int
arf_add(arf_ptr z, arf_srcptr x, arf_srcptr y, slong prec, arf_rnd_t rnd)
{
mp_size_t xn, yn;
mp_srcptr xptr, yptr;
slong shift;
if (arf_is_special(x) || arf_is_special(y))
{
return arf_add_special(z, x, y, prec, rnd);
}
shift = _fmpz_sub_small(ARF_EXPREF(x), ARF_EXPREF(y));
if (shift < 0)
{
arf_srcptr __t;
__t = x; x = y; y = __t;
shift = -shift;
}
ARF_GET_MPN_READONLY(xptr, xn, x);
ARF_GET_MPN_READONLY(yptr, yn, y);
return _arf_add_mpn(z, xptr, xn, ARF_SGNBIT(x), ARF_EXPREF(x),
yptr, yn, ARF_SGNBIT(y), shift, prec, rnd);
}
void
arf_ceil(arf_t z, const arf_t x)
{
if (arf_is_special(x) || arf_is_int(x))
{
arf_set(z, x);
}
else
{
slong exp = ARF_EXP(x);
/* now exp cannot be too large, as we would have
caught this in arf_is_int() */
if (COEFF_IS_MPZ(exp) || exp <= 0)
{
if (ARF_SGNBIT(x))
arf_zero(z);
else
arf_one(z);
}
else if (exp == 1)
{
arf_set_si(z, ARF_SGNBIT(x) ? -1 : 2);
}
else
{
arf_set_round(z, x, exp, ARF_RND_CEIL);
}
}
}
int
arf_sub_special(arf_t z, const arf_t x, const arf_t y, slong prec, arf_rnd_t rnd)
{
if (arf_is_zero(x))
{
if (arf_is_zero(y))
{
arf_zero(z);
return 0;
}
else
return arf_neg_round(z, y, prec, rnd);
}
else if (arf_is_zero(y))
{
return arf_set_round(z, x, prec, rnd);
}
else if (arf_is_nan(x) || arf_is_nan(y)
|| (arf_is_pos_inf(x) && arf_is_pos_inf(y))
|| (arf_is_neg_inf(x) && arf_is_neg_inf(y)))
{
arf_nan(z);
return 0;
}
else if (arf_is_special(x))
{
arf_set(z, x);
return 0;
}
else
{
arf_neg(z, y);
return 0;
}
}
void
arb_pow(arb_t z, const arb_t x, const arb_t y, slong prec)
{
if (arb_is_zero(y))
{
arb_one(z);
return;
}
if (arb_is_zero(x))
{
if (arb_is_positive(y))
arb_zero(z);
else
arb_indeterminate(z);
return;
}
if (arb_is_exact(y) && !arf_is_special(arb_midref(x)))
{
const arf_struct * ymid = arb_midref(y);
/* small half-integer or integer */
if (arf_cmpabs_2exp_si(ymid, BINEXP_LIMIT) < 0 &&
arf_is_int_2exp_si(ymid, -1))
{
fmpz_t e;
fmpz_init(e);
if (arf_is_int(ymid))
{
arf_get_fmpz_fixed_si(e, ymid, 0);
arb_pow_fmpz_binexp(z, x, e, prec);
}
else
{
arf_get_fmpz_fixed_si(e, ymid, -1);
arb_sqrt(z, x, prec + fmpz_bits(e));
arb_pow_fmpz_binexp(z, z, e, prec);
}
fmpz_clear(e);
return;
}
else if (arf_is_int(ymid) && arf_sgn(arb_midref(x)) < 0)
{
/* use (-x)^n = (-1)^n * x^n to avoid NaNs
at least at high enough precision */
int odd = !arf_is_int_2exp_si(ymid, 1);
_arb_pow_exp(z, x, 1, y, prec);
if (odd)
arb_neg(z, z);
return;
}
}
_arb_pow_exp(z, x, 0, y, prec);
}
static void
_arb_arf_div_rounded_den_add_err(arb_t res, const arf_t x, const arf_t y, int yinexact, slong prec)
{
int inexact = arf_div(arb_midref(res), x, y, prec, ARB_RND);
if (yinexact && !arf_is_special(arb_midref(res)))
arf_mag_add_ulp(arb_radref(res), arb_radref(res), arb_midref(res), prec - 1);
else if (inexact)
arf_mag_add_ulp(arb_radref(res), arb_radref(res), arb_midref(res), prec);
}
int
arf_cmpabs(const arf_t x, const arf_t y)
{
int ec, mc;
mp_size_t xn, yn;
mp_srcptr xp, yp;
if (arf_is_special(x) || arf_is_special(y))
{
if (arf_equal(x, y))
return 0;
if (arf_is_nan(x) || arf_is_nan(y))
return 0;
if (arf_is_zero(x)) return -1;
if (arf_is_zero(y)) return 1;
if (arf_is_inf(x)) return arf_is_inf(y) ? 0 : 1;
if (arf_is_inf(y)) return -1;
return -1;
}
ec = fmpz_cmp(ARF_EXPREF(x), ARF_EXPREF(y));
if (ec != 0)
return (ec < 0) ? -1 : 1;
ARF_GET_MPN_READONLY(xp, xn, x);
ARF_GET_MPN_READONLY(yp, yn, y);
if (xn >= yn)
mc = mpn_cmp(xp + xn - yn, yp, yn);
else
mc = mpn_cmp(xp, yp + yn - xn, xn);
if (mc != 0)
return (mc < 0) ? -1 : 1;
if (xn != yn)
return (xn < yn) ? -1 : 1;
return 0;
}
int
arf_set_round(arf_t y, const arf_t x, slong prec, arf_rnd_t rnd)
{
if (arf_is_special(x))
{
arf_set(y, x);
return 0;
}
else
{
int inexact;
slong fix;
mp_size_t xn;
mp_srcptr xptr;
if (y == x)
{
mp_ptr xtmp;
TMP_INIT;
ARF_GET_MPN_READONLY(xptr, xn, x);
/* exact */
if (xn * FLINT_BITS <= prec)
return 0;
if ((xn - 1) * FLINT_BITS < prec)
{
/* exact */
if ((xptr[0] << (prec - (xn-1) * FLINT_BITS)) == 0)
return 0;
}
/* inexact */
TMP_START;
xtmp = TMP_ALLOC(xn * sizeof(mp_limb_t));
flint_mpn_copyi(xtmp, xptr, xn);
inexact = _arf_set_round_mpn(y, &fix, xtmp, xn, ARF_SGNBIT(x), prec, rnd);
_fmpz_add_fast(ARF_EXPREF(y), ARF_EXPREF(x), fix);
TMP_END;
return inexact;
}
else
{
ARF_GET_MPN_READONLY(xptr, xn, x);
inexact = _arf_set_round_mpn(y, &fix, xptr, xn,
ARF_SGNBIT(x), prec, rnd);
_fmpz_add_fast(ARF_EXPREF(y), ARF_EXPREF(x), fix);
return inexact;
}
}
}
int
arf_sub_si(arf_ptr z, arf_srcptr x, slong y, slong prec, arf_rnd_t rnd)
{
mp_size_t xn, yn;
mp_srcptr xptr, yptr;
mp_limb_t ytmp;
int xsgnbit, ysgnbit;
fmpz yexp;
slong shift;
if (y == 0)
{
return arf_set_round(z, x, prec, rnd);
}
else if (arf_is_special(x))
{
if (arf_is_zero(x))
{
arf_set_si(z, y);
return arf_neg_round(z, z, prec, rnd);
}
else
{
arf_set(z, x);
return 0;
}
}
ysgnbit = (y < 0);
if (ysgnbit)
ytmp = -y;
else
ytmp = y;
yptr = &ytmp;
