Error in check_underflow on %s, c = %d\n",
mpfr_print_rnd_mode ((mpfr_rnd_t) r), c);
printf ("Expected ");
mpfr_dump (sum1);
printf ("with inex = %d\n", inex1);
printf ("Got ");
mpfr_dump (sum2);
printf ("with inex = %d\n", inex2);
exit (1);
}
}
}
for (i = 0; i < NUNFL; i++)
mpfr_clear (t[i]);
mpfr_clears (sum1, sum2, (mpfr_ptr) 0);
set_emin (emin);
set_emax (emax);
}
/* generic code */
__float128
mpfr_get_float128 (mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x)))
return (__float128) mpfr_get_d (x, rnd_mode);
else /* now x is a normal non-zero number */
{
__float128 r; /* result */
__float128 m;
double s; /* part of result */
mpfr_exp_t sh; /* exponent shift, so that x/2^sh is in the double range */
mpfr_t y, z;
int sign;
/* first round x to the target __float128 precision, so that
all subsequent operations are exact (this avoids double rounding
problems) */
mpfr_init2 (y, IEEE_FLOAT128_MANT_DIG);
mpfr_init2 (z, IEEE_FLOAT128_MANT_DIG);
mpfr_set (y, x, rnd_mode);
sh = MPFR_GET_EXP (y);
sign = MPFR_SIGN (y);
MPFR_SET_EXP (y, 0);
MPFR_SET_POS (y);
r = 0.0;
do
{
s = mpfr_get_d (y, MPFR_RNDN); /* high part of y */
r += (__float128) s;
mpfr_set_d (z, s, MPFR_RNDN); /* exact */
mpfr_sub (y, y, z, MPFR_RNDN); /* exact */
}
while (MPFR_NOTZERO (y));
mpfr_clear (z);
mpfr_clear (y);
/* we now have to multiply r by 2^sh */
MPFR_ASSERTD (r > 0);
if (sh != 0)
{
/* An overflow may occur (example: 0.5*2^1024) */
while (r < 1.0)
{
r += r;
sh--;
}
if (sh > 0)
m = 2.0;
else
{
m = 0.5;
sh = -sh;
}
for (;;)
{
if (sh % 2)
r = r * m;
sh >>= 1;
if (sh == 0)
break;
m = m * m;
}
}
if (sign < 0)
r = -r;
return r;
}
}
int
mpfr_yn (mpfr_ptr res, long n, mpfr_srcptr z, mp_rnd_t r)
{
int inex;
unsigned long absn;
MPFR_LOG_FUNC (("x[%#R]=%R n=%d rnd=%d", z, z, n, r),
("y[%#R]=%R", res, res));
absn = SAFE_ABS (unsigned long, n);
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (z)))
{
if (MPFR_IS_NAN (z))
{
MPFR_SET_NAN (res); /* y(n,NaN) = NaN */
MPFR_RET_NAN;
}
/* y(n,z) tends to zero when z goes to +Inf, oscillating around
0. We choose to return +0 in that case. */
else if (MPFR_IS_INF (z))
{
if (MPFR_SIGN(z) > 0)
return mpfr_set_ui (res, 0, r);
else /* y(n,-Inf) = NaN */
{
MPFR_SET_NAN (res);
MPFR_RET_NAN;
}
}
else /* y(n,z) tends to -Inf for n >= 0 or n even, to +Inf otherwise,
when z goes to zero */
{
MPFR_SET_INF(res);
if (n >= 0 || (n & 1) == 0)
MPFR_SET_NEG(res);
else
MPFR_SET_POS(res);
MPFR_RET(0);
}
}
/* for z < 0, y(n,z) is imaginary except when j(n,|z|) = 0, which we
assume does not happen for a rational z. */
if (MPFR_SIGN(z) < 0)
{
MPFR_SET_NAN (res);
MPFR_RET_NAN;
}
/* now z is not singular, and z > 0 */
/* Deal with tiny arguments. We have:
y0(z) = 2 log(z)/Pi + 2 (euler - log(2))/Pi + O(log(z)*z^2), more
precisely for 0 <= z <= 1/2, with g(z) = 2/Pi + 2(euler-log(2))/Pi/log(z),
g(z) - 0.41*z^2 < y0(z)/log(z) < g(z)
thus since log(z) is negative:
g(z)*log(z) < y0(z) < (g(z) - z^2/2)*log(z)
and since |g(z)| >= 0.63 for 0 <= z <= 1/2, the relative error on
y0(z)/log(z) is bounded by 0.41*z^2/0.63 <= 0.66*z^2.
Note: we use both the main term in log(z) and the constant term, because
otherwise the relative error would be only in 1/log(|log(z)|).
*/
if (n == 0 && MPFR_EXP(z) < - (mp_exp_t) (MPFR_PREC(res) / 2))
{
mpfr_t l, h, t, logz;
mp_prec_t prec;
int ok, inex2;
prec = MPFR_PREC(res) + 10;
mpfr_init2 (l, prec);
mpfr_init2 (h, prec);
mpfr_init2 (t, prec);
mpfr_init2 (logz, prec);
/* first enclose log(z) + euler - log(2) = log(z/2) + euler */
mpfr_log (logz, z, GMP_RNDD); /* lower bound of log(z) */
mpfr_set (h, logz, GMP_RNDU); /* exact */
mpfr_nextabove (h); /* upper bound of log(z) */
mpfr_const_euler (t, GMP_RNDD); /* lower bound of euler */
mpfr_add (l, logz, t, GMP_RNDD); /* lower bound of log(z) + euler */
mpfr_nextabove (t); /* upper bound of euler */
mpfr_add (h, h, t, GMP_RNDU); /* upper bound of log(z) + euler */
mpfr_const_log2 (t, GMP_RNDU); /* upper bound of log(2) */
mpfr_sub (l, l, t, GMP_RNDD); /* lower bound of log(z/2) + euler */
mpfr_nextbelow (t); /* lower bound of log(2) */
mpfr_sub (h, h, t, GMP_RNDU); /* upper bound of log(z/2) + euler */
mpfr_const_pi (t, GMP_RNDU); /* upper bound of Pi */
mpfr_div (l, l, t, GMP_RNDD); /* lower bound of (log(z/2)+euler)/Pi */
mpfr_nextbelow (t); /* lower bound of Pi */
mpfr_div (h, h, t, GMP_RNDD); /* upper bound of (log(z/2)+euler)/Pi */
mpfr_mul_2ui (l, l, 1, GMP_RNDD); /* lower bound on g(z)*log(z) */
mpfr_mul_2ui (h, h, 1, GMP_RNDU); /* upper bound on g(z)*log(z) */
/* we now have l <= g(z)*log(z) <= h, and we need to add -z^2/2*log(z)
to h */
mpfr_mul (t, z, z, GMP_RNDU); /* upper bound on z^2 */
/* since logz is negative, a lower bound corresponds to an upper bound
for its absolute value */
mpfr_neg (t, t, GMP_RNDD);
mpfr_div_2ui (t, t, 1, GMP_RNDD);
mpfr_mul (t, t, logz, GMP_RNDU); /* upper bound on z^2/2*log(z) */
/* an underflow may happen in the above instructions, clear flag */
mpfr_clear_underflow ();
mpfr_add (h, h, t, GMP_RNDU);
inex = mpfr_prec_round (l, MPFR_PREC(res), r);
inex2 = mpfr_prec_round (h, MPFR_PREC(res), r);
/* we need h=l and inex=inex2 */
ok = (inex == inex2) && (mpfr_cmp (l, h) == 0);
if (ok)
mpfr_set (res, h, r); /* exact */
mpfr_clear (l);
mpfr_clear (h);
mpfr_clear (t);
mpfr_clear (logz);
if (ok)
return inex;
}
/* small argument check for y1(z) = -2/Pi/z + O(log(z)):
for 0 <= z <= 1, |y1(z) + 2/Pi/z| <= 0.25 */
if (n == 1 && MPFR_EXP(z) + 1 < - (mp_exp_t) MPFR_PREC(res))
{
mpfr_t y;
mp_prec_t prec;
mp_exp_t err1;
int ok;
MPFR_BLOCK_DECL (flags);
/* since 2/Pi > 0.5, and |y1(z)| >= |2/Pi/z|, if z <= 2^(-emax-1),
then |y1(z)| > 2^emax */
prec = MPFR_PREC(res) + 10;
mpfr_init2 (y, prec);
mpfr_const_pi (y, GMP_RNDU); /* Pi*(1+u)^2, where here and below u
represents a quantity <= 1/2^prec */
mpfr_mul (y, y, z, GMP_RNDU); /* Pi*z * (1+u)^4, upper bound */
MPFR_BLOCK (flags, mpfr_ui_div (y, 2, y, GMP_RNDZ));
/* 2/Pi/z * (1+u)^6, lower bound, with possible overflow */
if (MPFR_OVERFLOW (flags))
{
mpfr_clear (y);
return mpfr_overflow (res, r, -1);
}
mpfr_neg (y, y, GMP_RNDN);
/* (1+u)^6 can be written 1+7u [for another value of u], thus the
error on 2/Pi/z is less than 7ulp(y). The truncation error is less
than 1/4, thus if ulp(y)>=1/4, the total error is less than 8ulp(y),
otherwise it is less than 1/4+7/8 <= 2. */
if (MPFR_EXP(y) + 2 >= MPFR_PREC(y)) /* ulp(y) >= 1/4 */
err1 = 3;
else /* ulp(y) <= 1/8 */
err1 = (mp_exp_t) MPFR_PREC(y) - MPFR_EXP(y) + 1;
ok = MPFR_CAN_ROUND (y, prec - err1, MPFR_PREC(res), r);
if (ok)
inex = mpfr_set (res, y, r);
mpfr_clear (y);
if (ok)
return inex;
}
/* we can use the asymptotic expansion as soon as z > p log(2)/2,
but to get some margin we use it for z > p/2 */
if (mpfr_cmp_ui (z, MPFR_PREC(res) / 2 + 3) > 0)
{
inex = mpfr_yn_asympt (res, n, z, r);
if (inex != 0)
return inex;
}
/* General case */
{
mp_prec_t prec;
mp_exp_t err1, err2, err3;
mpfr_t y, s1, s2, s3;
MPFR_ZIV_DECL (loop);
mpfr_init (y);
mpfr_init (s1);
mpfr_init (s2);
mpfr_init (s3);
prec = MPFR_PREC(res) + 2 * MPFR_INT_CEIL_LOG2 (MPFR_PREC (res)) + 13;
MPFR_ZIV_INIT (loop, prec);
for (;;)
{
mpfr_set_prec (y, prec);
mpfr_set_prec (s1, prec);
mpfr_set_prec (s2, prec);
mpfr_set_prec (s3, prec);
mpfr_mul (y, z, z, GMP_RNDN);
mpfr_div_2ui (y, y, 2, GMP_RNDN); /* z^2/4 */
/* store (z/2)^n temporarily in s2 */
mpfr_pow_ui (s2, z, absn, GMP_RNDN);
mpfr_div_2si (s2, s2, absn, GMP_RNDN);
/* compute S1 * (z/2)^(-n) */
if (n == 0)
{
mpfr_set_ui (s1, 0, GMP_RNDN);
err1 = 0;
}
else
err1 = mpfr_yn_s1 (s1, y, absn - 1);
mpfr_div (s1, s1, s2, GMP_RNDN); /* (z/2)^(-n) * S1 */
/* See algorithms.tex: the relative error on s1 is bounded by
(3n+3)*2^(e+1-prec). */
err1 = MPFR_INT_CEIL_LOG2 (3 * absn + 3) + err1 + 1;
/* rel_err(s1) <= 2^(err1-prec), thus err(s1) <= 2^err1 ulps */
/* compute (z/2)^n * S3 */
mpfr_neg (y, y, GMP_RNDN); /* -z^2/4 */
err3 = mpfr_yn_s3 (s3, y, s2, absn); /* (z/2)^n * S3 */
/* the error on s3 is bounded by 2^err3 ulps */
/* add s1+s3 */
err1 += MPFR_EXP(s1);
mpfr_add (s1, s1, s3, GMP_RNDN);
/* the error is bounded by 1/2 + 2^err1*2^(- EXP(s1))
+ 2^err3*2^(EXP(s3) - EXP(s1)) */
err3 += MPFR_EXP(s3);
err1 = (err3 > err1) ? err3 + 1 : err1 + 1;
err1 -= MPFR_EXP(s1);
err1 = (err1 >= 0) ? err1 + 1 : 1;
/* now the error on s1 is bounded by 2^err1*ulp(s1) */
/* compute S2 */
mpfr_div_2ui (s2, z, 1, GMP_RNDN); /* z/2 */
mpfr_log (s2, s2, GMP_RNDN); /* log(z/2) */
mpfr_const_euler (s3, GMP_RNDN);
err2 = MPFR_EXP(s2) > MPFR_EXP(s3) ? MPFR_EXP(s2) : MPFR_EXP(s3);
mpfr_add (s2, s2, s3, GMP_RNDN); /* log(z/2) + gamma */
err2 -= MPFR_EXP(s2);
mpfr_mul_2ui (s2, s2, 1, GMP_RNDN); /* 2*(log(z/2) + gamma) */
mpfr_jn (s3, absn, z, GMP_RNDN); /* Jn(z) */
mpfr_mul (s2, s2, s3, GMP_RNDN); /* 2*(log(z/2) + gamma)*Jn(z) */
err2 += 4; /* the error on s2 is bounded by 2^err2 ulps, see
algorithms.tex */
/* add all three sums */
err1 += MPFR_EXP(s1); /* the error on s1 is bounded by 2^err1 */
err2 += MPFR_EXP(s2); /* the error on s2 is bounded by 2^err2 */
mpfr_sub (s2, s2, s1, GMP_RNDN); /* s2 - (s1+s3) */
err2 = (err1 > err2) ? err1 + 1 : err2 + 1;
err2 -= MPFR_EXP(s2);
err2 = (err2 >= 0) ? err2 + 1 : 1;
/* now the error on s2 is bounded by 2^err2*ulp(s2) */
mpfr_const_pi (y, GMP_RNDN); /* error bounded by 1 ulp */
mpfr_div (s2, s2, y, GMP_RNDN); /* error bounded by
2^(err2+1)*ulp(s2) */
err2 ++;
if (MPFR_LIKELY (MPFR_CAN_ROUND (s2, prec - err2, MPFR_PREC(res), r)))
break;
MPFR_ZIV_NEXT (loop, prec);
}
MPFR_ZIV_FREE (loop);
inex = (n >= 0 || (n & 1) == 0)
? mpfr_set (res, s2, r)
: mpfr_neg (res, s2, r);
mpfr_clear (y);
mpfr_clear (s1);
mpfr_clear (s2);
mpfr_clear (s3);
}
return inex;
}
int
mpfr_sinh_cosh (mpfr_ptr sh, mpfr_ptr ch, mpfr_srcptr xt, mpfr_rnd_t rnd_mode)
{
mpfr_t x;
int inexact_sh, inexact_ch;
MPFR_ASSERTN (sh != ch);
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg rnd=%d",
mpfr_get_prec (xt), mpfr_log_prec, xt, rnd_mode),
("sh[%Pu]=%.*Rg ch[%Pu]=%.*Rg",
mpfr_get_prec (sh), mpfr_log_prec, sh,
mpfr_get_prec (ch), mpfr_log_prec, ch));
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (xt)))
{
if (MPFR_IS_NAN (xt))
{
MPFR_SET_NAN (ch);
MPFR_SET_NAN (sh);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (xt))
{
MPFR_SET_INF (sh);
MPFR_SET_SAME_SIGN (sh, xt);
MPFR_SET_INF (ch);
MPFR_SET_POS (ch);
MPFR_RET (0);
}
else /* xt is zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO (xt));
MPFR_SET_ZERO (sh); /* sinh(0) = 0 */
MPFR_SET_SAME_SIGN (sh, xt);
inexact_sh = 0;
inexact_ch = mpfr_set_ui (ch, 1, rnd_mode); /* cosh(0) = 1 */
return INEX(inexact_sh,inexact_ch);
}
}
/* Warning: if we use MPFR_FAST_COMPUTE_IF_SMALL_INPUT here, make sure
that the code also works in case of overlap (see sin_cos.c) */
