Пример #1
0
int
mpfr_rint (mpfr_ptr r, mpfr_srcptr u, mpfr_rnd_t rnd_mode)
{
  int sign;
  int rnd_away;
  mpfr_exp_t exp;

  if (MPFR_UNLIKELY( MPFR_IS_SINGULAR(u) ))
    {
      if (MPFR_IS_NAN(u))
        {
          MPFR_SET_NAN(r);
          MPFR_RET_NAN;
        }
      MPFR_SET_SAME_SIGN(r, u);
      if (MPFR_IS_INF(u))
        {
          MPFR_SET_INF(r);
          MPFR_RET(0);  /* infinity is exact */
        }
      else /* now u is zero */
        {
          MPFR_ASSERTD(MPFR_IS_ZERO(u));
          MPFR_SET_ZERO(r);
          MPFR_RET(0);  /* zero is exact */
        }
    }
  MPFR_SET_SAME_SIGN (r, u); /* Does nothing if r==u */

  sign = MPFR_INT_SIGN (u);
  exp = MPFR_GET_EXP (u);

  rnd_away =
    rnd_mode == MPFR_RNDD ? sign < 0 :
    rnd_mode == MPFR_RNDU ? sign > 0 :
    rnd_mode == MPFR_RNDZ ? 0        :
    rnd_mode == MPFR_RNDA ? 1        :
    -1; /* round to nearest-even (RNDN) or nearest-away (RNDNA) */

  /* rnd_away:
     1 if round away from zero,
     0 if round to zero,
     -1 if not decided yet.
   */

  if (MPFR_UNLIKELY (exp <= 0))  /* 0 < |u| < 1 ==> round |u| to 0 or 1 */
    {
      /* Note: in the MPFR_RNDN mode, 0.5 must be rounded to 0. */
      if (rnd_away != 0 &&
          (rnd_away > 0 ||
           (exp == 0 && (rnd_mode == MPFR_RNDNA ||
                         !mpfr_powerof2_raw (u)))))
        {
          mp_limb_t *rp;
          mp_size_t rm;

          rp = MPFR_MANT(r);
          rm = (MPFR_PREC(r) - 1) / GMP_NUMB_BITS;
          rp[rm] = MPFR_LIMB_HIGHBIT;
          MPN_ZERO(rp, rm);
          MPFR_SET_EXP (r, 1);  /* |r| = 1 */
          MPFR_RET(sign > 0 ? 2 : -2);
        }
      else
        {
          MPFR_SET_ZERO(r);  /* r = 0 */
          MPFR_RET(sign > 0 ? -2 : 2);
        }
    }
  else  /* exp > 0, |u| >= 1 */
    {
      mp_limb_t *up, *rp;
      mp_size_t un, rn, ui;
      int sh, idiff;
      int uflags;

      /*
       * uflags will contain:
       *   _ 0 if u is an integer representable in r,
       *   _ 1 if u is an integer not representable in r,
       *   _ 2 if u is not an integer.
       */

      up = MPFR_MANT(u);
      rp = MPFR_MANT(r);

      un = MPFR_LIMB_SIZE(u);
      rn = MPFR_LIMB_SIZE(r);
      MPFR_UNSIGNED_MINUS_MODULO (sh, MPFR_PREC (r));

      MPFR_SET_EXP (r, exp); /* Does nothing if r==u */

      if ((exp - 1) / GMP_NUMB_BITS >= un)
        {
          ui = un;
          idiff = 0;
          uflags = 0;  /* u is an integer, representable or not in r */
        }
      else
        {
          mp_size_t uj;

          ui = (exp - 1) / GMP_NUMB_BITS + 1;  /* #limbs of the int part */
          MPFR_ASSERTD (un >= ui);
          uj = un - ui;  /* lowest limb of the integer part */
          idiff = exp % GMP_NUMB_BITS;  /* #int-part bits in up[uj] or 0 */

          uflags = idiff == 0 || (up[uj] << idiff) == 0 ? 0 : 2;
          if (uflags == 0)
            while (uj > 0)
              if (up[--uj] != 0)
                {
                  uflags = 2;
                  break;
                }
        }

