static void
tst_isqrt (unsigned long n, unsigned long r)
{
unsigned long i;
i = __gmpfr_isqrt (n);
if (i != r)
{
printf ("Error in __gmpfr_isqrt (%lu): got %lu instead of %lu\n",
n, i, r);
exit (1);
}
}
int
mpfr_cos (mpfr_ptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
mpfr_prec_t K0, K, precy, m, k, l;
int inexact, reduce = 0;
mpfr_t r, s, xr, c;
mpfr_exp_t exps, cancel = 0, expx;
MPFR_ZIV_DECL (loop);
MPFR_SAVE_EXPO_DECL (expo);
MPFR_GROUP_DECL (group);
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,
inexact));
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, rnd_mode);
}
}
MPFR_SAVE_EXPO_MARK (expo);
/* cos(x) = 1-x^2/2 + ..., so error < 2^(2*EXP(x)-1) */
expx = MPFR_GET_EXP (x);
MPFR_SMALL_INPUT_AFTER_SAVE_EXPO (y, __gmpfr_one, -2 * expx,
1, 0, rnd_mode, expo, {});
/* Compute initial precision */
precy = MPFR_PREC (y);
if (precy >= MPFR_SINCOS_THRESHOLD)
{
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_cos_fast (y, x, rnd_mode);
}
K0 = __gmpfr_isqrt (precy / 3);
m = precy + 2 * MPFR_INT_CEIL_LOG2 (precy) + 2 * K0;
if (expx >= 3)
{
reduce = 1;
/* As expx + m - 1 will silently be converted into mpfr_prec_t
in the mpfr_init2 call, the assert below may be useful to
avoid undefined behavior. */
MPFR_ASSERTN (expx + m - 1 <= MPFR_PREC_MAX);
mpfr_init2 (c, expx + m - 1);
mpfr_init2 (xr, m);
}
MPFR_GROUP_INIT_2 (group, m, r, s);
MPFR_ZIV_INIT (loop, m);
for (;;)
{
/* If |x| >= 4, first reduce x cmod (2*Pi) into xr, using mpfr_remainder:
let e = EXP(x) >= 3, and m the target precision:
(1) c <- 2*Pi [precision e+m-1, nearest]
(2) xr <- remainder (x, c) [precision m, nearest]
We have |c - 2*Pi| <= 1/2ulp(c) = 2^(3-e-m)
|xr - x - k c| <= 1/2ulp(xr) <= 2^(1-m)
|k| <= |x|/(2*Pi) <= 2^(e-2)
Thus |xr - x - 2kPi| <= |k| |c - 2Pi| + 2^(1-m) <= 2^(2-m).
It follows |cos(xr) - cos(x)| <= 2^(2-m). */
if (reduce)
{
mpfr_const_pi (c, MPFR_RNDN);
mpfr_mul_2ui (c, c, 1, MPFR_RNDN); /* 2Pi */
mpfr_remainder (xr, x, c, MPFR_RNDN);
if (MPFR_IS_ZERO(xr))
goto ziv_next;
/* now |xr| <= 4, thus r <= 16 below */
mpfr_mul (r, xr, xr, MPFR_RNDU); /* err <= 1 ulp */
}
else
mpfr_mul (r, x, x, MPFR_RNDU); /* err <= 1 ulp */
/* now |x| < 4 (or xr if reduce = 1), thus |r| <= 16 */
/* we need |r| < 1/2 for mpfr_cos2_aux, i.e., EXP(r) - 2K <= -1 */
K = K0 + 1 + MAX(0, MPFR_GET_EXP(r)) / 2;
/* since K0 >= 0, if EXP(r) < 0, then K >= 1, thus EXP(r) - 2K <= -3;
otherwise if EXP(r) >= 0, then K >= 1/2 + EXP(r)/2, thus
EXP(r) - 2K <= -1 */
MPFR_SET_EXP (r, MPFR_GET_EXP (r) - 2 * K); /* Can't overflow! */
/* s <- 1 - r/2! + ... + (-1)^l r^l/(2l)! */
l = mpfr_cos2_aux (s, r);
/* l is the error bound in ulps on s */
MPFR_SET_ONE (r);
for (k = 0; k < K; k++)
{
mpfr_sqr (s, s, MPFR_RNDU); /* err <= 2*olderr */
MPFR_SET_EXP (s, MPFR_GET_EXP (s) + 1); /* Can't overflow */
mpfr_sub (s, s, r, MPFR_RNDN); /* err <= 4*olderr */
if (MPFR_IS_ZERO(s))
goto ziv_next;
MPFR_ASSERTD (MPFR_GET_EXP (s) <= 1);
}
/* The absolute error on s is bounded by (2l+1/3)*2^(2K-m)
2l+1/3 <= 2l+1.
If |x| >= 4, we need to add 2^(2-m) for the argument reduction
by 2Pi: if K = 0, this amounts to add 4 to 2l+1/3, i.e., to add
2 to l; if K >= 1, this amounts to add 1 to 2*l+1/3. */
l = 2 * l + 1;
if (reduce)
l += (K == 0) ? 4 : 1;
k = MPFR_INT_CEIL_LOG2 (l) + 2*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, precy, rnd_mode)))
break;
if (MPFR_UNLIKELY (exps == 1))
/* s = 1 or -1, and except x=0 which was already checked above,
cos(x) cannot be 1 or -1, so we can round if the error is less
than 2^(-precy) for directed rounding, or 2^(-precy-1) for rounding
to nearest. */
{
if (m > k && (m - k >= precy + (rnd_mode == MPFR_RNDN)))
{
/* If round to nearest or away, result is s = 1 or -1,
otherwise it is round(nexttoward (s, 0)). However in order to
have the inexact flag correctly set below, we set |s| to
1 - 2^(-m) in all cases. */
mpfr_nexttozero (s);
break;
}
}
if (exps < cancel)
{
m += cancel - exps;
cancel = exps;
}
ziv_next:
MPFR_ZIV_NEXT (loop, m);
MPFR_GROUP_REPREC_2 (group, m, r, s);
if (reduce)
{
mpfr_set_prec (xr, m);
mpfr_set_prec (c, expx + m - 1);
}
}
MPFR_ZIV_FREE (loop);
inexact = mpfr_set (y, s, rnd_mode);
MPFR_GROUP_CLEAR (group);
if (reduce)
{
mpfr_clear (xr);
mpfr_clear (c);
}
MPFR_SAVE_EXPO_FREE (expo);
return mpfr_check_range (y, inexact, rnd_mode);
}
/* 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;
}
/* 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;
}
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;
}