MPFR_TMP_INIT_ABS (x, xt);
{
mpfr_t s, c, ti;
mpfr_exp_t d;
mpfr_prec_t N; /* Precision of the intermediary variables */
long int err; /* Precision of error */
MPFR_ZIV_DECL (loop);
MPFR_SAVE_EXPO_DECL (expo);
MPFR_GROUP_DECL (group);
MPFR_SAVE_EXPO_MARK (expo);
/* compute the precision of intermediary variable */
N = MPFR_PREC (ch);
N = MAX (N, MPFR_PREC (sh));
/* the optimal number of bits : see algorithms.ps */
N = N + MPFR_INT_CEIL_LOG2 (N) + 4;
/* initialise of intermediary variables */
MPFR_GROUP_INIT_3 (group, N, s, c, ti);
/* First computation of sinh_cosh */
MPFR_ZIV_INIT (loop, N);
for (;;)
{
MPFR_BLOCK_DECL (flags);
/* compute sinh_cosh */
MPFR_BLOCK (flags, mpfr_exp (s, x, MPFR_RNDD));
if (MPFR_OVERFLOW (flags))
/* exp(x) does overflow */
{
/* since cosh(x) >= exp(x), cosh(x) overflows too */
inexact_ch = mpfr_overflow (ch, rnd_mode, MPFR_SIGN_POS);
/* sinh(x) may be representable */
inexact_sh = mpfr_sinh (sh, xt, rnd_mode);
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_OVERFLOW);
break;
}
d = MPFR_GET_EXP (s);
mpfr_ui_div (ti, 1, s, MPFR_RNDU); /* 1/exp(x) */
mpfr_add (c, s, ti, MPFR_RNDU); /* exp(x) + 1/exp(x) */
mpfr_sub (s, s, ti, MPFR_RNDN); /* exp(x) - 1/exp(x) */
mpfr_div_2ui (c, c, 1, MPFR_RNDN); /* 1/2(exp(x) + 1/exp(x)) */
mpfr_div_2ui (s, s, 1, MPFR_RNDN); /* 1/2(exp(x) - 1/exp(x)) */
/* it may be that s is zero (in fact, it can only occur when exp(x)=1,
and thus ti=1 too) */
if (MPFR_IS_ZERO (s))
err = N; /* double the precision */
else
{
/* calculation of the error */
d = d - MPFR_GET_EXP (s) + 2;
/* error estimate: err = N-(__gmpfr_ceil_log2(1+pow(2,d)));*/
err = N - (MAX (d, 0) + 1);
if (MPFR_LIKELY (MPFR_CAN_ROUND (s, err, MPFR_PREC (sh),
rnd_mode) && \
MPFR_CAN_ROUND (c, err, MPFR_PREC (ch),
rnd_mode)))
{
inexact_sh = mpfr_set4 (sh, s, rnd_mode, MPFR_SIGN (xt));
inexact_ch = mpfr_set (ch, c, rnd_mode);
break;
}
}
/* actualisation of the precision */
N += err;
MPFR_ZIV_NEXT (loop, N);
MPFR_GROUP_REPREC_3 (group, N, s, c, ti);
}
MPFR_ZIV_FREE (loop);
MPFR_GROUP_CLEAR (group);
MPFR_SAVE_EXPO_FREE (expo);
}
/* now, let's raise the flags if needed */
inexact_sh = mpfr_check_range (sh, inexact_sh, rnd_mode);
inexact_ch = mpfr_check_range (ch, inexact_ch, rnd_mode);
return INEX(inexact_sh,inexact_ch);
}
int
mpc_acos (mpc_ptr rop, mpc_srcptr op, mpc_rnd_t rnd)
{
int inex_re, inex_im, inex;
mpfr_prec_t p_re, p_im, p;
mpc_t z1;
mpfr_t pi_over_2;
mpfr_exp_t e1, e2;
mpfr_rnd_t rnd_im;
mpc_rnd_t rnd1;
inex_re = 0;
inex_im = 0;
/* special values */
if (mpfr_nan_p (mpc_realref (op)) || mpfr_nan_p (mpc_imagref (op)))
{
if (mpfr_inf_p (mpc_realref (op)) || mpfr_inf_p (mpc_imagref (op)))
{
mpfr_set_inf (mpc_imagref (rop), mpfr_signbit (mpc_imagref (op)) ? +1 : -1);
mpfr_set_nan (mpc_realref (rop));
}
else if (mpfr_zero_p (mpc_realref (op)))
{
inex_re = set_pi_over_2 (mpc_realref (rop), +1, MPC_RND_RE (rnd));
mpfr_set_nan (mpc_imagref (rop));
}
else
{
mpfr_set_nan (mpc_realref (rop));
mpfr_set_nan (mpc_imagref (rop));
}
return MPC_INEX (inex_re, 0);
}
if (mpfr_inf_p (mpc_realref (op)) || mpfr_inf_p (mpc_imagref (op)))
{
if (mpfr_inf_p (mpc_realref (op)))
{
if (mpfr_inf_p (mpc_imagref (op)))
{
if (mpfr_sgn (mpc_realref (op)) > 0)
{
inex_re =
set_pi_over_2 (mpc_realref (rop), +1, MPC_RND_RE (rnd));
mpfr_div_2ui (mpc_realref (rop), mpc_realref (rop), 1, GMP_RNDN);
}
else
{
/* the real part of the result is 3*pi/4
a = o(pi) error(a) < 1 ulp(a)
b = o(3*a) error(b) < 2 ulp(b)
c = b/4 exact
thus 1 bit is lost */
mpfr_t x;
mpfr_prec_t prec;
int ok;
mpfr_init (x);
prec = mpfr_get_prec (mpc_realref (rop));
p = prec;
do
{
p += mpc_ceil_log2 (p);
mpfr_set_prec (x, p);
mpfr_const_pi (x, GMP_RNDD);
mpfr_mul_ui (x, x, 3, GMP_RNDD);
ok =
mpfr_can_round (x, p - 1, GMP_RNDD, MPC_RND_RE (rnd),
prec+(MPC_RND_RE (rnd) == GMP_RNDN));
} while (ok == 0);
inex_re =
mpfr_div_2ui (mpc_realref (rop), x, 2, MPC_RND_RE (rnd));
mpfr_clear (x);
}
}
else
{
if (mpfr_sgn (mpc_realref (op)) > 0)
mpfr_set_ui (mpc_realref (rop), 0, GMP_RNDN);
else
inex_re = mpfr_const_pi (mpc_realref (rop), MPC_RND_RE (rnd));
}
}
else
inex_re = set_pi_over_2 (mpc_realref (rop), +1, MPC_RND_RE (rnd));
mpfr_set_inf (mpc_imagref (rop), mpfr_signbit (mpc_imagref (op)) ? +1 : -1);
return MPC_INEX (inex_re, 0);
}
/* pure real argument */
if (mpfr_zero_p (mpc_imagref (op)))
{
int s_im;
s_im = mpfr_signbit (mpc_imagref (op));
if (mpfr_cmp_ui (mpc_realref (op), 1) > 0)
{
if (s_im)
inex_im = mpfr_acosh (mpc_imagref (rop), mpc_realref (op),
MPC_RND_IM (rnd));
else
inex_im = -mpfr_acosh (mpc_imagref (rop), mpc_realref (op),
INV_RND (MPC_RND_IM (rnd)));
mpfr_set_ui (mpc_realref (rop), 0, GMP_RNDN);
}
else if (mpfr_cmp_si (mpc_realref (op), -1) < 0)
{
mpfr_t minus_op_re;
minus_op_re[0] = mpc_realref (op)[0];
MPFR_CHANGE_SIGN (minus_op_re);
if (s_im)
inex_im = mpfr_acosh (mpc_imagref (rop), minus_op_re,
MPC_RND_IM (rnd));
else
inex_im = -mpfr_acosh (mpc_imagref (rop), minus_op_re,
INV_RND (MPC_RND_IM (rnd)));
inex_re = mpfr_const_pi (mpc_realref (rop), MPC_RND_RE (rnd));
}
else
{
inex_re = mpfr_acos (mpc_realref (rop), mpc_realref (op), MPC_RND_RE (rnd));
mpfr_set_ui (mpc_imagref (rop), 0, MPC_RND_IM (rnd));
}
if (!s_im)
mpc_conj (rop, rop, MPC_RNDNN);
return MPC_INEX (inex_re, inex_im);
}
/* pure imaginary argument */
if (mpfr_zero_p (mpc_realref (op)))
{
inex_re = set_pi_over_2 (mpc_realref (rop), +1, MPC_RND_RE (rnd));
inex_im = -mpfr_asinh (mpc_imagref (rop), mpc_imagref (op),
INV_RND (MPC_RND_IM (rnd)));
mpc_conj (rop,rop, MPC_RNDNN);
return MPC_INEX (inex_re, inex_im);
}
/* regular complex argument: acos(z) = Pi/2 - asin(z) */
p_re = mpfr_get_prec (mpc_realref(rop));
p_im = mpfr_get_prec (mpc_imagref(rop));
p = p_re;
mpc_init3 (z1, p, p_im); /* we round directly the imaginary part to p_im,
with rounding mode opposite to rnd_im */
rnd_im = MPC_RND_IM(rnd);
/* the imaginary part of asin(z) has the same sign as Im(z), thus if
Im(z) > 0 and rnd_im = RNDZ, we want to round the Im(asin(z)) to -Inf
so that -Im(asin(z)) is rounded to zero */
if (rnd_im == GMP_RNDZ)
rnd_im = mpfr_sgn (mpc_imagref(op)) > 0 ? GMP_RNDD : GMP_RNDU;
else
rnd_im = rnd_im == GMP_RNDU ? GMP_RNDD
: rnd_im == GMP_RNDD ? GMP_RNDU
: rnd_im; /* both RNDZ and RNDA map to themselves for -asin(z) */
rnd1 = MPC_RND (GMP_RNDN, rnd_im);
mpfr_init2 (pi_over_2, p);
for (;;)
{
p += mpc_ceil_log2 (p) + 3;
mpfr_set_prec (mpc_realref(z1), p);
mpfr_set_prec (pi_over_2, p);
set_pi_over_2 (pi_over_2, +1, GMP_RNDN);
e1 = 1; /* Exp(pi_over_2) */
inex = mpc_asin (z1, op, rnd1); /* asin(z) */
MPC_ASSERT (mpfr_sgn (mpc_imagref(z1)) * mpfr_sgn (mpc_imagref(op)) > 0);
inex_im = MPC_INEX_IM(inex); /* inex_im is in {-1, 0, 1} */
e2 = mpfr_get_exp (mpc_realref(z1));
mpfr_sub (mpc_realref(z1), pi_over_2, mpc_realref(z1), GMP_RNDN);
if (!mpfr_zero_p (mpc_realref(z1)))
{
/* the error on x=Re(z1) is bounded by 1/2 ulp(x) + 2^(e1-p-1) +
2^(e2-p-1) */
e1 = e1 >= e2 ? e1 + 1 : e2 + 1;
/* the error on x is bounded by 1/2 ulp(x) + 2^(e1-p-1) */
e1 -= mpfr_get_exp (mpc_realref(z1));
/* the error on x is bounded by 1/2 ulp(x) [1 + 2^e1] */
e1 = e1 <= 0 ? 0 : e1;
/* the error on x is bounded by 2^e1 * ulp(x) */
mpfr_neg (mpc_imagref(z1), mpc_imagref(z1), GMP_RNDN); /* exact */
inex_im = -inex_im;
if (mpfr_can_round (mpc_realref(z1), p - e1, GMP_RNDN, GMP_RNDZ,
p_re + (MPC_RND_RE(rnd) == GMP_RNDN)))
break;
}
}
inex = mpc_set (rop, z1, rnd);
inex_re = MPC_INEX_RE(inex);
mpc_clear (z1);
mpfr_clear (pi_over_2);
return MPC_INEX(inex_re, inex_im);
}
/* we have x >= 1/2 here */
static int
mpfr_digamma_positive (mpfr_ptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_prec_t p = MPFR_PREC(y) + 10, q;
mpfr_t t, u, x_plus_j;
int inex;
mpfr_exp_t errt, erru, expt;
unsigned long j = 0, min;
MPFR_ZIV_DECL (loop);
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg rnd=%d", mpfr_get_prec(x), mpfr_log_prec, x, rnd_mode),
("y[%Pu]=%.*Rg inexact=%d", mpfr_get_prec(y), mpfr_log_prec, y, inex));
/* compute a precision q such that x+1 is exact */
if (MPFR_PREC(x) < MPFR_EXP(x))
q = MPFR_EXP(x);
else
q = MPFR_PREC(x) + 1;
/* for very large x, use |digamma(x) - log(x)| < 1/x < 2^(1-EXP(x)) */
if (MPFR_PREC(y) + 10 < MPFR_EXP(x))
{
/* this ensures EXP(x) >= 3, thus x >= 4, thus log(x) > 1 */
mpfr_init2 (t, MPFR_PREC(y) + 10);
mpfr_log (t, x, MPFR_RNDZ);
if (MPFR_CAN_ROUND (t, MPFR_PREC(y) + 10, MPFR_PREC(y), rnd_mode))
{
inex = mpfr_set (y, t, rnd_mode);
mpfr_clear (t);
return inex;
}
mpfr_clear (t);
}
mpfr_init2 (x_plus_j, q);
mpfr_init2 (t, p);
mpfr_init2 (u, p);
MPFR_ZIV_INIT (loop, p);
for(;;)
{
/* Lower bound for x+j in mpfr_digamma_approx call: since the smallest
term of the divergent series for Digamma(x) is about exp(-2*Pi*x), and
we want it to be less than 2^(-p), this gives x > p*log(2)/(2*Pi)
i.e., x >= 0.1103 p.
To be safe, we ensure x >= 0.25 * p.
*/
min = (p + 3) / 4;
if (min < 2)
min = 2;
mpfr_set (x_plus_j, x, MPFR_RNDN);
mpfr_set_ui (u, 0, MPFR_RNDN);
j = 0;
while (mpfr_cmp_ui (x_plus_j, min) < 0)
{
j ++;
mpfr_ui_div (t, 1, x_plus_j, MPFR_RNDN); /* err <= 1/2 ulp */
mpfr_add (u, u, t, MPFR_RNDN);
inex = mpfr_add_ui (x_plus_j, x_plus_j, 1, MPFR_RNDZ);
if (inex != 0) /* we lost one bit */
{
q ++;
mpfr_prec_round (x_plus_j, q, MPFR_RNDZ);
mpfr_nextabove (x_plus_j);
}
/* since all terms are positive, the error is bounded by j ulps */
}
for (erru = 0; j > 1; erru++, j = (j + 1) / 2);
errt = mpfr_digamma_approx (t, x_plus_j);
expt = MPFR_EXP(t);
mpfr_sub (t, t, u, MPFR_RNDN);
if (MPFR_EXP(t) < expt)
errt += expt - MPFR_EXP(t);
if (MPFR_EXP(t) < MPFR_EXP(u))
erru += MPFR_EXP(u) - MPFR_EXP(t);
if (errt > erru)
errt = errt + 1;
else if (errt == erru)
errt = errt + 2;
else
errt = erru + 1;
if (MPFR_CAN_ROUND (t, p - errt, MPFR_PREC(y), rnd_mode))
break;
MPFR_ZIV_NEXT (loop, p);
mpfr_set_prec (t, p);
mpfr_set_prec (u, p);
}
MPFR_ZIV_FREE (loop);
inex = mpfr_set (y, t, rnd_mode);
mpfr_clear (t);
mpfr_clear (u);
mpfr_clear (x_plus_j);
return inex;
}
/* Put in s an approximation of digamma(x).
Assumes x >= 2.
Assumes s does not overlap with x.
Returns an integer e such that the error is bounded by 2^e ulps
of the result s.
*/
static mpfr_exp_t
mpfr_digamma_approx (mpfr_ptr s, mpfr_srcptr x)
{
mpfr_prec_t p = MPFR_PREC (s);
mpfr_t t, u, invxx;
mpfr_exp_t e, exps, f, expu;
unsigned long n;
MPFR_ASSERTN(MPFR_IS_POS(x) && (MPFR_EXP(x) >= 2));
mpfr_init2 (t, p);
mpfr_init2 (u, p);
mpfr_init2 (invxx, p);
mpfr_log (s, x, MPFR_RNDN); /* error <= 1/2 ulp */
mpfr_ui_div (t, 1, x, MPFR_RNDN); /* error <= 1/2 ulp */
mpfr_div_2exp (t, t, 1, MPFR_RNDN); /* exact */
mpfr_sub (s, s, t, MPFR_RNDN);
/* error <= 1/2 + 1/2*2^(EXP(olds)-EXP(s)) + 1/2*2^(EXP(t)-EXP(s)).
