void
fmpz_poly_bit_unpack_unsigned(fmpz_poly_t poly, const fmpz_t f,
mp_bitcnt_t bit_size)
{
slong len;
mpz_t tmp;
if (fmpz_sgn(f) < 0)
{
flint_printf("Exception (fmpz_poly_bit_unpack_unsigned). Expected an unsigned value.\n");
abort();
}
if (bit_size == 0 || fmpz_is_zero(f))
{
fmpz_poly_zero(poly);
return;
}
len = (fmpz_bits(f) + bit_size - 1) / bit_size;
mpz_init2(tmp, bit_size*len);
flint_mpn_zero(tmp->_mp_d, tmp->_mp_alloc);
fmpz_get_mpz(tmp, f);
fmpz_poly_fit_length(poly, len);
_fmpz_poly_bit_unpack_unsigned(poly->coeffs, len, tmp->_mp_d, bit_size);
_fmpz_poly_set_length(poly, len);
_fmpz_poly_normalise(poly);
mpz_clear(tmp);
}
void _nmod_poly_div_newton(mp_ptr Q, mp_srcptr A, slong lenA,
mp_srcptr B, slong lenB, nmod_t mod)
{
const slong lenQ = lenA - lenB + 1;
mp_ptr Arev, Brev;
Arev = _nmod_vec_init(2 * lenQ);
Brev = Arev + lenQ;
_nmod_poly_reverse(Arev, A + (lenA - lenQ), lenQ, lenQ);
if (lenB >= lenQ)
{
_nmod_poly_reverse(Brev, B + (lenB - lenQ), lenQ, lenQ);
}
else
{
_nmod_poly_reverse(Brev, B, lenB, lenB);
flint_mpn_zero(Brev + lenB, lenQ - lenB);
}
_nmod_poly_div_series(Q, Arev, Brev, lenQ, mod);
_nmod_poly_reverse(Q, Q, lenQ, lenQ);
_nmod_vec_clear(Arev);
}
/* truncates, returns inexact */
int
_arf_get_integer_mpn(mp_ptr y, mp_srcptr x, mp_size_t xn, slong exp)
{
slong bot_exp = exp - xn * FLINT_BITS;
if (bot_exp >= 0)
{
mp_size_t bot_limbs;
mp_bitcnt_t bot_bits;
bot_limbs = bot_exp / FLINT_BITS;
bot_bits = bot_exp % FLINT_BITS;
flint_mpn_zero(y, bot_limbs);
if (bot_bits == 0)
flint_mpn_copyi(y + bot_limbs, x, xn);
else
y[bot_limbs + xn] = mpn_lshift(y + bot_limbs, x, xn, bot_bits);
/* exact */
return 0;
}
else if (exp <= 0)
{
/* inexact */
return 1;
}
else
{
mp_size_t top_limbs;
mp_bitcnt_t top_bits;
mp_limb_t cy;
top_limbs = exp / FLINT_BITS;
top_bits = exp % FLINT_BITS;
if (top_bits == 0)
{
flint_mpn_copyi(y, x + xn - top_limbs, top_limbs);
/* inexact */
return 1;
}
else
{
/* can be inexact */
cy = mpn_rshift(y, x + xn - top_limbs - 1,
top_limbs + 1, FLINT_BITS - top_bits);
return (cy != 0) || (top_limbs + 1 != xn);
}
}
}
/* computes x + y * 2^shift (optionally negated) */
slong
_fmpr_add_mpn(fmpr_t z,
mp_srcptr xman, mp_size_t xn, int xsign, const fmpz_t xexp,
mp_srcptr yman, mp_size_t yn, int ysign, const fmpz_t yexp,
slong shift, slong prec, fmpr_rnd_t rnd)
{
slong tn, zn, alloc, ret, shift_bits, shift_limbs;
int negative;
mp_limb_t tmp_stack[ADD_STACK_ALLOC];