yn = 1;
yexp = FLINT_BITS;
ysgnbit ^= 1;
shift = _fmpz_sub_small(ARF_EXPREF(x), &yexp);
xsgnbit = ARF_SGNBIT(x);
ARF_GET_MPN_READONLY(xptr, xn, x);
if (shift >= 0)
return _arf_add_mpn(z, xptr, xn, xsgnbit, ARF_EXPREF(x),
yptr, yn, ysgnbit, shift, prec, rnd);
else
return _arf_add_mpn(z, yptr, yn, ysgnbit, &yexp,
xptr, xn, xsgnbit, -shift, prec, rnd);
}
int
arf_sub(arf_ptr z, arf_srcptr x, arf_srcptr y, slong prec, arf_rnd_t rnd)
{
mp_size_t xn, yn;
mp_srcptr xptr, yptr;
slong shift;
if (arf_is_special(x) || arf_is_special(y))
{
return arf_sub_special(z, x, y, prec, rnd);
}
shift = _fmpz_sub_small(ARF_EXPREF(x), ARF_EXPREF(y));
ARF_GET_MPN_READONLY(xptr, xn, x);
ARF_GET_MPN_READONLY(yptr, yn, y);
if (shift >= 0)
return _arf_add_mpn(z, xptr, xn, ARF_SGNBIT(x), ARF_EXPREF(x),
yptr, yn, ARF_SGNBIT(y) ^ 1, shift, prec, rnd);
else
return _arf_add_mpn(z, yptr, yn, ARF_SGNBIT(y) ^ 1, ARF_EXPREF(y),
xptr, xn, ARF_SGNBIT(x), -shift, prec, rnd);
}
void
_arb_poly_get_scale(fmpz_t scale, arb_srcptr x, slong xlen,
arb_srcptr y, slong ylen)
{
slong xa, xb, ya, yb, den;
fmpz_zero(scale);
/* ignore zeros (and infs/nans!); find the first and last
finite nonzero entries to determine the scale */
xa = 0;
xb = xlen - 1;
while (xa < xlen && arf_is_special(arb_midref(x + xa))) xa++;
while (xb > xa && arf_is_special(arb_midref(x + xb))) xb--;
ya = 0;
yb = ylen - 1;
while (ya < ylen && arf_is_special(arb_midref(y + ya))) ya++;
while (yb > ya && arf_is_special(arb_midref(y + yb))) yb--;
/* compute average of exponent differences, weighted by the lengths */
if (xa <= xb && ya <= yb && (xa < xb || ya < yb))
{
fmpz_add(scale, scale, ARF_EXPREF(arb_midref(x + xb)));
fmpz_sub(scale, scale, ARF_EXPREF(arb_midref(x + xa)));
fmpz_add(scale, scale, ARF_EXPREF(arb_midref(y + yb)));
fmpz_sub(scale, scale, ARF_EXPREF(arb_midref(y + ya)));
den = (xb - xa) + (yb - ya);
/* scale = floor(scale / den + 1/2) = floor((2 scale + den) / (2 den)) */
fmpz_mul_2exp(scale, scale, 1);
fmpz_add_ui(scale, scale, den);
fmpz_fdiv_q_ui(scale, scale, 2 * den);
}
}
int
arf_add_fmpz_2exp(arf_ptr z, arf_srcptr x, const fmpz_t y, const fmpz_t exp, slong prec, arf_rnd_t rnd)
{
mp_size_t xn, yn;
mp_srcptr xptr, yptr;
mp_limb_t ytmp;
int xsgnbit, ysgnbit, inexact;
fmpz_t yexp;
slong shift;
if (fmpz_is_zero(y))
{
return arf_set_round(z, x, prec, rnd);
}
else if (arf_is_special(x))
{
if (arf_is_zero(x))
{
inexact = arf_set_round_fmpz(z, y, prec, rnd);
arf_mul_2exp_fmpz(z, z, exp);
return inexact;
}
else
{
arf_set(z, x);
return 0;
}
}
FMPZ_GET_MPN_READONLY(ysgnbit, yn, yptr, ytmp, *y)
fmpz_init(yexp);
fmpz_add_ui(yexp, exp, yn * FLINT_BITS);
shift = _fmpz_sub_small(ARF_EXPREF(x), yexp);
xsgnbit = ARF_SGNBIT(x);
ARF_GET_MPN_READONLY(xptr, xn, x);
if (shift >= 0)
inexact = _arf_add_mpn(z, xptr, xn, xsgnbit, ARF_EXPREF(x),
yptr, yn, ysgnbit, shift, prec, rnd);
else
inexact = _arf_add_mpn(z, yptr, yn, ysgnbit, yexp,
xptr, xn, xsgnbit, -shift, prec, rnd);
fmpz_clear(yexp);
return inexact;
}
void
arf_get_fmpr(fmpr_t y, const arf_t x)
{
if (arf_is_special(x))
{
if (arf_is_zero(x))
fmpr_zero(y);
else if (arf_is_pos_inf(x))
fmpr_pos_inf(y);
else if (arf_is_neg_inf(x))
fmpr_neg_inf(y);
else
fmpr_nan(y);
}
else
{
arf_get_fmpz_2exp(fmpr_manref(y), fmpr_expref(y), x);
}
}
int
arf_sub_fmpz(arf_ptr z, arf_srcptr x, const fmpz_t y, slong prec, arf_rnd_t rnd)
{
mp_size_t xn, yn;
mp_srcptr xptr, yptr;
mp_limb_t ytmp;
int xsgnbit, ysgnbit;
fmpz yexp;
slong shift;
if (fmpz_is_zero(y))
{
return arf_set_round(z, x, prec, rnd);
}
else if (arf_is_special(x))
{
if (arf_is_zero(x))
{
arf_set_fmpz(z, y);
return arf_neg_round(z, z, prec, rnd);
}
else
{
arf_set(z, x);
return 0;
}
}
FMPZ_GET_MPN_READONLY(ysgnbit, yn, yptr, ytmp, *y)
yexp = yn * FLINT_BITS;
shift = _fmpz_sub_small(ARF_EXPREF(x), &yexp);
ysgnbit ^= 1;
xsgnbit = ARF_SGNBIT(x);
ARF_GET_MPN_READONLY(xptr, xn, x);
if (shift >= 0)
return _arf_add_mpn(z, xptr, xn, xsgnbit, ARF_EXPREF(x),
yptr, yn, ysgnbit, shift, prec, rnd);
else
return _arf_add_mpn(z, yptr, yn, ysgnbit, &yexp,
xptr, xn, xsgnbit, -shift, prec, rnd);
}
void
arf_get_fmpq(fmpq_t y, const arf_t x)
{
if (arf_is_zero(x))
{
fmpq_zero(y);
}
else if (arf_is_special(x) || !ARF_IS_LAGOM(x))
{
flint_printf("exception: arf_get_fmpq: cannot convert to rational\n");
abort();
}
else
{
fmpz_t man, exp;
slong e;
fmpz_init(man);
fmpz_init(exp);
arf_get_fmpz_2exp(man, exp, x);
e = *exp;
fmpz_set_ui(fmpq_denref(y), UWORD(1));
if (e >= 0)
{
fmpz_mul_2exp(fmpq_numref(y), man, e);
}
else
{
fmpz_set(fmpq_numref(y), man);
fmpz_mul_2exp(fmpq_denref(y), fmpq_denref(y), -e);
}
fmpz_clear(man);
fmpz_clear(exp);
}
}
int
arf_cmp_2exp_si(const arf_t x, long e)
{
if (arf_is_special(x))
{
if (arf_is_zero(x)) return -1;
if (arf_is_pos_inf(x)) return 1;
if (arf_is_neg_inf(x)) return -1;
return 0;
}
if (ARF_SGNBIT(x))
return -1;
/* Fast path. */
if (!COEFF_IS_MPZ(ARF_EXP(x)))