      if (ui > rn)
        {
          /* More limbs in the integer part of u than in r.
             Just round u with the precision of r. */
          MPFR_ASSERTD (rp != up && un > rn);
          MPN_COPY (rp, up + (un - rn), rn); /* r != u */
          if (rnd_away < 0)
            {
              /* This is a rounding to nearest mode (MPFR_RNDN or MPFR_RNDNA).
                 Decide the rounding direction here. */
              if (rnd_mode == MPFR_RNDN &&
                  (rp[0] & (MPFR_LIMB_ONE << sh)) == 0)
                { /* halfway cases rounded toward zero */
                  mp_limb_t a, b;
                  /* a: rounding bit and some of the following bits */
                  /* b: boundary for a (weight of the rounding bit in a) */
                  if (sh != 0)
                    {
                      a = rp[0] & ((MPFR_LIMB_ONE << sh) - 1);
                      b = MPFR_LIMB_ONE << (sh - 1);
                    }
                  else
                    {
                      a = up[un - rn - 1];
                      b = MPFR_LIMB_HIGHBIT;
                    }
                  rnd_away = a > b;
                  if (a == b)
                    {
                      mp_size_t i;
                      for (i = un - rn - 1 - (sh == 0); i >= 0; i--)
                        if (up[i] != 0)
                          {
                            rnd_away = 1;
                            break;
                          }
                    }
                }
              else  /* halfway cases rounded away from zero */
                rnd_away =  /* rounding bit */
                  ((sh != 0 && (rp[0] & (MPFR_LIMB_ONE << (sh - 1))) != 0) ||
                   (sh == 0 && (up[un - rn - 1] & MPFR_LIMB_HIGHBIT) != 0));
            }
          if (uflags == 0)
            { /* u is an integer; determine if it is representable in r */
              if (sh != 0 && rp[0] << (GMP_NUMB_BITS - sh) != 0)
                uflags = 1;  /* u is not representable in r */
              else
                {
                  mp_size_t i;
                  for (i = un - rn - 1; i >= 0; i--)
                    if (up[i] != 0)
                      {
                        uflags = 1;  /* u is not representable in r */
                        break;
                      }
                }
            }
        }
      else  /* ui <= rn */
        {
          mp_size_t uj, rj;
          int ush;

          uj = un - ui;  /* lowest limb of the integer part in u */
          rj = rn - ui;  /* lowest limb of the integer part in r */

          if (MPFR_LIKELY (rp != up))
            MPN_COPY(rp + rj, up + uj, ui);

          /* Ignore the lowest rj limbs, all equal to zero. */
          rp += rj;
          rn = ui;

          /* number of fractional bits in whole rp[0] */
          ush = idiff == 0 ? 0 : GMP_NUMB_BITS - idiff;

          if (rj == 0 && ush < sh)
            {
              /* If u is an integer (uflags == 0), we need to determine
                 if it is representable in r, i.e. if its sh - ush bits
                 in the non-significant part of r are all 0. */
              if (uflags == 0 && (rp[0] & ((MPFR_LIMB_ONE << sh) -
                                           (MPFR_LIMB_ONE << ush))) != 0)
                uflags = 1;  /* u is an integer not representable in r */
            }
          else  /* The integer part of u fits in r, we'll round to it. */
            sh = ush;

          if (rnd_away < 0)
            {
              /* This is a rounding to nearest mode.
                 Decide the rounding direction here. */
              if (uj == 0 && sh == 0)
                rnd_away = 0; /* rounding bit = 0 (not represented in u) */
              else if (rnd_mode == MPFR_RNDN &&
                       (rp[0] & (MPFR_LIMB_ONE << sh)) == 0)
                { /* halfway cases rounded toward zero */
                  mp_limb_t a, b;
                  /* a: rounding bit and some of the following bits */
                  /* b: boundary for a (weight of the rounding bit in a) */
                  if (sh != 0)
                    {
                      a = rp[0] & ((MPFR_LIMB_ONE << sh) - 1);
                      b = MPFR_LIMB_ONE << (sh - 1);
                    }
                  else
                    {
                      MPFR_ASSERTD (uj >= 1);  /* see above */
                      a = up[uj - 1];
                      b = MPFR_LIMB_HIGHBIT;
                    }
                  rnd_away = a > b;
                  if (a == b)
                    {
                      mp_size_t i;
                      for (i = uj - 1 - (sh == 0); i >= 0; i--)
                        if (up[i] != 0)
                          {
                            rnd_away = 1;
                            break;
                          }
                    }
                }
              else  /* halfway cases rounded away from zero */
                rnd_away =  /* rounding bit */
                  ((sh != 0 && (rp[0] & (MPFR_LIMB_ONE << (sh - 1))) != 0) ||
                   (sh == 0 && (MPFR_ASSERTD (uj >= 1),
                                up[uj - 1] & MPFR_LIMB_HIGHBIT) != 0));
            }
          /* Now we can make the low rj limbs to 0 */
          MPN_ZERO (rp-rj, rj);
        }

