/* fix the sign of Re(z) or Im(z) in case it is zero,
and Re(x) is zero.
sign_eps is 0 if Re(x) = +0, 1 if Re(x) = -0
sign_a is the sign bit of Im(x).
Assume y is an integer (does nothing otherwise).
*/
static void
fix_sign (mpc_ptr z, int sign_eps, int sign_a, mpfr_srcptr y)
{
int ymod4 = -1;
mpfr_exp_t ey;
mpz_t my;
unsigned long int t;
mpz_init (my);
ey = mpfr_get_z_exp (my, y);
/* normalize so that my is odd */
t = mpz_scan1 (my, 0);
ey += (mpfr_exp_t) t;
mpz_tdiv_q_2exp (my, my, t);
/* y = my*2^ey */
/* compute y mod 4 (in case y is an integer) */
if (ey >= 2)
ymod4 = 0;
else if (ey == 1)
ymod4 = mpz_tstbit (my, 0) * 2; /* correct if my < 0 */
else if (ey == 0)
{
ymod4 = mpz_tstbit (my, 1) * 2 + mpz_tstbit (my, 0);
if (mpz_cmp_ui (my , 0) < 0)
ymod4 = 4 - ymod4;
}
else /* y is not an integer */
goto end;
if (mpfr_zero_p (mpc_realref(z)))
{
/* we assume y is always integer in that case (FIXME: prove it):
(eps+I*a)^y = +0 + I*a^y for y = 1 mod 4 and sign_eps = 0
(eps+I*a)^y = -0 - I*a^y for y = 3 mod 4 and sign_eps = 0 */
MPC_ASSERT (ymod4 == 1 || ymod4 == 3);
if ((ymod4 == 3 && sign_eps == 0) ||
(ymod4 == 1 && sign_eps == 1))
mpfr_neg (mpc_realref(z), mpc_realref(z), MPFR_RNDZ);
}
else if (mpfr_zero_p (mpc_imagref(z)))
{
/* we assume y is always integer in that case (FIXME: prove it):
(eps+I*a)^y = a^y - 0*I for y = 0 mod 4 and sign_a = sign_eps
(eps+I*a)^y = -a^y +0*I for y = 2 mod 4 and sign_a = sign_eps */
MPC_ASSERT (ymod4 == 0 || ymod4 == 2);
if ((ymod4 == 0 && sign_a == sign_eps) ||
(ymod4 == 2 && sign_a != sign_eps))
mpfr_neg (mpc_imagref(z), mpc_imagref(z), MPFR_RNDZ);
}
end:
mpz_clear (my);
}
/* return non zero iff x^y is exact.
Assumes x and y are ordinary numbers (neither NaN nor Inf),
and y is not zero.
*/
int
mpfr_pow_is_exact (mpfr_srcptr x, mpfr_srcptr y)
{
mp_exp_t d;
unsigned long i, c;
mp_limb_t *yp;
if ((mpfr_sgn (x) < 0) && (mpfr_isinteger (y) == 0))
return 0;
if (mpfr_sgn (y) < 0)
return mpfr_cmp_si_2exp (x, MPFR_SIGN(x), MPFR_EXP(x) - 1) == 0;
/* compute d such that y = c*2^d with c odd integer */
d = MPFR_EXP(y) - MPFR_PREC(y);
/* since y is not zero, necessarily one of the mantissa limbs is not zero,
thus we can simply loop until we find a non zero limb */
yp = MPFR_MANT(y);
for (i = 0; yp[i] == 0; i++, d += BITS_PER_MP_LIMB);
/* now yp[i] is not zero */
count_trailing_zeros (c, yp[i]);
d += c;
if (d < 0)
{
mpz_t a;
mp_exp_t b;
mpz_init (a);
b = mpfr_get_z_exp (a, x); /* x = a * 2^b */
c = mpz_scan1 (a, 0);
mpz_div_2exp (a, a, c);
b += c;
/* now a is odd */
while (d != 0)
{
if (mpz_perfect_square_p (a))
{
d++;
mpz_sqrt (a, a);
}
else
{
mpz_clear (a);
return 0;
}
}
mpz_clear (a);
}
return 1;
}
void
mpfr_get_z (mpz_ptr z, mpfr_srcptr f, mp_rnd_t rnd)
{
mpfr_t r;
mp_exp_t exp = MPFR_EXP (f);
/* if exp <= 0, then |f|<1, thus |o(f)|<=1 */
MPFR_ASSERTN (exp < 0 || exp <= MPFR_PREC_MAX);
mpfr_init2 (r, (exp < (mp_exp_t) MPFR_PREC_MIN ?
MPFR_PREC_MIN : (mpfr_prec_t) exp));
mpfr_rint (r, f, rnd);
MPFR_ASSERTN (MPFR_IS_FP (r) );
exp = mpfr_get_z_exp (z, r);
if (exp >= 0)
mpz_mul_2exp (z, z, exp);
else
mpz_div_2exp (z, z, -exp);
mpfr_clear (r);
}
/* s <- 1 + r/1! + r^2/2! + ... + r^l/l! while MPFR_EXP(r^l/l!)+MPFR_EXPR(r)>-q
using naive method with O(l) multiplications.
Return the number of iterations l.
The absolute error on s is less than 3*l*(l+1)*2^(-q).
Version using fixed-point arithmetic with mpz instead
of mpfr for internal computations.
s must have at least qn+1 limbs (qn should be enough, but currently fails
since mpz_mul_2exp(s, s, q-1) reallocates qn+1 limbs)
*/
static unsigned long
mpfr_exp2_aux (mpz_t s, mpfr_srcptr r, mp_prec_t q, mp_exp_t *exps)
{
unsigned long l;
mp_exp_t dif;
mp_size_t qn;
mpz_t t, rr;
mp_exp_t expt, expr;
TMP_DECL(marker);
TMP_MARK(marker);
qn = 1 + (q-1)/BITS_PER_MP_LIMB;
expt = 0;
*exps = 1 - (mp_exp_t) q; /* s = 2^(q-1) */
MY_INIT_MPZ(t, 2*qn+1);
MY_INIT_MPZ(rr, qn+1);
mpz_set_ui(t, 1);
mpz_set_ui(s, 1);
mpz_mul_2exp(s, s, q-1);
expr = mpfr_get_z_exp(rr, r); /* no error here */
l = 0;
do {
l++;
mpz_mul(t, t, rr);
expt += expr;
dif = *exps + mpz_sizeinbase(s, 2) - expt - mpz_sizeinbase(t, 2);
/* truncates the bits of t which are < ulp(s) = 2^(1-q) */
expt += mpz_normalize(t, t, (mp_exp_t) q-dif); /* error at most 2^(1-q) */
mpz_div_ui(t, t, l); /* error at most 2^(1-q) */
/* the error wrt t^l/l! is here at most 3*l*ulp(s) */
MPFR_ASSERTD (expt == *exps);
mpz_add(s, s, t); /* no error here: exact */
/* ensures rr has the same size as t: after several shifts, the error
on rr is still at most ulp(t)=ulp(s) */
expr += mpz_normalize(rr, rr, mpz_sizeinbase(t, 2));
} while (mpz_cmp_ui(t, 0));
TMP_FREE(marker);
return l;
}
/* s <- 1 + r/1! + r^2/2! + ... + r^l/l! while MPFR_EXP(r^l/l!)+MPFR_EXPR(r)>-q
using naive method with O(l) multiplications.
Return the number of iterations l.
The absolute error on s is less than 3*l*(l+1)*2^(-q).