mp_limb_t cy;
mp_ptr tmp, tmp2;
shift_limbs = shift / FLINT_BITS;
shift_bits = shift % FLINT_BITS;
/* x does not overlap with y or the result -- outcome is
equivalent to adding/subtracting a small number to/from y
and rounding */
if (shift > xn * FLINT_BITS &&
prec != FMPR_PREC_EXACT &&
xn * FLINT_BITS + prec - (FLINT_BITS * (yn - 1)) < shift)
{
zn = (prec + FLINT_BITS - 1) / FLINT_BITS;
zn = FLINT_MAX(zn, yn) + 2;
shift_limbs = zn - yn;
alloc = zn;
ADD_TMP_ALLOC
flint_mpn_zero(tmp, shift_limbs);
flint_mpn_copyi(tmp + shift_limbs, yman, yn);
if (xsign == ysign)
{
tmp[0] = 1;
}
else
{
mpn_sub_1(tmp, tmp, zn, 1);
while (tmp[zn-1] == 0)
zn--;
}
ret = _fmpr_set_round_mpn(&shift, fmpr_manref(z), tmp, zn, ysign, prec, rnd);
shift -= shift_limbs * FLINT_BITS;
fmpz_add_si_inline(fmpr_expref(z), yexp, shift);
ADD_TMP_FREE
return ret;
}
void
fmpz_poly_bit_unpack(fmpz_poly_t poly, const fmpz_t f, mp_bitcnt_t bit_size)
{
slong len;
mpz_t tmp;
int negate, borrow;
if (bit_size == 0 || fmpz_is_zero(f))
{
fmpz_poly_zero(poly);
return;
}
/* Round up */
len = (fmpz_bits(f) + bit_size - 1) / bit_size;
negate = (fmpz_sgn(f) < 0) ? -1 : 0;
mpz_init2(tmp, bit_size*len);
/* TODO: avoid all this wastefulness */
flint_mpn_zero(tmp->_mp_d, tmp->_mp_alloc);
fmpz_get_mpz(tmp, f);
fmpz_poly_fit_length(poly, len + 1);
borrow = _fmpz_poly_bit_unpack(poly->coeffs, len,
tmp->_mp_d, bit_size, negate);
if (borrow)
{
fmpz_set_si(poly->coeffs + len, negate ? WORD(-1) : WORD(1));
_fmpz_poly_set_length(poly, len + 1);
}
else
{
_fmpz_poly_set_length(poly, len);
_fmpz_poly_normalise(poly);
}
mpz_clear(tmp);
}
void
nmod_poly_asin_series(nmod_poly_t g, const nmod_poly_t h, slong n)
{
mp_ptr h_coeffs;
slong h_len = h->length;
if (h_len > 0 && h->coeffs[0] != UWORD(0))
{
flint_printf("Exception (nmod_poly_asin_series). Constant term != 0.\n");
abort();
}
if (h_len == 1 || n < 2)
{
nmod_poly_zero(g);
return;
}
nmod_poly_fit_length(g, n);
if (h_len < n)
{
h_coeffs = _nmod_vec_init(n);
flint_mpn_copyi(h_coeffs, h->coeffs, h_len);
flint_mpn_zero(h_coeffs + h_len, n - h_len);
}
else
h_coeffs = h->coeffs;
_nmod_poly_asin_series(g->coeffs, h_coeffs, n, h->mod);
if (h_len < n)
_nmod_vec_clear(h_coeffs);
g->length = n;
_nmod_poly_normalise(g);
}
slong _nmod_poly_xgcd_hgcd(mp_ptr G, mp_ptr S, mp_ptr T,
mp_srcptr A, slong lenA, mp_srcptr B, slong lenB,
nmod_t mod)
{
const slong cutoff = FLINT_BIT_COUNT(mod.n) <= 8 ?
