niceSOFT/mpfr | Gitverse

src/exp_2.c
456 строк15 KB
Vincent Lefèvre
03 янв 2018, 13:53
03 янв 2018, 13:532fce388
/* mpfr_exp_2 -- exponential of a floating-point number
                 using algorithms in O(n^(1/2)*M(n)) and O(n^(1/3)*M(n))

Copyright 1999-2018 Free Software Foundation, Inc.
Contributed by the AriC and Caramba projects, INRIA.

This file is part of the GNU MPFR Library.

The GNU MPFR Library is free software; you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as published by
the Free Software Foundation; either version 3 of the License, or (at your
option) any later version.

The GNU MPFR Library is distributed in the hope that it will be useful, but
WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
License for more details.

You should have received a copy of the GNU Lesser General Public License
along with the GNU MPFR Library; see the file COPYING.LESSER.  If not, see
http://www.gnu.org/licenses/ or write to the Free Software Foundation, Inc.,
51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. */

#define MPFR_NEED_LONGLONG_H /* for count_leading_zeros */
#include "mpfr-impl.h"

static unsigned long
mpfr_exp2_aux (mpz_t, mpfr_srcptr, mpfr_prec_t, mpfr_exp_t *);
static unsigned long
mpfr_exp2_aux2 (mpz_t, mpfr_srcptr, mpfr_prec_t, mpfr_exp_t *);
static mpfr_exp_t
mpz_normalize  (mpz_t, mpz_t, mpfr_exp_t);
static mpfr_exp_t
mpz_normalize2 (mpz_t, mpz_t, mpfr_exp_t, mpfr_exp_t);

/* if k = the number of bits of z > q, divides z by 2^(k-q) and returns k-q.
   Otherwise do nothing and return 0.
 */
static mpfr_exp_t
mpz_normalize (mpz_t rop, mpz_t z, mpfr_exp_t q)
{
  size_t k;

  MPFR_MPZ_SIZEINBASE2 (k, z);
  MPFR_ASSERTD (k == (mpfr_uexp_t) k);
  if (MPFR_LIKELY(q < 0 || (mpfr_uexp_t) k > (mpfr_uexp_t) q))
    {
      mpz_fdiv_q_2exp (rop, z, (unsigned long) ((mpfr_uexp_t) k - q));
      return (mpfr_exp_t) k - q;
    }
  mpz_set (rop, z);
  return 0;
}

/* if expz > target, shift z by (expz-target) bits to the left.
   if expz < target, shift z by (target-expz) bits to the right.
   Returns target.
*/
static mpfr_exp_t
mpz_normalize2 (mpz_t rop, mpz_t z, mpfr_exp_t expz, mpfr_exp_t target)
{
  if (MPFR_LIKELY(target > expz))
    mpz_fdiv_q_2exp (rop, z, target - expz);
  else
    mpz_mul_2exp (rop, z, expz - target);
  return target;
}

/* use Brent's formula exp(x) = (1+r+r^2/2!+r^3/3!+...)^(2^K)*2^n
   where x = n*log(2)+(2^K)*r
   together with the Paterson-Stockmeyer O(t^(1/2)) algorithm for the
   evaluation of power series. The resulting complexity is O(n^(1/3)*M(n)).
   This function returns with the exact flags due to exp.
*/
int
mpfr_exp_2 (mpfr_ptr y, mpfr_srcptr x, mpfr_rnd_t rnd_mode)
{
  long n;
  unsigned long K, k, l, err; /* FIXME: Which type ? */
  int error_r;
  mpfr_exp_t exps, expx;
  mpfr_prec_t q, precy;
  int inexact;
  mpfr_t s, r;
  mpz_t ss;
  MPFR_GROUP_DECL(group);
  MPFR_ZIV_DECL (loop);

  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));

  expx = MPFR_GET_EXP (x);
  precy = MPFR_PREC(y);