Version using fixed-point arithmetic with mpz instead
of mpfr for internal computations.
s must have at least qn+1 limbs (qn should be enough, but currently fails
since mpz_mul_2exp(s, s, q-1) reallocates qn+1 limbs)
*/
static int
mpfr_exp2_aux (mpz_t s, mpfr_srcptr r, int q, int *exps)
{
int l, dif, qn;
mpz_t t, rr; mp_exp_t expt, expr;
TMP_DECL(marker);
TMP_MARK(marker);
qn = 1 + (q-1)/BITS_PER_MP_LIMB;
MY_INIT_MPZ(t, 2*qn+1); /* 2*qn+1 is neeeded since mpz_div_2exp may
need an extra limb */
MY_INIT_MPZ(rr, qn+1);
mpz_set_ui(t, 1); expt=0;
mpz_set_ui(s, 1); mpz_mul_2exp(s, s, q-1); *exps = 1-q; /* s = 2^(q-1) */
expr = mpfr_get_z_exp(rr, r); /* no error here */
l = 0;
do {
l++;
mpz_mul(t, t, rr); expt=expt+expr;
dif = *exps + mpz_sizeinbase(s, 2) - expt - mpz_sizeinbase(t, 2);
/* truncates the bits of t which are < ulp(s) = 2^(1-q) */
expt += mpz_normalize(t, t, q-dif); /* error at most 2^(1-q) */
mpz_div_ui(t, t, l); /* error at most 2^(1-q) */
/* the error wrt t^l/l! is here at most 3*l*ulp(s) */
#ifdef DEBUG
if (expt != *exps) {
fprintf(stderr, "Error: expt != exps %d %d\n", expt, *exps); exit(1);
}
#endif
mpz_add(s, s, t); /* no error here: exact */
/* ensures rr has the same size as t: after several shifts, the error
on rr is still at most ulp(t)=ulp(s) */
expr += mpz_normalize(rr, rr, mpz_sizeinbase(t, 2));
} while (mpz_cmp_ui(t, 0));
TMP_FREE(marker);
return l;
}
/* If x^y is exactly representable (with maybe a larger precision than z),
round it in z and return the (mpc) inexact flag in [0, 10].
If x^y is not exactly representable, return -1.
If intermediate computations lead to numbers of more than maxprec bits,
then abort and return -2 (in that case, to avoid loops, mpc_pow_exact
should be called again with a larger value of maxprec).
Assume one of Re(x) or Im(x) is non-zero, and y is non-zero (y is real).
Warning: z and x might be the same variable, same for Re(z) or Im(z) and y.
In case -1 or -2 is returned, z is not modified.
*/
static int
mpc_pow_exact (mpc_ptr z, mpc_srcptr x, mpfr_srcptr y, mpc_rnd_t rnd,
mpfr_prec_t maxprec)
{
mpfr_exp_t ec, ed, ey;
mpz_t my, a, b, c, d, u;
unsigned long int t;
int ret = -2;
int sign_rex = mpfr_signbit (mpc_realref(x));
int sign_imx = mpfr_signbit (mpc_imagref(x));
int x_imag = mpfr_zero_p (mpc_realref(x));
int z_is_y = 0;
mpfr_t copy_of_y;
if (mpc_realref (z) == y || mpc_imagref (z) == y)
{
z_is_y = 1;
mpfr_init2 (copy_of_y, mpfr_get_prec (y));
mpfr_set (copy_of_y, y, MPFR_RNDN);
}
mpz_init (my);
mpz_init (a);
mpz_init (b);
mpz_init (c);
mpz_init (d);
mpz_init (u);
ey = mpfr_get_z_exp (my, y);
/* normalize so that my is odd */
t = mpz_scan1 (my, 0);
ey += (mpfr_exp_t) t;
mpz_tdiv_q_2exp (my, my, t);
/* y = my*2^ey with my odd */
if (x_imag)
{
mpz_set_ui (c, 0);
ec = 0;
}
else
ec = mpfr_get_z_exp (c, mpc_realref(x));
if (mpfr_zero_p (mpc_imagref(x)))
{
mpz_set_ui (d, 0);
ed = ec;
}
else
{
ed = mpfr_get_z_exp (d, mpc_imagref(x));
if (x_imag)
ec = ed;
}
/* x = c*2^ec + I * d*2^ed */
/* equalize the exponents of x */
if (ec < ed)
{
mpz_mul_2exp (d, d, (unsigned long int) (ed - ec));
if ((mpfr_prec_t) mpz_sizeinbase (d, 2) > maxprec)
goto end;
}
else if (ed < ec)
{
mpz_mul_2exp (c, c, (unsigned long int) (ec - ed));
if ((mpfr_prec_t) mpz_sizeinbase (c, 2) > maxprec)
goto end;
ec = ed;
}
/* now ec=ed and x = (c + I * d) * 2^ec */
/* divide by two if possible */
if (mpz_cmp_ui (c, 0) == 0)
{
t = mpz_scan1 (d, 0);
mpz_tdiv_q_2exp (d, d, t);
ec += (mpfr_exp_t) t;
}
else if (mpz_cmp_ui (d, 0) == 0)
{
t = mpz_scan1 (c, 0);
mpz_tdiv_q_2exp (c, c, t);
ec += (mpfr_exp_t) t;
}
else /* neither c nor d is zero */
{
unsigned long v;
t = mpz_scan1 (c, 0);
v = mpz_scan1 (d, 0);
if (v < t)
t = v;
mpz_tdiv_q_2exp (c, c, t);
mpz_tdiv_q_2exp (d, d, t);
ec += (mpfr_exp_t) t;
}
/* now either one of c, d is odd */
while (ey < 0)
{
/* check if x is a square */
if (ec & 1)
{
mpz_mul_2exp (c, c, 1);
mpz_mul_2exp (d, d, 1);
ec --;
}
/* now ec is even */
if (mpc_perfect_square_p (a, b, c, d) == 0)
break;
mpz_swap (a, c);
mpz_swap (b, d);
ec /= 2;
ey ++;
}
if (ey < 0)
{
ret = -1; /* not representable */
goto end;
}
/* Now ey >= 0, it thus suffices to check that x^my is representable.
If my > 0, this is always true. If my < 0, we first try to invert
(c+I*d)*2^ec.
*/
if (mpz_cmp_ui (my, 0) < 0)
{
/* If my < 0, 1 / (c + I*d) = (c - I*d)/(c^2 + d^2), thus a sufficient
condition is that c^2 + d^2 is a power of two, assuming |c| <> |d|.
Assume a prime p <> 2 divides c^2 + d^2,
then if p does not divide c or d, 1 / (c + I*d) cannot be exact.
If p divides both c and d, then we can write c = p*c', d = p*d',
and 1 / (c + I*d) = 1/p * 1/(c' + I*d'). This shows that if 1/(c+I*d)
is exact, then 1/(c' + I*d') is exact too, and we are back to the
previous case. In conclusion, a necessary and sufficient condition
is that c^2 + d^2 is a power of two.
*/
/* FIXME: we could first compute c^2+d^2 mod a limb for example */
mpz_mul (a, c, c);
mpz_addmul (a, d, d);
t = mpz_scan1 (a, 0);
if (mpz_sizeinbase (a, 2) != 1 + t) /* a is not a power of two */
{
ret = -1; /* not representable */
goto end;
}
/* replace (c,d) by (c/(c^2+d^2), -d/(c^2+d^2)) */
mpz_neg (d, d);
ec = -ec - (mpfr_exp_t) t;
mpz_neg (my, my);
}
/* now ey >= 0 and my >= 0, and we want to compute
[(c + I * d) * 2^ec] ^ (my * 2^ey).