NMOD_POLY_SMALL_GCD_CUTOFF : NMOD_POLY_GCD_CUTOFF;
slong lenG, lenS, lenT;
if (lenB == 1)
{
G[0] = B[0];
T[0] = 1;
lenG = 1;
lenS = 0;
lenT = 1;
}
else
{
mp_ptr q = _nmod_vec_init(lenA + lenB);
mp_ptr r = q + lenA;
slong lenq, lenr;
__divrem(q, lenq, r, lenr, A, lenA, B, lenB);
if (lenr == 0)
{
__set(G, lenG, B, lenB);
T[0] = 1;
lenS = 0;
lenT = 1;
}
else
{
mp_ptr h, j, v, w, R[4], X;
slong lenh, lenj, lenv, lenw, lenR[4];
int sgnR;
lenh = lenj = lenB;
lenv = lenw = lenA + lenB - 2;
lenR[0] = lenR[1] = lenR[2] = lenR[3] = (lenB + 1) / 2;
X = _nmod_vec_init(2 * lenh + 2 * lenv + 4 * lenR[0]);
h = X;
j = h + lenh;
v = j + lenj;
w = v + lenv;
R[0] = w + lenw;
R[1] = R[0] + lenR[0];
R[2] = R[1] + lenR[1];
R[3] = R[2] + lenR[2];
sgnR = _nmod_poly_hgcd(R, lenR, h, &lenh, j, &lenj, B, lenB, r, lenr, mod);
if (sgnR > 0)
{
_nmod_vec_neg(S, R[1], lenR[1], mod);
_nmod_vec_set(T, R[0], lenR[0]);
}
else
{
_nmod_vec_set(S, R[1], lenR[1]);
_nmod_vec_neg(T, R[0], lenR[0], mod);
}
lenS = lenR[1];
lenT = lenR[0];
while (lenj != 0)
{
__divrem(q, lenq, r, lenr, h, lenh, j, lenj);
__mul(v, lenv, q, lenq, T, lenT);
{
slong l;
_nmod_vec_swap(S, T, FLINT_MAX(lenS, lenT));
l = lenS; lenS = lenT; lenT = l;
}
__sub(T, lenT, T, lenT, v, lenv);
if (lenr == 0)
{
__set(G, lenG, j, lenj);
goto cofactor;
}
if (lenj < cutoff)
{
mp_ptr u0 = R[0], u1 = R[1];
slong lenu0 = lenr - 1, lenu1 = lenj - 1;
lenG = _nmod_poly_xgcd_euclidean(G, u0, u1, j, lenj, r, lenr, mod);
MPN_NORM(u0, lenu0);
MPN_NORM(u1, lenu1);
__mul(v, lenv, S, lenS, u0, lenu0);
__mul(w, lenw, T, lenT, u1, lenu1);
__add(S, lenS, v, lenv, w, lenw);
goto cofactor;
}
sgnR = _nmod_poly_hgcd(R, lenR, h, &lenh, j, &lenj, j,lenj, r, lenr, mod);
__mul(v, lenv, R[1], lenR[1], T, lenT);
__mul(w, lenw, R[2], lenR[2], S, lenS);
__mul(q, lenq, S, lenS, R[3], lenR[3]);
if (sgnR > 0)
__sub(S, lenS, q, lenq, v, lenv);
else
__sub(S, lenS, v, lenv, q, lenq);
__mul(q, lenq, T, lenT, R[0], lenR[0]);
if (sgnR > WORD(0))
__sub(T, lenT, q, lenq, w, lenw);
else
__sub(T, lenT, w, lenw, q, lenq);
}
__set(G, lenG, h, lenh);
cofactor:
__mul(v, lenv, S, lenS, A, lenA);
__sub(w, lenw, G, lenG, v, lenv);
__div(T, lenT, w, lenw, B, lenB);
_nmod_vec_clear(X);
}
_nmod_vec_clear(q);
}
flint_mpn_zero(S + lenS, lenB - 1 - lenS);
flint_mpn_zero(T + lenT, lenA - 1 - lenT);
return lenG;
}
void
_nmod_poly_divrem_divconquer_recursive(mp_ptr Q, mp_ptr BQ, mp_ptr W, mp_ptr V,