{
if (ARF_IS_POW2(x) && (ARF_EXP(x) - 1 == e))
return 0;
else
return (ARF_EXP(x) <= e) ? -1 : 1;
}
if (ARF_IS_POW2(x))
{
fmpz_t t;
fmpz_init(t);
fmpz_one(t);
fmpz_add_si(t, t, e);
if (fmpz_equal(ARF_EXPREF(x), t))
{
fmpz_clear(t);
return 0;
}
fmpz_clear(t);
}
return (fmpz_cmp_si(ARF_EXPREF(x), e) <= 0) ? -1 : 1;
}
static __inline__ void
_arb_vec_get_fmpz_2exp_blocks(fmpz * coeffs, fmpz * exps,
slong * blocks, const fmpz_t scale, arb_srcptr x, slong len, slong prec)
{
fmpz_t top, bot, t, b, v, block_top, block_bot;
slong i, j, s, block, bits, maxheight;
int in_zero;
fmpz_init(top);
fmpz_init(bot);
fmpz_init(t);
fmpz_init(b);
fmpz_init(v);
fmpz_init(block_top);
fmpz_init(block_bot);
blocks[0] = 0;
block = 0;
in_zero = 1;
if (prec == ARF_PREC_EXACT)
maxheight = ARF_PREC_EXACT;
else
maxheight = ALPHA * prec + BETA;
for (i = 0; i < len; i++)
{
bits = arf_bits(arb_midref(x + i));
/* Skip (must be zero, since we assume there are no Infs/NaNs). */
if (bits == 0)
continue;
/* Bottom and top exponent of current number */
fmpz_set(top, ARF_EXPREF(arb_midref(x + i)));
fmpz_submul_ui(top, scale, i);
fmpz_sub_ui(bot, top, bits);
/* Extend current block. */
if (in_zero)
{
fmpz_swap(block_top, top);
fmpz_swap(block_bot, bot);
}
else
{
fmpz_max(t, top, block_top);
fmpz_min(b, bot, block_bot);
fmpz_sub(v, t, b);
/* extend current block */
if (fmpz_cmp_ui(v, maxheight) < 0)
{
fmpz_swap(block_top, t);
fmpz_swap(block_bot, b);
}
else /* start new block */
{
/* write exponent for previous block */
fmpz_set(exps + block, block_bot);
block++;
blocks[block] = i;
fmpz_swap(block_top, top);
fmpz_swap(block_bot, bot);
}
}
in_zero = 0;
}
/* write exponent for last block */
fmpz_set(exps + block, block_bot);
/* end marker */
blocks[block + 1] = len;
/* write the block data */
for (i = 0; blocks[i] != len; i++)
{
for (j = blocks[i]; j < blocks[i + 1]; j++)
{
if (arf_is_special(arb_midref(x + j)))
{
fmpz_zero(coeffs + j);
}
else
{
/* TODO: make this a single operation */
arf_get_fmpz_2exp(coeffs + j, bot, arb_midref(x + j));
fmpz_mul_ui(t, scale, j);
fmpz_sub(t, bot, t);
s = _fmpz_sub_small(t, exps + i);
if (s < 0) abort(); /* Bug catcher */
fmpz_mul_2exp(coeffs + j, coeffs + j, s);
}
}
}
fmpz_clear(top);
fmpz_clear(bot);
fmpz_clear(t);
fmpz_clear(b);
fmpz_clear(v);
fmpz_clear(block_top);
fmpz_clear(block_bot);
}
int
arf_submul_mpz(arf_ptr z, arf_srcptr x, const mpz_t y, slong prec, arf_rnd_t rnd)
{
mp_size_t xn, yn, zn, tn, alloc;
mp_srcptr xptr, yptr, zptr;
mp_ptr tptr, tptr2;
fmpz_t texp, yexp;
slong shift;
int tsgnbit, ysgnbit, inexact;
ARF_MUL_TMP_DECL
yn = FLINT_ABS(y->_mp_size);
if (arf_is_special(x) || yn == 0 || arf_is_special(z))
{
if (arf_is_zero(z))
{
/* TODO: make more efficient */
arf_mul_mpz(z, x, y, ARF_PREC_EXACT, rnd);
return arf_neg_round(z, z, prec, rnd);
}
else if (arf_is_finite(x))
{
return arf_set_round(z, z, prec, rnd);
}
else
{
/* todo: speed up */
arf_t t;
arf_init(t);
arf_mul_mpz(t, x, y, ARF_PREC_EXACT, ARF_RND_DOWN);
inexact = arf_sub(z, z, t, prec, rnd);
arf_clear(t);
return inexact;
}
}
ARF_GET_MPN_READONLY(xptr, xn, x);
yptr = y->_mp_d;
ysgnbit = (y->_mp_size > 0);
*yexp = yn * FLINT_BITS;
ARF_GET_MPN_READONLY(zptr, zn, z);
fmpz_init(texp);
tsgnbit = ARF_SGNBIT(x) ^ ysgnbit;
alloc = tn = xn + yn;
ARF_MUL_TMP_ALLOC(tptr2, alloc)
tptr = tptr2;
ARF_MPN_MUL(tptr, xptr, xn, yptr, yn);
shift = (tptr[tn - 1] == 0) * FLINT_BITS;
tn -= (tptr[tn - 1] == 0);
_fmpz_add2_fast(texp, ARF_EXPREF(x), yexp, -shift);
shift = _fmpz_sub_small(ARF_EXPREF(z), texp);
if (shift >= 0)
inexact = _arf_add_mpn(z, zptr, zn, ARF_SGNBIT(z), ARF_EXPREF(z),
tptr, tn, tsgnbit, shift, prec, rnd);
else
inexact = _arf_add_mpn(z, tptr, tn, tsgnbit, texp,
zptr, zn, ARF_SGNBIT(z), -shift, prec, rnd);
ARF_MUL_TMP_FREE(tptr2, alloc)
fmpz_clear(texp);
return inexact;
}
int
arf_submul(arf_ptr z, arf_srcptr x, arf_srcptr y, slong prec, arf_rnd_t rnd)
{
mp_size_t xn, yn, zn, tn, alloc;
mp_srcptr xptr, yptr, zptr;
mp_ptr tptr, tptr2;
fmpz_t texp;
slong shift;
int tsgnbit, inexact;
ARF_MUL_TMP_DECL
if (arf_is_special(x) || arf_is_special(y) || arf_is_special(z))
{
if (arf_is_zero(z))
{
return arf_neg_mul(z, x, y, prec, rnd);
}
else if (arf_is_finite(x) && arf_is_finite(y))
{
return arf_set_round(z, z, prec, rnd);
}
else
{
/* todo: speed up */
arf_t t;
arf_init(t);
arf_mul(t, x, y, ARF_PREC_EXACT, ARF_RND_DOWN);
inexact = arf_sub(z, z, t, prec, rnd);
arf_clear(t);
return inexact;
}
}
tsgnbit = ARF_SGNBIT(x) ^ ARF_SGNBIT(y) ^ 1;
ARF_GET_MPN_READONLY(xptr, xn, x);
ARF_GET_MPN_READONLY(yptr, yn, y);
ARF_GET_MPN_READONLY(zptr, zn, z);
fmpz_init(texp);
_fmpz_add2_fast(texp, ARF_EXPREF(x), ARF_EXPREF(y), 0);
shift = _fmpz_sub_small(ARF_EXPREF(z), texp);
alloc = tn = xn + yn;
ARF_MUL_TMP_ALLOC(tptr2, alloc)
tptr = tptr2;
ARF_MPN_MUL(tptr, xptr, xn, yptr, yn);
tn -= (tptr[0] == 0);
tptr += (tptr[0] == 0);