We first compute [(c + I * d) * 2^ec]^my, then square ey times. */
t = mpz_sizeinbase (my, 2) - 1;
mpz_set (a, c);
mpz_set (b, d);
ed = ec;
/* invariant: (a + I*b) * 2^ed = ((c + I*d) * 2^ec)^trunc(my/2^t) */
while (t-- > 0)
{
unsigned long int v, w;
/* square a + I*b */
mpz_mul (u, a, b);
mpz_mul (a, a, a);
mpz_submul (a, b, b);
mpz_mul_2exp (b, u, 1);
ed *= 2;
if (mpz_tstbit (my, t)) /* multiply by c + I*d */
{
mpz_mul (u, a, c);
mpz_submul (u, b, d); /* ac-bd */
mpz_mul (b, b, c);
mpz_addmul (b, a, d); /* bc+ad */
mpz_swap (a, u);
ed += ec;
}
/* remove powers of two in (a,b) */
if (mpz_cmp_ui (a, 0) == 0)
{
w = mpz_scan1 (b, 0);
mpz_tdiv_q_2exp (b, b, w);
ed += (mpfr_exp_t) w;
}
else if (mpz_cmp_ui (b, 0) == 0)
{
w = mpz_scan1 (a, 0);
mpz_tdiv_q_2exp (a, a, w);
ed += (mpfr_exp_t) w;
}
else
{
w = mpz_scan1 (a, 0);
v = mpz_scan1 (b, 0);
if (v < w)
w = v;
mpz_tdiv_q_2exp (a, a, w);
mpz_tdiv_q_2exp (b, b, w);
ed += (mpfr_exp_t) w;
}
if ( (mpfr_prec_t) mpz_sizeinbase (a, 2) > maxprec
|| (mpfr_prec_t) mpz_sizeinbase (b, 2) > maxprec)
goto end;
}
/* now a+I*b = (c+I*d)^my */
while (ey-- > 0)
{
unsigned long sa, sb;
/* square a + I*b */
mpz_mul (u, a, b);
mpz_mul (a, a, a);
mpz_submul (a, b, b);
mpz_mul_2exp (b, u, 1);
ed *= 2;
/* divide by largest 2^n possible, to avoid many loops for e.g.,
(2+2*I)^16777216 */
sa = mpz_scan1 (a, 0);
sb = mpz_scan1 (b, 0);
sa = (sa <= sb) ? sa : sb;
mpz_tdiv_q_2exp (a, a, sa);
mpz_tdiv_q_2exp (b, b, sa);
ed += (mpfr_exp_t) sa;
if ( (mpfr_prec_t) mpz_sizeinbase (a, 2) > maxprec
|| (mpfr_prec_t) mpz_sizeinbase (b, 2) > maxprec)
goto end;
}
ret = mpfr_set_z (mpc_realref(z), a, MPC_RND_RE(rnd));
ret = MPC_INEX(ret, mpfr_set_z (mpc_imagref(z), b, MPC_RND_IM(rnd)));
mpfr_mul_2si (mpc_realref(z), mpc_realref(z), ed, MPC_RND_RE(rnd));
mpfr_mul_2si (mpc_imagref(z), mpc_imagref(z), ed, MPC_RND_IM(rnd));
end:
mpz_clear (my);
mpz_clear (a);
mpz_clear (b);
mpz_clear (c);
mpz_clear (d);
mpz_clear (u);
if (ret >= 0 && x_imag)
fix_sign (z, sign_rex, sign_imx, (z_is_y) ? copy_of_y : y);
if (z_is_y)
mpfr_clear (copy_of_y);
return ret;
}
/* use Brent's formula exp(x) = (1+r+r^2/2!+r^3/3!+...)^(2^K)*2^n
where x = n*log(2)+(2^K)*r
together with Brent-Kung O(t^(1/2)) algorithm for the evaluation of
power series. The resulting complexity is O(n^(1/3)*M(n)).
*/
int
mpfr_exp_2 (mpfr_ptr y, mpfr_srcptr x, mp_rnd_t rnd_mode)
{
long n;
unsigned long K, k, l, err; /* FIXME: Which type ? */
int error_r;
mp_exp_t exps;
mp_prec_t q, precy;
int inexact;
mpfr_t r, s, t;
mpz_t ss;
TMP_DECL(marker);
precy = MPFR_PREC(y);
MPFR_TRACE ( printf("Py=%d Px=%d", MPFR_PREC(y), MPFR_PREC(x)) );
MPFR_TRACE ( MPFR_DUMP (x) );
n = (long) (mpfr_get_d1 (x) / LOG2);
/* error bounds the cancelled bits in x - n*log(2) */
if (MPFR_UNLIKELY(n == 0))
error_r = 0;
else
count_leading_zeros (error_r, (mp_limb_t) (n < 0) ? -n : n);
error_r = BITS_PER_MP_LIMB - error_r + 2;
/* for the O(n^(1/2)*M(n)) method, the Taylor series computation of
n/K terms costs about n/(2K) multiplications when computed in fixed
point */
K = (precy < SWITCH) ? __gmpfr_isqrt ((precy + 1) / 2)
: __gmpfr_cuberoot (4*precy);
l = (precy - 1) / K + 1;
err = K + MPFR_INT_CEIL_LOG2 (2 * l + 18);
/* add K extra bits, i.e. failure probability <= 1/2^K = O(1/precy) */
q = precy + err + K + 5;
/*q = ( (q-1)/BITS_PER_MP_LIMB + 1) * BITS_PER_MP_LIMB; */
mpfr_init2 (r, q + error_r);
mpfr_init2 (s, q + error_r);
mpfr_init2 (t, q);
/* the algorithm consists in computing an upper bound of exp(x) using
a precision of q bits, and see if we can round to MPFR_PREC(y) taking
into account the maximal error. Otherwise we increase q. */
for (;;)
{
MPFR_TRACE ( printf("n=%d K=%d l=%d q=%d\n",n,K,l,q) );
/* if n<0, we have to get an upper bound of log(2)
in order to get an upper bound of r = x-n*log(2) */
mpfr_const_log2 (s, (n >= 0) ? GMP_RNDZ : GMP_RNDU);
/* s is within 1 ulp of log(2) */
mpfr_mul_ui (r, s, (n < 0) ? -n : n, (n >= 0) ? GMP_RNDZ : GMP_RNDU);
/* r is within 3 ulps of n*log(2) */
if (n < 0)
mpfr_neg (r, r, GMP_RNDD); /* exact */
/* r = floor(n*log(2)), within 3 ulps */
MPFR_TRACE ( MPFR_DUMP (x) );
MPFR_TRACE ( MPFR_DUMP (r) );
mpfr_sub (r, x, r, GMP_RNDU);
/* possible cancellation here: the error on r is at most
3*2^(EXP(old_r)-EXP(new_r)) */
while (MPFR_IS_NEG (r))
{ /* initial approximation n was too large */
n--;
mpfr_add (r, r, s, GMP_RNDU);
}
mpfr_prec_round (r, q, GMP_RNDU);
MPFR_TRACE ( MPFR_DUMP (r) );
MPFR_ASSERTD (MPFR_IS_POS (r));
mpfr_div_2ui (r, r, K, GMP_RNDU); /* r = (x-n*log(2))/2^K, exact */
TMP_MARK(marker);
MY_INIT_MPZ(ss, 3 + 2*((q-1)/BITS_PER_MP_LIMB));
exps = mpfr_get_z_exp (ss, s);
/* s <- 1 + r/1! + r^2/2! + ... + r^l/l! */
l = (precy < SWITCH) ?