For x >= 2, log(x) >= 2*(1/(2x)), thus olds >= 2t, and olds - t >= olds/2,
thus 0 <= EXP(olds)-EXP(s) <= 1, and EXP(t)-EXP(s) <= 0, thus
error <= 1/2 + 1/2*2 + 1/2 <= 2 ulps. */
e = 2; /* initial error */
mpfr_mul (invxx, x, x, MPFR_RNDZ); /* invxx = x^2 * (1 + theta)
for |theta| <= 2^(-p) */
mpfr_ui_div (invxx, 1, invxx, MPFR_RNDU); /* invxx = 1/x^2 * (1 + theta)^2 */
/* in the following we note err=xxx when the ratio between the approximation
and the exact result can be written (1 + theta)^xxx for |theta| <= 2^(-p),
following Higham's method */
mpfr_set_ui (t, 1, MPFR_RNDN); /* err = 0 */
for (n = 1;; n++)
{
/* The main term is Bernoulli[2n]/(2n)/x^(2n) = B[n]/(2n+1)!(2n)/x^(2n)
= B[n]*t[n]/(2n) where t[n]/t[n-1] = 1/(2n)/(2n+1)/x^2. */
mpfr_mul (t, t, invxx, MPFR_RNDU); /* err = err + 3 */
mpfr_div_ui (t, t, 2 * n, MPFR_RNDU); /* err = err + 1 */
mpfr_div_ui (t, t, 2 * n + 1, MPFR_RNDU); /* err = err + 1 */
/* we thus have err = 5n here */
mpfr_div_ui (u, t, 2 * n, MPFR_RNDU); /* err = 5n+1 */
mpfr_mul_z (u, u, mpfr_bernoulli_cache(n), MPFR_RNDU);/* err = 5n+2, and the
absolute error is bounded
by 10n+4 ulp(u) [Rule 11] */
/* if the terms 'u' are decreasing by a factor two at least,
then the error coming from those is bounded by
sum((10n+4)/2^n, n=1..infinity) = 24 */
exps = mpfr_get_exp (s);
expu = mpfr_get_exp (u);
if (expu < exps - (mpfr_exp_t) p)
break;
mpfr_sub (s, s, u, MPFR_RNDN); /* error <= 24 + n/2 */
if (mpfr_get_exp (s) < exps)
e <<= exps - mpfr_get_exp (s);
e ++; /* error in mpfr_sub */
f = 10 * n + 4;
while (expu < exps)
{
f = (1 + f) / 2;
expu ++;
}
e += f; /* total rounding error coming from 'u' term */
}
mpfr_clear (t);
mpfr_clear (u);
mpfr_clear (invxx);
f = 0;
while (e > 1)
{
f++;
e = (e + 1) / 2;
/* Invariant: 2^f * e does not decrease */
}
return f;
}
int
main (int argc, char *argv[])
{
mpfr_t x, y;
mp_exp_t emin, emax;
tests_start_mpfr ();
test_set_underflow ();
test_set_overflow ();
check_default_rnd();
mpfr_init (x);
mpfr_init (y);
emin = mpfr_get_emin ();
emax = mpfr_get_emax ();
if (emin >= emax)
{
printf ("Error: emin >= emax\n");
exit (1);
}
mpfr_set_ui (x, 1, GMP_RNDN);
mpfr_mul_2exp (x, x, 1024, GMP_RNDN);
mpfr_set_double_range ();
mpfr_check_range (x, 0, GMP_RNDN);
if (!mpfr_inf_p (x) || (mpfr_sgn(x) <= 0))
{
printf ("Error: 2^1024 rounded to nearest should give +Inf\n");
exit (1);
}
set_emax (1025);
mpfr_set_ui (x, 1, GMP_RNDN);
mpfr_mul_2exp (x, x, 1024, GMP_RNDN);
mpfr_set_double_range ();
mpfr_check_range (x, 0, GMP_RNDD);
if (!mpfr_number_p (x))
{
printf ("Error: 2^1024 rounded down should give a normal number\n");
exit (1);
}
mpfr_set_ui (x, 1, GMP_RNDN);
mpfr_mul_2exp (x, x, 1023, GMP_RNDN);
mpfr_add (x, x, x, GMP_RNDN);
if (!mpfr_inf_p (x) || (mpfr_sgn(x) <= 0))
{
printf ("Error: x+x rounded to nearest for x=2^1023 should give +Inf\n");
printf ("emax = %ld\n", mpfr_get_emax ());
printf ("got "); mpfr_print_binary (x); puts ("");
exit (1);
}
mpfr_set_ui (x, 1, GMP_RNDN);
mpfr_mul_2exp (x, x, 1023, GMP_RNDN);
mpfr_add (x, x, x, GMP_RNDD);
if (!mpfr_number_p (x))
{
printf ("Error: x+x rounded down for x=2^1023 should give"
" a normal number\n");
exit (1);
}
mpfr_set_ui (x, 1, GMP_RNDN);
mpfr_div_2exp (x, x, 1022, GMP_RNDN);
mpfr_set_str_binary (y, "1.1e-1022"); /* y = 3/2*x */
mpfr_sub (y, y, x, GMP_RNDZ);
if (mpfr_cmp_ui (y, 0))
{
printf ("Error: y-x rounded to zero should give 0"
" for y=3/2*2^(-1022), x=2^(-1022)\n");
printf ("y="); mpfr_print_binary (y); puts ("");
exit (1);
}
set_emin (-1026);
mpfr_set_ui (x, 1, GMP_RNDN);
mpfr_div_2exp (x, x, 1025, GMP_RNDN);
mpfr_set_double_range ();
mpfr_check_range (x, 0, GMP_RNDN);
if (!MPFR_IS_ZERO (x) )
{
printf ("Error: x rounded to nearest for x=2^-1024 should give Zero\n");
printf ("emin = %ld\n", mpfr_get_emin ());
printf ("got "); mpfr_dump (x);
exit (1);
}
mpfr_clear (x);
mpfr_clear (y);
check_emin_emax();
check_flags();
tests_end_mpfr ();
return 0;
}
/* Use the reflection formula Digamma(1-x) = Digamma(x) + Pi * cot(Pi*x),
i.e., Digamma(x) = Digamma(1-x) - Pi * cot(Pi*x).
Assume x < 1/2. */
static int
mpfr_digamma_reflection (mpfr_ptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_prec_t p = MPFR_PREC(y) + 10, q;
mpfr_t t, u, v;
mpfr_exp_t e1, expv;
int inex;
MPFR_ZIV_DECL (loop);
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg rnd=%d", mpfr_get_prec(x), mpfr_log_prec, x, rnd_mode),
("y[%Pu]=%.*Rg inexact=%d", mpfr_get_prec(y), mpfr_log_prec, y, inex));
/* we want that 1-x is exact with precision q: if 0 < x < 1/2, then
q = PREC(x)-EXP(x) is ok, otherwise if -1 <= x < 0, q = PREC(x)-EXP(x)
is ok, otherwise for x < -1, PREC(x) is ok if EXP(x) <= PREC(x),
otherwise we need EXP(x) */
if (MPFR_EXP(x) < 0)
q = MPFR_PREC(x) + 1 - MPFR_EXP(x);
else if (MPFR_EXP(x) <= MPFR_PREC(x))
q = MPFR_PREC(x) + 1;
else
q = MPFR_EXP(x);
mpfr_init2 (u, q);
MPFR_DBGRES(inex = mpfr_ui_sub (u, 1, x, MPFR_RNDN));
MPFR_ASSERTN(inex == 0);
/* if x is half an integer, cot(Pi*x) = 0, thus Digamma(x) = Digamma(1-x) */
mpfr_mul_2exp (u, u, 1, MPFR_RNDN);
inex = mpfr_integer_p (u);
mpfr_div_2exp (u, u, 1, MPFR_RNDN);
if (inex)
{
inex = mpfr_digamma (y, u, rnd_mode);
goto end;
}
mpfr_init2 (t, p);
mpfr_init2 (v, p);
MPFR_ZIV_INIT (loop, p);
for (;;)
{
mpfr_const_pi (v, MPFR_RNDN); /* v = Pi*(1+theta) for |theta|<=2^(-p) */
mpfr_mul (t, v, x, MPFR_RNDN); /* (1+theta)^2 */
e1 = MPFR_EXP(t) - (mpfr_exp_t) p + 1; /* bound for t: err(t) <= 2^e1 */
mpfr_cot (t, t, MPFR_RNDN);
/* cot(t * (1+h)) = cot(t) - theta * (1 + cot(t)^2) with |theta|<=t*h */
if (MPFR_EXP(t) > 0)
e1 = e1 + 2 * MPFR_EXP(t) + 1;
else
e1 = e1 + 1;
/* now theta * (1 + cot(t)^2) <= 2^e1 */
e1 += (mpfr_exp_t) p - MPFR_EXP(t); /* error is now 2^e1 ulps */
mpfr_mul (t, t, v, MPFR_RNDN);
e1 ++;
mpfr_digamma (v, u, MPFR_RNDN); /* error <= 1/2 ulp */
expv = MPFR_EXP(v);
mpfr_sub (v, v, t, MPFR_RNDN);
if (MPFR_EXP(v) < MPFR_EXP(t))
e1 += MPFR_EXP(t) - MPFR_EXP(v); /* scale error for t wrt new v */
/* now take into account the 1/2 ulp error for v */
if (expv - MPFR_EXP(v) - 1 > e1)
e1 = expv - MPFR_EXP(v) - 1;
else
e1 ++;
e1 ++; /* rounding error for mpfr_sub */
if (MPFR_CAN_ROUND (v, p - e1, MPFR_PREC(y), rnd_mode))
break;
MPFR_ZIV_NEXT (loop, p);
mpfr_set_prec (t, p);
mpfr_set_prec (v, p);
}
MPFR_ZIV_FREE (loop);
inex = mpfr_set (y, v, rnd_mode);
mpfr_clear (t);
mpfr_clear (v);
end:
mpfr_clear (u);
return inex;
}
static inline void
adjust_lunar_phase_to_zero(mpfr_t *result) {
mpfr_t ll, delta;
int mode = -1;
int loop = 1;
int count = 0;
/* Adjust values so that it's as close as possible to 0 degrees.
* if we have a delta of 1 degree, then we're about
* 1 / ( 360 / MEAN_SYNODIC_MONTH )
* days apart
*/
mpfr_init(ll);
mpfr_init_set_d(delta, 0.0001, GMP_RNDN);
while (loop) {
int flipped = mode;
mpfr_t new_moment;
count++;
mpfr_init(new_moment);
lunar_phase(&ll, result);
#if (TRACE)
mpfr_fprintf(stderr,
"Adjusting ll from (%.30RNf) moment is %.5RNf delta is %.30RNf\n", ll, *result, delta);
#endif
/* longitude was greater than 180, so we're looking to add a few
* degrees to make it close to 360 ( 0 )
*/
if (mpfr_cmp_ui( ll, 180 ) > 0) {
mode = 1;
mpfr_sub_ui(delta, ll, 360, GMP_RNDN);
mpfr_div_d(delta, delta, 360 / MEAN_SYNODIC_MONTH, GMP_RNDN);
mpfr_add( new_moment, *result, delta, GMP_RNDN );
#if (TRACE)
mpfr_fprintf(stderr, "add %.30RNf -> %.30RNf\n", *result, new_moment);
#endif
mpfr_set(*result, new_moment, GMP_RNDN);
if (mpfr_cmp(new_moment, *result) == 0) {
loop = 0;
}
} else if (mpfr_cmp_ui( ll, 180 ) < 0 ) {
if ( mpfr_cmp_d( ll, 0.000000000000000000001 ) < 0) {
loop = 0;
} else {
mode = 0;
mpfr_sub_ui(delta, ll, 0, GMP_RNDN);
mpfr_div_d(delta, delta, 360 / MEAN_SYNODIC_MONTH, GMP_RNDN);
mpfr_sub( new_moment, *result, delta, GMP_RNDN );
#if (TRACE)
mpfr_fprintf(stderr, "sub %.120RNf -> %.120RNf\n", *result, new_moment);
#endif
if (mpfr_cmp(new_moment, *result) == 0) {
loop = 0;
}
mpfr_set(*result, new_moment, GMP_RNDN);
}
} else {
loop = 0;
}
if (flipped != -1 && flipped != mode) {
mpfr_div_d(delta, delta, 1.1, GMP_RNDN);
}
mpfr_clear(new_moment);
}
mpfr_clear(delta);
mpfr_clear(ll);
}
int
lunar_longitude( mpfr_t *result, mpfr_t *moment ) {
mpfr_t C, mean_moon, elongation, solar_anomaly, lunar_anomaly, moon_node, E, correction, venus, jupiter, flat_earth, N, fullangle;
mpfr_init(C);
julian_centuries( &C, moment );
{
mpfr_t a, b, c, d, e;
mpfr_init(mean_moon);
mpfr_init_set_d(a, 218.316591, GMP_RNDN);
mpfr_init_set_d(b, 481267.88134236, GMP_RNDN);
mpfr_init_set_d(c, -0.0013268, GMP_RNDN);
mpfr_init_set_ui(d, 1, GMP_RNDN);
mpfr_div_ui(d, d, 538841, GMP_RNDN);
mpfr_init_set_si(e, -1, GMP_RNDN);
mpfr_div_ui(e, e, 65194000, GMP_RNDN);
polynomial( &mean_moon, &C, 5, &a, &b, &c, &d, &e );
mpfr_clear(a);
mpfr_clear(b);
mpfr_clear(c);
mpfr_clear(d);
mpfr_clear(e);
}
{
mpfr_t a, b, c, d, e;
mpfr_init(elongation);
mpfr_init_set_d(a, 297.8502042, GMP_RNDN);
mpfr_init_set_d(b, 445267.1115168, GMP_RNDN);
mpfr_init_set_d(c, -0.00163, GMP_RNDN);
mpfr_init_set_ui(d, 1, GMP_RNDN);
mpfr_div_ui(d, d, 545868, GMP_RNDN);
mpfr_init_set_si(e, -1, GMP_RNDN);
mpfr_div_ui(e, e, 113065000, GMP_RNDN);
polynomial( &elongation, &C, 5, &a, &b, &c, &d, &e );
mpfr_clear(a);
mpfr_clear(b);
mpfr_clear(c);
mpfr_clear(d);
mpfr_clear(e);
}
{
mpfr_t a, b, c, d;
mpfr_init(solar_anomaly);
mpfr_init_set_d(a, 357.5291092, GMP_RNDN);
mpfr_init_set_d(b, 35999.0502909, GMP_RNDN);
mpfr_init_set_d(c, -0.0001536, GMP_RNDN);
mpfr_init_set_ui(d, 1, GMP_RNDN);
mpfr_div_ui(d, d, 24490000, GMP_RNDN);
polynomial( &solar_anomaly, &C, 4, &a, &b, &c, &d );
mpfr_clear(a);
mpfr_clear(b);
mpfr_clear(c);
mpfr_clear(d);
}
{
mpfr_t a, b, c, d, e;
mpfr_init(lunar_anomaly);
mpfr_init_set_d(a, 134.9634114, GMP_RNDN);
mpfr_init_set_d(b, 477198.8676313, GMP_RNDN);
mpfr_init_set_d(c, 0.0008997, GMP_RNDN);
mpfr_init_set_ui(d, 1, GMP_RNDN);
mpfr_div_ui(d, d, 69699, GMP_RNDN);
mpfr_init_set_si(e, -1, GMP_RNDN);
mpfr_div_ui(e, e, 14712000, GMP_RNDN);
polynomial( &lunar_anomaly, &C, 5, &a, &b, &c, &d, &e);
mpfr_clear(a);
mpfr_clear(b);
mpfr_clear(c);
mpfr_clear(d);
mpfr_clear(e);
}
{
mpfr_t a, b, c, d, e;
mpfr_init(moon_node);
mpfr_init_set_d(a, 93.2720993, GMP_RNDN);
mpfr_init_set_d(b, 483202.0175273, GMP_RNDN);
mpfr_init_set_d(c, -0.0034029, GMP_RNDN);
mpfr_init_set_si(d, -1, GMP_RNDN);
mpfr_div_ui(d, d, 3526000, GMP_RNDN);
mpfr_init_set_ui(e, 1, GMP_RNDN);
mpfr_div_ui(e, e, 863310000, GMP_RNDN);
polynomial(&moon_node, &C, 5, &a, &b, &c, &d, &e);
mpfr_clear(a);
mpfr_clear(b);
mpfr_clear(c);
mpfr_clear(d);
mpfr_clear(e);
}
{
mpfr_t a, b, c;
mpfr_init(E);
mpfr_init_set_ui(a, 1, GMP_RNDN);
mpfr_init_set_d(b, -0.002516, GMP_RNDN);
mpfr_init_set_d(c, -0.0000074, GMP_RNDN);
polynomial( &E, &C, 3, &a, &b, &c );
mpfr_clear(a);
mpfr_clear(b);
mpfr_clear(c);
}
{
int i;
mpfr_t fugly;
mpfr_init_set_ui(fugly, 0, GMP_RNDN);
for(i = 0; i < LUNAR_LONGITUDE_ARGS_SIZE; i++) {
mpfr_t a, b, v, w, x, y, z;
mpfr_init_set_d( v, LUNAR_LONGITUDE_ARGS[i][0], GMP_RNDN );
mpfr_init_set_d( w, LUNAR_LONGITUDE_ARGS[i][1], GMP_RNDN );
mpfr_init_set_d( x, LUNAR_LONGITUDE_ARGS[i][2], GMP_RNDN );
mpfr_init_set_d( y, LUNAR_LONGITUDE_ARGS[i][3], GMP_RNDN );
mpfr_init_set_d( z, LUNAR_LONGITUDE_ARGS[i][4], GMP_RNDN );
mpfr_init(b);
mpfr_pow(b, E, x, GMP_RNDN);
mpfr_mul(w, w, elongation, GMP_RNDN);
mpfr_mul(x, x, solar_anomaly, GMP_RNDN);
mpfr_mul(y, y, lunar_anomaly, GMP_RNDN);
mpfr_mul(z, z, moon_node, GMP_RNDN);
mpfr_init_set(a, w, GMP_RNDN);
mpfr_add(a, a, x, GMP_RNDN);
mpfr_add(a, a, y, GMP_RNDN);
mpfr_add(a, a, z, GMP_RNDN);
dt_astro_sin(&a, &a);
mpfr_mul(a, a, v, GMP_RNDN);
mpfr_mul(a, a, b, GMP_RNDN);
mpfr_add(fugly, fugly, a, GMP_RNDN);
mpfr_clear(a);
mpfr_clear(b);
mpfr_clear(v);
mpfr_clear(w);
mpfr_clear(x);
mpfr_clear(y);
mpfr_clear(z);
}
mpfr_init_set_d( correction, 0.000001, GMP_RNDN );
mpfr_mul( correction, correction, fugly, GMP_RNDN);
mpfr_clear(fugly);
}
{
mpfr_t a, b;
mpfr_init(venus);
mpfr_init_set_d(a, 119.75, GMP_RNDN);
mpfr_init_set(b, C, GMP_RNDN);
mpfr_mul_d(b, b, 131.849, GMP_RNDN);
mpfr_add(a, a, b, GMP_RNDN);
dt_astro_sin(&a, &a);
mpfr_mul_d(venus, a, 0.003957, GMP_RNDN );
mpfr_clear(a);
mpfr_clear(b);
}
{
mpfr_t a, b;
mpfr_init(jupiter);
mpfr_init_set_d(a, 53.09, GMP_RNDN);
mpfr_init_set(b, C, GMP_RNDN);
mpfr_mul_d(b, b, 479264.29, GMP_RNDN);
mpfr_add(a, a, b, GMP_RNDN);
dt_astro_sin(&a, &a);
mpfr_mul_d(jupiter, a, 0.000318, GMP_RNDN );
mpfr_clear(a);
mpfr_clear(b);
}
{
mpfr_t a;
mpfr_init(flat_earth);
mpfr_init_set(a, mean_moon, GMP_RNDN);
mpfr_sub(a, a, moon_node, GMP_RNDN);
dt_astro_sin(&a, &a);
mpfr_mul_d(flat_earth, a, 0.001962, GMP_RNDN);
mpfr_clear(a);
}
mpfr_set(*result, mean_moon, GMP_RNDN);
mpfr_add(*result, *result, correction, GMP_RNDN);
mpfr_add(*result, *result, venus, GMP_RNDN);
mpfr_add(*result, *result, jupiter, GMP_RNDN);
mpfr_add(*result, *result, flat_earth, GMP_RNDN);
#ifdef ANNOYING_DEBUG
#if (ANNOYING_DEBUG)
mpfr_fprintf(stderr,
"mean_moon = %.10RNf\ncorrection = %.10RNf\nvenus = %.10RNf\njupiter = %.10RNf\nflat_earth = %.10RNf\n",
mean_moon,
correction,
venus,
jupiter,
flat_earth);
#endif
#endif
mpfr_init(N);
nutation(&N, moment);
mpfr_add(*result, *result, N, GMP_RNDN);
mpfr_init_set_ui(fullangle, 360, GMP_RNDN);
#ifdef ANNOYING_DEBUG
#if (ANNOYING_DEBUG)
mpfr_fprintf(stderr, "lunar = mod(%.10RNf) = ", *result );
#endif
#endif
dt_astro_mod(result, result, &fullangle);
#ifdef ANNOYING_DEBUG
#if (ANNOYING_DEBUG)
mpfr_fprintf(stderr, "%.10RNf\n", *result );
#endif
#endif
mpfr_clear(C);
mpfr_clear(mean_moon);
mpfr_clear(elongation);
mpfr_clear(solar_anomaly);
mpfr_clear(lunar_anomaly);
mpfr_clear(moon_node);
mpfr_clear(E);
mpfr_clear(correction);
mpfr_clear(venus);
mpfr_clear(jupiter);
mpfr_clear(flat_earth);
mpfr_clear(N);
mpfr_clear(fullangle);
return 1;
}
void mp_ICsub (mp_interval_t *rop, mpfr_t op1, mp_interval_t op2) {
// LEFT BOUNDARY, ROUNDING DOWNWARDS
mpfr_sub (rop->a, op1, op2.b, MPFR_RNDD);
// RIGHT BOUNDARY, ROUNDING UPWARDS
mpfr_sub (rop->b, op1, op2.a, MPFR_RNDU);
}
int
mpfr_cos (mpfr_ptr y, mpfr_srcptr x, mp_rnd_t rnd_mode)
{
int K0, K, precy, m, k, l;
int inexact;
mpfr_t r, s;
mp_limb_t *rp, *sp;
mp_size_t sm;
mp_exp_t exps, cancel = 0;
TMP_DECL (marker);
if (MPFR_UNLIKELY(MPFR_IS_SINGULAR(x)))
{
if (MPFR_IS_NAN(x) || MPFR_IS_INF(x))
{
MPFR_SET_NAN(y);
MPFR_RET_NAN;
}
else
{
MPFR_ASSERTD(MPFR_IS_ZERO(x));
return mpfr_set_ui (y, 1, GMP_RNDN);
}
}
mpfr_save_emin_emax ();
precy = MPFR_PREC(y);
K0 = __gmpfr_isqrt(precy / 2); /* Need K + log2(precy/K) extra bits */
m = precy + 3 * (K0 + 2 * MAX(MPFR_GET_EXP (x), 0)) + 3;
TMP_MARK(marker);
sm = (m + BITS_PER_MP_LIMB - 1) / BITS_PER_MP_LIMB;
MPFR_TMP_INIT(rp, r, m, sm);
MPFR_TMP_INIT(sp, s, m, sm);
for (;;)
{
mpfr_mul (r, x, x, GMP_RNDU); /* err <= 1 ulp */
/* we need that |r| < 1 for mpfr_cos2_aux, i.e. up(x^2)/2^(2K) < 1 */
K = K0 + MAX (MPFR_GET_EXP (r), 0);
mpfr_div_2ui (r, r, 2 * K, GMP_RNDN); /* r = (x/2^K)^2, err <= 1 ulp */
/* s <- 1 - r/2! + ... + (-1)^l r^l/(2l)! */
l = mpfr_cos2_aux (s, r);
MPFR_SET_ONE (r);
for (k = 0; k < K; k++)
{
mpfr_mul (s, s, s, GMP_RNDU); /* err <= 2*olderr */
mpfr_mul_2ui (s, s, 1, GMP_RNDU); /* err <= 4*olderr */
mpfr_sub (s, s, r, GMP_RNDN);
}
/* absolute error on s is bounded by (2l+1/3)*2^(2K-m) */
for (k = 2 * K, l = 2 * l + 1; l > 1; l = (l + 1) >> 1)
k++;
/* now the error is bounded by 2^(k-m) = 2^(EXP(s)-err) */
exps = MPFR_GET_EXP(s);
if (MPFR_LIKELY(mpfr_can_round (s, exps + m - k, GMP_RNDN, GMP_RNDZ,
precy + (rnd_mode == GMP_RNDN))))
break;
m += BITS_PER_MP_LIMB;
if (exps < cancel)
{
m += cancel - exps;
cancel = exps;
}
sm = (m + BITS_PER_MP_LIMB - 1) / BITS_PER_MP_LIMB;
MPFR_TMP_INIT(rp, r, m, sm);
MPFR_TMP_INIT(sp, s, m, sm);
}
mpfr_restore_emin_emax ();
inexact = mpfr_set (y, s, rnd_mode); /* FIXME: Dont' need check range? */
TMP_FREE(marker);
return inexact;
}
/* Input: s - a floating-point number >= 1/2.