if (shift >= 0)
inexact = _arf_add_mpn(z, zptr, zn, ARF_SGNBIT(z), ARF_EXPREF(z),
tptr, tn, tsgnbit, shift, prec, rnd);
else
inexact = _arf_add_mpn(z, tptr, tn, tsgnbit, texp,
zptr, zn, ARF_SGNBIT(z), -shift, prec, rnd);
ARF_MUL_TMP_FREE(tptr2, alloc)
fmpz_clear(texp);
return inexact;
}
void
acb_inv(acb_t res, const acb_t z, slong prec)
{
mag_t am, bm;
slong hprec;
#define a arb_midref(acb_realref(z))
#define b arb_midref(acb_imagref(z))
#define x arb_radref(acb_realref(z))
#define y arb_radref(acb_imagref(z))
/* choose precision for the floating-point approximation of a^2+b^2 so
that the double rounding result in less than
2 ulp error; also use at least MAG_BITS bits since the
value will be recycled for error bounds */
hprec = FLINT_MAX(prec + 3, MAG_BITS);
if (arb_is_zero(acb_imagref(z)))
{
arb_inv(acb_realref(res), acb_realref(z), prec);
arb_zero(acb_imagref(res));
return;
}
if (arb_is_zero(acb_realref(z)))
{
arb_inv(acb_imagref(res), acb_imagref(z), prec);
arb_neg(acb_imagref(res), acb_imagref(res));
arb_zero(acb_realref(res));
return;
}
if (!acb_is_finite(z))
{
acb_indeterminate(res);
return;
}
if (mag_is_zero(x) && mag_is_zero(y))
{
int inexact;
arf_t a2b2;
arf_init(a2b2);
inexact = arf_sosq(a2b2, a, b, hprec, ARF_RND_DOWN);
if (arf_is_special(a2b2))
{
acb_indeterminate(res);
}
else
{
_arb_arf_div_rounded_den(acb_realref(res), a, a2b2, inexact, prec);
_arb_arf_div_rounded_den(acb_imagref(res), b, a2b2, inexact, prec);
arf_neg(arb_midref(acb_imagref(res)), arb_midref(acb_imagref(res)));
}
arf_clear(a2b2);
return;
}
mag_init(am);
mag_init(bm);
/* first bound |a|-x, |b|-y */
arb_get_mag_lower(am, acb_realref(z));
arb_get_mag_lower(bm, acb_imagref(z));
if ((mag_is_zero(am) && mag_is_zero(bm)))
{
acb_indeterminate(res);
}
else
{
/*
The propagated error in the real part is given exactly by
(a+x')/((a+x')^2+(b+y'))^2 - a/(a^2+b^2) = P / Q,
P = [(b^2-a^2) x' - a (x'^2+y'^2 + 2y'b)]
Q = [(a^2+b^2)((a+x')^2+(b+y')^2)]
where |x'| <= x and |y'| <= y, and analogously for the imaginary part.
*/
mag_t t, u, v, w;
arf_t a2b2;
int inexact;
mag_init(t);
mag_init(u);
mag_init(v);
mag_init(w);
arf_init(a2b2);
inexact = arf_sosq(a2b2, a, b, hprec, ARF_RND_DOWN);
/* compute denominator */
/* t = (|a|-x)^2 + (|b|-x)^2 (lower bound) */
mag_mul_lower(t, am, am);
mag_mul_lower(u, bm, bm);
mag_add_lower(t, t, u);
/* u = a^2 + b^2 (lower bound) */
arf_get_mag_lower(u, a2b2);
/* t = ((|a|-x)^2 + (|b|-x)^2)(a^2 + b^2) (lower bound) */
mag_mul_lower(t, t, u);
/* compute numerator */
/* real: |a^2-b^2| x + |a| ((x^2 + y^2) + 2 |b| y)) */
/* imag: |a^2-b^2| y + |b| ((x^2 + y^2) + 2 |a| x)) */
/* am, bm = upper bounds for a, b */
arf_get_mag(am, a);
arf_get_mag(bm, b);
/* v = x^2 + y^2 */
mag_mul(v, x, x);
mag_addmul(v, y, y);
/* u = |a| ((x^2 + y^2) + 2 |b| y) */
mag_mul_2exp_si(u, bm, 1);
mag_mul(u, u, y);
mag_add(u, u, v);
mag_mul(u, u, am);
/* v = |b| ((x^2 + y^2) + 2 |a| x) */
mag_mul_2exp_si(w, am, 1);
mag_addmul(v, w, x);
mag_mul(v, v, bm);
/* w = |b^2 - a^2| (upper bound) */
if (arf_cmpabs(a, b) >= 0)
mag_mul(w, am, am);
else
mag_mul(w, bm, bm);
mag_addmul(u, w, x);
mag_addmul(v, w, y);
mag_div(arb_radref(acb_realref(res)), u, t);
mag_div(arb_radref(acb_imagref(res)), v, t);
_arb_arf_div_rounded_den_add_err(acb_realref(res), a, a2b2, inexact, prec);
_arb_arf_div_rounded_den_add_err(acb_imagref(res), b, a2b2, inexact, prec);
arf_neg(arb_midref(acb_imagref(res)), arb_midref(acb_imagref(res)));
mag_clear(t);
mag_clear(u);
mag_clear(v);
mag_clear(w);
arf_clear(a2b2);
}
mag_clear(am);
mag_clear(bm);
#undef a
#undef b
#undef x
#undef y
}
void
arb_atan_arf(arb_t z, const arf_t x, slong prec)
{
if (arf_is_special(x))
{
if (arf_is_zero(x))
{
arb_zero(z);
}
else if (arf_is_pos_inf(x))
{
arb_const_pi(z, prec);
arb_mul_2exp_si(z, z, -1);
}
else if (arf_is_neg_inf(x))
{
arb_const_pi(z, prec);
arb_mul_2exp_si(z, z, -1);
arb_neg(z, z);
}
else
{
arb_indeterminate(z);
}
}
else if (COEFF_IS_MPZ(*ARF_EXPREF(x)))
{
if (fmpz_sgn(ARF_EXPREF(x)) < 0)
arb_atan_eps(z, x, prec);
else
arb_atan_inf_eps(z, x, prec);
}
else
{
slong exp, wp, wn, N, r;
mp_srcptr xp;
mp_size_t xn, tn;
mp_ptr tmp, w, t, u;
mp_limb_t p1, q1bits, p2, q2bits, error, error2;
int negative, inexact, reciprocal;
TMP_INIT;
exp = ARF_EXP(x);
negative = ARF_SGNBIT(x);
if (exp < -(prec/2) - 2 || exp > prec + 2)
{
if (exp < 0)
arb_atan_eps(z, x, prec);
else
arb_atan_inf_eps(z, x, prec);
return;
}
ARF_GET_MPN_READONLY(xp, xn, x);
/* Special case: +/- 1 (we require |x| != 1 later on) */
if (exp == 1 && xn == 1 && xp[xn-1] == LIMB_TOP)
{
arb_const_pi(z, prec);
arb_mul_2exp_si(z, z, -2);
if (negative)
arb_neg(z, z);
return;
}
/* Absolute working precision (NOT rounded to a limb multiple) */
wp = prec - FLINT_MIN(0, exp) + 4;
/* Too high precision to use table */
if (wp > ARB_ATAN_TAB2_PREC)
{
arb_atan_arf_bb(z, x, prec);
return;
}