mpfr_exp2_aux (ss, r, q, &exps) /* naive method */
: mpfr_exp2_aux2 (ss, r, q, &exps); /* Brent/Kung method */
MPFR_TRACE(printf("l=%d q=%d (K+l)*q^2=%1.3e\n", l, q, (K+l)*(double)q*q));
for (k = 0; k < K; k++)
{
mpz_mul (ss, ss, ss);
exps <<= 1;
exps += mpz_normalize (ss, ss, q);
}
mpfr_set_z (s, ss, GMP_RNDN);
MPFR_SET_EXP(s, MPFR_GET_EXP (s) + exps);
TMP_FREE(marker); /* don't need ss anymore */
if (n>0)
mpfr_mul_2ui(s, s, n, GMP_RNDU);
else
mpfr_div_2ui(s, s, -n, GMP_RNDU);
/* error is at most 2^K*(3l*(l+1)) ulp for mpfr_exp2_aux */
l = (precy < SWITCH) ? 3*l*(l+1) : l*(l+4) ;
k = MPFR_INT_CEIL_LOG2 (l);
/* k = 0; while (l) { k++; l >>= 1; } */
/* now k = ceil(log(error in ulps)/log(2)) */
K += k;
MPFR_TRACE ( printf("after mult. by 2^n:\n") );
MPFR_TRACE ( MPFR_DUMP (s) );
MPFR_TRACE ( printf("err=%d bits\n", K) );
if (mpfr_can_round (s, q - K, GMP_RNDN, GMP_RNDZ,
precy + (rnd_mode == GMP_RNDN)) )
break;
MPFR_TRACE (printf("prec++, use %d\n", q+BITS_PER_MP_LIMB) );
MPFR_TRACE (printf("q=%d q-K=%d precy=%d\n",q,q-K,precy) );
q += BITS_PER_MP_LIMB;
mpfr_set_prec (r, q);
mpfr_set_prec (s, q);
mpfr_set_prec (t, q);
}
inexact = mpfr_set (y, s, rnd_mode);
mpfr_clear (r);
mpfr_clear (s);
mpfr_clear (t);
return inexact;
}
/* s <- 1 + r/1! + r^2/2! + ... + r^l/l! while MPFR_EXP(r^l/l!)+MPFR_EXPR(r)>-q
using Brent/Kung method with O(sqrt(l)) multiplications.
Return l.
Uses m multiplications of full size and 2l/m of decreasing size,
i.e. a total equivalent to about m+l/m full multiplications,
i.e. 2*sqrt(l) for m=sqrt(l).
Version using mpz. ss must have at least (sizer+1) limbs.
The error is bounded by (l^2+4*l) ulps where l is the return value.
*/
static unsigned long
mpfr_exp2_aux2 (mpz_t s, mpfr_srcptr r, mp_prec_t q, mp_exp_t *exps)
{
mp_exp_t expr, *expR, expt;
mp_size_t sizer;
mp_prec_t ql;
unsigned long l, m, i;
mpz_t t, *R, rr, tmp;
TMP_DECL(marker);
/* estimate value of l */
MPFR_ASSERTD (MPFR_GET_EXP (r) < 0);
l = q / (- MPFR_GET_EXP (r));
m = __gmpfr_isqrt (l);
/* we access R[2], thus we need m >= 2 */
if (m < 2)
m = 2;
TMP_MARK(marker);
R = (mpz_t*) TMP_ALLOC((m+1)*sizeof(mpz_t)); /* R[i] is r^i */
expR = (mp_exp_t*) TMP_ALLOC((m+1)*sizeof(mp_exp_t)); /* exponent for R[i] */
sizer = 1 + (MPFR_PREC(r)-1)/BITS_PER_MP_LIMB;
mpz_init(tmp);
MY_INIT_MPZ(rr, sizer+2);
MY_INIT_MPZ(t, 2*sizer); /* double size for products */
mpz_set_ui(s, 0);
*exps = 1-q; /* 1 ulp = 2^(1-q) */
for (i = 0 ; i <= m ; i++)
MY_INIT_MPZ(R[i], sizer+2);
expR[1] = mpfr_get_z_exp(R[1], r); /* exact operation: no error */
expR[1] = mpz_normalize2(R[1], R[1], expR[1], 1-q); /* error <= 1 ulp */
mpz_mul(t, R[1], R[1]); /* err(t) <= 2 ulps */
mpz_div_2exp(R[2], t, q-1); /* err(R[2]) <= 3 ulps */
expR[2] = 1-q;
for (i = 3 ; i <= m ; i++)
{
mpz_mul(t, R[i-1], R[1]); /* err(t) <= 2*i-2 */
mpz_div_2exp(R[i], t, q-1); /* err(R[i]) <= 2*i-1 ulps */
expR[i] = 1-q;
}
mpz_set_ui (R[0], 1);
mpz_mul_2exp (R[0], R[0], q-1);
expR[0] = 1-q; /* R[0]=1 */
mpz_set_ui (rr, 1);
expr = 0; /* rr contains r^l/l! */
/* by induction: err(rr) <= 2*l ulps */
l = 0;
ql = q; /* precision used for current giant step */
do
{
/* all R[i] must have exponent 1-ql */
if (l != 0)
for (i = 0 ; i < m ; i++)
expR[i] = mpz_normalize2 (R[i], R[i], expR[i], 1-ql);
/* the absolute error on R[i]*rr is still 2*i-1 ulps */
expt = mpz_normalize2 (t, R[m-1], expR[m-1], 1-ql);
/* err(t) <= 2*m-1 ulps */
/* computes t = 1 + r/(l+1) + ... + r^(m-1)*l!/(l+m-1)!