mp_srcptr A, mp_srcptr B, slong lenB, nmod_t mod)
{
if (lenB <= NMOD_DIVREM_DIVCONQUER_CUTOFF)
{
mp_ptr t = V;
mp_ptr w = t + 2*lenB - 1;
flint_mpn_copyi(t + lenB - 1, A + lenB - 1, lenB);
flint_mpn_zero(t, lenB - 1);
_nmod_poly_divrem_basecase(Q, BQ, w, t, 2 * lenB - 1, B, lenB, mod);
/* BQ = A - R */
_nmod_vec_neg(BQ, BQ, lenB - 1, mod);
}
else
{
const slong n2 = lenB / 2;
const slong n1 = lenB - n2;
mp_ptr W1 = W;
mp_ptr W2 = W + n2;
mp_srcptr p1 = A + 2 * n2;
mp_srcptr p2;
mp_srcptr d1 = B + n2;
mp_srcptr d2 = B;
mp_srcptr d3 = B + n1;
mp_srcptr d4 = B;
mp_ptr q1 = Q + n2;
mp_ptr q2 = Q;
mp_ptr dq1 = BQ + n2;
mp_ptr d1q1 = BQ + n2 - (n1 - 1);
mp_ptr d2q1, d3q2, d4q2, t;
/*
Set q1 to p1 div d1, a 2 n1 - 1 by n1 division so q1 ends up
being of length n1; low(d1q1) = d1 q1 is of length n1 - 1
*/
_nmod_poly_divrem_divconquer_recursive(q1, d1q1, W1, V, p1, d1, n1, mod);
/*
Compute bottom n1 + n2 - 1 coeffs of d2q1 = d2 q1
*/
d2q1 = W1;
_nmod_poly_mullow(d2q1, q1, n1, d2, n2, n1 + n2 - 1, mod);
/*
Compute dq1 = d1 q1 x^n2 + d2 q1, of length n1 + n2 - 1
Split it into a segment of length n1 - 1 at dq1 and a piece
of length n2 at BQ.
*/
flint_mpn_copyi(dq1, d2q1, n1 - 1);
if (n2 > n1 - 1)
BQ[0] = d2q1[n1 - 1];
_nmod_vec_add(d1q1, d1q1, d2q1 + n2, n1 - 1, mod);
/*
Compute t = A/x^n2 - dq1, which has length 2 n1 + n2 - 1, but we
are not interested in the top n1 coeffs as they will be zero, so
this has effective length n1 + n2 - 1
For the following division, we want to set {p2, 2 n2 - 1} to the
top 2 n2 - 1 coeffs of this
Since the bottom n2 - 1 coeffs of p2 are irrelevant for the
division, we in fact set {t, n2} to the relevant coeffs
*/
t = W1;
_nmod_vec_sub(t, A + n2 + (n1 - 1), BQ, n2, mod);
p2 = t - (n2 - 1);
/*
Compute q2 = t div d3, a 2 n2 - 1 by n2 division, so q2 will have
length n2; let low(d3q2) = d3 q2, of length n2 - 1
*/
d3q2 = BQ;
_nmod_poly_divrem_divconquer_recursive(q2, d3q2, W2, V, p2, d3, n2, mod);
/*
Compute d4q2 = d4 q2, of length n1 + n2 - 1
*/
d4q2 = W1;
_nmod_poly_mullow(d4q2, d4, n1, q2, n2, n1 + n2 - 1, mod);
/*
Compute dq2 = d3q2 x^n1 + d4q2, of length n1 + n2 - 1
*/
_nmod_vec_add(BQ + n1, BQ + n1, d3q2, n2 - 1, mod);
flint_mpn_copyi(BQ, d4q2, n2);
_nmod_vec_add(BQ + n2, BQ + n2, d4q2 + n2, n1 - 1, mod);
/*
Note Q = q1 x^n2 + q2, and BQ = dq1 x^n2 + dq2
*/
}
}
void _arb_sin_cos_taylor_rs(mp_ptr ysin, mp_ptr ycos,
mp_limb_t * error, mp_srcptr x, mp_size_t xn, ulong N,