/* Working precision in limbs */
wn = (wp + FLINT_BITS - 1) / FLINT_BITS;
TMP_START;
tmp = TMP_ALLOC_LIMBS(4 * wn + 3);
w = tmp; /* requires wn+1 limbs */
t = w + wn + 1; /* requires wn+1 limbs */
u = t + wn + 1; /* requires 2wn+1 limbs */
/* ----------------------------------------------------------------- */
/* Convert x or 1/x to a fixed-point number |w| < 1 */
/* ----------------------------------------------------------------- */
if (exp <= 0) /* |x| < 1 */
{
reciprocal = 0;
/* todo: just zero top */
flint_mpn_zero(w, wn);
/* w = x as a fixed-point number */
error = _arf_get_integer_mpn(w, xp, xn, exp + wn * FLINT_BITS);
}
else /* |x| > 1 */
{
slong one_exp, one_limbs, one_bits;
mp_ptr one;
reciprocal = 1;
one_exp = xn * FLINT_BITS + wn * FLINT_BITS - exp;
flint_mpn_zero(w, wn);
/* 1/x becomes zero */
if (one_exp >= FLINT_BITS - 1)
{
/* w = 1/x */
one_limbs = one_exp / FLINT_BITS;
one_bits = one_exp % FLINT_BITS;
if (one_limbs + 1 >= xn)
{
one = TMP_ALLOC_LIMBS(one_limbs + 1);
flint_mpn_zero(one, one_limbs);
one[one_limbs] = UWORD(1) << one_bits;
/* todo: only zero necessary part */
flint_mpn_zero(w, wn);
mpn_tdiv_q(w, one, one_limbs + 1, xp, xn);
/* Now w must be < 1 since x > 1 and we rounded down; thus
w[wn] must be zero */
}
}
/* todo: moderate powers of two would be exact... */
error = 1;
}
/* ----------------------------------------------------------------- */
/* Table-based argument reduction */
/* ----------------------------------------------------------------- */
/* choose p such that p/q <= x < (p+1)/q */
if (wp <= ARB_ATAN_TAB1_PREC)
q1bits = ARB_ATAN_TAB1_BITS;
else
q1bits = ARB_ATAN_TAB21_BITS;
p1 = w[wn-1] >> (FLINT_BITS - q1bits);
/* atan(w) = atan(p/q) + atan(w2) */
/* where w2 = (q*w-p)/(q+p*w) */
if (p1 != 0)
{
t[wn] = (UWORD(1) << q1bits) + mpn_mul_1(t, w, wn, p1);
flint_mpn_zero(u, wn);
u[2 * wn] = mpn_lshift(u + wn, w, wn, q1bits) - p1;
mpn_tdiv_q(w, u, 2 * wn + 1, t, wn + 1);
error++; /* w2 is computed with 1 ulp error */
}
/* Do a second round of argument reduction */
if (wp <= ARB_ATAN_TAB1_PREC)
{
p2 = 0;
}
else
{
q2bits = ARB_ATAN_TAB21_BITS + ARB_ATAN_TAB22_BITS;
p2 = w[wn-1] >> (FLINT_BITS - q2bits);
if (p2 != 0)
{
t[wn] = (UWORD(1) << q2bits) + mpn_mul_1(t, w, wn, p2);
flint_mpn_zero(u, wn);
u[2 * wn] = mpn_lshift(u + wn, w, wn, q2bits) - p2;
mpn_tdiv_q(w, u, 2 * wn + 1, t, wn + 1);
error++;
}
}
/* |w| <= 2^-r */
r = _arb_mpn_leading_zeros(w, wn);
/* N >= (wp-r)/(2r) */
N = (wp - r + (2*r-1)) / (2*r);
/* Evaluate Taylor series */
_arb_atan_taylor_rs(t, &error2, w, wn, N, 1);
/* Taylor series evaluation error */
error += error2;
/* Size of output number */
tn = wn;
/* First table lookup */
if (p1 != 0)
{
if (wp <= ARB_ATAN_TAB1_PREC)
mpn_add_n(t, t, arb_atan_tab1[p1] + ARB_ATAN_TAB1_LIMBS - tn, tn);
else
mpn_add_n(t, t, arb_atan_tab21[p1] + ARB_ATAN_TAB2_LIMBS - tn, tn);
error++;
}
/* Second table lookup */
if (p2 != 0)
{
mpn_add_n(t, t, arb_atan_tab22[p2] + ARB_ATAN_TAB2_LIMBS - tn, tn);
error++;
}
/* pi/2 - atan(1/x) */
if (reciprocal)
{
t[tn] = LIMB_ONE - mpn_sub_n(t,
arb_atan_pi2_minus_one + ARB_ATAN_TAB2_LIMBS - tn, t, tn);
/* result can be >= 1 */
tn += (t[tn] != 0);
/* error of pi/2 */
error++;
}
/* The accumulated arithmetic error */
mag_set_ui_2exp_si(arb_radref(z), error, -wn * FLINT_BITS);
/* Truncation error from the Taylor series */
mag_add_ui_2exp_si(arb_radref(z), arb_radref(z), 1, -r*(2*N+1));
/* Set the midpoint */
inexact = _arf_set_mpn_fixed(arb_midref(z), t, tn, wn, negative, prec, ARB_RND);
if (inexact)
arf_mag_add_ulp(arb_radref(z), arb_radref(z), arb_midref(z), prec);
TMP_END;
}
}
int main()
{
slong iter;
flint_rand_t state;
flint_printf("atan_arf_bb....");
fflush(stdout);
flint_randinit(state);
for (iter = 0; iter < 5000; iter++)
{
arb_t x, y, z;
slong prec, prec2;
arb_init(x);
arb_init(y);
arb_init(z);
prec = 2 + n_randint(state, 8000);
arb_randtest(x, state, 1 + n_randint(state, 8000), 3);
mag_zero(arb_radref(x));
if (n_randint(state, 2))
arb_mul_2exp_si(x, x, 1 + n_randint(state, 40));
else
arb_mul_2exp_si(x, x, -n_randint(state, 1.5 * prec));
if (!arf_is_special(arb_midref(x)))
prec2 = prec + 100 + 2 * (-ARF_EXP(arb_midref(x)));
else
prec2 = prec + 100;
arb_atan_arf_via_mpfr(y, arb_midref(x), prec2);
arb_atan_arf_bb(z, arb_midref(x), prec);
if (!arb_contains(z, y))
{
flint_printf("FAIL: containment\n\n");
flint_printf("prec = %wd\n\n", prec);
flint_printf("x = "); arb_printd(x, 50); flint_printf("\n\n");
flint_printf("y = "); arb_printd(y, 50); flint_printf("\n\n");
flint_printf("z = "); arb_printd(z, 50); flint_printf("\n\n");
abort();
}
if (arb_rel_accuracy_bits(z) < prec - 2)
{
flint_printf("FAIL: poor accuracy\n\n");
flint_printf("prec = %wd, acc = %wd\n\n", prec, arb_rel_accuracy_bits(z));
flint_printf("x = "); arb_printd(x, 50); flint_printf("\n\n");
flint_printf("y = "); arb_printd(y, 50); flint_printf("\n\n");