using Horner's scheme */
for (i = m-1 ; i-- != 0 ; )
{
mpz_div_ui(t, t, l+i+1); /* err(t) += 1 ulp */
mpz_add(t, t, R[i]);
}
/* now err(t) <= (3m-2) ulps */
/* now multiplies t by r^l/l! and adds to s */
mpz_mul(t, t, rr);
expt += expr;
expt = mpz_normalize2(t, t, expt, *exps);
/* err(t) <= (3m-1) + err_rr(l) <= (3m-2) + 2*l */
MPFR_ASSERTD (expt == *exps);
mpz_add(s, s, t); /* no error here */
/* updates rr, the multiplication of the factors l+i could be done
using binary splitting too, but it is not sure it would save much */
mpz_mul(t, rr, R[m]); /* err(t) <= err(rr) + 2m-1 */
expr += expR[m];
mpz_set_ui (tmp, 1);
for (i = 1 ; i <= m ; i++)
mpz_mul_ui (tmp, tmp, l + i);
mpz_fdiv_q(t, t, tmp); /* err(t) <= err(rr) + 2m */
expr += mpz_normalize(rr, t, ql); /* err_rr(l+1) <= err_rr(l) + 2m+1 */
ql = q - *exps - mpz_sizeinbase(s, 2) + expr + mpz_sizeinbase(rr, 2);
l += m;
}
while ((size_t) expr+mpz_sizeinbase(rr, 2) > (size_t)((int)-q));
TMP_FREE(marker);
mpz_clear(tmp);
return l;
}
/* compute in y an approximation of sum(x^k/k/k!, k=1..infinity),
and return e such that the absolute error is bound by 2^e ulp(y) */
static mp_exp_t
mpfr_eint_aux (mpfr_t y, mpfr_srcptr x)
{
mpfr_t eps; /* dynamic (absolute) error bound on t */
mpfr_t erru, errs;
mpz_t m, s, t, u;
mp_exp_t e, sizeinbase;
mp_prec_t w = MPFR_PREC(y);
unsigned long k;
MPFR_GROUP_DECL (group);
/* for |x| <= 1, we have S := sum(x^k/k/k!, k=1..infinity) = x + R(x)
where |R(x)| <= (x/2)^2/(1-x/2) <= 2*(x/2)^2
thus |R(x)/x| <= |x|/2
thus if |x| <= 2^(-PREC(y)) we have |S - o(x)| <= ulp(y) */
if (MPFR_GET_EXP(x) <= - (mp_exp_t) w)
{
mpfr_set (y, x, GMP_RNDN);
return 0;
}
mpz_init (s); /* initializes to 0 */
mpz_init (t);
mpz_init (u);
mpz_init (m);
MPFR_GROUP_INIT_3 (group, 31, eps, erru, errs);
e = mpfr_get_z_exp (m, x); /* x = m * 2^e */
MPFR_ASSERTD (mpz_sizeinbase (m, 2) == MPFR_PREC (x));
if (MPFR_PREC (x) > w)
{
e += MPFR_PREC (x) - w;
mpz_tdiv_q_2exp (m, m, MPFR_PREC (x) - w);
}
/* remove trailing zeroes from m: this will speed up much cases where
x is a small integer divided by a power of 2 */
k = mpz_scan1 (m, 0);
mpz_tdiv_q_2exp (m, m, k);
e += k;
/* initialize t to 2^w */
mpz_set_ui (t, 1);
mpz_mul_2exp (t, t, w);
mpfr_set_ui (eps, 0, GMP_RNDN); /* eps[0] = 0 */
mpfr_set_ui (errs, 0, GMP_RNDN);
for (k = 1;; k++)
{
/* let eps[k] be the absolute error on t[k]:
since t[k] = trunc(t[k-1]*m*2^e/k), we have
eps[k+1] <= 1 + eps[k-1]*m*2^e/k + t[k-1]*m*2^(1-w)*2^e/k
= 1 + (eps[k-1] + t[k-1]*2^(1-w))*m*2^e/k
= 1 + (eps[k-1]*2^(w-1) + t[k-1])*2^(1-w)*m*2^e/k */
mpfr_mul_2ui (eps, eps, w - 1, GMP_RNDU);
mpfr_add_z (eps, eps, t, GMP_RNDU);
MPFR_MPZ_SIZEINBASE2 (sizeinbase, m);
mpfr_mul_2si (eps, eps, sizeinbase - (w - 1) + e, GMP_RNDU);
mpfr_div_ui (eps, eps, k, GMP_RNDU);
mpfr_add_ui (eps, eps, 1, GMP_RNDU);
mpz_mul (t, t, m);
if (e < 0)
mpz_tdiv_q_2exp (t, t, -e);
else
mpz_mul_2exp (t, t, e);
mpz_tdiv_q_ui (t, t, k);
mpz_tdiv_q_ui (u, t, k);
mpz_add (s, s, u);
/* the absolute error on u is <= 1 + eps[k]/k */
mpfr_div_ui (erru, eps, k, GMP_RNDU);
mpfr_add_ui (erru, erru, 1, GMP_RNDU);
/* and that on s is the sum of all errors on u */
mpfr_add (errs, errs, erru, GMP_RNDU);
/* we are done when t is smaller than errs */
if (mpz_sgn (t) == 0)
sizeinbase = 0;
else
MPFR_MPZ_SIZEINBASE2 (sizeinbase, t);
if (sizeinbase < MPFR_GET_EXP (errs))
break;
}
/* the truncation error is bounded by (|t|+eps)/k*(|x|/k + |x|^2/k^2 + ...)
<= (|t|+eps)/k*|x|/(k-|x|) */
mpz_abs (t, t);
mpfr_add_z (eps, eps, t, GMP_RNDU);
mpfr_div_ui (eps, eps, k, GMP_RNDU);
mpfr_abs (erru, x, GMP_RNDU); /* |x| */
mpfr_mul (eps, eps, erru, GMP_RNDU);
mpfr_ui_sub (erru, k, erru, GMP_RNDD);
if (MPFR_IS_NEG (erru))
{
/* the truncated series does not converge, return fail */
e = w;
}
else
{
mpfr_div (eps, eps, erru, GMP_RNDU);
mpfr_add (errs, errs, eps, GMP_RNDU);
mpfr_set_z (y, s, GMP_RNDN);
mpfr_div_2ui (y, y, w, GMP_RNDN);
/* errs was an absolute error bound on s. We must convert it to an error
in terms of ulp(y). Since ulp(y) = 2^(EXP(y)-PREC(y)), we must
divide the error by 2^(EXP(y)-PREC(y)), but since we divided also
y by 2^w = 2^PREC(y), we must simply divide by 2^EXP(y). */
e = MPFR_GET_EXP (errs) - MPFR_GET_EXP (y);
}
MPFR_GROUP_CLEAR (group);
mpz_clear (s);
mpz_clear (t);
mpz_clear (u);
mpz_clear (m);
return e;
}
int
mpfr_cbrt (mpfr_ptr y, mpfr_srcptr x, mp_rnd_t rnd_mode)
{
mpz_t m;
mp_exp_t e, r, sh;
mp_prec_t n, size_m, tmp;
int inexact, negative;
MPFR_SAVE_EXPO_DECL (expo);
/* special values */
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x)))
{
if (MPFR_IS_NAN (x))
{
MPFR_SET_NAN (y);
MPFR_RET_NAN;
}
else if (MPFR_IS_INF (x))
{
MPFR_SET_INF (y);
MPFR_SET_SAME_SIGN (y, x);
MPFR_RET (0);
}
/* case 0: cbrt(+/- 0) = +/- 0 */
else /* x is necessarily 0 */
{
MPFR_ASSERTD (MPFR_IS_ZERO (x));
MPFR_SET_ZERO (y);
MPFR_SET_SAME_SIGN (y, x);
MPFR_RET (0);
}
}
/* General case */
MPFR_SAVE_EXPO_MARK (expo);
mpz_init (m);
e = mpfr_get_z_exp (m, x); /* x = m * 2^e */
if ((negative = MPFR_IS_NEG(x)))
mpz_neg (m, m);
r = e % 3;
if (r < 0)
r += 3;
/* x = (m*2^r) * 2^(e-r) = (m*2^r) * 2^(3*q) */
MPFR_MPZ_SIZEINBASE2 (size_m, m);
n = MPFR_PREC (y) + (rnd_mode == GMP_RNDN);
/* we want 3*n-2 <= size_m + 3*sh + r <= 3*n
i.e. 3*sh + size_m + r <= 3*n */
sh = (3 * (mp_exp_t) n - (mp_exp_t) size_m - r) / 3;
sh = 3 * sh + r;
if (sh >= 0)
{
mpz_mul_2exp (m, m, sh);