  /* First perform argument reduction: if x' = x - n*log(2) we have
     exp(x) = exp(x)*2^n. We should take n near from x/log(2) but it does not
     need to be exact.
     Warning: we cannot use the 'double' type here, since on 64-bit machines
     x may be as large as 2^62*log(2) without overflow, and then x/log(2)
     is about 2^62: not every integer of that size can be represented as a
     'double', thus the argument reduction would fail. */
  if (MPFR_UNLIKELY(expx <= -2))
    /* |x| <= 0.25, thus n = round(x/log(2)) = 0 */
    n = 0;
  else
    {
      mp_limb_t r_limb[(sizeof (long) -1) / sizeof(mp_limb_t) + 1];
      /* Note: we use precision sizeof (long) * CHAR_BIT - 1 here since it is
         more efficient that full limb precision. */
      MPFR_TMP_INIT1(r_limb, r, sizeof (long) * CHAR_BIT - 1);
      mpfr_div (r, x, __gmpfr_const_log2_RNDD, MPFR_RNDN);
#ifdef MPFR_LONG_WITHIN_LIMB
      /* The following code assume an unsigned long can fit in a mp_limb_t */
      {
        mp_limb_t a;
        mpfr_exp_t exp;
        MPFR_STAT_STATIC_ASSERT (MPFR_LIMB_MAX >= ULONG_MAX);
        /* Read the long directly (faster than using mpfr_get_si
           since it fits, it is not singular, it can't be zero
           and there is no conversion to do) */
        MPFR_ASSERTD (MPFR_NOTZERO (r));
        exp = MPFR_GET_EXP (r);
        MPFR_ASSERTD (exp <= GMP_NUMB_BITS);
        if (exp >= 1)
          {
            a = MPFR_MANT(r)[0] >> (GMP_NUMB_BITS - exp);
            n = MPFR_IS_POS (r) ? a : a <= LONG_MAX ? - (long) a : LONG_MIN;
          }
        else
          n = 0;
      }
#else
      /* Use generic way to get the long */
      n = mpfr_get_si (r, MPFR_RNDN);
#endif
    }
  /* we have |x| <= (|n|+1)*log(2) */
  MPFR_LOG_MSG (("d(x)=%1.30e n=%ld\n", mpfr_get_d1(x), n));

  /* error_r bounds the cancelled bits in x - n*log(2) */
  if (MPFR_UNLIKELY (n == 0))
    error_r = 0;
  else
    {
      count_leading_zeros (error_r,
                           (mp_limb_t) SAFE_ABS (unsigned long, n) + 1);
      error_r = GMP_NUMB_BITS - error_r;
      /* we have |x| <= 2^error_r * log(2) */
    }

  /* for the O(n^(1/2)*M(n)) method, the Taylor series computation of
     n/K terms costs about n/(2K) multiplications when computed in fixed
     point */
  K = (precy < MPFR_EXP_2_THRESHOLD) ? __gmpfr_isqrt ((precy + 1) / 2) + 3
    : __gmpfr_cuberoot (4*precy);
  l = (precy - 1) / K + 1;
  err = K + MPFR_INT_CEIL_LOG2 (2 * l + 18);
  /* add K extra bits, i.e. failure probability <= 1/2^K = O(1/precy) */
  q = precy + err + K + 10;
  /* if |x| >> 1, take into account the cancelled bits */
  if (expx > 0)
    q += expx;

  /* Even with to the mpfr_prec_round below, it is possible to use
     the MPFR_GROUP_* macros here because mpfr_prec_round is only
     called in a special case. */
  MPFR_GROUP_INIT_2(group, q + error_r, r, s);
  mpz_init (ss);

  /* the algorithm consists in computing an upper bound of exp(x) using
     a precision of q bits, and see if we can round to MPFR_PREC(y) taking
     into account the maximal error. Otherwise we increase q. */
  MPFR_ZIV_INIT (loop, q);
  for (;;)
    {
      MPFR_LOG_MSG (("n=%ld K=%lu l=%lu q=%lu error_r=%d\n",
                     n, K, l, (unsigned long) q, error_r));

      /* First reduce the argument to r = x - n * log(2),
         so that r is small in absolute value. We want an upper
         bound on r to get an upper bound on exp(x). */

      /* if n<0, we have to get an upper bound of log(2)
         in order to get an upper bound of r = x-n*log(2) */
      mpfr_const_log2 (s, (n >= 0) ? MPFR_RNDZ : MPFR_RNDU);
      /* s is within 1 ulp(s) of log(2) */

      mpfr_mul_ui (r, s, (n < 0) ? -n : n, (n >= 0) ? MPFR_RNDZ : MPFR_RNDU);
      /* r is within 3 ulps of |n|*log(2) */
      if (n < 0)
        MPFR_CHANGE_SIGN (r);
      /* r <= n*log(2), within 3 ulps */

      MPFR_LOG_VAR (x);
      MPFR_LOG_VAR (r);

      mpfr_sub (r, x, r, MPFR_RNDU);

      while (MPFR_IS_PURE_FP(r) && MPFR_IS_NEG (r))
        { /* initial approximation n was too large */
          n--;
          mpfr_add (r, r, s, MPFR_RNDU);
        }