      if (sh != 0)
        rp[0] &= MP_LIMB_T_MAX << sh;

      /* If u is a representable integer, there is no rounding. */
      if (uflags == 0)
        MPFR_RET(0);

      MPFR_ASSERTD (rnd_away >= 0);  /* rounding direction is defined */
      if (rnd_away && mpn_add_1(rp, rp, rn, MPFR_LIMB_ONE << sh))
        {
          if (exp == __gmpfr_emax)
            return mpfr_overflow(r, rnd_mode, MPFR_SIGN(r)) >= 0 ?
              uflags : -uflags;
          else
            {
              MPFR_SET_EXP(r, exp + 1);
              rp[rn-1] = MPFR_LIMB_HIGHBIT;
            }
        }

      MPFR_RET (rnd_away ^ (sign < 0) ? uflags : -uflags);
    }  /* exp > 0, |u| >= 1 */
}
Пример #2
0
int
mpfr_frac (mpfr_ptr r, mpfr_srcptr u, mpfr_rnd_t rnd_mode)
{
  mpfr_exp_t re, ue;
  mpfr_prec_t uq;
  mp_size_t un, tn, t0;
  mp_limb_t *up, *tp, k;
  int sh;
  mpfr_t tmp;
  mpfr_ptr t;
  int inex;
  MPFR_SAVE_EXPO_DECL (expo);

  /* Special cases */
  if (MPFR_UNLIKELY(MPFR_IS_NAN(u)))
    {
      MPFR_SET_NAN(r);
      MPFR_RET_NAN;
    }
  else if (MPFR_UNLIKELY(MPFR_IS_INF(u) || mpfr_integer_p (u)))
    {
      MPFR_SET_SAME_SIGN(r, u);
      MPFR_SET_ZERO(r);
      MPFR_RET(0);  /* zero is exact */
    }

  ue = MPFR_GET_EXP (u);
  if (ue <= 0)  /* |u| < 1 */
    return mpfr_set (r, u, rnd_mode);

  /* Now |u| >= 1, meaning that an overflow is not possible. */

  uq = MPFR_PREC(u);
  un = (uq - 1) / GMP_NUMB_BITS;  /* index of most significant limb */
  un -= (mp_size_t) (ue / GMP_NUMB_BITS);
  /* now the index of the MSL containing bits of the fractional part */

  up = MPFR_MANT(u);
  sh = ue % GMP_NUMB_BITS;
  k = up[un] << sh;
  /* the first bit of the fractional part is the MSB of k */

  if (k != 0)
    {
      int cnt;

      count_leading_zeros(cnt, k);
      /* first bit 1 of the fractional part -> MSB of the number */
      re = -cnt;
      sh += cnt;
      MPFR_ASSERTN (sh < GMP_NUMB_BITS);
      k <<= cnt;
    }
  else
    {
      re = sh - GMP_NUMB_BITS;
      /* searching for the first bit 1 (exists since u isn't an integer) */
      while (up[--un] == 0)
        re -= GMP_NUMB_BITS;
      MPFR_ASSERTN(un >= 0);
      k = up[un];
      count_leading_zeros(sh, k);
      re -= sh;
      k <<= sh;
    }
  /* The exponent of r will be re */
  /* un: index of the limb of u that contains the first bit 1 of the FP */

  t = (mp_size_t) (MPFR_PREC(r) - 1) / GMP_NUMB_BITS < un ?
    (mpfr_init2 (tmp, (un + 1) * GMP_NUMB_BITS), tmp) : r;
  /* t has enough precision to contain the fractional part of u */
  /* If we use a temporary variable, we take the non-significant bits
     of u into account, because of the mpn_lshift below. */
  MPFR_SET_SAME_SIGN(t, u);