flint_printf("z = "); arb_printd(z, 50); flint_printf("\n\n");
abort();
}
arb_atan_arf_bb(x, arb_midref(x), prec);
if (!arb_overlaps(x, z))
{
flint_printf("FAIL: aliasing\n\n");
abort();
}
arb_clear(x);
arb_clear(y);
arb_clear(z);
}
flint_randclear(state);
flint_cleanup();
flint_printf("PASS\n");
return EXIT_SUCCESS;
}
int
arf_root(arf_ptr z, arf_srcptr x, ulong k, slong prec, arf_rnd_t rnd)
{
mp_size_t xn, zn, val;
mp_srcptr xptr;
mp_ptr tmp, zptr;
mpfr_t xf, zf;
fmpz_t q, r;
int inexact;
if (k == 0)
{
arf_nan(z);
return 0;
}
if (k == 1)
return arf_set_round(z, x, prec, rnd);
if (k == 2)
return arf_sqrt(z, x, prec, rnd);
if (arf_is_special(x))
{
if (arf_is_neg_inf(x))
arf_nan(z);
else
arf_set(z, x);
return 0;
}
if (ARF_SGNBIT(x))
{
arf_nan(z);
return 0;
}
fmpz_init(q);
fmpz_init(r);
/* x = m * 2^e where e = qk + r */
/* x^(1/k) = (m * 2^(qk+r))^(1/k) */
/* x^(1/k) = (m * 2^r)^(1/k) * 2^q */
fmpz_set_ui(r, k);
fmpz_fdiv_qr(q, r, ARF_EXPREF(x), r);
ARF_GET_MPN_READONLY(xptr, xn, x);
zn = (prec + FLINT_BITS - 1) / FLINT_BITS;
zf->_mpfr_d = tmp = flint_malloc(zn * sizeof(mp_limb_t));
zf->_mpfr_prec = prec;
zf->_mpfr_sign = 1;
zf->_mpfr_exp = 0;
xf->_mpfr_d = (mp_ptr) xptr;
xf->_mpfr_prec = xn * FLINT_BITS;
xf->_mpfr_sign = 1;
xf->_mpfr_exp = fmpz_get_ui(r);
inexact = mpfr_root(zf, xf, k, arf_rnd_to_mpfr(rnd));
inexact = (inexact != 0);
val = 0;
while (tmp[val] == 0)
val++;
ARF_GET_MPN_WRITE(zptr, zn - val, z);
flint_mpn_copyi(zptr, tmp + val, zn - val);
fmpz_add_si(ARF_EXPREF(z), q, zf->_mpfr_exp);
flint_free(tmp);
fmpz_clear(q);
fmpz_clear(r);
return inexact;
}
int main()
{
slong iter;
flint_rand_t state;
flint_printf("set_round....");
fflush(stdout);
flint_randinit(state);
{
arf_t x, y, z;
arf_init(x);
arf_init(y);
arf_init(z);
for (iter = 0; iter < 100000 * arb_test_multiplier(); iter++)
{
slong bits1, bits2;
int ret1, ret2;
mpfr_t g1, g2;
fmpz_t e;
arf_rnd_t rnd;
bits1 = 1 + n_randint(state, 1000);
bits2 = 2 + n_randint(state, 1000);
if (n_randint(state, 100) == 0)
bits2 = ARF_PREC_EXACT;
switch (n_randint(state, 5))
{
case 0: rnd = ARF_RND_DOWN; break;
case 1: rnd = ARF_RND_UP; break;
case 2: rnd = ARF_RND_FLOOR; break;
case 3: rnd = ARF_RND_CEIL; break;
default: rnd = ARF_RND_NEAR; break;
}
fmpz_init(e);
mpfr_init2(g1, FLINT_MAX(2, bits1));
mpfr_init2(g2, FLINT_MIN(bits2, 10000));
if (n_randint(state, 100) == 0)
{
arf_clear(x); arf_clear(y); arf_clear(z);
arf_init(x); arf_init(y); arf_init(z);
}
/* dirty output variables */
if (n_randint(state, 2))
{
arf_randtest_special(y, state, 1 + n_randint(state, 1000),
1 + n_randint(state, 100));
arf_randtest_special(z, state, 1 + n_randint(state, 1000),
1 + n_randint(state, 100));
}
arf_randtest_special(x, state, bits1, 1 + n_randint(state, 10));
arf_get_mpfr(g1, x, MPFR_RNDD); /* exact */
/* test large exponents */
if (n_randint(state, 4) == 0)
fmpz_randtest(e, state, 1 + n_randint(state, 100));
if (!arf_is_special(x))
fmpz_add(ARF_EXPREF(x), ARF_EXPREF(x), e);
ret1 = arf_set_round(y, x, bits2, rnd);
ret2 = mpfr_set(g2, g1, arf_rnd_to_mpfr(rnd));
arf_set_mpfr(z, g2);
if (!arf_is_special(y))
fmpz_sub(ARF_EXPREF(y), ARF_EXPREF(y), e);
if (!arf_equal(y, z) || ((ret1 == ARF_RESULT_EXACT) != (ret2 == 0)))
{
flint_printf("FAIL\n\n");
flint_printf("bits1: %wd\n", bits1);
flint_printf("bits2: %wd\n", bits2);
flint_printf("x = "); arf_print(x); flint_printf("\n\n");
flint_printf("y = "); arf_print(y); flint_printf("\n\n");
flint_printf("z = "); arf_print(z); flint_printf("\n\n");
flint_printf("ret1 = %d, ret2 = %d\n\n", ret1, ret2);
flint_abort();
}
if (!arf_is_special(x))
fmpz_add(ARF_EXPREF(x), ARF_EXPREF(x), e);
ret1 = arf_set_round(y, x, bits2, rnd);
arf_set(z, x);
ret2 = arf_set_round(z, z, bits2, rnd);
if (!arf_equal(y, z) || ret1 != ret2)
{
flint_printf("FAIL (aliasing)\n\n");
flint_printf("bits1: %wd\n", bits1);
flint_printf("bits2: %wd\n", bits2);
flint_printf("x = "); arf_print(x); flint_printf("\n\n");
flint_printf("y = "); arf_print(y); flint_printf("\n\n");
flint_printf("z = "); arf_print(z); flint_printf("\n\n");
flint_printf("ret1 = %d, ret2 = %d\n\n", ret1, ret2);
flint_abort();
}
mpfr_clear(g1);
mpfr_clear(g2);
fmpz_clear(e);
}
arf_clear(x);
arf_clear(y);
arf_clear(z);
}
flint_randclear(state);
flint_cleanup();
flint_printf("PASS\n");
return EXIT_SUCCESS;
}
void
arb_log_arf(arb_t z, const arf_t x, slong prec)
{
if (arf_is_special(x))
{
if (arf_is_pos_inf(x))
arb_pos_inf(z);
else
arb_indeterminate(z);
}
else if (ARF_SGNBIT(x))
{
arb_indeterminate(z);
}
else if (ARF_IS_POW2(x))
{
if (fmpz_is_one(ARF_EXPREF(x)))
{
arb_zero(z);
}
else
{
fmpz_t exp;
fmpz_init(exp);
_fmpz_add_fast(exp, ARF_EXPREF(x), -1);
arb_const_log2(z, prec + 2);
arb_mul_fmpz(z, z, exp, prec);
fmpz_clear(exp);
}
}
else if (COEFF_IS_MPZ(*ARF_EXPREF(x)))
{
arb_log_arf_huge(z, x, prec);
}
else
{
slong exp, wp, wn, N, r, closeness_to_one;
