/* 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;
}
/* 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;
}