int sinonly, int alternating)
{
mp_ptr s, t, xpow;
mp_limb_t new_denom, old_denom, c;
slong power, k, m;
int cosorsin;
TMP_INIT;
TMP_START;
if (2 * N >= FACTORIAL_TAB_SIZE - 1)
{
flint_printf("_arb_sin_cos_taylor_rs: N too large!\n");
abort();
}
if (N <= 1)
{
if (N == 0)
{
flint_mpn_zero(ysin, xn);
if (!sinonly) flint_mpn_zero(ycos, xn);
error[0] = 0;
}
else if (N == 1)
{
flint_mpn_copyi(ysin, x, xn);
if (!sinonly) flint_mpn_store(ycos, xn, LIMB_ONES);
error[0] = 1;
}
}
else
{
/* Choose m ~= sqrt(num_terms) (m must be even, >= 2) */
m = 2;
while (m * m < N)
m += 2;
/* todo: merge allocations */
xpow = TMP_ALLOC_LIMBS((m + 1) * xn);
s = TMP_ALLOC_LIMBS(xn + 2);
t = TMP_ALLOC_LIMBS(2 * xn + 2); /* todo: 1 limb too much? */
/* higher index ---> */
/* | ---xn--- | */
/* xpow = | <temp> | x^m | x^(m-1) | ... | x^2 | x | */
#define XPOW_WRITE(__k) (xpow + (m - (__k)) * xn)
#define XPOW_READ(__k) (xpow + (m - (__k) + 1) * xn)
mpn_sqr(XPOW_WRITE(1), x, xn);
mpn_sqr(XPOW_WRITE(2), XPOW_READ(1), xn);
for (k = 4; k <= m; k += 2)
{
mpn_mul_n(XPOW_WRITE(k - 1), XPOW_READ(k / 2), XPOW_READ(k / 2 - 1), xn);
mpn_sqr(XPOW_WRITE(k), XPOW_READ(k / 2), xn);
}
for (cosorsin = sinonly; cosorsin < 2; cosorsin++)
{
flint_mpn_zero(s, xn + 1);
/* todo: skip one nonscalar multiplication (use x^m)
when starting on x^0 */
power = (N - 1) % m;
for (k = N - 1; k >= 0; k--)
{
c = factorial_tab_numer[2 * k + cosorsin];
new_denom = factorial_tab_denom[2 * k + cosorsin];
old_denom = factorial_tab_denom[2 * k + cosorsin + 2];
/* change denominators */
if (new_denom != old_denom && k < N - 1)
{
if (alternating && (k % 2 == 0))
s[xn] += old_denom;
mpn_divrem_1(s, 0, s, xn + 1, old_denom);
if (alternating && (k % 2 == 0))
s[xn] -= 1;
}
if (power == 0)
{
/* add c * x^0 -- only top limb is affected */
if (alternating & k)
s[xn] -= c;
else
s[xn] += c;
/* Outer polynomial evaluation: multiply by x^m */
if (k != 0)
{
mpn_mul(t, s, xn + 1, XPOW_READ(m), xn);
flint_mpn_copyi(s, t + xn, xn + 1);
}
power = m - 1;
}
else
{
if (alternating & k)
s[xn] -= mpn_submul_1(s, XPOW_READ(power), xn, c);
else
s[xn] += mpn_addmul_1(s, XPOW_READ(power), xn, c);
power--;
}
}
/* finally divide by denominator */
if (cosorsin == 0)
{
mpn_divrem_1(t, 0, s, xn + 1, factorial_tab_denom[0]);
/* perturb down to a number < 1 if necessary. note that this
does not invalidate the error bound: 1 - ulp is either