  /* Put the fractional part of u into t */
  tn = (MPFR_PREC(t) - 1) / GMP_NUMB_BITS;
  MPFR_ASSERTN(tn >= un);
  t0 = tn - un;
  tp = MPFR_MANT(t);
  if (sh == 0)
    MPN_COPY_DECR(tp + t0, up, un + 1);
  else /* warning: un may be 0 here */
    tp[tn] = k | ((un) ? mpn_lshift (tp + t0, up, un, sh) : (mp_limb_t) 0);
  if (t0 > 0)
    MPN_ZERO(tp, t0);

  MPFR_SAVE_EXPO_MARK (expo);

  if (t != r)
    { /* t is tmp */
      MPFR_EXP (t) = 0;  /* should be re, but not necessarily in the range */
      inex = mpfr_set (r, t, rnd_mode);  /* no underflow */
      mpfr_clear (t);
      MPFR_EXP (r) += re;
    }
  else
    { /* There may be remaining non-significant bits in t (= r). */
      int carry;

      MPFR_EXP (r) = re;
      carry = mpfr_round_raw (tp, tp,
                              (mpfr_prec_t) (tn + 1) * GMP_NUMB_BITS,
                              MPFR_IS_NEG (r), MPFR_PREC (r), rnd_mode,
                              &inex);
      if (carry)
        {
          tp[tn] = MPFR_LIMB_HIGHBIT;
          MPFR_EXP (r) ++;
        }
    }

  MPFR_SAVE_EXPO_FREE (expo);
  return mpfr_check_range (r, inex, rnd_mode);
}
Пример #3
0
int
mpfr_atanh (mpfr_ptr y, mpfr_srcptr xt , mpfr_rnd_t rnd_mode)
{
  int inexact;
  mpfr_t x, t, te;
  mpfr_prec_t Nx, Ny, Nt;
  mpfr_exp_t err;
  MPFR_ZIV_DECL (loop);
  MPFR_SAVE_EXPO_DECL (expo);

  MPFR_LOG_FUNC
    (("x[%Pu]=%.*Rg rnd=%d", mpfr_get_prec (xt), mpfr_log_prec, xt, rnd_mode),
    ("y[%Pu]=%.*Rg inexact=%d",
     mpfr_get_prec (y), mpfr_log_prec, y, inexact));

  /* Special cases */
  if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (xt)))
    {
      /* atanh(NaN) = NaN, and atanh(+/-Inf) = NaN since tanh gives a result
         between -1 and 1 */
      if (MPFR_IS_NAN (xt) || MPFR_IS_INF (xt))
        {
          MPFR_SET_NAN (y);
          MPFR_RET_NAN;
        }
      else /* necessarily xt is 0 */
        {
          MPFR_ASSERTD (MPFR_IS_ZERO (xt));
          MPFR_SET_ZERO (y);   /* atanh(0) = 0 */
          MPFR_SET_SAME_SIGN (y,xt);
          MPFR_RET (0);
        }
    }

  /* atanh (x) = NaN as soon as |x| > 1, and arctanh(+/-1) = +/-Inf */
  if (MPFR_UNLIKELY (MPFR_GET_EXP (xt) > 0))
    {
      if (MPFR_GET_EXP (xt) == 1 && mpfr_powerof2_raw (xt))
        {
          MPFR_SET_INF (y);
          MPFR_SET_SAME_SIGN (y, xt);
          MPFR_SET_DIVBY0 ();
          MPFR_RET (0);
        }
      MPFR_SET_NAN (y);
      MPFR_RET_NAN;
    }

  /* atanh(x) = x + x^3/3 + ... so the error is < 2^(3*EXP(x)-1) */
  MPFR_FAST_COMPUTE_IF_SMALL_INPUT (y, xt, -2 * MPFR_GET_EXP (xt), 1, 1,
                                    rnd_mode, {});