rnd_mode - a rounding mode.
Assumes s is neither NaN nor Infinite.
Output: z - Zeta(s) rounded to the precision of z with direction rnd_mode
*/
static int
mpfr_zeta_pos (mpfr_t z, mpfr_srcptr s, mpfr_rnd_t rnd_mode)
{
mpfr_t b, c, z_pre, f, s1;
double beta, sd, dnep;
mpfr_t *tc1;
mpfr_prec_t precz, precs, d, dint;
int p, n, l, add;
int inex;
MPFR_GROUP_DECL (group);
MPFR_ZIV_DECL (loop);
MPFR_ASSERTD (MPFR_IS_POS (s) && MPFR_GET_EXP (s) >= 0);
precz = MPFR_PREC (z);
precs = MPFR_PREC (s);
/* Zeta(x) = 1+1/2^x+1/3^x+1/4^x+1/5^x+O(1/6^x)
so with 2^(EXP(x)-1) <= x < 2^EXP(x)
So for x > 2^3, k^x > k^8, so 2/k^x < 2/k^8
Zeta(x) = 1 + 1/2^x*(1+(2/3)^x+(2/4)^x+...)
= 1 + 1/2^x*(1+sum((2/k)^x,k=3..infinity))
<= 1 + 1/2^x*(1+sum((2/k)^8,k=3..infinity))
And sum((2/k)^8,k=3..infinity) = -257+128*Pi^8/4725 ~= 0.0438035
So Zeta(x) <= 1 + 1/2^x*2 for x >= 8
The error is < 2^(-x+1) <= 2^(-2^(EXP(x)-1)+1) */
if (MPFR_GET_EXP (s) > 3)
{
mpfr_exp_t err;
err = MPFR_GET_EXP (s) - 1;
if (err > (mpfr_exp_t) (sizeof (mpfr_exp_t)*CHAR_BIT-2))
err = MPFR_EMAX_MAX;
else
err = ((mpfr_exp_t)1) << err;
err = 1 - (-err+1); /* GET_EXP(one) - (-err+1) = err :) */
MPFR_FAST_COMPUTE_IF_SMALL_INPUT (z, __gmpfr_one, err, 0, 1,
rnd_mode, {});
}
d = precz + MPFR_INT_CEIL_LOG2(precz) + 10;
/* we want that s1 = s-1 is exact, i.e. we should have PREC(s1) >= EXP(s) */
dint = (mpfr_uexp_t) MPFR_GET_EXP (s);
mpfr_init2 (s1, MAX (precs, dint));
inex = mpfr_sub (s1, s, __gmpfr_one, MPFR_RNDN);
MPFR_ASSERTD (inex == 0);
/* case s=1 should have already been handled */
MPFR_ASSERTD (!MPFR_IS_ZERO (s1));
MPFR_GROUP_INIT_4 (group, MPFR_PREC_MIN, b, c, z_pre, f);
MPFR_ZIV_INIT (loop, d);
for (;;)
{
/* Principal loop: we compute, in z_pre,
an approximation of Zeta(s), that we send to can_round */
if (MPFR_GET_EXP (s1) <= -(mpfr_exp_t) ((mpfr_prec_t) (d-3)/2))
/* Branch 1: when s-1 is very small, one
uses the approximation Zeta(s)=1/(s-1)+gamma,
where gamma is Euler's constant */
{
dint = MAX (d + 3, precs);
/* branch 1, with internal precision dint */
MPFR_GROUP_REPREC_4 (group, dint, b, c, z_pre, f);
mpfr_div (z_pre, __gmpfr_one, s1, MPFR_RNDN);
mpfr_const_euler (f, MPFR_RNDN);
mpfr_add (z_pre, z_pre, f, MPFR_RNDN);
}
else /* Branch 2 */
{
size_t size;
/* branch 2 */
/* Computation of parameters n, p and working precision */
dnep = (double) d * LOG2;
sd = mpfr_get_d (s, MPFR_RNDN);
/* beta = dnep + 0.61 + sd * log (6.2832 / sd);
but a larger value is OK */
#define LOG6dot2832 1.83787940484160805532
beta = dnep + 0.61 + sd * (LOG6dot2832 - LOG2 *
__gmpfr_floor_log2 (sd));
if (beta <= 0.0)
{
p = 0;
/* n = 1 + (int) (exp ((dnep - LOG2) / sd)); */
n = 1 + (int) __gmpfr_ceil_exp2 ((d - 1.0) / sd);
}
else
{
p = 1 + (int) beta / 2;
n = 1 + (int) ((sd + 2.0 * (double) p - 1.0) / 6.2832);
}
/* add = 4 + floor(1.5 * log(d) / log (2)).
We should have add >= 10, which is always fulfilled since
d = precz + 11 >= 12, thus ceil(log2(d)) >= 4 */
add = 4 + (3 * MPFR_INT_CEIL_LOG2 (d)) / 2;
MPFR_ASSERTD(add >= 10);
dint = d + add;
if (dint < precs)
dint = precs;
/* internal precision is dint */
size = (p + 1) * sizeof(mpfr_t);
tc1 = (mpfr_t*) mpfr_allocate_func (size);
for (l=1; l<=p; l++)
mpfr_init2 (tc1[l], dint);
MPFR_GROUP_REPREC_4 (group, dint, b, c, z_pre, f);
/* precision of z is precz */
/* Computation of the coefficients c_k */
mpfr_zeta_c (p, tc1);
/* Computation of the 3 parts of the function Zeta. */
mpfr_zeta_part_a (z_pre, s, n);
mpfr_zeta_part_b (b, s, n, p, tc1);
/* s1 = s-1 is already computed above */
mpfr_div (c, __gmpfr_one, s1, MPFR_RNDN);
mpfr_ui_pow (f, n, s1, MPFR_RNDN);
mpfr_div (c, c, f, MPFR_RNDN);
mpfr_add (z_pre, z_pre, c, MPFR_RNDN);
mpfr_add (z_pre, z_pre, b, MPFR_RNDN);
for (l=1; l<=p; l++)
mpfr_clear (tc1[l]);
mpfr_free_func (tc1, size);
/* End branch 2 */
}
if (MPFR_LIKELY (MPFR_CAN_ROUND (z_pre, d-3, precz, rnd_mode)))
break;
MPFR_ZIV_NEXT (loop, d);
}
MPFR_ZIV_FREE (loop);
inex = mpfr_set (z, z_pre, rnd_mode);
MPFR_GROUP_CLEAR (group);
mpfr_clear (s1);
return inex;
}
static void
underflow_up1 (void)
{
mpfr_t delta, x, y, z, z0;
mpfr_exp_t n;
int inex;
int rnd;
int i;
n = mpfr_get_emin ();
if (n < LONG_MIN)
return;
mpfr_init2 (delta, 2);
inex = mpfr_set_ui_2exp (delta, 1, -2, MPFR_RNDN);
MPFR_ASSERTN (inex == 0);
mpfr_init2 (x, 8);
inex = mpfr_set_ui (x, 2, MPFR_RNDN);
MPFR_ASSERTN (inex == 0);
mpfr_init2 (y, sizeof (long) * CHAR_BIT + 4);
inex = mpfr_set_si (y, n, MPFR_RNDN);
MPFR_ASSERTN (inex == 0);
mpfr_init2 (z0, 2);
mpfr_set_ui (z0, 0, MPFR_RNDN);
mpfr_init2 (z, 32);
for (i = 0; i <= 12; i++)
{
unsigned int flags = 0;
char sy[16];
/* Test 2^(emin - i/4).
* --> Underflow iff i > 4.
* --> Zero in MPFR_RNDN iff i >= 8.
*/
if (i != 0 && i != 4)
flags |= MPFR_FLAGS_INEXACT;
if (i > 4)
flags |= MPFR_FLAGS_UNDERFLOW;
sprintf (sy, "emin - %d/4", i);
RND_LOOP (rnd)
{
int zero;
zero = (i > 4 && (rnd == MPFR_RNDZ || rnd == MPFR_RNDD)) ||
(i >= 8 && rnd == MPFR_RNDN);
mpfr_clear_flags ();
inex = mpfr_pow (z, x, y, (mpfr_rnd_t) rnd);
cmpres (1, "2", sy, (mpfr_rnd_t) rnd, zero ? z0 : (mpfr_ptr) NULL,
-1, z, inex, flags, "underflow_up1", "mpfr_pow");
test_others ("2", sy, (mpfr_rnd_t) rnd, x, y, z, inex, flags,
"underflow_up1");
}
inex = mpfr_sub (y, y, delta, MPFR_RNDN);
MPFR_ASSERTN (inex == 0);
}
mpfr_clears (delta, x, y, z, z0, (mpfr_ptr) 0);
}
/* return in z a lower bound (for rnd = RNDD) or upper bound (for rnd = RNDU)
of |zeta(s)|/2, using:
log(|zeta(s)|/2) = (s-1)*log(2*Pi) + lngamma(1-s)
+ log(|sin(Pi*s/2)| * zeta(1-s)).
Assumes s < 1/2 and s1 = 1-s exactly, thus s1 > 1/2.
y and p are temporary variables.
At input, p is Pi rounded down.
The comments in the code are for rnd = RNDD. */
static void
mpfr_reflection_overflow (mpfr_t z, mpfr_t s1, const mpfr_t s, mpfr_t y,
mpfr_t p, mpfr_rnd_t rnd)
{
mpz_t sint;
MPFR_ASSERTD (rnd == MPFR_RNDD || rnd == MPFR_RNDU);
/* Since log is increasing, we want lower bounds on |sin(Pi*s/2)| and
zeta(1-s). */
mpz_init (sint);
mpfr_get_z (sint, s, MPFR_RNDD); /* sint = floor(s) */
/* We first compute a lower bound of |sin(Pi*s/2)|, which is a periodic
function of period 2. Thus:
if 2k < s < 2k+1, then |sin(Pi*s/2)| is increasing;
if 2k-1 < s < 2k, then |sin(Pi*s/2)| is decreasing.
These cases are distinguished by testing bit 0 of floor(s) as if
represented in two's complement (or equivalently, as an unsigned
integer mod 2):
0: sint = 0 mod 2, thus 2k < s < 2k+1 and |sin(Pi*s/2)| is increasing;
1: sint = 1 mod 2, thus 2k-1 < s < 2k and |sin(Pi*s/2)| is decreasing.
Let's recall that the comments are for rnd = RNDD. */
if (mpz_tstbit (sint, 0) == 0) /* |sin(Pi*s/2)| is increasing: round down
Pi*s to get a lower bound. */
{
mpfr_mul (y, p, s, rnd);
if (rnd == MPFR_RNDD)
mpfr_nextabove (p); /* we will need p rounded above afterwards */
}
else /* |sin(Pi*s/2)| is decreasing: round up Pi*s to get a lower bound. */
{
if (rnd == MPFR_RNDD)
mpfr_nextabove (p);
mpfr_mul (y, p, s, MPFR_INVERT_RND(rnd));
}
mpfr_div_2ui (y, y, 1, MPFR_RNDN); /* exact, rounding mode doesn't matter */
/* The rounding direction of sin depends on its sign. We have:
if -4k-2 < s < -4k, then -2k-1 < s/2 < -2k, thus sin(Pi*s/2) < 0;
if -4k < s < -4k+2, then -2k < s/2 < -2k+1, thus sin(Pi*s/2) > 0.
These cases are distinguished by testing bit 1 of floor(s) as if
represented in two's complement (or equivalently, as an unsigned
integer mod 4):
0: sint = {0,1} mod 4, thus -2k < s/2 < -2k+1 and sin(Pi*s/2) > 0;
1: sint = {2,3} mod 4, thus -2k-1 < s/2 < -2k and sin(Pi*s/2) < 0.
Let's recall that the comments are for rnd = RNDD. */
if (mpz_tstbit (sint, 1) == 0) /* -2k < s/2 < -2k+1; sin(Pi*s/2) > 0 */
{
/* Round sin down to get a lower bound of |sin(Pi*s/2)|. */
mpfr_sin (y, y, rnd);
}
else /* -2k-1 < s/2 < -2k; sin(Pi*s/2) < 0 */
{
/* Round sin up to get a lower bound of |sin(Pi*s/2)|. */
mpfr_sin (y, y, MPFR_INVERT_RND(rnd));
mpfr_abs (y, y, MPFR_RNDN); /* exact, rounding mode doesn't matter */
}
mpz_clear (sint);
/* now y <= |sin(Pi*s/2)| when rnd=RNDD, y >= |sin(Pi*s/2)| when rnd=RNDU */
mpfr_zeta_pos (z, s1, rnd); /* zeta(1-s) */
mpfr_mul (z, z, y, rnd);
/* now z <= |sin(Pi*s/2)|*zeta(1-s) */
mpfr_log (z, z, rnd);
/* now z <= log(|sin(Pi*s/2)|*zeta(1-s)) */
mpfr_lngamma (y, s1, rnd);
mpfr_add (z, z, y, rnd);
/* z <= lngamma(1-s) + log(|sin(Pi*s/2)|*zeta(1-s)) */
/* since s-1 < 0, we want to round log(2*pi) upwards */
mpfr_mul_2ui (y, p, 1, MPFR_INVERT_RND(rnd));
mpfr_log (y, y, MPFR_INVERT_RND(rnd));
mpfr_mul (y, y, s1, MPFR_INVERT_RND(rnd));
mpfr_sub (z, z, y, rnd);
mpfr_exp (z, z, rnd);
if (rnd == MPFR_RNDD)
mpfr_nextbelow (p); /* restore original p */
}
/* Implements asymptotic expansion for jn or yn (formulae 9.2.5 and 9.2.6
from Abramowitz & Stegun).
