/* (y, z) <- (sin(x), cos(x)), return value is 0 iff both results are exact
ie, iff x = 0 */
int
mpfr_sin_cos (mpfr_ptr y, mpfr_ptr z, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_prec_t prec, m;
int neg, reduce;
mpfr_t c, xr;
mpfr_srcptr xx;
mpfr_exp_t err, expx;
int inexy, inexz;
MPFR_ZIV_DECL (loop);
MPFR_SAVE_EXPO_DECL (expo);
MPFR_ASSERTN (y != z);
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x)))
{
if (MPFR_IS_NAN(x) || MPFR_IS_INF(x))
{
MPFR_SET_NAN (y);
MPFR_SET_NAN (z);
MPFR_RET_NAN;
}
else /* x is zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO (x));
MPFR_SET_ZERO (y);
MPFR_SET_SAME_SIGN (y, x);
/* y = 0, thus exact, but z is inexact in case of underflow
or overflow */
inexy = 0; /* y is exact */
inexz = mpfr_set_ui (z, 1, rnd_mode);
return INEX(inexy,inexz);
}
}
MPFR_LOG_FUNC
(("x[%Pu]=%.*Rg rnd=%d", mpfr_get_prec (x), mpfr_log_prec, x, rnd_mode),
("sin[%Pu]=%.*Rg cos[%Pu]=%.*Rg", mpfr_get_prec(y), mpfr_log_prec, y,
mpfr_get_prec (z), mpfr_log_prec, z));
MPFR_SAVE_EXPO_MARK (expo);
prec = MAX (MPFR_PREC (y), MPFR_PREC (z));
m = prec + MPFR_INT_CEIL_LOG2 (prec) + 13;
expx = MPFR_GET_EXP (x);
/* When x is close to 0, say 2^(-k), then there is a cancellation of about
2k bits in 1-cos(x)^2. FIXME: in that case, it would be more efficient
to compute sin(x) directly. VL: This is partly done by using
MPFR_FAST_COMPUTE_IF_SMALL_INPUT from the mpfr_sin and mpfr_cos
functions. Moreover, any overflow on m is avoided. */
if (expx < 0)
{
/* Warning: in case y = x, and the first call to
MPFR_FAST_COMPUTE_IF_SMALL_INPUT succeeds but the second fails,
we will have clobbered the original value of x.
The workaround is to first compute z = cos(x) in that case, since
y and z are different. */
if (y != x)
/* y and x differ, thus we can safely try to compute y first */
{
MPFR_FAST_COMPUTE_IF_SMALL_INPUT (
y, x, -2 * expx, 2, 0, rnd_mode,
{ inexy = _inexact;
goto small_input; });
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
mpfr_grandom (mpfr_ptr rop1, mpfr_ptr rop2, gmp_randstate_t rstate,
mpfr_rnd_t rnd)
{
int inex1, inex2, s1, s2;
mpz_t x, y, xp, yp, t, a, b, s;
mpfr_t sfr, l, r1, r2;
mpfr_prec_t tprec, tprec0;
inex2 = inex1 = 0;
if (rop2 == NULL) /* only one output requested. */
{
tprec0 = MPFR_PREC (rop1);
}
else
{
tprec0 = MAX (MPFR_PREC (rop1), MPFR_PREC (rop2));
}
tprec0 += 11;
/* We use "Marsaglia polar method" here (cf.
George Marsaglia, Normal (Gaussian) random variables for supercomputers
The Journal of Supercomputing, Volume 5, Number 1, 49–55
DOI: 10.1007/BF00155857).
First we draw uniform x and y in [0,1] using mpz_urandomb (in
fixed precision), and scale them to [-1, 1].
*/
mpz_init (xp);
mpz_init (yp);
mpz_init (x);
mpz_init (y);
mpz_init (t);
mpz_init (s);
mpz_init (a);
mpz_init (b);
mpfr_init2 (sfr, MPFR_PREC_MIN);
mpfr_init2 (l, MPFR_PREC_MIN);
mpfr_init2 (r1, MPFR_PREC_MIN);
if (rop2 != NULL)
mpfr_init2 (r2, MPFR_PREC_MIN);
mpz_set_ui (xp, 0);
mpz_set_ui (yp, 0);
for (;;)
{
tprec = tprec0;
do
{
mpz_urandomb (xp, rstate, tprec);
mpz_urandomb (yp, rstate, tprec);
mpz_mul (a, xp, xp);
mpz_mul (b, yp, yp);
mpz_add (s, a, b);
}
while (mpz_sizeinbase (s, 2) > tprec * 2); /* x^2 + y^2 <= 2^{2tprec} */
for (;;)
{
/* FIXME: compute s as s += 2x + 2y + 2 */
mpz_add_ui (a, xp, 1);
mpz_add_ui (b, yp, 1);
mpz_mul (a, a, a);
mpz_mul (b, b, b);
mpz_add (s, a, b);
if ((mpz_sizeinbase (s, 2) <= 2 * tprec) ||
((mpz_sizeinbase (s, 2) == 2 * tprec + 1) &&
(mpz_scan1 (s, 0) == 2 * tprec)))
goto yeepee;
/* Extend by 32 bits */
mpz_mul_2exp (xp, xp, 32);