e = e - sh;
}
else if (r > 0)
{
mpz_mul_2exp (m, m, r);
e = e - r;
}
/* invariant: x = m*2^e, with e divisible by 3 */
/* we reuse the variable m to store the cube root, since it is not needed
any more: we just need to know if the root is exact */
inexact = mpz_root (m, m, 3) == 0;
MPFR_MPZ_SIZEINBASE2 (tmp, m);
sh = tmp - n;
if (sh > 0) /* we have to flush to 0 the last sh bits from m */
{
inexact = inexact || ((mp_exp_t) mpz_scan1 (m, 0) < sh);
mpz_div_2exp (m, m, sh);
e += 3 * sh;
}
if (inexact)
{
if (negative)
rnd_mode = MPFR_INVERT_RND (rnd_mode);
if (rnd_mode == GMP_RNDU
|| (rnd_mode == GMP_RNDN && mpz_tstbit (m, 0)))
inexact = 1, mpz_add_ui (m, m, 1);
else
inexact = -1;
}
/* either inexact is not zero, and the conversion is exact, i.e. inexact
is not changed; or inexact=0, and inexact is set only when
rnd_mode=GMP_RNDN and bit (n+1) from m is 1 */
inexact += mpfr_set_z (y, m, GMP_RNDN);
MPFR_SET_EXP (y, MPFR_GET_EXP (y) + e / 3);
if (negative)
{
MPFR_CHANGE_SIGN (y);
inexact = -inexact;
}
mpz_clear (m);
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (y, inexact, rnd_mode);
}
static int
mpfr_rem1 (mpfr_ptr rem, long *quo, mp_rnd_t rnd_q,
mpfr_srcptr x, mpfr_srcptr y, mp_rnd_t rnd)
{
mp_exp_t ex, ey;
int compare, inex, q_is_odd, sign, signx = MPFR_SIGN (x);
mpz_t mx, my, r;
MPFR_ASSERTD (rnd_q == GMP_RNDN || rnd_q == GMP_RNDZ);
if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x) || MPFR_IS_SINGULAR (y)))
{
if (MPFR_IS_NAN (x) || MPFR_IS_NAN (y) || MPFR_IS_INF (x)
|| MPFR_IS_ZERO (y))
{
/* for remquo, quo is undefined */
MPFR_SET_NAN (rem);
MPFR_RET_NAN;
}
else /* either y is Inf and x is 0 or non-special,
or x is 0 and y is non-special,
in both cases the quotient is zero. */
{
if (quo)
*quo = 0;
return mpfr_set (rem, x, rnd);
}
}
/* now neither x nor y is NaN, Inf or zero */
mpz_init (mx);
mpz_init (my);
mpz_init (r);
ex = mpfr_get_z_exp (mx, x); /* x = mx*2^ex */
ey = mpfr_get_z_exp (my, y); /* y = my*2^ey */
/* to get rid of sign problems, we compute it separately:
quo(-x,-y) = quo(x,y), rem(-x,-y) = -rem(x,y)
quo(-x,y) = -quo(x,y), rem(-x,y) = -rem(x,y)
thus quo = sign(x/y)*quo(|x|,|y|), rem = sign(x)*rem(|x|,|y|) */
sign = (signx == MPFR_SIGN (y)) ? 1 : -1;
mpz_abs (mx, mx);
mpz_abs (my, my);
q_is_odd = 0;
/* divide my by 2^k if possible to make operations mod my easier */
{
unsigned long k = mpz_scan1 (my, 0);
ey += k;
mpz_div_2exp (my, my, k);
}
if (ex <= ey)
{
/* q = x/y = mx/(my*2^(ey-ex)) */
mpz_mul_2exp (my, my, ey - ex); /* divide mx by my*2^(ey-ex) */
if (rnd_q == GMP_RNDZ)
/* 0 <= |r| <= |my|, r has the same sign as mx */
mpz_tdiv_qr (mx, r, mx, my);
else
/* 0 <= |r| <= |my|, r has the same sign as my */
mpz_fdiv_qr (mx, r, mx, my);
if (rnd_q == GMP_RNDN)
q_is_odd = mpz_tstbit (mx, 0);
if (quo) /* mx is the quotient */
{
mpz_tdiv_r_2exp (mx, mx, WANTED_BITS);
*quo = mpz_get_si (mx);
}
}
else /* ex > ey */
{
if (quo)
/* for remquo, to get the low WANTED_BITS more bits of the quotient,
we first compute R = X mod Y*2^WANTED_BITS, where X and Y are
defined below. Then the low WANTED_BITS of the quotient are
floor(R/Y). */
mpz_mul_2exp (my, my, WANTED_BITS); /* 2^WANTED_BITS*Y */
else
/* Let X = mx*2^(ex-ey) and Y = my. Then both X and Y are integers.
Assume X = R mod Y, then x = X*2^ey = R*2^ey mod (Y*2^ey=y).
To be able to perform the rounding, we need the least significant
bit of the quotient, i.e., one more bit in the remainder,
which is obtained by dividing by 2Y. */
mpz_mul_2exp (my, my, 1); /* 2Y */
mpz_set_ui (r, 2);
mpz_powm_ui (r, r, ex - ey, my); /* 2^(ex-ey) mod my */
mpz_mul (r, r, mx);
mpz_mod (r, r, my);
if (quo) /* now 0 <= r < 2^WANTED_BITS*Y */
{
mpz_div_2exp (my, my, WANTED_BITS); /* back to Y */
mpz_tdiv_qr (mx, r, r, my);
/* oldr = mx*my + newr */
*quo = mpz_get_si (mx);
q_is_odd = *quo & 1;
}
else /* now 0 <= r < 2Y */
{
mpz_div_2exp (my, my, 1); /* back to Y */
if (rnd_q == GMP_RNDN)
{
/* least significant bit of q */
q_is_odd = mpz_cmpabs (r, my) >= 0;
if (q_is_odd)
mpz_sub (r, r, my);
}
}
/* now 0 <= |r| < |my|, and if needed,
q_is_odd is the least significant bit of q */
}
if (mpz_cmp_ui (r, 0) == 0)
inex = mpfr_set_ui (rem, 0, GMP_RNDN);
else
{
if (rnd_q == GMP_RNDN)
{
/* FIXME: the comparison 2*r < my could be done more efficiently
at the mpn level */
mpz_mul_2exp (r, r, 1);
compare = mpz_cmpabs (r, my);
mpz_div_2exp (r, r, 1);
compare = ((compare > 0) ||
((rnd_q == GMP_RNDN) && (compare == 0) && q_is_odd));
/* if compare != 0, we need to subtract my to r, and add 1 to quo */
if (compare)
{
mpz_sub (r, r, my);
if (quo && (rnd_q == GMP_RNDN))
*quo += 1;
}
}
inex = mpfr_set_z (rem, r, rnd);
/* if ex > ey, rem should be multiplied by 2^ey, else by 2^ex */
MPFR_EXP (rem) += (ex > ey) ? ey : ex;
}
if (quo)
*quo *= sign;
/* take into account sign of x */
if (signx < 0)
{
mpfr_neg (rem, rem, GMP_RNDN);
inex = -inex;
}
mpz_clear (mx);
mpz_clear (my);
mpz_clear (r);
return inex;
}
/* s <- 1 + r/1! + r^2/2! + ... + r^l/l! while MPFR_EXP(r^l/l!)+MPFR_EXPR(r)>-q
using Brent/Kung method with O(sqrt(l)) multiplications.
Return l.
Uses m multiplications of full size and 2l/m of decreasing size,
i.e. a total equivalent to about m+l/m full multiplications,
i.e. 2*sqrt(l) for m=sqrt(l).
Version using mpz. ss must have at least (sizer+1) limbs.
The error is bounded by (l^2+4*l) ulps where l is the return value.