  MPFR_SAVE_EXPO_MARK (expo);

  /* Compute initial precision */
  Nx = MPFR_PREC (xt);
  MPFR_TMP_INIT_ABS (x, xt);
  Ny = MPFR_PREC (y);
  Nt = MAX (Nx, Ny);
  /* the optimal number of bits : see algorithms.ps */
  Nt = Nt + MPFR_INT_CEIL_LOG2 (Nt) + 4;

  /* initialize of intermediary variable */
  mpfr_init2 (t, Nt);
  mpfr_init2 (te, Nt);

  /* First computation of cosh */
  MPFR_ZIV_INIT (loop, Nt);
  for (;;)
    {
      /* compute atanh */
      mpfr_ui_sub (te, 1, x, MPFR_RNDU);   /* (1-xt)*/
      mpfr_add_ui (t,  x, 1, MPFR_RNDD);   /* (xt+1)*/
      mpfr_div (t, t, te, MPFR_RNDN);      /* (1+xt)/(1-xt)*/
      mpfr_log (t, t, MPFR_RNDN);          /* ln((1+xt)/(1-xt))*/
      mpfr_div_2ui (t, t, 1, MPFR_RNDN);   /* (1/2)*ln((1+xt)/(1-xt))*/

      /* error estimate: see algorithms.tex */
      /* FIXME: this does not correspond to the value in algorithms.tex!!! */
      /* err=Nt-__gmpfr_ceil_log2(1+5*pow(2,1-MPFR_EXP(t)));*/
      err = Nt - (MAX (4 - MPFR_GET_EXP (t), 0) + 1);

      if (MPFR_LIKELY (MPFR_IS_ZERO (t)
                       || MPFR_CAN_ROUND (t, err, Ny, rnd_mode)))
        break;

      /* reactualisation of the precision */
      MPFR_ZIV_NEXT (loop, Nt);
      mpfr_set_prec (t, Nt);
      mpfr_set_prec (te, Nt);
    }
  MPFR_ZIV_FREE (loop);

  inexact = mpfr_set4 (y, t, rnd_mode, MPFR_SIGN (xt));

  mpfr_clear(t);
  mpfr_clear(te);

  MPFR_SAVE_EXPO_FREE (expo);
  return mpfr_check_range (y, inexact, rnd_mode);
}
Пример #4
0
Файл: sin.c Проект: epowers/mpfr
int
mpfr_sin (mpfr_ptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
  mpfr_t c, xr;
  mpfr_srcptr xx;
  mpfr_exp_t expx, err;
  mpfr_prec_t precy, m;
  int inexact, sign, reduce;
  MPFR_ZIV_DECL (loop);
  MPFR_SAVE_EXPO_DECL (expo);

  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 /* x is zero */
        {
          MPFR_ASSERTD (MPFR_IS_ZERO (x));
          MPFR_SET_ZERO (y);
          MPFR_SET_SAME_SIGN (y, x);
          MPFR_RET (0);
        }
    }

  /* sin(x) = x - x^3/6 + ... so the error is < 2^(3*EXP(x)-2) */
  MPFR_FAST_COMPUTE_IF_SMALL_INPUT (y, x, -2 * MPFR_GET_EXP (x), 2, 0,
                                    rnd_mode, {});

  MPFR_SAVE_EXPO_MARK (expo);

  /* Compute initial precision */
  precy = MPFR_PREC (y);

  if (precy >= MPFR_SINCOS_THRESHOLD)
    return mpfr_sin_fast (y, x, rnd_mode);

  m = precy + MPFR_INT_CEIL_LOG2 (precy) + 13;
  expx = MPFR_GET_EXP (x);

  mpfr_init (c);
  mpfr_init (xr);