mp_srcptr xp;
mp_size_t xn, tn;
mp_ptr tmp, w, t, u;
mp_limb_t p1, q1bits, p2, q2bits, error, error2, cy;
int negative, inexact, used_taylor_series;
TMP_INIT;
exp = ARF_EXP(x);
negative = 0;
ARF_GET_MPN_READONLY(xp, xn, x);
/* compute a c >= 0 such that |x-1| <= 2^(-c) if c > 0 */
closeness_to_one = 0;
if (exp == 0)
{
slong i;
closeness_to_one = FLINT_BITS - FLINT_BIT_COUNT(~xp[xn - 1]);
if (closeness_to_one == FLINT_BITS)
{
for (i = xn - 2; i > 0 && xp[i] == LIMB_ONES; i--)
closeness_to_one += FLINT_BITS;
closeness_to_one += (FLINT_BITS - FLINT_BIT_COUNT(~xp[i]));
}
}
else if (exp == 1)
{
closeness_to_one = FLINT_BITS - FLINT_BIT_COUNT(xp[xn - 1] & (~LIMB_TOP));
if (closeness_to_one == FLINT_BITS)
{
slong i;
for (i = xn - 2; xp[i] == 0; i--)
closeness_to_one += FLINT_BITS;
closeness_to_one += (FLINT_BITS - FLINT_BIT_COUNT(xp[i]));
}
closeness_to_one--;
}
/* if |t-1| <= 0.5 */
/* |log(1+t) - t| <= t^2 */
/* |log(1+t) - (t-t^2/2)| <= t^3 */
if (closeness_to_one > prec + 1)
{
inexact = arf_sub_ui(arb_midref(z), x, 1, prec, ARB_RND);
mag_set_ui_2exp_si(arb_radref(z), 1, -2 * closeness_to_one);
if (inexact)
arf_mag_add_ulp(arb_radref(z), arb_radref(z), arb_midref(z), prec);
return;
}
else if (2 * closeness_to_one > prec + 1)
{
arf_t t, u;
arf_init(t);
arf_init(u);
arf_sub_ui(t, x, 1, ARF_PREC_EXACT, ARF_RND_DOWN);
arf_mul(u, t, t, ARF_PREC_EXACT, ARF_RND_DOWN);
arf_mul_2exp_si(u, u, -1);
inexact = arf_sub(arb_midref(z), t, u, prec, ARB_RND);
mag_set_ui_2exp_si(arb_radref(z), 1, -3 * closeness_to_one);
if (inexact)
arf_mag_add_ulp(arb_radref(z), arb_radref(z), arb_midref(z), prec);
arf_clear(t);
arf_clear(u);
return;
}
/* Absolute working precision (NOT rounded to a limb multiple) */
wp = prec + closeness_to_one + 5;
/* Too high precision to use table */
if (wp > ARB_LOG_TAB2_PREC)
{
arf_log_via_mpfr(arb_midref(z), x, prec, ARB_RND);
arf_mag_set_ulp(arb_radref(z), arb_midref(z), prec);
return;
}
/* Working precision in limbs */
wn = (wp + FLINT_BITS - 1) / FLINT_BITS;
TMP_START;
tmp = TMP_ALLOC_LIMBS(4 * wn + 3);
w = tmp; /* requires wn+1 limbs */
t = w + wn + 1; /* requires wn+1 limbs */
u = t + wn + 1; /* requires 2wn+1 limbs */
/* read x-1 */
if (xn <= wn)
{
flint_mpn_zero(w, wn - xn);
mpn_lshift(w + wn - xn, xp, xn, 1);
error = 0;
}
else
{
mpn_lshift(w, xp + xn - wn, wn, 1);
error = 1;
}
/* First table-based argument reduction */
if (wp <= ARB_LOG_TAB1_PREC)
q1bits = ARB_LOG_TAB11_BITS;
else
q1bits = ARB_LOG_TAB21_BITS;
p1 = w[wn-1] >> (FLINT_BITS - q1bits);
/* Special case: covers logarithms of small integers */
if (xn == 1 && (w[wn-1] == (p1 << (FLINT_BITS - q1bits))))
{
p2 = 0;
flint_mpn_zero(t, wn);
used_taylor_series = 0;
N = r = 0; /* silence compiler warning */
}
else
{
/* log(1+w) = log(1+p/q) + log(1 + (qw-p)/(p+q)) */
w[wn] = mpn_mul_1(w, w, wn, UWORD(1) << q1bits) - p1;
mpn_divrem_1(w, 0, w, wn + 1, p1 + (UWORD(1) << q1bits));
error += 1;
/* Second table-based argument reduction (fused with log->atanh
conversion) */
if (wp <= ARB_LOG_TAB1_PREC)
q2bits = ARB_LOG_TAB11_BITS + ARB_LOG_TAB12_BITS;
else
q2bits = ARB_LOG_TAB21_BITS + ARB_LOG_TAB22_BITS;
p2 = w[wn-1] >> (FLINT_BITS - q2bits);
u[2 * wn] = mpn_lshift(u + wn, w, wn, q2bits);
flint_mpn_zero(u, wn);
flint_mpn_copyi(t, u + wn, wn + 1);
t[wn] += p2 + (UWORD(1) << (q2bits + 1));
u[2 * wn] -= p2;
mpn_tdiv_q(w, u, 2 * wn + 1, t, wn + 1);
/* propagated error from 1 ulp error: 2 atanh'(1/3) = 2.25 */
error += 3;
/* |w| <= 2^-r */
r = _arb_mpn_leading_zeros(w, wn);
/* N >= (wp-r)/(2r) */
N = (wp - r + (2*r-1)) / (2*r);
N = FLINT_MAX(N, 0);
/* Evaluate Taylor series */
_arb_atan_taylor_rs(t, &error2, w, wn, N, 0);
/* Multiply by 2 */
mpn_lshift(t, t, wn, 1);
/* Taylor series evaluation error (multiply by 2) */
error += error2 * 2;
used_taylor_series = 1;
}
/* Size of output number */
tn = wn;
/* First table lookup */
if (p1 != 0)
{
if (wp <= ARB_LOG_TAB1_PREC)
mpn_add_n(t, t, arb_log_tab11[p1] + ARB_LOG_TAB1_LIMBS - tn, tn);
else
mpn_add_n(t, t, arb_log_tab21[p1] + ARB_LOG_TAB2_LIMBS - tn, tn);
error++;
}
/* Second table lookup */
if (p2 != 0)
{
if (wp <= ARB_LOG_TAB1_PREC)
mpn_add_n(t, t, arb_log_tab12[p2] + ARB_LOG_TAB1_LIMBS - tn, tn);
else
mpn_add_n(t, t, arb_log_tab22[p2] + ARB_LOG_TAB2_LIMBS - tn, tn);
error++;
}
/* add exp * log(2) */
exp--;
if (exp > 0)
{
cy = mpn_addmul_1(t, arb_log_log2_tab + ARB_LOG_TAB2_LIMBS - tn, tn, exp);
t[tn] = cy;
tn += (cy != 0);
error += exp;
}
else if (exp < 0)
{
t[tn] = 0;
u[tn] = mpn_mul_1(u, arb_log_log2_tab + ARB_LOG_TAB2_LIMBS - tn, tn, -exp);
if (mpn_cmp(t, u, tn + 1) >= 0)
{
mpn_sub_n(t, t, u, tn + 1);
}
else
{
mpn_sub_n(t, u, t, tn + 1);
negative = 1;
}
error += (-exp);
tn += (t[tn] != 0);
}
/* The accumulated arithmetic error */
mag_set_ui_2exp_si(arb_radref(z), error, -wn * FLINT_BITS);
/* Truncation error from the Taylor series */
if (used_taylor_series)