      /* if r is 0, we cannot round correctly */
      if (MPFR_LIKELY(MPFR_IS_PURE_FP (r)))
        {
          /* since there was a cancellation in x - n*log(2), the low error_r
             bits from r are zero and thus non significant, thus we can reduce
             the working precision */
          if (MPFR_LIKELY(error_r > 0))
            {
              MPFR_ASSERTD( MPFR_PREC(r) > q);
              /* since MPFR_PREC(r) > q, there is no reallocation to do,
                 and it is safe to use it with MPFR_GROUP functions */
              mpfr_prec_round (r, q, MPFR_RNDU);
            }
          /* the error on r is at most 3 ulps (3 ulps if error_r = 0,
             and 1 + 3/2 if error_r > 0) */
          MPFR_LOG_VAR (r);
          MPFR_ASSERTD (MPFR_IS_POS (r));
          mpfr_div_2ui (r, r, K, MPFR_RNDU); /* r = (x-n*log(2))/2^K, exact */

          /* s <- 1 + r/1! + r^2/2! + ... + r^l/l! */
          MPFR_ASSERTD (MPFR_IS_PURE_FP (r) && MPFR_EXP (r) < 0);
          l = (precy < MPFR_EXP_2_THRESHOLD)
            ? mpfr_exp2_aux (ss, r, q, &exps)   /* naive method */
            : mpfr_exp2_aux2 (ss, r, q, &exps); /* Paterson/Stockmeyer meth */

          MPFR_LOG_MSG (("l=%lu q=%lu (K+l)*q^2=%1.3e\n",
                         l, (unsigned long) q, (K + l) * (double) q * q));

          for (k = 0; k < K; k++)
            {
              mpz_mul (ss, ss, ss);
              exps *= 2;
              exps += mpz_normalize (ss, ss, q);
            }
          mpfr_set_z_2exp (s, ss, exps, MPFR_RNDN);

          /* error is at most 2^K*l, plus 2 to take into account of
             the error of 3 ulps on r */
          err = K + MPFR_INT_CEIL_LOG2 (l) + 2;

          MPFR_LOG_MSG (("before mult. by 2^n:\n", 0));
          MPFR_LOG_VAR (s);
          MPFR_LOG_MSG (("err=%lu bits\n", K));

          if (MPFR_LIKELY (MPFR_CAN_ROUND (s, q - err, precy, rnd_mode)))
            {
              MPFR_CLEAR_FLAGS ();
              inexact = mpfr_mul_2si (y, s, n, rnd_mode);
              break;
            }
        }
      MPFR_ZIV_NEXT (loop, q);
      MPFR_GROUP_REPREC_2(group, q+error_r, r, s);
    }
  MPFR_ZIV_FREE (loop);
  mpz_clear (ss);
  MPFR_GROUP_CLEAR (group);

  return inexact;
}

/* s <- 1 + r/1! + r^2/2! + ... + r^l/l! while MPFR_EXP(r^l/l!)+MPFR_EXPR(r)>-q
   using naive method with O(l) multiplications.
   Return the number of iterations l.
   The absolute error on s is less than 3*l*(l+1)*2^(-q).
   Version using fixed-point arithmetic with mpz instead
   of mpfr for internal computations.
*/
static unsigned long
mpfr_exp2_aux (mpz_t s, mpfr_srcptr r, mpfr_prec_t q, mpfr_exp_t *exps)
{
  unsigned long l;
  mpfr_exp_t dif, expt, expr;
  mpz_t t, rr;
  mp_size_t sbit, tbit;

  MPFR_ASSERTD (MPFR_IS_PURE_FP (r));

  expt = 0;
  *exps = 1 - (mpfr_exp_t) q;                   /* s = 2^(q-1) */
  mpz_init (t);
  mpz_init (rr);
  mpz_set_ui (t, 1);
  mpz_set_ui (s, 1);
  mpz_mul_2exp (s, s, q-1);
  expr = mpfr_get_z_2exp (rr, r);               /* no error here */