  MPFR_ZIV_INIT (loop, m);
  for (;;)
    {
      /* first perform argument reduction modulo 2*Pi (if needed),
         also helps to determine the sign of sin(x) */
      if (expx >= 2) /* If Pi < x < 4, we need to reduce too, to determine
                        the sign of sin(x). For 2 <= |x| < Pi, we could avoid
                        the reduction. */
        {
          reduce = 1;
          /* As expx + m - 1 will silently be converted into mpfr_prec_t
             in the mpfr_set_prec call, the assert below may be useful to
             avoid undefined behavior. */
          MPFR_ASSERTN (expx + m - 1 <= MPFR_PREC_MAX);
          mpfr_set_prec (c, expx + m - 1);
          mpfr_set_prec (xr, m);
          mpfr_const_pi (c, MPFR_RNDN);
          mpfr_mul_2ui (c, c, 1, MPFR_RNDN);
          mpfr_remainder (xr, x, c, MPFR_RNDN);
          /* The analysis is similar to that of cos.c:
             |xr - x - 2kPi| <= 2^(2-m). Thus we can decide the sign
             of sin(x) if xr is at distance at least 2^(2-m) of both
             0 and +/-Pi. */
          mpfr_div_2ui (c, c, 1, MPFR_RNDN);
          /* Since c approximates Pi with an error <= 2^(2-expx-m) <= 2^(-m),
             it suffices to check that c - |xr| >= 2^(2-m). */
          if (MPFR_SIGN (xr) > 0)
            mpfr_sub (c, c, xr, MPFR_RNDZ);
          else
            mpfr_add (c, c, xr, MPFR_RNDZ);
          if (MPFR_IS_ZERO(xr)
              || MPFR_GET_EXP(xr) < (mpfr_exp_t) 3 - (mpfr_exp_t) m
              || MPFR_IS_ZERO(c)
              || MPFR_GET_EXP(c) < (mpfr_exp_t) 3 - (mpfr_exp_t) m)
            goto ziv_next;

          /* |xr - x - 2kPi| <= 2^(2-m), thus |sin(xr) - sin(x)| <= 2^(2-m) */
          xx = xr;
        }
      else /* the input argument is already reduced */
        {
          reduce = 0;
          xx = x;
        }

      sign = MPFR_SIGN(xx);
      /* now that the argument is reduced, precision m is enough */
      mpfr_set_prec (c, m);
      mpfr_cos (c, xx, MPFR_RNDZ);    /* can't be exact */
      mpfr_nexttoinf (c);           /* now c = cos(x) rounded away */
      mpfr_mul (c, c, c, MPFR_RNDU); /* away */
      mpfr_ui_sub (c, 1, c, MPFR_RNDZ);
      mpfr_sqrt (c, c, MPFR_RNDZ);
      if (MPFR_IS_NEG_SIGN(sign))
        MPFR_CHANGE_SIGN(c);

      /* Warning: c may be 0! */
      if (MPFR_UNLIKELY (MPFR_IS_ZERO (c)))
        {
          /* Huge cancellation: increase prec a lot! */
          m = MAX (m, MPFR_PREC (x));
          m = 2 * m;
        }
      else
        {
          /* the absolute error on c is at most 2^(3-m-EXP(c)),
             plus 2^(2-m) if there was an argument reduction.
             Since EXP(c) <= 1, 3-m-EXP(c) >= 2-m, thus the error
             is at most 2^(3-m-EXP(c)) in case of argument reduction. */
          err = 2 * MPFR_GET_EXP (c) + (mpfr_exp_t) m - 3 - (reduce != 0);
          if (MPFR_CAN_ROUND (c, err, precy, rnd_mode))
            break;

          /* check for huge cancellation (Near 0) */
          if (err < (mpfr_exp_t) MPFR_PREC (y))
            m += MPFR_PREC (y) - err;
          /* Check if near 1 */
          if (MPFR_GET_EXP (c) == 1)
            m += m;
        }

    ziv_next:
      /* Else generic increase */
      MPFR_ZIV_NEXT (loop, m);
    }
  MPFR_ZIV_FREE (loop);

  inexact = mpfr_set (y, c, rnd_mode);
  /* inexact cannot be 0, since this would mean that c was representable
     within the target precision, but in that case mpfr_can_round will fail */

  mpfr_clear (c);
  mpfr_clear (xr);

  MPFR_SAVE_EXPO_FREE (expo);
  return mpfr_check_range (y, inexact, rnd_mode);
}