Assumes z > p log(2)/2, where p is the target precision.
Return 0 if the expansion does not converge enough (the value 0 as inexact
flag should not happen for normal input).
*/
static int
FUNCTION (mpfr_ptr res, long n, mpfr_srcptr z, mp_rnd_t r)
{
mpfr_t s, c, P, Q, t, iz, err_t, err_s, err_u;
mp_prec_t w;
long k;
int inex, stop, diverge = 0;
mp_exp_t err2, err;
MPFR_ZIV_DECL (loop);
mpfr_init (c);
w = MPFR_PREC(res) + MPFR_INT_CEIL_LOG2(MPFR_PREC(res)) + 4;
MPFR_ZIV_INIT (loop, w);
for (;;)
{
mpfr_set_prec (c, w);
mpfr_init2 (s, w);
mpfr_init2 (P, w);
mpfr_init2 (Q, w);
mpfr_init2 (t, w);
mpfr_init2 (iz, w);
mpfr_init2 (err_t, 31);
mpfr_init2 (err_s, 31);
mpfr_init2 (err_u, 31);
/* Approximate sin(z) and cos(z). In the following, err <= k means that
the approximate value y and the true value x are related by
y = x * (1 + u)^k with |u| <= 2^(-w), following Higham's method. */
mpfr_sin_cos (s, c, z, GMP_RNDN);
if (MPFR_IS_NEG(z))
mpfr_neg (s, s, GMP_RNDN); /* compute jn/yn(|z|), fix sign later */
/* The absolute error on s/c is bounded by 1/2 ulp(1/2) <= 2^(-w-1). */
mpfr_add (t, s, c, GMP_RNDN);
mpfr_sub (c, s, c, GMP_RNDN);
mpfr_swap (s, t);
/* now s approximates sin(z)+cos(z), and c approximates sin(z)-cos(z),
with total absolute error bounded by 2^(1-w). */
/* precompute 1/(8|z|) */
mpfr_si_div (iz, MPFR_IS_POS(z) ? 1 : -1, z, GMP_RNDN); /* err <= 1 */
mpfr_div_2ui (iz, iz, 3, GMP_RNDN);
/* compute P and Q */
mpfr_set_ui (P, 1, GMP_RNDN);
mpfr_set_ui (Q, 0, GMP_RNDN);
mpfr_set_ui (t, 1, GMP_RNDN); /* current term */
mpfr_set_ui (err_t, 0, GMP_RNDN); /* error on t */
mpfr_set_ui (err_s, 0, GMP_RNDN); /* error on P and Q (sum of errors) */
for (k = 1, stop = 0; stop < 4; k++)
{
/* compute next term: t(k)/t(k-1) = (2n+2k-1)(2n-2k+1)/(8kz) */
mpfr_mul_si (t, t, 2 * (n + k) - 1, GMP_RNDN); /* err <= err_k + 1 */
mpfr_mul_si (t, t, 2 * (n - k) + 1, GMP_RNDN); /* err <= err_k + 2 */
mpfr_div_ui (t, t, k, GMP_RNDN); /* err <= err_k + 3 */
mpfr_mul (t, t, iz, GMP_RNDN); /* err <= err_k + 5 */
/* the relative error on t is bounded by (1+u)^(5k)-1, which is
bounded by 6ku for 6ku <= 0.02: first |5 log(1+u)| <= |5.5u|
for |u| <= 0.15, then |exp(5.5u)-1| <= 6u for |u| <= 0.02. */
mpfr_mul_ui (err_t, t, 6 * k, MPFR_IS_POS(t) ? GMP_RNDU : GMP_RNDD);
mpfr_abs (err_t, err_t, GMP_RNDN); /* exact */
/* the absolute error on t is bounded by err_t * 2^(-w) */
mpfr_abs (err_u, t, GMP_RNDU);
mpfr_mul_2ui (err_u, err_u, w, GMP_RNDU); /* t * 2^w */
mpfr_add (err_u, err_u, err_t, GMP_RNDU); /* max|t| * 2^w */
if (stop >= 2)
{
/* take into account the neglected terms: t * 2^w */
mpfr_div_2ui (err_s, err_s, w, GMP_RNDU);
if (MPFR_IS_POS(t))
mpfr_add (err_s, err_s, t, GMP_RNDU);
else
mpfr_sub (err_s, err_s, t, GMP_RNDU);
mpfr_mul_2ui (err_s, err_s, w, GMP_RNDU);
stop ++;
}
/* if k is odd, add to Q, otherwise to P */
else if (k & 1)
{
/* if k = 1 mod 4, add, otherwise subtract */
if ((k & 2) == 0)
mpfr_add (Q, Q, t, GMP_RNDN);
else
mpfr_sub (Q, Q, t, GMP_RNDN);
/* check if the next term is smaller than ulp(Q): if EXP(err_u)
<= EXP(Q), since the current term is bounded by
err_u * 2^(-w), it is bounded by ulp(Q) */
if (MPFR_EXP(err_u) <= MPFR_EXP(Q))
stop ++;
else
stop = 0;
}
else
{
/* if k = 0 mod 4, add, otherwise subtract */
if ((k & 2) == 0)
mpfr_add (P, P, t, GMP_RNDN);
else
mpfr_sub (P, P, t, GMP_RNDN);
/* check if the next term is smaller than ulp(P) */
if (MPFR_EXP(err_u) <= MPFR_EXP(P))
stop ++;
else
stop = 0;
}
mpfr_add (err_s, err_s, err_t, GMP_RNDU);
/* the sum of the rounding errors on P and Q is bounded by
err_s * 2^(-w) */
/* stop when start to diverge */
if (stop < 2 &&
((MPFR_IS_POS(z) && mpfr_cmp_ui (z, (k + 1) / 2) < 0) ||
(MPFR_IS_NEG(z) && mpfr_cmp_si (z, - ((k + 1) / 2)) > 0)))
{
/* if we have to stop the series because it diverges, then
increasing the precision will most probably fail, since
we will stop to the same point, and thus compute a very
similar approximation */
diverge = 1;
stop = 2; /* force stop */
}
}
/* the sum of the total errors on P and Q is bounded by err_s * 2^(-w) */
/* Now combine: the sum of the rounding errors on P and Q is bounded by
err_s * 2^(-w), and the absolute error on s/c is bounded by 2^(1-w) */
if ((n & 1) == 0) /* n even: P * (sin + cos) + Q (cos - sin) for jn
Q * (sin + cos) + P (sin - cos) for yn */
{
#ifdef MPFR_JN
mpfr_mul (c, c, Q, GMP_RNDN); /* Q * (sin - cos) */
mpfr_mul (s, s, P, GMP_RNDN); /* P * (sin + cos) */
#else
mpfr_mul (c, c, P, GMP_RNDN); /* P * (sin - cos) */
mpfr_mul (s, s, Q, GMP_RNDN); /* Q * (sin + cos) */
#endif
err = MPFR_EXP(c);
if (MPFR_EXP(s) > err)
err = MPFR_EXP(s);
#ifdef MPFR_JN
mpfr_sub (s, s, c, GMP_RNDN);
#else
mpfr_add (s, s, c, GMP_RNDN);
#endif
}
else /* n odd: P * (sin - cos) + Q (cos + sin) for jn,
Q * (sin - cos) - P (cos + sin) for yn */
{
#ifdef MPFR_JN
mpfr_mul (c, c, P, GMP_RNDN); /* P * (sin - cos) */
mpfr_mul (s, s, Q, GMP_RNDN); /* Q * (sin + cos) */
#else
mpfr_mul (c, c, Q, GMP_RNDN); /* Q * (sin - cos) */
mpfr_mul (s, s, P, GMP_RNDN); /* P * (sin + cos) */
#endif
err = MPFR_EXP(c);
if (MPFR_EXP(s) > err)
err = MPFR_EXP(s);
#ifdef MPFR_JN
mpfr_add (s, s, c, GMP_RNDN);
#else
mpfr_sub (s, c, s, GMP_RNDN);
#endif
}
if ((n & 2) != 0)
mpfr_neg (s, s, GMP_RNDN);
if (MPFR_EXP(s) > err)
err = MPFR_EXP(s);
/* the absolute error on s is bounded by P*err(s/c) + Q*err(s/c)
+ err(P)*(s/c) + err(Q)*(s/c) + 3 * 2^(err - w - 1)
<= (|P|+|Q|) * 2^(1-w) + err_s * 2^(1-w) + 2^err * 2^(1-w),
since |c|, |old_s| <= 2. */
err2 = (MPFR_EXP(P) >= MPFR_EXP(Q)) ? MPFR_EXP(P) + 2 : MPFR_EXP(Q) + 2;
/* (|P| + |Q|) * 2^(1 - w) <= 2^(err2 - w) */
err = MPFR_EXP(err_s) >= err ? MPFR_EXP(err_s) + 2 : err + 2;
/* err_s * 2^(1-w) + 2^old_err * 2^(1-w) <= 2^err * 2^(-w) */
err2 = (err >= err2) ? err + 1 : err2 + 1;
/* now the absolute error on s is bounded by 2^(err2 - w) */
/* multiply by sqrt(1/(Pi*z)) */
mpfr_const_pi (c, GMP_RNDN); /* Pi, err <= 1 */
mpfr_mul (c, c, z, GMP_RNDN); /* err <= 2 */
mpfr_si_div (c, MPFR_IS_POS(z) ? 1 : -1, c, GMP_RNDN); /* err <= 3 */
mpfr_sqrt (c, c, GMP_RNDN); /* err<=5/2, thus the absolute error is
bounded by 3*u*|c| for |u| <= 0.25 */
mpfr_mul (err_t, c, s, MPFR_SIGN(c)==MPFR_SIGN(s) ? GMP_RNDU : GMP_RNDD);
mpfr_abs (err_t, err_t, GMP_RNDU);
mpfr_mul_ui (err_t, err_t, 3, GMP_RNDU);
/* 3*2^(-w)*|old_c|*|s| [see below] is bounded by err_t * 2^(-w) */
err2 += MPFR_EXP(c);
/* |old_c| * 2^(err2 - w) [see below] is bounded by 2^(err2-w) */
mpfr_mul (c, c, s, GMP_RNDN); /* the absolute error on c is bounded by
1/2 ulp(c) + 3*2^(-w)*|old_c|*|s|
+ |old_c| * 2^(err2 - w) */
/* compute err_t * 2^(-w) + 1/2 ulp(c) = (err_t + 2^EXP(c)) * 2^(-w) */
err = (MPFR_EXP(err_t) > MPFR_EXP(c)) ? MPFR_EXP(err_t) + 1 : MPFR_EXP(c) + 1;
/* err_t * 2^(-w) + 1/2 ulp(c) <= 2^(err - w) */
/* now err_t * 2^(-w) bounds 1/2 ulp(c) + 3*2^(-w)*|old_c|*|s| */
err = (err >= err2) ? err + 1 : err2 + 1;
/* the absolute error on c is bounded by 2^(err - w) */
mpfr_clear (s);
mpfr_clear (P);
mpfr_clear (Q);
mpfr_clear (t);
mpfr_clear (iz);
mpfr_clear (err_t);
mpfr_clear (err_s);
mpfr_clear (err_u);
err -= MPFR_EXP(c);
if (MPFR_LIKELY (MPFR_CAN_ROUND (c, w - err, MPFR_PREC(res), r)))
break;
if (diverge != 0)
{
mpfr_set (c, z, r); /* will force inex=0 below, which means the
asymptotic expansion failed */
break;
}
MPFR_ZIV_NEXT (loop, w);
}
MPFR_ZIV_FREE (loop);
inex = MPFR_IS_POS(z) ? mpfr_set (res, c, r) : mpfr_neg (res, c, r);
mpfr_clear (c);
return inex;
}
int
mpfr_zeta (mpfr_t z, mpfr_srcptr s, mpfr_rnd_t rnd_mode)
{
mpfr_t z_pre, s1, y, p;
long add;
mpfr_prec_t precz, prec1, precs, precs1;
int inex;
MPFR_GROUP_DECL (group);
MPFR_ZIV_DECL (loop);
MPFR_SAVE_EXPO_DECL (expo);
MPFR_LOG_FUNC (
("s[%Pu]=%.*Rg rnd=%d", mpfr_get_prec (s), mpfr_log_prec, s, rnd_mode),
("z[%Pu]=%.*Rg inexact=%d", mpfr_get_prec (z), mpfr_log_prec, z, inex));
/* Zero, Nan or Inf ? */
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (s)))
{
if (MPFR_IS_NAN (s))
{
MPFR_SET_NAN (z);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (s))
{
if (MPFR_IS_POS (s))
return mpfr_set_ui (z, 1, MPFR_RNDN); /* Zeta(+Inf) = 1 */
MPFR_SET_NAN (z); /* Zeta(-Inf) = NaN */
MPFR_RET_NAN;
}
else /* s iz zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO (s));
return mpfr_set_si_2exp (z, -1, -1, rnd_mode);
}
}
/* s is neither Nan, nor Inf, nor Zero */
/* check tiny s: we have zeta(s) = -1/2 - 1/2 log(2 Pi) s + ... around s=0,
and for |s| <= 2^(-4), we have |zeta(s) + 1/2| <= |s|.
EXP(s) + 1 < -PREC(z) is a sufficient condition to be able to round
correctly, for any PREC(z) >= 1 (see algorithms.tex for details). */
if (MPFR_GET_EXP (s) + 1 < - (mpfr_exp_t) MPFR_PREC(z))
{
int signs = MPFR_SIGN(s);
MPFR_SAVE_EXPO_MARK (expo);
mpfr_set_si_2exp (z, -1, -1, rnd_mode); /* -1/2 */
if (rnd_mode == MPFR_RNDA)
rnd_mode = MPFR_RNDD; /* the result is around -1/2, thus negative */
if ((rnd_mode == MPFR_RNDU || rnd_mode == MPFR_RNDZ) && signs < 0)
{
mpfr_nextabove (z); /* z = -1/2 + epsilon */
inex = 1;
}
else if (rnd_mode == MPFR_RNDD && signs > 0)
{
mpfr_nextbelow (z); /* z = -1/2 - epsilon */
inex = -1;
}
else
{
if (rnd_mode == MPFR_RNDU) /* s > 0: z = -1/2 */
inex = 1;
else if (rnd_mode == MPFR_RNDD)
inex = -1; /* s < 0: z = -1/2 */
else /* (MPFR_RNDZ and s > 0) or MPFR_RNDN: z = -1/2 */
inex = (signs > 0) ? 1 : -1;
}
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (z, inex, rnd_mode);
}
/* Check for case s= -2n */
if (MPFR_IS_NEG (s))
{
mpfr_t tmp;
tmp[0] = *s;
MPFR_EXP (tmp) = MPFR_GET_EXP (s) - 1;
if (mpfr_integer_p (tmp))
{
MPFR_SET_ZERO (z);
MPFR_SET_POS (z);
MPFR_RET (0);
}
}
/* Check for case s=1 before changing the exponent range */
if (mpfr_cmp (s, __gmpfr_one) == 0)
{
MPFR_SET_INF (z);
MPFR_SET_POS (z);
MPFR_SET_DIVBY0 ();
MPFR_RET (0);
}
MPFR_SAVE_EXPO_MARK (expo);
/* Compute Zeta */
if (MPFR_IS_POS (s) && MPFR_GET_EXP (s) >= 0) /* Case s >= 1/2 */
inex = mpfr_zeta_pos (z, s, rnd_mode);
else /* use reflection formula
zeta(s) = 2^s*Pi^(s-1)*sin(Pi*s/2)*gamma(1-s)*zeta(1-s) */
{
int overflow = 0;
precz = MPFR_PREC (z);
precs = MPFR_PREC (s);
/* Precision precs1 needed to represent 1 - s, and s + 2,
without any truncation */
precs1 = precs + 2 + MAX (0, - MPFR_GET_EXP (s));
/* Precision prec1 is the precision on elementary computations;
it ensures a final precision prec1 - add for zeta(s) */
add = compute_add (s, precz);
prec1 = precz + add;
/* FIXME: To avoid that the working precision (prec1) depends on the
input precision, one would need to take into account the error made
when s1 is not exactly 1-s when computing zeta(s1) and gamma(s1)
below, and also in the case y=Inf (i.e. when gamma(s1) overflows).
Make sure that underflows do not occur in intermediate computations.
Due to the limited precision, they are probably not possible
in practice; add some MPFR_ASSERTN's to be sure that problems
do not remain undetected? */
prec1 = MAX (prec1, precs1) + 10;
MPFR_GROUP_INIT_4 (group, prec1, z_pre, s1, y, p);
MPFR_ZIV_INIT (loop, prec1);
for (;;)
{
mpfr_exp_t ey;
mpfr_t z_up;
mpfr_const_pi (p, MPFR_RNDD); /* p is Pi */
mpfr_sub (s1, __gmpfr_one, s, MPFR_RNDN); /* s1 = 1-s */
mpfr_gamma (y, s1, MPFR_RNDN); /* gamma(1-s) */
if (MPFR_IS_INF (y)) /* zeta(s) < 0 for -4k-2 < s < -4k,
zeta(s) > 0 for -4k < s < -4k+2 */
{
/* FIXME: An overflow in gamma(s1) does not imply that
zeta(s) will overflow. A solution:
1. Compute
log(|zeta(s)|/2) = (s-1)*log(2*pi) + lngamma(1-s)
+ log(abs(sin(Pi*s/2)) * zeta(1-s))
(possibly sharing computations with the normal case)
with a rather good accuracy (see (2)).
Memorize the sign of sin(...) for the final sign.
2. Take the exponential, ~= |zeta(s)|/2. If there is an
overflow, then this means an overflow on the final result
(due to the multiplication by 2, which has not been done
yet).