mpz_mul_2exp (yp, yp, 32);
mpz_urandomb (x, rstate, 32);
mpz_urandomb (y, rstate, 32);
mpz_add (xp, xp, x);
mpz_add (yp, yp, y);
tprec += 32;
mpz_mul (a, xp, xp);
mpz_mul (b, yp, yp);
mpz_add (s, a, b);
if (mpz_sizeinbase (s, 2) > tprec * 2)
break;
}
}
yeepee:
/* FIXME: compute s with s -= 2x + 2y + 2 */
mpz_mul (a, xp, xp);
mpz_mul (b, yp, yp);
mpz_add (s, a, b);
/* Compute the signs of the output */
mpz_urandomb (x, rstate, 2);
s1 = mpz_tstbit (x, 0);
s2 = mpz_tstbit (x, 1);
for (;;)
{
/* s = xp^2 + yp^2 (loop invariant) */
mpfr_set_prec (sfr, 2 * tprec);
mpfr_set_prec (l, tprec);
mpfr_set_z (sfr, s, MPFR_RNDN); /* exact */
mpfr_mul_2si (sfr, sfr, -2 * tprec, MPFR_RNDN); /* exact */
mpfr_log (l, sfr, MPFR_RNDN);
mpfr_neg (l, l, MPFR_RNDN);
mpfr_mul_2si (l, l, 1, MPFR_RNDN);
mpfr_div (l, l, sfr, MPFR_RNDN);
mpfr_sqrt (l, l, MPFR_RNDN);
mpfr_set_prec (r1, tprec);
mpfr_mul_z (r1, l, xp, MPFR_RNDN);
mpfr_div_2ui (r1, r1, tprec, MPFR_RNDN); /* exact */
if (s1)
mpfr_neg (r1, r1, MPFR_RNDN);
if (MPFR_CAN_ROUND (r1, tprec - 2, MPFR_PREC (rop1), rnd))
{
if (rop2 != NULL)
{
mpfr_set_prec (r2, tprec);
mpfr_mul_z (r2, l, yp, MPFR_RNDN);
mpfr_div_2ui (r2, r2, tprec, MPFR_RNDN); /* exact */
if (s2)
mpfr_neg (r2, r2, MPFR_RNDN);
if (MPFR_CAN_ROUND (r2, tprec - 2, MPFR_PREC (rop2), rnd))
break;
}
else
break;
}
/* Extend by 32 bits */
mpz_mul_2exp (xp, xp, 32);
mpz_mul_2exp (yp, yp, 32);
mpz_urandomb (x, rstate, 32);
mpz_urandomb (y, rstate, 32);
mpz_add (xp, xp, x);
mpz_add (yp, yp, y);
tprec += 32;
mpz_mul (a, xp, xp);
mpz_mul (b, yp, yp);
mpz_add (s, a, b);
}
inex1 = mpfr_set (rop1, r1, rnd);
if (rop2 != NULL)
{
inex2 = mpfr_set (rop2, r2, rnd);
inex2 = mpfr_check_range (rop2, inex2, rnd);
}
inex1 = mpfr_check_range (rop1, inex1, rnd);
if (rop2 != NULL)
mpfr_clear (r2);
mpfr_clear (r1);
mpfr_clear (l);
mpfr_clear (sfr);
mpz_clear (b);
mpz_clear (a);
mpz_clear (s);
mpz_clear (t);
mpz_clear (y);
mpz_clear (x);
mpz_clear (yp);
mpz_clear (xp);
return INEX (inex1, inex2);
}
/* Set iop to the integral part of op and fop to its fractional part */
int
mpfr_modf (mpfr_ptr iop, mpfr_ptr fop, mpfr_srcptr op, mpfr_rnd_t rnd_mode)
{
mpfr_exp_t ope;
mpfr_prec_t opq;
int inexi, inexf;
MPFR_LOG_FUNC
(("op[%Pu]=%.*Rg rnd=%d",
mpfr_get_prec (op), mpfr_log_prec, op, rnd_mode),
("iop[%Pu]=%.*Rg fop[%Pu]=%.*Rg",
mpfr_get_prec (iop), mpfr_log_prec, iop,
mpfr_get_prec (fop), mpfr_log_prec, fop));
MPFR_ASSERTN (iop != fop);
if ( MPFR_UNLIKELY (MPFR_IS_SINGULAR (op)) )
{
if (MPFR_IS_NAN (op))
{
MPFR_SET_NAN (iop);
MPFR_SET_NAN (fop);
MPFR_RET_NAN;
}
MPFR_SET_SAME_SIGN (iop, op);
MPFR_SET_SAME_SIGN (fop, op);
if (MPFR_IS_INF (op))
{
MPFR_SET_INF (iop);
MPFR_SET_ZERO (fop);
MPFR_RET (0);
}
else /* op is zero */
{
MPFR_ASSERTD (MPFR_IS_ZERO (op));
MPFR_SET_ZERO (iop);
MPFR_SET_ZERO (fop);
MPFR_RET (0);
}
}
ope = MPFR_GET_EXP (op);
opq = MPFR_PREC (op);
if (ope <= 0) /* 0 < |op| < 1 */
{
inexf = (fop != op) ? mpfr_set (fop, op, rnd_mode) : 0;
MPFR_SET_SAME_SIGN (iop, op);
MPFR_SET_ZERO (iop);
MPFR_RET (INEX(0, inexf));
}
else if (ope >= opq) /* op has no fractional part */
{
inexi = (iop != op) ? mpfr_set (iop, op, rnd_mode) : 0;
MPFR_SET_SAME_SIGN (fop, op);
MPFR_SET_ZERO (fop);
MPFR_RET (INEX(inexi, 0));
}
else /* op has both integral and fractional parts */
{
if (iop != op)
{
inexi = mpfr_rint_trunc (iop, op, rnd_mode);
inexf = mpfr_frac (fop, op, rnd_mode);
}
else
{
MPFR_ASSERTN (fop != op);
inexf = mpfr_frac (fop, op, rnd_mode);
inexi = mpfr_rint_trunc (iop, op, rnd_mode);
}
MPFR_RET (INEX(inexi, inexf));
}
}