1 ulp too small or must be closer to the exact value */
if (t[xn] == 0)
flint_mpn_copyi(ycos, t, xn);
else
flint_mpn_store(ycos, xn, LIMB_ONES);
}
else
{
mpn_divrem_1(s, 0, s, xn + 1, factorial_tab_denom[0]);
mpn_mul(t, s, xn + 1, x, xn);
flint_mpn_copyi(ysin, t + xn, xn);
}
}
/* error bound (ulp) */
error[0] = 2;
}
TMP_END;
}
void _arb_exp_taylor_rs(mp_ptr y, mp_limb_t * error,
mp_srcptr x, mp_size_t xn, ulong N)
{
mp_ptr s, t, xpow;
mp_limb_t new_denom, old_denom, c;
slong power, k, m;
TMP_INIT;
TMP_START;
if (N >= FACTORIAL_TAB_SIZE - 1)
{
flint_printf("_arb_exp_taylor_rs: N too large!\n");
abort();
}
if (N <= 3)
{
if (N <= 1)
{
flint_mpn_zero(y, xn);
y[xn] = N;
error[0] = 0;
}
else if (N == 2)
{
flint_mpn_copyi(y, x, xn);
y[xn] = 1;
error[0] = 0;
}
else
{
/* 1 + x + x^2 / 2 */
t = TMP_ALLOC_LIMBS(2 * xn);
mpn_sqr(t, x, xn);
mpn_rshift(t + xn, t + xn, xn, 1);
y[xn] = mpn_add_n(y, x, t + xn, xn) + 1;
error[0] = 2;
}
}
else
{
/* Choose m ~= sqrt(num_terms) (m must be even, >= 2) */
/* TODO: drop evenness assumption since we don't have sign issues here? */
/* TODO: then just need to fix power construction below... */
m = 2;
while (m * m < N)
m += 2;
/* todo: merge allocations */
xpow = TMP_ALLOC_LIMBS((m + 1) * xn);
s = TMP_ALLOC_LIMBS(xn + 2);
t = TMP_ALLOC_LIMBS(2 * xn + 2); /* todo: 1 limb too much? */
/* higher index ---> */
/* | ---xn--- | */
/* xpow = | <temp> | x^m | x^(m-1) | ... | x^2 | x | */
#define XPOW_WRITE(__k) (xpow + (m - (__k)) * xn)
#define XPOW_READ(__k) (xpow + (m - (__k) + 1) * xn)
flint_mpn_copyi(XPOW_READ(1), x, xn);
mpn_sqr(XPOW_WRITE(2), XPOW_READ(1), xn);
for (k = 4; k <= m; k += 2)
{
mpn_mul_n(XPOW_WRITE(k - 1), XPOW_READ(k / 2), XPOW_READ(k / 2 - 1), xn);
mpn_sqr(XPOW_WRITE(k), XPOW_READ(k / 2), xn);
}
flint_mpn_zero(s, xn + 1);
/* todo: skip one nonscalar multiplication (use x^m)
when starting on x^0 */
power = (N - 1) % m;
for (k = N - 1; k >= 0; k--)
{
c = factorial_tab_numer[k];
new_denom = factorial_tab_denom[k];
old_denom = factorial_tab_denom[k+1];
/* change denominators */
if (new_denom != old_denom && k < N - 1)
{
mpn_divrem_1(s, 0, s, xn + 1, old_denom);
}
if (power == 0)
{
/* add c * x^0 -- only top limb is affected */
s[xn] += c;
/* Outer polynomial evaluation: multiply by x^m */
if (k != 0)
{
mpn_mul(t, s, xn + 1, XPOW_READ(m), xn);
flint_mpn_copyi(s, t + xn, xn + 1);
}
power = m - 1;
}
else
{
s[xn] += mpn_addmul_1(s, XPOW_READ(power), xn, c);
power--;
}