mag_add_ui_2exp_si(arb_radref(z), arb_radref(z), 1, -r*(2*N+1) + 1);
/* Set the midpoint */
inexact = _arf_set_mpn_fixed(arb_midref(z), t, tn, wn, negative, prec);
if (inexact)
arf_mag_add_ulp(arb_radref(z), arb_radref(z), arb_midref(z), prec);
TMP_END;
}
}
void
arf_get_fmpz(fmpz_t z, const arf_t x, arf_rnd_t rnd)
{
if (arf_is_special(x))
{
if (arf_is_zero(x))
{
fmpz_zero(z);
}
else
{
flint_printf("arf_get_fmpz: cannot convert infinity or nan to integer\n");
abort();
}
}
else if (COEFF_IS_MPZ(*ARF_EXPREF(x)))
{
/* tiny */
if (fmpz_sgn(ARF_EXPREF(x)) < 0)
{
int negative = ARF_SGNBIT(x);
if (rnd == ARF_RND_NEAR
|| rnd == ARF_RND_DOWN
|| (rnd == ARF_RND_FLOOR && !negative)
|| (rnd == ARF_RND_CEIL && negative))
{
fmpz_zero(z);
}
else
{
fmpz_set_si(z, negative ? -1 : 1);
}
}
else
{
flint_printf("arf_get_fmpz: number too large to convert to integer\n");
abort();
}
}
else
{
slong exp;
int negative, inexact;
mp_size_t xn, zn;
mp_srcptr xp;
__mpz_struct * zz;
/* TBD: implement efficiently */
if (rnd == ARF_RND_NEAR)
{
fmpr_t t;
fmpr_init(t);
arf_get_fmpr(t, x);
fmpr_get_fmpz(z, t, rnd);
fmpr_clear(t);
return;
}
exp = ARF_EXP(x);
negative = ARF_SGNBIT(x);
/* |x| < 1 */
if (exp <= 0)
{
if (rnd == ARF_RND_DOWN ||
(rnd == ARF_RND_FLOOR && !negative) ||
(rnd == ARF_RND_CEIL && negative))
{
fmpz_zero(z);
}
else
{
fmpz_set_si(z, negative ? -1 : 1);
}
return;
}
ARF_GET_MPN_READONLY(xp, xn, x);
/* |x| < 2^31 or 2^63 (must save 1 bit for rounding up!) */
if (exp < FLINT_BITS)
{
mp_limb_t v, v2;
v = xp[xn - 1];
v2 = v >> (FLINT_BITS - exp);
inexact = (xn > 1) || ((v2 << (FLINT_BITS - exp)) != v);
if (inexact && rnd != ARF_RND_DOWN)
{
if (negative && (rnd == ARF_RND_UP || rnd == ARF_RND_FLOOR))
v2++;
if (!negative && (rnd == ARF_RND_UP || rnd == ARF_RND_CEIL))
v2++;
}
if (negative)
fmpz_neg_ui(z, v2);
else
fmpz_set_ui(z, v2);
return;
}
/* |x| >= 1 */
zn = (exp + FLINT_BITS - 1) / FLINT_BITS;
zz = _fmpz_promote(z);
if (zz->_mp_alloc < zn)
mpz_realloc2(zz, zn * FLINT_BITS);
inexact = _arf_get_integer_mpn(zz->_mp_d, xp, xn, exp);
zz->_mp_size = negative ? -zn : zn;
_fmpz_demote_val(z);
if (inexact && rnd != ARF_RND_DOWN)
{
if (negative && (rnd == ARF_RND_UP || rnd == ARF_RND_FLOOR))
fmpz_sub_ui(z, z, 1);
if (!negative && (rnd == ARF_RND_UP || rnd == ARF_RND_CEIL))
fmpz_add_ui(z, z, 1);
}
}
void
arb_exp_arf_bb(arb_t z, const arf_t x, slong prec, int minus_one)
{
slong k, iter, bits, r, mag, q, wp, N;
slong argred_bits, start_bits;
mp_bitcnt_t Qexp[1];
int inexact;
fmpz_t t, u, T, Q;
arb_t w;
if (arf_is_zero(x))
{
if (minus_one)
arb_zero(z);
else
arb_one(z);
return;
}
if (arf_is_special(x))
{
abort();
}
mag = arf_abs_bound_lt_2exp_si(x);
/* We assume that this function only gets called with something
reasonable as input (huge/tiny input will be handled by
the main exp wrapper). */
if (mag > 200 || mag < -2 * prec - 100)
{
flint_printf("arb_exp_arf_bb: unexpectedly large/small input\n");
abort();
}
if (prec < 100000000)
{
argred_bits = 16;
start_bits = 32;
}
else
{
argred_bits = 32;
start_bits = 64;
}
/* Argument reduction: exp(x) -> exp(x/2^q). This improves efficiency
of the first iteration in the bit-burst algorithm. */
q = FLINT_MAX(0, mag + argred_bits);
/* Determine working precision. */
wp = prec + 10 + 2 * q + 2 * FLINT_BIT_COUNT(prec);
if (minus_one && mag < 0)
wp += (-mag);
fmpz_init(t);
fmpz_init(u);
fmpz_init(Q);
fmpz_init(T);
arb_init(w);
/* Convert x/2^q to a fixed-point number. */
inexact = arf_get_fmpz_fixed_si(t, x, -wp + q);
/* Aliasing of z and x is safe now that only use t. */
/* Start with z = 1. */
arb_one(z);
/* Bit-burst loop. */
for (iter = 0, bits = start_bits; !fmpz_is_zero(t);
iter++, bits *= 2)
{
/* Extract bits. */
r = FLINT_MIN(bits, wp);
fmpz_tdiv_q_2exp(u, t, wp - r);
/* Binary splitting (+1 fixed-point ulp truncation error). */
mag = fmpz_bits(u) - r;
N = bs_num_terms(mag, wp);
_arb_exp_sum_bs_powtab(T, Q, Qexp, u, r, N);
/* T = T / Q (+1 fixed-point ulp error). */
if (*Qexp >= wp)
{
fmpz_tdiv_q_2exp(T, T, *Qexp - wp);
fmpz_tdiv_q(T, T, Q);
}
else
{
fmpz_mul_2exp(T, T, wp - *Qexp);
fmpz_tdiv_q(T, T, Q);
}
/* T = 1 + T */
fmpz_one(Q);
fmpz_mul_2exp(Q, Q, wp);
fmpz_add(T, T, Q);
/* Now T = exp(u) with at most 2 fixed-point ulp error. */
/* Set z = z * T. */
arf_set_fmpz(arb_midref(w), T);
arf_mul_2exp_si(arb_midref(w), arb_midref(w), -wp);
mag_set_ui_2exp_si(arb_radref(w), 2, -wp);
arb_mul(z, z, w, wp);
/* Remove used bits. */
fmpz_mul_2exp(u, u, wp - r);
fmpz_sub(t, t, u);
}
/* We have exp(x + eps) - exp(x) < 2*eps (by assumption that the argument
reduction is large enough). */
if (inexact)
arb_add_error_2exp_si(z, -wp + 1);
fmpz_clear(t);
fmpz_clear(u);
fmpz_clear(Q);
fmpz_clear(T);
arb_clear(w);
/* exp(x) = exp(x/2^q)^(2^q) */
for (k = 0; k < q; k++)
arb_mul(z, z, z, wp);
if (minus_one)
arb_sub_ui(z, z, 1, wp);
arb_set_round(z, z, prec);
}