slong
fmpr_rsqrt(fmpr_t y, const fmpr_t x, slong prec, fmpr_rnd_t rnd)
{
slong r;
if (fmpr_is_special(x))
{
if (fmpr_is_zero(x))
fmpr_pos_inf(y);
else if (fmpr_is_pos_inf(x))
fmpr_zero(y);
else
fmpr_nan(y);
return FMPR_RESULT_EXACT;
}
if (fmpr_sgn(x) < 0)
{
fmpr_nan(y);
return FMPR_RESULT_EXACT;
}
/* special case: 4^n */
if (fmpz_is_one(fmpr_manref(x)) && fmpz_is_even(fmpr_expref(x)))
{
r = fmpr_set_round(y, x, prec, rnd);
fmpz_tdiv_q_2exp(fmpr_expref(y), fmpr_expref(y), 1);
fmpz_neg(fmpr_expref(y), fmpr_expref(y));
return r;
}
{
fmpr_t t;
fmpz_t e;
fmpr_init(t);
fmpz_init(e);
fmpz_neg(e, fmpr_expref(x));
if (fmpz_is_odd(e))
fmpz_add_ui(e, e, 1);
fmpr_mul_2exp_fmpz(t, x, e);
CALL_MPFR_FUNC(r, mpfr_rec_sqrt, y, t, prec, rnd);
fmpz_tdiv_q_2exp(e, e, 1);
fmpr_mul_2exp_fmpz(y, y, e);
fmpr_clear(t);
fmpz_clear(e);
return r;
}
}
void
fmpr_divappr_abs_ubound(fmpr_t z, const fmpr_t x, const fmpr_t y, slong prec)
{
if (fmpr_is_special(x) || fmpr_is_special(y) || fmpz_is_pm1(fmpr_manref(y)))
{
fmpr_div(z, x, y, prec, FMPR_RND_UP);
fmpr_abs(z, z);
}
else
{
fmpz_t t, u;
slong xbits, ybits, tbits, ubits, shift;
xbits = fmpz_bits(fmpr_manref(x));
ybits = fmpz_bits(fmpr_manref(y));
fmpz_init(t);
fmpz_init(u);
ubits = FLINT_MIN(ybits, prec);
tbits = prec + ubits + 1;
/* upper bound for |x|, shifted */
if (xbits <= tbits)
{
fmpz_mul_2exp(t, fmpr_manref(x), tbits - xbits);
fmpz_abs(t, t);
}
else if (fmpz_sgn(fmpr_manref(x)) > 0)
{
fmpz_cdiv_q_2exp(t, fmpr_manref(x), xbits - tbits);
}
else
{
fmpz_fdiv_q_2exp(t, fmpr_manref(x), xbits - tbits);
fmpz_neg(t, t);
}
/* lower bound for |y|, shifted */
if (ybits <= ubits)
fmpz_mul_2exp(u, fmpr_manref(y), ubits - ybits);
else
fmpz_tdiv_q_2exp(u, fmpr_manref(y), ybits - ubits);
fmpz_abs(u, u);
fmpz_cdiv_q(fmpr_manref(z), t, u);
shift = (ubits - ybits) - (tbits - xbits);
fmpz_sub(fmpr_expref(z), fmpr_expref(x), fmpr_expref(y));
if (shift >= 0)
fmpz_add_ui(fmpr_expref(z), fmpr_expref(z), shift);
else
fmpz_sub_ui(fmpr_expref(z), fmpr_expref(z), -shift);
_fmpr_normalise(fmpr_manref(z), fmpr_expref(z), prec, FMPR_RND_UP);
fmpz_clear(t);
fmpz_clear(u);
}
}
void
_fmpz_vec_scalar_tdiv_q_2exp(fmpz * vec1, const fmpz * vec2, slong len2,
ulong exp)
{
slong i;
for (i = 0; i < len2; i++)
fmpz_tdiv_q_2exp(vec1 + i, vec2 + i, exp);
}
void
mag_expinv(mag_t res, const mag_t x)
{
if (mag_is_zero(x))
{
mag_one(res);
}
else if (mag_is_inf(x))
{
mag_zero(res);
}
else if (fmpz_sgn(MAG_EXPREF(x)) <= 0)
{
mag_one(res);
}