3. Ziv test.
4. Correct the sign from the sign of sin(...).
5. Round then multiply by 2. Here, an overflow in either
operation means a real overflow. */
mpfr_reflection_overflow (z_pre, s1, s, y, p, MPFR_RNDD);
/* z_pre is a lower bound of |zeta(s)|/2, thus if it overflows,
or has exponent emax, then |zeta(s)| overflows too. */
if (MPFR_IS_INF (z_pre) || MPFR_GET_EXP(z_pre) == __gmpfr_emax)
{ /* determine the sign of overflow */
mpfr_div_2ui (s1, s, 2, MPFR_RNDN); /* s/4, exact */
mpfr_frac (s1, s1, MPFR_RNDN); /* exact, -1 < s1 < 0 */
overflow = (mpfr_cmp_si_2exp (s1, -1, -1) > 0) ? -1 : 1;
break;
}
else /* EXP(z_pre) < __gmpfr_emax */
{
int ok = 0;
mpfr_t z_down;
mpfr_init2 (z_up, mpfr_get_prec (z_pre));
mpfr_reflection_overflow (z_up, s1, s, y, p, MPFR_RNDU);
/* if the lower approximation z_pre does not overflow, but
z_up does, we need more precision */
if (MPFR_IS_INF (z_up) || MPFR_GET_EXP(z_up) == __gmpfr_emax)
goto next_loop;
/* check if z_pre and z_up round to the same number */
mpfr_init2 (z_down, precz);
mpfr_set (z_down, z_pre, rnd_mode);
/* Note: it might be that EXP(z_down) = emax here, in that
case we will have overflow below when we multiply by 2 */
mpfr_prec_round (z_up, precz, rnd_mode);
ok = mpfr_cmp (z_down, z_up) == 0;
mpfr_clear (z_up);
mpfr_clear (z_down);
if (ok)
{
/* get correct sign and multiply by 2 */
mpfr_div_2ui (s1, s, 2, MPFR_RNDN); /* s/4, exact */
mpfr_frac (s1, s1, MPFR_RNDN); /* exact, -1 < s1 < 0 */
if (mpfr_cmp_si_2exp (s1, -1, -1) > 0)
mpfr_neg (z_pre, z_pre, rnd_mode);
mpfr_mul_2ui (z_pre, z_pre, 1, rnd_mode);
break;
}
else
goto next_loop;
}
}
mpfr_zeta_pos (z_pre, s1, MPFR_RNDN); /* zeta(1-s) */
mpfr_mul (z_pre, z_pre, y, MPFR_RNDN); /* gamma(1-s)*zeta(1-s) */
/* multiply z_pre by 2^s*Pi^(s-1) where p=Pi, s1=1-s */
mpfr_mul_2ui (y, p, 1, MPFR_RNDN); /* 2*Pi */
mpfr_neg (s1, s1, MPFR_RNDN); /* s-1 */
mpfr_pow (y, y, s1, MPFR_RNDN); /* (2*Pi)^(s-1) */
mpfr_mul (z_pre, z_pre, y, MPFR_RNDN);
mpfr_mul_2ui (z_pre, z_pre, 1, MPFR_RNDN);
/* multiply z_pre by sin(Pi*s/2) */
mpfr_mul (y, s, p, MPFR_RNDN);
mpfr_div_2ui (p, y, 1, MPFR_RNDN); /* p = s*Pi/2 */
/* FIXME: sinpi will be available, we should replace the mpfr_sin
call below by mpfr_sinpi(s/2), where s/2 will be exact.
Can mpfr_sin underflow? Moreover, the code below should be
improved so that the "if" condition becomes unlikely, e.g.
by taking a slightly larger working precision. */
mpfr_sin (y, p, MPFR_RNDN); /* y = sin(Pi*s/2) */
ey = MPFR_GET_EXP (y);
if (ey < 0) /* take account of cancellation in sin(p) */
{
mpfr_t t;
MPFR_ASSERTN (- ey < MPFR_PREC_MAX - prec1);
mpfr_init2 (t, prec1 - ey);
mpfr_const_pi (t, MPFR_RNDD);
mpfr_mul (t, s, t, MPFR_RNDN);
mpfr_div_2ui (t, t, 1, MPFR_RNDN);
mpfr_sin (y, t, MPFR_RNDN);
mpfr_clear (t);
}
mpfr_mul (z_pre, z_pre, y, MPFR_RNDN);
if (MPFR_LIKELY (MPFR_CAN_ROUND (z_pre, prec1 - add, precz,
rnd_mode)))
break;
next_loop:
MPFR_ZIV_NEXT (loop, prec1);
MPFR_GROUP_REPREC_4 (group, prec1, z_pre, s1, y, p);
}
MPFR_ZIV_FREE (loop);
if (overflow != 0)
{
inex = mpfr_overflow (z, rnd_mode, overflow);
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_OVERFLOW);
}
else
inex = mpfr_set (z, z_pre, rnd_mode);
MPFR_GROUP_CLEAR (group);
}
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (z, inex, rnd_mode);
}
static PyObject *
GMPy_Real_DivMod_1(PyObject *x, PyObject *y, CTXT_Object *context)
{
MPFR_Object *tempx = NULL, *tempy = NULL, *quo, *rem;
PyObject *result;
CHECK_CONTEXT(context);
result = PyTuple_New(2);
rem = GMPy_MPFR_New(0, context);
quo = GMPy_MPFR_New(0, context);
if (!result || !rem || !quo) {
Py_XDECREF((PyObject*)result);
Py_XDECREF((PyObject*)quo);
Py_XDECREF((PyObject*)rem);
return NULL;
}
if (IS_REAL(x) && IS_REAL(y)) {
tempx = GMPy_MPFR_From_Real(x, 1, context);
tempy = GMPy_MPFR_From_Real(y, 1, context);
if (!tempx || !tempy) {
goto error;
}
if (mpfr_zero_p(tempy->f)) {
context->ctx.divzero = 1;
if (context->ctx.traps & TRAP_DIVZERO) {
GMPY_DIVZERO("divmod() division by zero");
goto error;
}
}
if (mpfr_nan_p(tempx->f) || mpfr_nan_p(tempy->f) || mpfr_inf_p(tempx->f)) {
context->ctx.invalid = 1;
if (context->ctx.traps & TRAP_INVALID) {
GMPY_INVALID("divmod() invalid operation");
goto error;
}
else {
mpfr_set_nan(quo->f);
mpfr_set_nan(rem->f);
}
}
else if (mpfr_inf_p(tempy->f)) {
context->ctx.invalid = 1;
if (context->ctx.traps & TRAP_INVALID) {
GMPY_INVALID("divmod() invalid operation");
goto error;
}
if (mpfr_zero_p(tempx->f)) {
mpfr_set_zero(quo->f, mpfr_sgn(tempy->f));
mpfr_set_zero(rem->f, mpfr_sgn(tempy->f));
}
else if ((mpfr_signbit(tempx->f)) != (mpfr_signbit(tempy->f))) {
mpfr_set_si(quo->f, -1, MPFR_RNDN);
mpfr_set_inf(rem->f, mpfr_sgn(tempy->f));
}
else {
mpfr_set_si(quo->f, 0, MPFR_RNDN);
rem->rc = mpfr_set(rem->f, tempx->f, MPFR_RNDN);
}
}
else {
MPFR_Object *temp;
if (!(temp = GMPy_MPFR_New(0, context))) {
goto error;
}
mpfr_fmod(rem->f, tempx->f, tempy->f, MPFR_RNDN);
mpfr_sub(temp->f, tempx->f, rem->f, MPFR_RNDN);
mpfr_div(quo->f, temp->f, tempy->f, MPFR_RNDN);
if (!mpfr_zero_p(rem->f)) {
if ((mpfr_sgn(tempy->f) < 0) != (mpfr_sgn(rem->f) < 0)) {
mpfr_add(rem->f, rem->f, tempy->f, MPFR_RNDN);
mpfr_sub_ui(quo->f, quo->f, 1, MPFR_RNDN);
}
}
else {
mpfr_copysign(rem->f, rem->f, tempy->f, MPFR_RNDN);
}
if (!mpfr_zero_p(quo->f)) {
mpfr_round(quo->f, quo->f);
}
else {
mpfr_setsign(quo->f, quo->f, mpfr_sgn(tempx->f) * mpfr_sgn(tempy->f) - 1, MPFR_RNDN);
}
Py_DECREF((PyObject*)temp);
}
GMPY_MPFR_CHECK_RANGE(quo, context);
GMPY_MPFR_CHECK_RANGE(rem, context);
GMPY_MPFR_SUBNORMALIZE(quo, context);
GMPY_MPFR_SUBNORMALIZE(rem, context);
Py_DECREF((PyObject*)tempx);
Py_DECREF((PyObject*)tempy);
PyTuple_SET_ITEM(result, 0, (PyObject*)quo);
PyTuple_SET_ITEM(result, 1, (PyObject*)rem);
return (PyObject*)result;
}
Py_DECREF((PyObject*)rem);
Py_DECREF((PyObject*)quo);
Py_DECREF(result);
Py_RETURN_NOTIMPLEMENTED;
error:
Py_XDECREF((PyObject*)tempx);
Py_XDECREF((PyObject*)tempy);
Py_DECREF((PyObject*)rem);
Py_DECREF((PyObject*)quo);
Py_DECREF(result);
return NULL;
}
/* Put in y an approximation of erfc(x) for large x, using formulae 7.1.23 and
7.1.24 from Abramowitz and Stegun.
Returns e such that the error is bounded by 2^e ulp(y),
or returns 0 in case of underflow.
*/
static mpfr_exp_t
mpfr_erfc_asympt (mpfr_ptr y, mpfr_srcptr x)
{
mpfr_t t, xx, err;
unsigned long k;
mpfr_prec_t prec = MPFR_PREC(y);
mpfr_exp_t exp_err;
mpfr_init2 (t, prec);
mpfr_init2 (xx, prec);
mpfr_init2 (err, 31);
/* let u = 2^(1-p), and let us represent the error as (1+u)^err
with a bound for err */
mpfr_mul (xx, x, x, MPFR_RNDD); /* err <= 1 */
mpfr_ui_div (xx, 1, xx, MPFR_RNDU); /* upper bound for 1/(2x^2), err <= 2 */
mpfr_div_2ui (xx, xx, 1, MPFR_RNDU); /* exact */
mpfr_set_ui (t, 1, MPFR_RNDN); /* current term, exact */
mpfr_set (y, t, MPFR_RNDN); /* current sum */
mpfr_set_ui (err, 0, MPFR_RNDN);
for (k = 1; ; k++)
{
mpfr_mul_ui (t, t, 2 * k - 1, MPFR_RNDU); /* err <= 4k-3 */
mpfr_mul (t, t, xx, MPFR_RNDU); /* err <= 4k */
/* for -1 < x < 1, and |nx| < 1, we have |(1+x)^n| <= 1+7/4|nx|.
Indeed, for x>=0: log((1+x)^n) = n*log(1+x) <= n*x. Let y=n*x < 1,
then exp(y) <= 1+7/4*y.
For x<=0, let x=-x, we can prove by induction that (1-x)^n >= 1-n*x.*/
mpfr_mul_2si (err, err, MPFR_GET_EXP (y) - MPFR_GET_EXP (t), MPFR_RNDU);
mpfr_add_ui (err, err, 14 * k, MPFR_RNDU); /* 2^(1-p) * t <= 2 ulp(t) */
mpfr_div_2si (err, err, MPFR_GET_EXP (y) - MPFR_GET_EXP (t), MPFR_RNDU);
if (MPFR_GET_EXP (t) + (mpfr_exp_t) prec <= MPFR_GET_EXP (y))
{
/* the truncation error is bounded by |t| < ulp(y) */
mpfr_add_ui (err, err, 1, MPFR_RNDU);
break;
}
if (k & 1)
mpfr_sub (y, y, t, MPFR_RNDN);
else
mpfr_add (y, y, t, MPFR_RNDN);
}
/* the error on y is bounded by err*ulp(y) */
mpfr_mul (t, x, x, MPFR_RNDU); /* rel. err <= 2^(1-p) */
mpfr_div_2ui (err, err, 3, MPFR_RNDU); /* err/8 */
mpfr_add (err, err, t, MPFR_RNDU); /* err/8 + xx */
mpfr_mul_2ui (err, err, 3, MPFR_RNDU); /* err + 8*xx */
mpfr_exp (t, t, MPFR_RNDU); /* err <= 1/2*ulp(t) + err(x*x)*t
<= 1/2*ulp(t)+2*|x*x|*ulp(t)
<= (2*|x*x|+1/2)*ulp(t) */
mpfr_mul (t, t, x, MPFR_RNDN); /* err <= 1/2*ulp(t) + (4*|x*x|+1)*ulp(t)
<= (4*|x*x|+3/2)*ulp(t) */
mpfr_const_pi (xx, MPFR_RNDZ); /* err <= ulp(Pi) */
mpfr_sqrt (xx, xx, MPFR_RNDN); /* err <= 1/2*ulp(xx) + ulp(Pi)/2/sqrt(Pi)
<= 3/2*ulp(xx) */
mpfr_mul (t, t, xx, MPFR_RNDN); /* err <= (8 |xx| + 13/2) * ulp(t) */
mpfr_div (y, y, t, MPFR_RNDN); /* the relative error on input y is bounded
by (1+u)^err with u = 2^(1-p), that on
t is bounded by (1+u)^(8 |xx| + 13/2),
thus that on output y is bounded by
8 |xx| + 7 + err. */
if (MPFR_IS_ZERO(y))
{
/* If y is zero, most probably we have underflow. We check it directly
using the fact that erfc(x) <= exp(-x^2)/sqrt(Pi)/x for x >= 0.
We compute an upper approximation of exp(-x^2)/sqrt(Pi)/x.