*/
static int
mpfr_exp2_aux2 (mpz_t s, mpfr_srcptr r, int q, int *exps)
{
int expr, l, m, i, sizer, *expR, expt, ql;
unsigned long int c;
mpz_t t, *R, rr, tmp;
TMP_DECL(marker);
/* estimate value of l */
l = q / (-MPFR_EXP(r));
m = (int) _mpfr_isqrt (l);
/* we access R[2], thus we need m >= 2 */
if (m < 2) m = 2;
TMP_MARK(marker);
R = (mpz_t*) TMP_ALLOC((m+1)*sizeof(mpz_t)); /* R[i] stands for r^i */
expR = (int*) TMP_ALLOC((m+1)*sizeof(int)); /* exponent for R[i] */
sizer = 1 + (MPFR_PREC(r)-1)/BITS_PER_MP_LIMB;
mpz_init(tmp);
MY_INIT_MPZ(rr, sizer+2);
MY_INIT_MPZ(t, 2*sizer); /* double size for products */
mpz_set_ui(s, 0); *exps = 1-q; /* initialize s to zero, 1 ulp = 2^(1-q) */
for (i=0;i<=m;i++) MY_INIT_MPZ(R[i], sizer+2);
expR[1] = mpfr_get_z_exp(R[1], r); /* exact operation: no error */
expR[1] = mpz_normalize2(R[1], R[1], expR[1], 1-q); /* error <= 1 ulp */
mpz_mul(t, R[1], R[1]); /* err(t) <= 2 ulps */
mpz_div_2exp(R[2], t, q-1); /* err(R[2]) <= 3 ulps */
expR[2] = 1-q;
for (i=3;i<=m;i++) {
mpz_mul(t, R[i-1], R[1]); /* err(t) <= 2*i-2 */
mpz_div_2exp(R[i], t, q-1); /* err(R[i]) <= 2*i-1 ulps */
expR[i] = 1-q;
}
mpz_set_ui(R[0], 1); mpz_mul_2exp(R[0], R[0], q-1); expR[0]=1-q; /* R[0]=1 */
mpz_set_ui(rr, 1); expr=0; /* rr contains r^l/l! */
/* by induction: err(rr) <= 2*l ulps */
l = 0;
ql = q; /* precision used for current giant step */
do {
/* all R[i] must have exponent 1-ql */
if (l) for (i=0;i<m;i++) {
expR[i] = mpz_normalize2(R[i], R[i], expR[i], 1-ql);
}
/* the absolute error on R[i]*rr is still 2*i-1 ulps */
expt = mpz_normalize2(t, R[m-1], expR[m-1], 1-ql);
/* err(t) <= 2*m-1 ulps */
/* computes t = 1 + r/(l+1) + ... + r^(m-1)*l!/(l+m-1)!
using Horner's scheme */
for (i=m-2;i>=0;i--) {
mpz_div_ui(t, t, l+i+1); /* err(t) += 1 ulp */
mpz_add(t, t, R[i]);
}
/* now err(t) <= (3m-2) ulps */
/* now multiplies t by r^l/l! and adds to s */
mpz_mul(t, t, rr); expt += expr;
expt = mpz_normalize2(t, t, expt, *exps);
/* err(t) <= (3m-1) + err_rr(l) <= (3m-2) + 2*l */
#ifdef DEBUG
if (expt != *exps) {
fprintf(stderr, "Error: expt != exps %d %d\n", expt, *exps); exit(1);
}
#endif
mpz_add(s, s, t); /* no error here */
/* updates rr, the multiplication of the factors l+i could be done
using binary splitting too, but it is not sure it would save much */
mpz_mul(t, rr, R[m]); /* err(t) <= err(rr) + 2m-1 */
expr += expR[m];
mpz_set_ui (tmp, 1);
for (i=1, c=1; i<=m; i++) {
if (l+i > ~((unsigned long int) 0)/c) {
mpz_mul_ui(tmp, tmp, c);
c = l+i;
}
else c *= (unsigned long int) l+i;
}
if (c != 1) mpz_mul_ui (tmp, tmp, c); /* tmp is exact */
mpz_fdiv_q(t, t, tmp); /* err(t) <= err(rr) + 2m */
expr += mpz_normalize(rr, t, ql); /* err_rr(l+1) <= err_rr(l) + 2m+1 */
ql = q - *exps - mpz_sizeinbase(s, 2) + expr + mpz_sizeinbase(rr, 2);
l+=m;
} while (expr+mpz_sizeinbase(rr, 2) > -q);
TMP_FREE(marker);
mpz_clear(tmp);
return l;
}
/* use Brent's formula exp(x) = (1+r+r^2/2!+r^3/3!+...)^(2^K)*2^n
where x = n*log(2)+(2^K)*r
together with Brent-Kung O(t^(1/2)) algorithm for the evaluation of
power series. The resulting complexity is O(n^(1/3)*M(n)).
*/
int
mpfr_exp_2 (mpfr_ptr y, mpfr_srcptr x, mp_rnd_t rnd_mode)
{
int n, K, precy, q, k, l, err, exps, inexact;
mpfr_t r, s, t; mpz_t ss;
TMP_DECL(marker);
precy = MPFR_PREC(y);
n = (int) (mpfr_get_d1 (x) / LOG2);
/* for the O(n^(1/2)*M(n)) method, the Taylor series computation of
n/K terms costs about n/(2K) multiplications when computed in fixed
point */
K = (precy<SWITCH) ? _mpfr_isqrt((precy + 1) / 2) : _mpfr_cuberoot (4*precy);
l = (precy-1)/K + 1;
err = K + (int) _mpfr_ceil_log2 (2.0 * (double) l + 18.0);
/* add K extra bits, i.e. failure probability <= 1/2^K = O(1/precy) */
q = precy + err + K + 3;
mpfr_init2 (r, q);
mpfr_init2 (s, q);
mpfr_init2 (t, q);
/* the algorithm consists in computing an upper bound of exp(x) using
a precision of q bits, and see if we can round to MPFR_PREC(y) taking
into account the maximal error. Otherwise we increase q. */
do {
#ifdef DEBUG
printf("n=%d K=%d l=%d q=%d\n",n,K,l,q);
#endif
/* if n<0, we have to get an upper bound of log(2)
in order to get an upper bound of r = x-n*log(2) */
mpfr_const_log2 (s, (n>=0) ? GMP_RNDZ : GMP_RNDU);
#ifdef DEBUG
printf("n=%d log(2)=",n); mpfr_print_binary(s); putchar('\n');
#endif
mpfr_mul_ui (r, s, (n<0) ? -n : n, (n>=0) ? GMP_RNDZ : GMP_RNDU);
if (n<0) mpfr_neg(r, r, GMP_RNDD);
/* r = floor(n*log(2)) */
#ifdef DEBUG
printf("x=%1.20e\n", mpfr_get_d1 (x));
printf(" ="); mpfr_print_binary(x); putchar('\n');
printf("r=%1.20e\n", mpfr_get_d1 (r));
printf(" ="); mpfr_print_binary(r); putchar('\n');
#endif
mpfr_sub(r, x, r, GMP_RNDU);
if (MPFR_SIGN(r)<0) { /* initial approximation n was too large */
n--;
mpfr_mul_ui(r, s, (n<0) ? -n : n, GMP_RNDZ);
if (n<0) mpfr_neg(r, r, GMP_RNDD);
mpfr_sub(r, x, r, GMP_RNDU);