else if (fmpz_cmp_ui(MAG_EXPREF(x), 2 * MAG_BITS) > 0)
{
fmpz_t t;
fmpz_init(t);
/* If x > 2^60, exp(-x) < 2^(-2^60 / log(2)) */
/* -1/log(2) < -369/256 */
fmpz_set_si(t, -369);
fmpz_mul_2exp(t, t, 2 * MAG_BITS - 8);
mag_one(res);
mag_mul_2exp_fmpz(res, res, t);
fmpz_clear(t);
}
else
{
fmpz_t t;
slong e = MAG_EXP(x);
fmpz_init(t);
fmpz_set_ui(t, MAG_MAN(x));
if (e >= MAG_BITS)
fmpz_mul_2exp(t, t, e - MAG_BITS);
else
fmpz_tdiv_q_2exp(t, t, MAG_BITS - e);
/* upper bound for 1/e */
mag_set_ui_2exp_si(res, 395007543, -30);
mag_pow_fmpz(res, res, t);
fmpz_clear(t);
}
}
int
main(void)
{
long iter;
int result;
flint_rand_t state;
printf("abs_ubound_ui_2exp....");
fflush(stdout);
flint_randinit(state);
for (iter = 0; iter < 100000; iter++)
{
fmpz_t x, y;
long bits, yexp;
long exp;
mp_limb_t man;
fmpz_init(x);
fmpz_init(y);
fmpz_randtest_not_zero(x, state, 1 + n_randint(state, 400));
bits = 1 + n_randint(state, FLINT_BITS - 1);
/* compute an exactly rounded mantissa */
fmpz_abs(y, x);
if (fmpz_is_zero(y))
{
yexp = 0;
}
else
{
yexp = fmpz_bits(y) - bits;
if (yexp >= 0)
{
fmpz_cdiv_q_2exp(y, y, yexp);
if (fmpz_bits(y) == bits + 1)
{
fmpz_tdiv_q_2exp(y, y, 1);
yexp--;
}
}
else
{
fmpz_mul_2exp(y, y, -yexp);
}
}
man = fmpz_abs_ubound_ui_2exp(&exp, x, bits);
if (FLINT_BIT_COUNT(man) != bits)
{
printf("wrong number of bits!\n");
printf("bits = %ld, count = %u\n\n", bits, FLINT_BIT_COUNT(man));
printf("x = "); fmpz_print(x); printf("\n\n");
printf("bits(x) = %ld\n\n", fmpz_bits(x));
printf("y = "); fmpz_print(y); printf("\n\n");
printf("yexp = %ld\n\n", yexp);
printf("man = %lu, exp = %ld\n", man, exp);
abort();
}
/* ok if equal */
result = (fmpz_cmp_ui(y, man) == 0);
/* ok if mantissa is 1 larger */
if (!result)
{
result = ((exp == yexp) && (fmpz_cmp_ui(y, man - 1) == 0));
}
/* ok if the exact mantissa is 2^r-1 and overflow to 2^r happened */
if (!result)
{
fmpz_t t;
fmpz_init(t);
fmpz_set_ui(t, man);
fmpz_mul_ui(t, t, 2);
fmpz_sub_ui(t, t, 1);
result = (exp == yexp + 1) && fmpz_equal(t, y);
fmpz_clear(t);
}
if (!result)
{
printf("different from exact ceiling division\n");
printf("bits = %ld\n\n", bits);
printf("x = "); fmpz_print(x); printf("\n\n");
printf("bits(x) = %ld\n\n", fmpz_bits(x));
printf("y = "); fmpz_print(y); printf(", yexp = %ld\n\n", yexp);
printf("man = %lu, exp = %ld\n", man, exp);
abort();
}
fmpz_clear(x);
fmpz_clear(y);
}
flint_randclear(state);
_fmpz_cleanup();
printf("PASS\n");
return 0;
}
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);
}