*/
mpfr_mul (t, x, x, MPFR_RNDD); /* t <= x^2 */
mpfr_neg (t, t, MPFR_RNDU); /* -x^2 <= t */
mpfr_exp (t, t, MPFR_RNDU); /* exp(-x^2) <= t */
mpfr_const_pi (xx, MPFR_RNDD); /* xx <= sqrt(Pi), cached */
mpfr_mul (xx, xx, x, MPFR_RNDD); /* xx <= sqrt(Pi)*x */
mpfr_div (y, t, xx, MPFR_RNDN); /* if y is zero, this means that the upper
approximation of exp(-x^2)/sqrt(Pi)/x
is nearer from 0 than from 2^(-emin-1),
thus we have underflow. */
exp_err = 0;
}
else
{
mpfr_add_ui (err, err, 7, MPFR_RNDU);
exp_err = MPFR_GET_EXP (err);
}
mpfr_clear (t);
mpfr_clear (xx);
mpfr_clear (err);
return exp_err;
}
void subtract(elem &result, elem a, elem b) const { mpfr_sub(&result, &a, &b, GMP_RNDN); }
int my_mpfr_lbeta(mpfr_t R, mpfr_t a, mpfr_t b, mpfr_rnd_t RND)
{
mpfr_prec_t p_a = mpfr_get_prec(a), p_b = mpfr_get_prec(b);
if(p_a < p_b) p_a = p_b;// p_a := max(p_a, p_b)
if(mpfr_get_prec(R) < p_a)
mpfr_prec_round(R, p_a, RND);// so prec(R) = max( prec(a), prec(b) )
int ans;
mpfr_t s;
mpfr_init2(s, p_a);
/* "FIXME": check each 'ans' below, and return when not ok ... */
ans = mpfr_add(s, a, b, RND);
if(mpfr_integer_p(s) && mpfr_sgn(s) <= 0) { // (a + b) is integer <= 0
if(!mpfr_integer_p(a) && !mpfr_integer_p(b)) {
// but a,b not integer ==> R = ln(finite / +-Inf) = ln(0) = -Inf :
mpfr_set_inf (R, -1);
mpfr_clear (s);
return ans;
}// else: sum is integer; at least one integer ==> both integer
int sX = mpfr_sgn(a), sY = mpfr_sgn(b);
if(sX * sY < 0) { // one negative, one positive integer
// ==> special treatment here :
if(sY < 0) // ==> sX > 0; swap the two
mpfr_swap(a, b);
/* now have --- a < 0 < b <= |a| integer ------------------
* ================
* --> see my_mpfr_beta() above */
unsigned long b_ = 0;// -Wall
Rboolean
b_fits_ulong = mpfr_fits_ulong_p(b, RND),
small_b = b_fits_ulong && (b_ = mpfr_get_ui(b, RND)) < b_large;
if(small_b) {
//----------------- small b ------------------
// use GMP big integer arithmetic:
mpz_t S; mpz_init(S); mpfr_get_z(S, s, RND); // S := s
mpz_sub_ui (S, S, (unsigned long) 1); // S = s - 1 = (a+b-1)
/* binomial coefficient choose(N, k) requires k a 'long int';
* here, b must fit into a long: */
mpz_bin_ui (S, S, b_); // S = choose(S, b) = choose(a+b-1, b)
mpz_mul_ui (S, S, b_); // S = S*b = b * choose(a+b-1, b)
// back to mpfr: R = log(|1 / S|) = - log(|S|)
mpz_abs(S, S);
mpfr_set_z(s, S, RND); // <mpfr> s := |S|
mpfr_log(R, s, RND); // R := log(s) = log(|S|)
mpfr_neg(R, R, RND); // R = -R = -log(|S|)
mpz_clear(S);
}
else { // b is "large", use direct B(.,.) formula
// a := (-1)^b -- not needed here, neither 'neg': want log( |.| )
// s' := 1-s = 1-a-b
mpfr_ui_sub(s, 1, s, RND);
// R := log(|B(1-a-b, b)|) = log(|B(s', b)|)
my_mpfr_lbeta (R, s, b, RND);
}
mpfr_clear(s);
return ans;
}
}
ans = mpfr_lngamma(s, s, RND); // s = lngamma(a + b)
ans = mpfr_lngamma(a, a, RND);
ans = mpfr_lngamma(b, b, RND);
ans = mpfr_add (b, b, a, RND); // b' = lngamma(a) + lngamma(b)
ans = mpfr_sub (R, b, s, RND);
mpfr_clear (s);
return ans;
}
int
mpfr_sinh (mpfr_ptr y, mpfr_srcptr xt, mpfr_rnd_t rnd_mode)
{
mpfr_t x;
int inexact;
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg rnd=%d", mpfr_get_prec (xt), mpfr_log_prec, xt, rnd_mode),
("y[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (y), mpfr_log_prec, y, inexact));
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (xt)))
{
if (MPFR_IS_NAN (xt))
{
MPFR_SET_NAN (y);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (xt))
{
MPFR_SET_INF (y);
MPFR_SET_SAME_SIGN (y, xt);
MPFR_RET (0);
}
else /* xt is zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO (xt));
MPFR_SET_ZERO (y); /* sinh(0) = 0 */
MPFR_SET_SAME_SIGN (y, xt);
MPFR_RET (0);
}
}
/* sinh(x) = x + x^3/6 + ... so the error is < 2^(3*EXP(x)-2) */
MPFR_FAST_COMPUTE_IF_SMALL_INPUT (y, xt, -2 * MPFR_GET_EXP(xt), 2, 1,
rnd_mode, {});
MPFR_TMP_INIT_ABS (x, xt);
{
mpfr_t t, ti;
mpfr_exp_t d;
mpfr_prec_t Nt; /* Precision of the intermediary variable */
long int err; /* Precision of error */
MPFR_ZIV_DECL (loop);
MPFR_SAVE_EXPO_DECL (expo);
MPFR_GROUP_DECL (group);
MPFR_SAVE_EXPO_MARK (expo);
/* compute the precision of intermediary variable */
Nt = MAX (MPFR_PREC (x), MPFR_PREC (y));
/* the optimal number of bits : see algorithms.ps */
Nt = Nt + MPFR_INT_CEIL_LOG2 (Nt) + 4;
/* If x is near 0, exp(x) - 1/exp(x) = 2*x+x^3/3+O(x^5) */
if (MPFR_GET_EXP (x) < 0)
Nt -= 2*MPFR_GET_EXP (x);
/* initialize of intermediary variables */
MPFR_GROUP_INIT_2 (group, Nt, t, ti);
/* First computation of sinh */
MPFR_ZIV_INIT (loop, Nt);
for (;;)
{
MPFR_BLOCK_DECL (flags);
/* compute sinh */
MPFR_BLOCK (flags, mpfr_exp (t, x, MPFR_RNDD));
if (MPFR_OVERFLOW (flags))
/* exp(x) does overflow */
{
/* sinh(x) = 2 * sinh(x/2) * cosh(x/2) */
mpfr_div_2ui (ti, x, 1, MPFR_RNDD); /* exact */
/* t <- cosh(x/2): error(t) <= 1 ulp(t) */
MPFR_BLOCK (flags, mpfr_cosh (t, ti, MPFR_RNDD));
if (MPFR_OVERFLOW (flags))
/* when x>1 we have |sinh(x)| >= cosh(x/2), so sinh(x)
overflows too */
{
inexact = mpfr_overflow (y, rnd_mode, MPFR_SIGN (xt));
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_OVERFLOW);
break;
}
/* ti <- sinh(x/2): , error(ti) <= 1 ulp(ti)
cannot overflow because 0 < sinh(x) < cosh(x) when x > 0 */
mpfr_sinh (ti, ti, MPFR_RNDD);
/* multiplication below, error(t) <= 5 ulp(t) */
MPFR_BLOCK (flags, mpfr_mul (t, t, ti, MPFR_RNDD));
if (MPFR_OVERFLOW (flags))
{
inexact = mpfr_overflow (y, rnd_mode, MPFR_SIGN (xt));
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_OVERFLOW);
break;
}
/* doubling below, exact */
MPFR_BLOCK (flags, mpfr_mul_2ui (t, t, 1, MPFR_RNDN));
if (MPFR_OVERFLOW (flags))
{
inexact = mpfr_overflow (y, rnd_mode, MPFR_SIGN (xt));
MPFR_SAVE_EXPO_UPDATE_FLAGS (expo, MPFR_FLAGS_OVERFLOW);
break;
}
/* we have lost at most 3 bits of precision */
err = Nt - 3;
if (MPFR_LIKELY (MPFR_CAN_ROUND (t, err, MPFR_PREC (y),
rnd_mode)))
{
inexact = mpfr_set4 (y, t, rnd_mode, MPFR_SIGN (xt));
break;
}
err = Nt; /* double the precision */
}
else
{
d = MPFR_GET_EXP (t);
mpfr_ui_div (ti, 1, t, MPFR_RNDU); /* 1/exp(x) */
mpfr_sub (t, t, ti, MPFR_RNDN); /* exp(x) - 1/exp(x) */
mpfr_div_2ui (t, t, 1, MPFR_RNDN); /* 1/2(exp(x) - 1/exp(x)) */
/* it may be that t is zero (in fact, it can only occur when te=1,
and thus ti=1 too) */
if (MPFR_IS_ZERO (t))
err = Nt; /* double the precision */
else
{
/* calculation of the error */
d = d - MPFR_GET_EXP (t) + 2;
/* error estimate: err = Nt-(__gmpfr_ceil_log2(1+pow(2,d)));*/
err = Nt - (MAX (d, 0) + 1);
if (MPFR_LIKELY (MPFR_CAN_ROUND (t, err, MPFR_PREC (y),
rnd_mode)))
{
inexact = mpfr_set4 (y, t, rnd_mode, MPFR_SIGN (xt));
break;
}
}
}
/* actualization of the precision */
Nt += err;
MPFR_ZIV_NEXT (loop, Nt);
MPFR_GROUP_REPREC_2 (group, Nt, t, ti);
}
MPFR_ZIV_FREE (loop);
MPFR_GROUP_CLEAR (group);
MPFR_SAVE_EXPO_FREE (expo);
}
return mpfr_check_range (y, inexact, rnd_mode);
}
void subtract(ElementType &result, const ElementType& a, const ElementType& b) const
{
mpfr_sub(&result, &a, &b, GMP_RNDN);
}
/* evaluates erf(x) using the expansion at x=0:
erf(x) = 2/sqrt(Pi) * sum((-1)^k*x^(2k+1)/k!/(2k+1), k=0..infinity)
Assumes x is neither NaN nor infinite nor zero.
Assumes also that e*x^2 <= n (target precision).
*/
static int
mpfr_erf_0 (mpfr_ptr res, mpfr_srcptr x, double xf2, mpfr_rnd_t rnd_mode)
{
mpfr_prec_t n, m;
mpfr_exp_t nuk, sigmak;
double tauk;
mpfr_t y, s, t, u;
unsigned int k;
int log2tauk;
int inex;
MPFR_GROUP_DECL (group);
MPFR_ZIV_DECL (loop);
n = MPFR_PREC (res); /* target precision */
/* initial working precision */
m = n + (mpfr_prec_t) (xf2 / LOG2) + 8 + MPFR_INT_CEIL_LOG2 (n);
MPFR_GROUP_INIT_4(group, m, y, s, t, u);
MPFR_ZIV_INIT (loop, m);
for (;;)
{
mpfr_mul (y, x, x, MPFR_RNDU); /* err <= 1 ulp */
mpfr_set_ui (s, 1, MPFR_RNDN);
mpfr_set_ui (t, 1, MPFR_RNDN);
tauk = 0.0;
for (k = 1; ; k++)
{
mpfr_mul (t, y, t, MPFR_RNDU);
mpfr_div_ui (t, t, k, MPFR_RNDU);
mpfr_div_ui (u, t, 2 * k + 1, MPFR_RNDU);
sigmak = MPFR_GET_EXP (s);
if (k % 2)
mpfr_sub (s, s, u, MPFR_RNDN);
else
mpfr_add (s, s, u, MPFR_RNDN);
sigmak -= MPFR_GET_EXP(s);
nuk = MPFR_GET_EXP(u) - MPFR_GET_EXP(s);
if ((nuk < - (mpfr_exp_t) m) && ((double) k >= xf2))
break;
/* tauk <- 1/2 + tauk * 2^sigmak + (1+8k)*2^nuk */
tauk = 0.5 + mul_2exp (tauk, sigmak)
+ mul_2exp (1.0 + 8.0 * (double) k, nuk);
}
mpfr_mul (s, x, s, MPFR_RNDU);
MPFR_SET_EXP (s, MPFR_GET_EXP (s) + 1);
mpfr_const_pi (t, MPFR_RNDZ);
mpfr_sqrt (t, t, MPFR_RNDZ);
mpfr_div (s, s, t, MPFR_RNDN);
tauk = 4.0 * tauk + 11.0; /* final ulp-error on s */
log2tauk = __gmpfr_ceil_log2 (tauk);
if (MPFR_LIKELY (MPFR_CAN_ROUND (s, m - log2tauk, n, rnd_mode)))
break;
/* Actualisation of the precision */
MPFR_ZIV_NEXT (loop, m);
MPFR_GROUP_REPREC_4 (group, m, y, s, t, u);
}
MPFR_ZIV_FREE (loop);
inex = mpfr_set (res, s, rnd_mode);
MPFR_GROUP_CLEAR (group);
return inex;
}
int
main (void)
{
mpfr_t x;
tests_start_mpfr ();
mpfr_init (x);
/* check +infinity gives non-zero for mpfr_inf_p only */
mpfr_set_ui (x, 1L, MPFR_RNDZ);
mpfr_div_ui (x, x, 0L, MPFR_RNDZ);
if (mpfr_nan_p (x) || (mpfr_nan_p) (x) )
{
printf ("Error: mpfr_nan_p(+Inf) gives non-zero\n");
exit (1);
}
if (mpfr_inf_p (x) == 0)
{
printf ("Error: mpfr_inf_p(+Inf) gives zero\n");
exit (1);
}
if (mpfr_number_p (x) || (mpfr_number_p) (x) )
{
printf ("Error: mpfr_number_p(+Inf) gives non-zero\n");
exit (1);
}
if (mpfr_zero_p (x) || (mpfr_zero_p) (x) )
{
printf ("Error: mpfr_zero_p(+Inf) gives non-zero\n");
exit (1);
}
if (mpfr_regular_p (x) || (mpfr_regular_p) (x) )
{
printf ("Error: mpfr_regular_p(+Inf) gives non-zero\n");
exit (1);
}
/* same for -Inf */
mpfr_neg (x, x, MPFR_RNDN);
if (mpfr_nan_p (x) || (mpfr_nan_p(x)))
{
printf ("Error: mpfr_nan_p(-Inf) gives non-zero\n");
exit (1);
}
if (mpfr_inf_p (x) == 0)
{
printf ("Error: mpfr_inf_p(-Inf) gives zero\n");
exit (1);
}
if (mpfr_number_p (x) || (mpfr_number_p)(x) )
{
printf ("Error: mpfr_number_p(-Inf) gives non-zero\n");
exit (1);
}
if (mpfr_zero_p (x) || (mpfr_zero_p)(x) )
{
printf ("Error: mpfr_zero_p(-Inf) gives non-zero\n");
exit (1);
}
if (mpfr_regular_p (x) || (mpfr_regular_p) (x) )
{
printf ("Error: mpfr_regular_p(-Inf) gives non-zero\n");
exit (1);
}
/* same for NaN */
mpfr_sub (x, x, x, MPFR_RNDN);
if (mpfr_nan_p (x) == 0)
{
printf ("Error: mpfr_nan_p(NaN) gives zero\n");
exit (1);
}
if (mpfr_inf_p (x) || (mpfr_inf_p)(x) )
{
printf ("Error: mpfr_inf_p(NaN) gives non-zero\n");
exit (1);
}
if (mpfr_number_p (x) || (mpfr_number_p) (x) )
{
printf ("Error: mpfr_number_p(NaN) gives non-zero\n");
exit (1);
}
if (mpfr_zero_p (x) || (mpfr_zero_p)(x) )
{
printf ("Error: mpfr_number_p(NaN) gives non-zero\n");
exit (1);
}
if (mpfr_regular_p (x) || (mpfr_regular_p) (x) )
{
printf ("Error: mpfr_regular_p(NaN) gives non-zero\n");
exit (1);
}
/* same for a regular number */
mpfr_set_ui (x, 1, MPFR_RNDN);
if (mpfr_nan_p (x) || (mpfr_nan_p)(x))
{
printf ("Error: mpfr_nan_p(1) gives non-zero\n");
exit (1);
}
if (mpfr_inf_p (x) || (mpfr_inf_p)(x) )
{
printf ("Error: mpfr_inf_p(1) gives non-zero\n");
exit (1);
}
if (mpfr_number_p (x) == 0)
{
printf ("Error: mpfr_number_p(1) gives zero\n");
exit (1);
}
if (mpfr_zero_p (x) || (mpfr_zero_p) (x) )
{
printf ("Error: mpfr_zero_p(1) gives non-zero\n");
exit (1);
}
if (mpfr_regular_p (x) == 0 || (mpfr_regular_p) (x) == 0)
{
printf ("Error: mpfr_regular_p(1) gives zero\n");
exit (1);
}
/* Same for +0 */
mpfr_set_ui (x, 0, MPFR_RNDN);
if (mpfr_nan_p (x) || (mpfr_nan_p)(x))
{
printf ("Error: mpfr_nan_p(+0) gives non-zero\n");
exit (1);
}
if (mpfr_inf_p (x) || (mpfr_inf_p)(x) )
{
printf ("Error: mpfr_inf_p(+0) gives non-zero\n");
exit (1);
}
if (mpfr_number_p (x) == 0)
{
printf ("Error: mpfr_number_p(+0) gives zero\n");
exit (1);
}
if (mpfr_zero_p (x) == 0 )
{
printf ("Error: mpfr_zero_p(+0) gives zero\n");
exit (1);
}
if (mpfr_regular_p (x) || (mpfr_regular_p) (x) )
{
printf ("Error: mpfr_regular_p(+0) gives non-zero\n");
exit (1);
}
/* Same for -0 */
mpfr_set_ui (x, 0, MPFR_RNDN);
mpfr_neg (x, x, MPFR_RNDN);
if (mpfr_nan_p (x) || (mpfr_nan_p)(x))
{
printf ("Error: mpfr_nan_p(-0) gives non-zero\n");
exit (1);
}
if (mpfr_inf_p (x) || (mpfr_inf_p)(x) )
{
printf ("Error: mpfr_inf_p(-0) gives non-zero\n");
exit (1);
}
if (mpfr_number_p (x) == 0)
{
printf ("Error: mpfr_number_p(-0) gives zero\n");
exit (1);
}
if (mpfr_zero_p (x) == 0 )
{
printf ("Error: mpfr_zero_p(-0) gives zero\n");
exit (1);
}
if (mpfr_regular_p (x) || (mpfr_regular_p) (x) )
{
printf ("Error: mpfr_regular_p(-0) gives non-zero\n");
exit (1);
}
mpfr_clear (x);
tests_end_mpfr ();
return 0;
}
/* Assumes that the exponent range has already been extended and if y is
an integer, then the result is not exact in unbounded exponent range.
If x < 0, assumes y is an integer.
*/
int
mpfr_pow_general (mpfr_ptr z, mpfr_srcptr x, mpfr_srcptr y,
mpfr_rnd_t rnd_mode, int y_is_integer, mpfr_save_expo_t *expo)
{
mpfr_t t, u, k, absx;
int neg_result = 0;
int k_non_zero = 0;
int check_exact_case = 0;
int inexact;
/* Declaration of the size variable */
mpfr_prec_t Nz = MPFR_PREC(z); /* target precision */
mpfr_prec_t Nt; /* working precision */
mpfr_exp_t err; /* error */
MPFR_ZIV_DECL (ziv_loop);
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg y[%Pu]=%.*Rg rnd=%d",
mpfr_get_prec (x), mpfr_log_prec, x,
mpfr_get_prec (y), mpfr_log_prec, y, rnd_mode),
("z[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (z), mpfr_log_prec, z, inexact));
/* We put the absolute value of x in absx, pointing to the significand
of x to avoid allocating memory for the significand of absx. */
MPFR_ALIAS(absx, x, /*sign=*/ 1, /*EXP=*/ MPFR_EXP(x));
/* We will compute the absolute value of the result. So, let's
invert the rounding mode if the result is negative (in which case
y not an integer was already filtered out). */
if (MPFR_IS_NEG (x))
{
MPFR_ASSERTD (y_is_integer);
if (is_odd (y))
{
neg_result = 1;
rnd_mode = MPFR_INVERT_RND (rnd_mode);
}
}
/* Compute the precision of intermediary variable. */
/* The increment 9 + MPFR_INT_CEIL_LOG2 (Nz) gives few Ziv failures
in binary64 and binary128 formats:
mfv5 -p53 -e1 mpfr_pow: 5903 / 6469.59 / 6686
mfv5 -p113 -e1 mpfr_pow: 10913 / 11989.46 / 12321 */
Nt = Nz + 9 + MPFR_INT_CEIL_LOG2 (Nz);
/* initialize of intermediary variable */
mpfr_init2 (t, Nt);
MPFR_ZIV_INIT (ziv_loop, Nt);
for (;;)
{
MPFR_BLOCK_DECL (flags1);
/* compute exp(y*ln|x|), using MPFR_RNDU to get an upper bound, so
that we can detect underflows. */
mpfr_log (t, absx, MPFR_IS_NEG (y) ? MPFR_RNDD : MPFR_RNDU); /* ln|x| */
mpfr_mul (t, y, t, MPFR_RNDU); /* y*ln|x| */
if (k_non_zero)
{
MPFR_LOG_MSG (("subtract k * ln(2)\n", 0));
mpfr_const_log2 (u, MPFR_RNDD);
mpfr_mul (u, u, k, MPFR_RNDD);
/* Error on u = k * log(2): < k * 2^(-Nt) < 1. */
mpfr_sub (t, t, u, MPFR_RNDU);
MPFR_LOG_MSG (("t = y * ln|x| - k * ln(2)\n", 0));
MPFR_LOG_VAR (t);
}
/* estimate of the error -- see pow function in algorithms.tex.