}
/* finally divide by denominator */
mpn_divrem_1(y, 0, s, xn + 1, factorial_tab_denom[0]);
/* error bound (ulp) */
error[0] = 2;
}
TMP_END;
}
void
_arb_atan_taylor_naive(mp_ptr y, mp_limb_t * error,
mp_srcptr x, mp_size_t xn, ulong N, int alternating)
{
ulong k;
mp_ptr s, t, x1, x2, u;
mp_size_t nn = xn + 1;
if (N == 0)
{
flint_mpn_zero(y, xn);
error[0] = 0;
return;
}
if (N == 1)
{
flint_mpn_copyi(y, x, xn);
error[0] = 0;
}
s = flint_malloc(sizeof(mp_limb_t) * nn);
t = flint_malloc(sizeof(mp_limb_t) * nn);
u = flint_malloc(sizeof(mp_limb_t) * 2 * nn);
x1 = flint_malloc(sizeof(mp_limb_t) * nn);
x2 = flint_malloc(sizeof(mp_limb_t) * nn);
flint_mpn_zero(s, nn);
flint_mpn_zero(t, nn);
flint_mpn_zero(u, 2 * nn);
flint_mpn_zero(x1, nn);
flint_mpn_zero(x2, nn);
/* x1 = x */
flint_mpn_copyi(x1 + 1, x, xn);
/* x2 = x * x */
mpn_mul_n(u, x1, x1, nn);
flint_mpn_copyi(x2, u + nn, nn);
/* s = t = x */
flint_mpn_copyi(s, x1, nn);
flint_mpn_copyi(t, x1, nn);
for (k = 1; k < N; k++)
{
/* t = t * x2 */
mpn_mul_n(u, t, x2, nn);
flint_mpn_copyi(t, u + nn, nn);
/* u = t / (2k+1) */
mpn_divrem_1(u, 0, t, nn, 2 * k + 1);
if (alternating & k)
mpn_sub_n(s, s, u, nn);
else
mpn_add_n(s, s, u, nn);
}
flint_mpn_copyi(y, s + 1, xn);
error[0] = 2;
flint_free(s);
flint_free(t);
flint_free(u);
flint_free(x1);
flint_free(x2);
}
int
_arb_get_mpn_fixed_mod_log2(mp_ptr w, fmpz_t q, mp_limb_t * error,
const arf_t x, mp_size_t wn)
{
mp_srcptr xp;
mp_size_t xn;
int negative;
long exp;
ARF_GET_MPN_READONLY(xp, xn, x);
exp = ARF_EXP(x);
negative = ARF_SGNBIT(x);
if (exp <= -1)
{
/* todo: just zero top */
flint_mpn_zero(w, wn);
*error = _arf_get_integer_mpn(w, xp, xn, exp + wn * FLINT_BITS);
if (!negative)
{
fmpz_zero(q);
}
else
{
if (wn > ARB_LOG_TAB2_LIMBS)
return 0;
mpn_sub_n(w, arb_log_log2_tab + ARB_LOG_TAB2_LIMBS - wn, w, wn);
*error += 1; /* log(2) has 1 ulp error */
fmpz_set_si(q, -1);
}
return 1; /* success */
}
else
{
mp_ptr qp, rp, np;
mp_srcptr dp;
mp_size_t qn, rn, nn, dn, tn, alloc;
TMP_INIT;
tn = ((exp + 2) + FLINT_BITS - 1) / FLINT_BITS;
dn = wn + tn; /* denominator */
nn = wn + 2 * tn; /* numerator */
qn = nn - dn + 1; /* quotient */
rn = dn; /* remainder */
if (dn > ARB_LOG_TAB2_LIMBS)
return 0;
TMP_START;
alloc = qn + rn + nn;
qp = TMP_ALLOC_LIMBS(alloc);
rp = qp + qn;
np = rp + rn;
dp = arb_log_log2_tab + ARB_LOG_TAB2_LIMBS - dn;
/* todo: prove that zeroing is unnecessary */
flint_mpn_zero(np, nn);