}
#ifdef DEBUG
printf("x-r=%1.20e\n", mpfr_get_d1 (r));
printf(" ="); mpfr_print_binary(r); putchar('\n');
if (MPFR_SIGN(r)<0) { fprintf(stderr,"Error in mpfr_exp: r<0\n"); exit(1); }
#endif
mpfr_div_2ui(r, r, K, GMP_RNDU); /* r = (x-n*log(2))/2^K */
TMP_MARK(marker);
MY_INIT_MPZ(ss, 3 + 2*((q-1)/BITS_PER_MP_LIMB));
exps = mpfr_get_z_exp(ss, s);
/* s <- 1 + r/1! + r^2/2! + ... + r^l/l! */
l = (precy<SWITCH) ? mpfr_exp2_aux(ss, r, q, &exps) /* naive method */
: mpfr_exp2_aux2(ss, r, q, &exps); /* Brent/Kung method */
#ifdef DEBUG
printf("l=%d q=%d (K+l)*q^2=%1.3e\n", l, q, (K+l)*(double)q*q);
#endif
for (k=0;k<K;k++) {
mpz_mul(ss, ss, ss); exps <<= 1;
exps += mpz_normalize(ss, ss, q);
}
mpfr_set_z(s, ss, GMP_RNDN); MPFR_EXP(s) += exps;
TMP_FREE(marker); /* don't need ss anymore */
if (n>0) mpfr_mul_2ui(s, s, n, GMP_RNDU);
else mpfr_div_2ui(s, s, -n, GMP_RNDU);
/* error is at most 2^K*(3l*(l+1)) ulp for mpfr_exp2_aux */
if (precy<SWITCH) l = 3*l*(l+1);
else l = l*(l+4);
k=0; while (l) { k++; l >>= 1; }
/* now k = ceil(log(error in ulps)/log(2)) */
K += k;
#ifdef DEBUG
printf("after mult. by 2^n:\n");
if (MPFR_EXP(s) > -1024)
printf("s=%1.20e\n", mpfr_get_d1 (s));
printf(" ="); mpfr_print_binary(s); putchar('\n');
printf("err=%d bits\n", K);
#endif
l = mpfr_can_round(s, q-K, GMP_RNDN, rnd_mode, precy);
if (l==0) {
#ifdef DEBUG
printf("not enough precision, use %d\n", q+BITS_PER_MP_LIMB);
printf("q=%d q-K=%d precy=%d\n",q,q-K,precy);
#endif
q += BITS_PER_MP_LIMB;
mpfr_set_prec(r, q); mpfr_set_prec(s, q); mpfr_set_prec(t, q);
}
} while (l==0);
inexact = mpfr_set (y, s, rnd_mode);
mpfr_clear(r); mpfr_clear(s); mpfr_clear(t);
return inexact;
}
/* The computation of z = pow(x,y) is done by
z = exp(y * log(x)) = x^y */
int
mpfr_pow (mpfr_ptr z, mpfr_srcptr x, mpfr_srcptr y, mp_rnd_t rnd_mode)
{
int inexact = 0;
if (MPFR_IS_NAN(x) || MPFR_IS_NAN(y))
{
MPFR_SET_NAN(z);
MPFR_RET_NAN;
}
if (MPFR_IS_INF(y))
{
mpfr_t one;
int cmp;
if (MPFR_SIGN(y) > 0)
{
if (MPFR_IS_INF(x))
{
MPFR_CLEAR_FLAGS(z);
if (MPFR_SIGN(x) > 0)
MPFR_SET_INF(z);
else
MPFR_SET_ZERO(z);
MPFR_SET_POS(z);
MPFR_RET(0);
}
MPFR_CLEAR_FLAGS(z);
if (MPFR_IS_ZERO(x))
{
MPFR_SET_ZERO(z);
MPFR_SET_POS(z);
MPFR_RET(0);
}
mpfr_init2(one, BITS_PER_MP_LIMB);
mpfr_set_ui(one, 1, GMP_RNDN);
cmp = mpfr_cmp_abs(x, one);
mpfr_clear(one);
if (cmp > 0)
{
MPFR_SET_INF(z);
MPFR_SET_POS(z);
MPFR_RET(0);
}
else if (cmp < 0)
{
MPFR_SET_ZERO(z);
MPFR_SET_POS(z);
MPFR_RET(0);
}
else
{
MPFR_SET_NAN(z);
MPFR_RET_NAN;
}
}
else
{
if (MPFR_IS_INF(x))
{
MPFR_CLEAR_FLAGS(z);
if (MPFR_SIGN(x) > 0)
MPFR_SET_ZERO(z);
else
MPFR_SET_INF(z);
MPFR_SET_POS(z);
MPFR_RET(0);
}
if (MPFR_IS_ZERO(x))
{
MPFR_SET_INF(z);
MPFR_SET_POS(z);
MPFR_RET(0);
}
mpfr_init2(one, BITS_PER_MP_LIMB);
mpfr_set_ui(one, 1, GMP_RNDN);
cmp = mpfr_cmp_abs(x, one);
mpfr_clear(one);
MPFR_CLEAR_FLAGS(z);
if (cmp > 0)
{
MPFR_SET_ZERO(z);
MPFR_SET_POS(z);
MPFR_RET(0);
}
else if (cmp < 0)
{
MPFR_SET_INF(z);
MPFR_SET_POS(z);
MPFR_RET(0);
}
else
{
MPFR_SET_NAN(z);
MPFR_RET_NAN;
}
}
}
if (MPFR_IS_ZERO(y))
{
return mpfr_set_ui (z, 1, GMP_RNDN);
}
if (mpfr_isinteger (y))
{
mpz_t zi;
long int zii;
int exptol;
mpz_init(zi);
exptol = mpfr_get_z_exp (zi, y);
if (exptol>0)
mpz_mul_2exp(zi, zi, exptol);
else
mpz_tdiv_q_2exp(zi, zi, (unsigned long int) (-exptol));
zii=mpz_get_ui(zi);
mpz_clear(zi);
return mpfr_pow_si (z, x, zii, rnd_mode);
}
if (MPFR_IS_INF(x))
{
if (MPFR_SIGN(x) > 0)
{
MPFR_CLEAR_FLAGS(z);
if (MPFR_SIGN(y) > 0)
MPFR_SET_INF(z);
else
MPFR_SET_ZERO(z);
MPFR_SET_POS(z);
MPFR_RET(0);
}
else
{
MPFR_SET_NAN(z);
MPFR_RET_NAN;
}
}
if (MPFR_IS_ZERO(x))
{
MPFR_CLEAR_FLAGS(z);
MPFR_SET_ZERO(z);
MPFR_SET_SAME_SIGN(z, x);
MPFR_RET(0);
}
if (MPFR_SIGN(x) < 0)
{
MPFR_SET_NAN(z);
MPFR_RET_NAN;
}
MPFR_CLEAR_FLAGS(z);
/* General case */
{
/* Declaration of the intermediary variable */
mpfr_t t, te, ti;
int loop = 0, ok;
/* Declaration of the size variable */
mp_prec_t Nx = MPFR_PREC(x); /* 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+5+_mpfr_ceil_log2(Nt);
/* initialise of intermediary variable */
mpfr_init(t);
mpfr_init(ti);
mpfr_init(te);
do
{
loop ++;
/* reactualisation of the precision */
mpfr_set_prec (t, Nt);
mpfr_set_prec (ti, Nt);
mpfr_set_prec (te, Nt);
/* compute exp(y*ln(x)) */
mpfr_log (ti, x, GMP_RNDU); /* ln(n) */
mpfr_mul (te, y, ti, GMP_RNDU); /* y*ln(n) */
mpfr_exp (t, te, GMP_RNDN); /* exp(x*ln(n))*/
/* estimation of the error -- see pow function in algorithms.ps*/
err = Nt - (MPFR_EXP(te) + 3);
/* actualisation of the precision */
Nt += 10;
ok = mpfr_can_round (t, err, GMP_RNDN, rnd_mode, Ny);
/* check exact power */
if (ok == 0 && loop == 1)
ok = mpfr_pow_is_exact (x, y);
}
while (err < 0 || ok == 0);
inexact = mpfr_set (z, t, rnd_mode);
mpfr_clear (t);
mpfr_clear (ti);
mpfr_clear (te);
}
return inexact;
}