The error on t is at most 1/2 + 3*2^(EXP(t)+1) ulps, which is
<= 2^(EXP(t)+3) for EXP(t) >= -1, and <= 2 ulps for EXP(t) <= -2.
Additional error if k_no_zero: treal = t * errk, with
1 - |k| * 2^(-Nt) <= exp(-|k| * 2^(-Nt)) <= errk <= 1,
i.e., additional absolute error <= 2^(EXP(k)+EXP(t)-Nt).
Total error <= 2^err1 + 2^err2 <= 2^(max(err1,err2)+1). */
err = MPFR_NOTZERO (t) && MPFR_GET_EXP (t) >= -1 ?
MPFR_GET_EXP (t) + 3 : 1;
if (k_non_zero)
{
if (MPFR_GET_EXP (k) > err)
err = MPFR_GET_EXP (k);
err++;
}
MPFR_BLOCK (flags1, mpfr_exp (t, t, MPFR_RNDN)); /* exp(y*ln|x|)*/
/* We need to test */
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (t) || MPFR_UNDERFLOW (flags1)))
{
mpfr_prec_t Ntmin;
MPFR_BLOCK_DECL (flags2);
MPFR_ASSERTN (!k_non_zero);
MPFR_ASSERTN (!MPFR_IS_NAN (t));
/* Real underflow? */
if (MPFR_IS_ZERO (t))
{
/* Underflow. We computed rndn(exp(t)), where t >= y*ln|x|.
Therefore rndn(|x|^y) = 0, and we have a real underflow on
|x|^y. */
inexact = mpfr_underflow (z, rnd_mode == MPFR_RNDN ? MPFR_RNDZ
: rnd_mode, MPFR_SIGN_POS);
if (expo != NULL)
MPFR_SAVE_EXPO_UPDATE_FLAGS (*expo, MPFR_FLAGS_INEXACT
| MPFR_FLAGS_UNDERFLOW);
break;
}
/* Real overflow? */
if (MPFR_IS_INF (t))
{
/* Note: we can probably use a low precision for this test. */
mpfr_log (t, absx, MPFR_IS_NEG (y) ? MPFR_RNDU : MPFR_RNDD);
mpfr_mul (t, y, t, MPFR_RNDD); /* y * ln|x| */
MPFR_BLOCK (flags2, mpfr_exp (t, t, MPFR_RNDD));
/* t = lower bound on exp(y * ln|x|) */
if (MPFR_OVERFLOW (flags2))
{
/* We have computed a lower bound on |x|^y, and it
overflowed. Therefore we have a real overflow
on |x|^y. */
inexact = mpfr_overflow (z, rnd_mode, MPFR_SIGN_POS);
if (expo != NULL)
MPFR_SAVE_EXPO_UPDATE_FLAGS (*expo, MPFR_FLAGS_INEXACT
| MPFR_FLAGS_OVERFLOW);
break;
}
}
k_non_zero = 1;
Ntmin = sizeof(mpfr_exp_t) * CHAR_BIT;
if (Ntmin > Nt)
{
Nt = Ntmin;
mpfr_set_prec (t, Nt);
}
mpfr_init2 (u, Nt);
mpfr_init2 (k, Ntmin);
mpfr_log2 (k, absx, MPFR_RNDN);
mpfr_mul (k, y, k, MPFR_RNDN);
mpfr_round (k, k);
MPFR_LOG_VAR (k);
/* |y| < 2^Ntmin, therefore |k| < 2^Nt. */
continue;
}
if (MPFR_LIKELY (MPFR_CAN_ROUND (t, Nt - err, Nz, rnd_mode)))
{
inexact = mpfr_set (z, t, rnd_mode);
break;
}
/* check exact power, except when y is an integer (since the
exact cases for y integer have already been filtered out) */
if (check_exact_case == 0 && ! y_is_integer)
{
if (mpfr_pow_is_exact (z, absx, y, rnd_mode, &inexact))
break;
check_exact_case = 1;
}
/* reactualisation of the precision */
MPFR_ZIV_NEXT (ziv_loop, Nt);
mpfr_set_prec (t, Nt);
if (k_non_zero)
mpfr_set_prec (u, Nt);
}
MPFR_ZIV_FREE (ziv_loop);
if (k_non_zero)
{
int inex2;
long lk;
/* The rounded result in an unbounded exponent range is z * 2^k. As
* MPFR chooses underflow after rounding, the mpfr_mul_2si below will
* correctly detect underflows and overflows. However, in rounding to
* nearest, if z * 2^k = 2^(emin - 2), then the double rounding may
* affect the result. We need to cope with that before overwriting z.
* This can occur only if k < 0 (this test is necessary to avoid a
* potential integer overflow).
* If inexact >= 0, then the real result is <= 2^(emin - 2), so that
* o(2^(emin - 2)) = +0 is correct. If inexact < 0, then the real
* result is > 2^(emin - 2) and we need to round to 2^(emin - 1).
*/
MPFR_ASSERTN (MPFR_EXP_MAX <= LONG_MAX);
lk = mpfr_get_si (k, MPFR_RNDN);
/* Due to early overflow detection, |k| should not be much larger than
* MPFR_EMAX_MAX, and as MPFR_EMAX_MAX <= MPFR_EXP_MAX/2 <= LONG_MAX/2,
* an overflow should not be possible in mpfr_get_si (and lk is exact).
* And one even has the following assertion. TODO: complete proof.
*/
MPFR_ASSERTD (lk > LONG_MIN && lk < LONG_MAX);
/* Note: even in case of overflow (lk inexact), the code is correct.
* Indeed, for the 3 occurrences of lk:
* - The test lk < 0 is correct as sign(lk) = sign(k).
* - In the test MPFR_GET_EXP (z) == __gmpfr_emin - 1 - lk,
* if lk is inexact, then lk = LONG_MIN <= MPFR_EXP_MIN
* (the minimum value of the mpfr_exp_t type), and
* __gmpfr_emin - 1 - lk >= MPFR_EMIN_MIN - 1 - 2 * MPFR_EMIN_MIN
* >= - MPFR_EMIN_MIN - 1 = MPFR_EMAX_MAX - 1. However, from the
* choice of k, z has been chosen to be around 1, so that the
* result of the test is false, as if lk were exact.
* - In the mpfr_mul_2si (z, z, lk, rnd_mode), if lk is inexact,
* then |lk| >= LONG_MAX >= MPFR_EXP_MAX, and as z is around 1,
* mpfr_mul_2si underflows or overflows in the same way as if
* lk were exact.
* TODO: give a bound on |t|, then on |EXP(z)|.
*/
if (rnd_mode == MPFR_RNDN && inexact < 0 && lk < 0 &&
MPFR_GET_EXP (z) == __gmpfr_emin - 1 - lk && mpfr_powerof2_raw (z))
{
/* Rounding to nearest, real result > z * 2^k = 2^(emin - 2),
* underflow case:
* (a) if the precision of z is > 1, we will obtain the correct
* result and exceptions by replacing z by nextabove(z).
* (b) if the precision of z is 1, we first copy z to zcopy of
* precision 2 bits and perform nextabove(zcopy).
*/
if (MPFR_PREC(z) >= 2)
mpfr_nextabove (z);
else
{
mpfr_t zcopy;
mpfr_init2 (zcopy, MPFR_PREC(z) + 1);
mpfr_set (zcopy, z, MPFR_RNDZ);
mpfr_nextabove (zcopy);
inex2 = mpfr_mul_2si (z, zcopy, lk, rnd_mode);
mpfr_clear (zcopy);
goto under_over;
}
}
MPFR_CLEAR_FLAGS ();
inex2 = mpfr_mul_2si (z, z, lk, rnd_mode);
under_over:
if (inex2) /* underflow or overflow */
{
inexact = inex2;
if (expo != NULL)
MPFR_SAVE_EXPO_UPDATE_FLAGS (*expo, __gmpfr_flags);
}
mpfr_clears (u, k, (mpfr_ptr) 0);
}
mpfr_clear (t);
/* update the sign of the result if x was negative */
if (neg_result)
{
MPFR_SET_NEG(z);
inexact = -inexact;
}
return inexact;
}
int
mpfr_atan2 (mpfr_ptr dest, mpfr_srcptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_t tmp, pi;
int inexact;
mpfr_prec_t prec;
mpfr_exp_t e;
MPFR_SAVE_EXPO_DECL (expo);
MPFR_ZIV_DECL (loop);
MPFR_LOG_FUNC
(("y[%Pu]=%.*Rg x[%Pu]=%.*Rg rnd=%d",
mpfr_get_prec (y), mpfr_log_prec, y,
mpfr_get_prec (x), mpfr_log_prec, x, rnd_mode),
("atan[%Pu]=%.*Rg inexact=%d",
mpfr_get_prec (dest), mpfr_log_prec, dest, inexact));
/* Special cases */
if (MPFR_ARE_SINGULAR (x, y))
{
/* atan2(0, 0) does not raise the "invalid" floating-point
exception, nor does atan2(y, 0) raise the "divide-by-zero"
floating-point exception.
-- atan2(±0, -0) returns ±pi.313)
-- atan2(±0, +0) returns ±0.
-- atan2(±0, x) returns ±pi, for x < 0.
-- atan2(±0, x) returns ±0, for x > 0.
-- atan2(y, ±0) returns -pi/2 for y < 0.
-- atan2(y, ±0) returns pi/2 for y > 0.
-- atan2(±oo, -oo) returns ±3pi/4.
-- atan2(±oo, +oo) returns ±pi/4.
-- atan2(±oo, x) returns ±pi/2, for finite x.
-- atan2(±y, -oo) returns ±pi, for finite y > 0.
-- atan2(±y, +oo) returns ±0, for finite y > 0.
*/
if (MPFR_IS_NAN (x) || MPFR_IS_NAN (y))
{
MPFR_SET_NAN (dest);
MPFR_RET_NAN;
}
if (MPFR_IS_ZERO (y))
{
if (MPFR_IS_NEG (x)) /* +/- PI */
{
set_pi:
if (MPFR_IS_NEG (y))
{
inexact = mpfr_const_pi (dest, MPFR_INVERT_RND (rnd_mode));
MPFR_CHANGE_SIGN (dest);
return -inexact;
}
else
return mpfr_const_pi (dest, rnd_mode);
}
else /* +/- 0 */
{
set_zero:
MPFR_SET_ZERO (dest);
MPFR_SET_SAME_SIGN (dest, y);
return 0;
}
}
if (MPFR_IS_ZERO (x))
{
return pi_div_2ui (dest, 1, MPFR_IS_NEG (y), rnd_mode);
}
if (MPFR_IS_INF (y))
{
if (!MPFR_IS_INF (x)) /* +/- PI/2 */
return pi_div_2ui (dest, 1, MPFR_IS_NEG (y), rnd_mode);
else if (MPFR_IS_POS (x)) /* +/- PI/4 */
return pi_div_2ui (dest, 2, MPFR_IS_NEG (y), rnd_mode);
else /* +/- 3*PI/4: Ugly since we have to round properly */
{
mpfr_t tmp2;
MPFR_ZIV_DECL (loop2);
mpfr_prec_t prec2 = MPFR_PREC (dest) + 10;
MPFR_SAVE_EXPO_MARK (expo);
mpfr_init2 (tmp2, prec2);
MPFR_ZIV_INIT (loop2, prec2);
for (;;)
{
mpfr_const_pi (tmp2, MPFR_RNDN);
mpfr_mul_ui (tmp2, tmp2, 3, MPFR_RNDN); /* Error <= 2 */
mpfr_div_2ui (tmp2, tmp2, 2, MPFR_RNDN);
if (mpfr_round_p (MPFR_MANT (tmp2), MPFR_LIMB_SIZE (tmp2),
MPFR_PREC (tmp2) - 2,
MPFR_PREC (dest) + (rnd_mode == MPFR_RNDN)))
break;
MPFR_ZIV_NEXT (loop2, prec2);
mpfr_set_prec (tmp2, prec2);
}
MPFR_ZIV_FREE (loop2);
if (MPFR_IS_NEG (y))
MPFR_CHANGE_SIGN (tmp2);
inexact = mpfr_set (dest, tmp2, rnd_mode);
mpfr_clear (tmp2);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (dest, inexact, rnd_mode);
}
}
MPFR_ASSERTD (MPFR_IS_INF (x));
if (MPFR_IS_NEG (x))
goto set_pi;
else
goto set_zero;
}
/* When x is a power of two, we call directly atan(y/x) since y/x is
exact. */
if (MPFR_UNLIKELY (MPFR_IS_POWER_OF_2 (x)))
{
int r;
mpfr_t yoverx;
unsigned int saved_flags = __gmpfr_flags;
mpfr_init2 (yoverx, MPFR_PREC (y));
if (MPFR_LIKELY (mpfr_div_2si (yoverx, y, MPFR_GET_EXP (x) - 1,
MPFR_RNDN) == 0))
{
/* Here the flags have not changed due to mpfr_div_2si. */
r = mpfr_atan (dest, yoverx, rnd_mode);
mpfr_clear (yoverx);
return r;
}
else
{
/* Division is inexact because of a small exponent range */
mpfr_clear (yoverx);
__gmpfr_flags = saved_flags;
}
}
MPFR_SAVE_EXPO_MARK (expo);
/* Set up initial prec */
prec = MPFR_PREC (dest) + 3 + MPFR_INT_CEIL_LOG2 (MPFR_PREC (dest));
mpfr_init2 (tmp, prec);
MPFR_ZIV_INIT (loop, prec);
if (MPFR_IS_POS (x))
/* use atan2(y,x) = atan(y/x) */
for (;;)
{
int div_inex;
MPFR_BLOCK_DECL (flags);
MPFR_BLOCK (flags, div_inex = mpfr_div (tmp, y, x, MPFR_RNDN));
if (div_inex == 0)
{
/* Result is exact. */
inexact = mpfr_atan (dest, tmp, rnd_mode);
goto end;
}
/* Error <= ulp (tmp) except in case of underflow or overflow. */
/* If the division underflowed, since |atan(z)/z| < 1, we have
an underflow. */
if (MPFR_UNDERFLOW (flags))
{
int sign;
/* In the case MPFR_RNDN with 2^(emin-2) < |y/x| < 2^(emin-1):
The smallest significand value S > 1 of |y/x| is:
* 1 / (1 - 2^(-px)) if py <= px,
* (1 - 2^(-px) + 2^(-py)) / (1 - 2^(-px)) if py >= px.
Therefore S - 1 > 2^(-pz), where pz = max(px,py). We have:
atan(|y/x|) > atan(z), where z = 2^(emin-2) * (1 + 2^(-pz)).
> z - z^3 / 3.
> 2^(emin-2) * (1 + 2^(-pz) - 2^(2 emin - 5))
Assuming pz <= -2 emin + 5, we can round away from zero
(this is what mpfr_underflow always does on MPFR_RNDN).
In the case MPFR_RNDN with |y/x| <= 2^(emin-2), we round
toward zero, as |atan(z)/z| < 1. */
MPFR_ASSERTN (MPFR_PREC_MAX <=
2 * (mpfr_uexp_t) - MPFR_EMIN_MIN + 5);
if (rnd_mode == MPFR_RNDN && MPFR_IS_ZERO (tmp))
rnd_mode = MPFR_RNDZ;
sign = MPFR_SIGN (tmp);
mpfr_clear (tmp);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_underflow (dest, rnd_mode, sign);
}
mpfr_atan (tmp, tmp, MPFR_RNDN); /* Error <= 2*ulp (tmp) since
abs(D(arctan)) <= 1 */
/* TODO: check that the error bound is correct in case of overflow. */
/* FIXME: Error <= ulp(tmp) ? */
if (MPFR_LIKELY (MPFR_CAN_ROUND (tmp, prec - 2, MPFR_PREC (dest),
rnd_mode)))
break;
MPFR_ZIV_NEXT (loop, prec);
mpfr_set_prec (tmp, prec);
}
else /* x < 0 */
/* Use sign(y)*(PI - atan (|y/x|)) */
{
mpfr_init2 (pi, prec);
for (;;)
{
mpfr_div (tmp, y, x, MPFR_RNDN); /* Error <= ulp (tmp) */
/* If tmp is 0, we have |y/x| <= 2^(-emin-2), thus
atan|y/x| < 2^(-emin-2). */
MPFR_SET_POS (tmp); /* no error */
mpfr_atan (tmp, tmp, MPFR_RNDN); /* Error <= 2*ulp (tmp) since
abs(D(arctan)) <= 1 */
mpfr_const_pi (pi, MPFR_RNDN); /* Error <= ulp(pi) /2 */
e = MPFR_NOTZERO(tmp) ? MPFR_GET_EXP (tmp) : __gmpfr_emin - 1;
mpfr_sub (tmp, pi, tmp, MPFR_RNDN); /* see above */
if (MPFR_IS_NEG (y))
MPFR_CHANGE_SIGN (tmp);
/* Error(tmp) <= (1/2+2^(EXP(pi)-EXP(tmp)-1)+2^(e-EXP(tmp)+1))*ulp
<= 2^(MAX (MAX (EXP(PI)-EXP(tmp)-1, e-EXP(tmp)+1),
-1)+2)*ulp(tmp) */
e = MAX (MAX (MPFR_GET_EXP (pi)-MPFR_GET_EXP (tmp) - 1,
e - MPFR_GET_EXP (tmp) + 1), -1) + 2;
if (MPFR_LIKELY (MPFR_CAN_ROUND (tmp, prec - e, MPFR_PREC (dest),
rnd_mode)))
break;
MPFR_ZIV_NEXT (loop, prec);
mpfr_set_prec (tmp, prec);
mpfr_set_prec (pi, prec);
}
mpfr_clear (pi);
}
inexact = mpfr_set (dest, tmp, rnd_mode);
end:
MPFR_ZIV_FREE (loop);
mpfr_clear (tmp);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (dest, inexact, rnd_mode);
}