_arf_get_integer_mpn(np, xp, xn, exp + dn * FLINT_BITS);
mpn_tdiv_qr(qp, rp, 0, np, nn, dp, dn);
if (!negative)
{
flint_mpn_copyi(w, rp + tn, wn);
*error = 2;
}
else
{
if (mpn_add_1(qp, qp, qn, 1))
{
/* I believe this cannot happen (should prove it) */
printf("mod log(2): unexpected carry\n");
abort();
}
mpn_sub_n(w, dp + tn, rp + tn, wn);
*error = 3;
}
/* read the exponent */
while (qn > 1 && qp[qn-1] == 0)
qn--;
if (qn == 1)
{
if (!negative)
fmpz_set_ui(q, qp[0]);
else
fmpz_neg_ui(q, qp[0]);
}
else
{
fmpz_set_mpn_large(q, qp, qn, negative);
}
TMP_END;
return 1;
}
}
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;
}
}
void mul_mfa_truncate_sqrt2(mp_ptr r1, mp_srcptr i1, mp_size_t n1,
mp_srcptr i2, mp_size_t n2, mp_bitcnt_t depth, mp_bitcnt_t w)
{
mp_size_t n = (UWORD(1)<<depth);
mp_bitcnt_t bits1 = (n*w - (depth+1))/2;
mp_size_t sqrt = (UWORD(1)<<(depth/2));
mp_size_t r_limbs = n1 + n2;
mp_size_t limbs = (n*w)/FLINT_BITS;
mp_size_t size = limbs + 1;
mp_size_t j1 = (n1*FLINT_BITS - 1)/bits1 + 1;
mp_size_t j2 = (n2*FLINT_BITS - 1)/bits1 + 1;
mp_size_t i, j, trunc;
mp_limb_t ** ii, ** jj, * t1, * t2, * s1, * ptr;
mp_limb_t * tt;
ii = flint_malloc((4*(n + n*size) + 5*size)*sizeof(mp_limb_t));
for (i = 0, ptr = (mp_limb_t *) ii + 4*n; i < 4*n; i++, ptr += size)
{
ii[i] = ptr;
}
t1 = ptr;
t2 = t1 + size;
s1 = t2 + size;
tt = s1 + size;
if (i1 != i2)
{
jj = flint_malloc(4*(n + n*size)*sizeof(mp_limb_t));
for (i = 0, ptr = (mp_limb_t *) jj + 4*n; i < 4*n; i++, ptr += size)
{
jj[i] = ptr;
}
} else jj = ii;
trunc = j1 + j2 - 1;
if (trunc <= 2*n) trunc = 2*n + 1;
trunc = 2*sqrt*((trunc + 2*sqrt - 1)/(2*sqrt)); /* trunc must be divisible by 2*sqrt */
j1 = fft_split_bits(ii, i1, n1, bits1, limbs);
for (j = j1 ; j < 4*n; j++)
flint_mpn_zero(ii[j], limbs + 1);
fft_mfa_truncate_sqrt2_outer(ii, n, w, &t1, &t2, &s1, sqrt, trunc);
if (i1 != i2)
{
j2 = fft_split_bits(jj, i2, n2, bits1, limbs);
for (j = j2 ; j < 4*n; j++)
flint_mpn_zero(jj[j], limbs + 1);
fft_mfa_truncate_sqrt2_outer(jj, n, w, &t1, &t2, &s1, sqrt, trunc);
} else j2 = j1;
fft_mfa_truncate_sqrt2_inner(ii, jj, n, w, &t1, &t2, &s1, sqrt, trunc, tt);
ifft_mfa_truncate_sqrt2_outer(ii, n, w, &t1, &t2, &s1, sqrt, trunc);
flint_mpn_zero(r1, r_limbs);
fft_combine_bits(r1, ii, j1 + j2 - 1, bits1, limbs, r_limbs);
flint_free(ii);
if (i1 != i2)
flint_free(jj);
}
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;
}
}