  l = 0;
  for (;;)
    {
      l++;
      mpz_mul(t, t, rr);
      expt += expr;
      MPFR_MPZ_SIZEINBASE2 (sbit, s);
      MPFR_MPZ_SIZEINBASE2 (tbit, t);
      dif = *exps + sbit - expt - tbit;
      /* truncates the bits of t which are < ulp(s) = 2^(1-q) */
      expt += mpz_normalize (t, t, (mpfr_exp_t) q - dif);
      /* error at most 2^(1-q) */
      if (l > 1)
        {
          /* GMP doesn't optimize the case of power of 2 */
          if (IS_POW2(l))
            {
              int bits = MPFR_INT_CEIL_LOG2(l);
              mpz_fdiv_q_2exp (t, t, bits);     /* error at most 2^(1-q) */
            }
          else
            {
              mpz_fdiv_q_ui (t, t, l);          /* error at most 2^(1-q) */
            }
          /* the error wrt t^l/l! is here at most 3*l*ulp(s) */
          MPFR_ASSERTD (expt == *exps);
        }
      if (mpz_sgn (t) == 0)
        break;
      mpz_add (s, s, t);                        /* no error here: exact */
      /* ensures rr has the same size as t: after several shifts, the error
         on rr is still at most ulp(t)=ulp(s) */
      MPFR_MPZ_SIZEINBASE2 (tbit, t);
      expr += mpz_normalize (rr, rr, tbit);
    }

  mpz_clear (t);
  mpz_clear (rr);

  return 3 * l * (l + 1);
}

/* s <- 1 + r/1! + r^2/2! + ... + r^l/l! while MPFR_EXP(r^l/l!)+MPFR_EXPR(r)>-q
   using Paterson-Stockmeyer algorithm with O(sqrt(l)) multiplications.
   Return l.
   Uses m multiplications of full size and 2l/m of decreasing size,
   i.e. a total equivalent to about m+l/m full multiplications,
   i.e. 2*sqrt(l) for m=sqrt(l).
   NOTE[VL]: The following sentence seems to be obsolete since MY_INIT_MPZ
   is no longer used (r6919); sizer was the number of limbs of r.
   Version using mpz. ss must have at least (sizer+1) limbs.
   The error is bounded by (l^2+4*l) ulps where l is the return value.
*/
static unsigned long
mpfr_exp2_aux2 (mpz_t s, mpfr_srcptr r, mpfr_prec_t q, mpfr_exp_t *exps)
{
  mpfr_exp_t expr, *expR, expt;
  mpfr_prec_t ql;
  unsigned long l, m, i;
  mpz_t t, *R, rr, tmp;
  mp_size_t sbit, rrbit;
  MPFR_TMP_DECL(marker);

  /* estimate value of l */
  MPFR_ASSERTD (MPFR_GET_EXP (r) < 0);
  l = q / (- MPFR_GET_EXP (r));
  m = __gmpfr_isqrt (l);
  /* we access R[2], thus we need m >= 2 */
  if (m < 2)
    m = 2;

  MPFR_TMP_MARK(marker);
  R = (mpz_t*) MPFR_TMP_ALLOC ((m + 1) * sizeof (mpz_t));     /* R[i] is r^i */
  expR = (mpfr_exp_t*) MPFR_TMP_ALLOC((m + 1) * sizeof (mpfr_exp_t));
  /* expR[i] is the exponent for R[i] */
  mpz_init (tmp);
  mpz_init (rr);
  mpz_init (t);
  mpz_set_ui (s, 0);
  *exps = 1 - q;                        /* 1 ulp = 2^(1-q) */
  for (i = 0 ; i <= m ; i++)
    mpz_init (R[i]);
  expR[1] = mpfr_get_z_2exp (R[1], r); /* exact operation: no error */
  expR[1] = mpz_normalize2 (R[1], R[1], expR[1], 1 - q); /* error <= 1 ulp */
  mpz_mul (t, R[1], R[1]); /* err(t) <= 2 ulps */
  mpz_fdiv_q_2exp (R[2], t, q - 1); /* err(R[2]) <= 3 ulps */
  expR[2] = 1 - q;
  for (i = 3 ; i <= m ; i++)
    {
      if ((i & 1) == 1)
        mpz_mul (t, R[i-1], R[1]); /* err(t) <= 2*i-2 */
      else
        mpz_mul (t, R[i/2], R[i/2]);
      mpz_fdiv_q_2exp (R[i], t, q - 1); /* err(R[i]) <= 2*i-1 ulps */
      expR[i] = 1 - q;
    }
  mpz_set_ui (R[0], 1);
  mpz_mul_2exp (R[0], R[0], q-1);
  expR[0] = 1-q; /* R[0]=1 */
  mpz_set_ui (rr, 1);
  expr = 0; /* rr contains r^l/l! */
  /* by induction: err(rr) <= 2*l ulps */

  l = 0;
  ql = q; /* precision used for current giant step */
  do
    {
      /* all R[i] must have exponent 1-ql */
      if (l != 0)
        for (i = 0 ; i < m ; i++)
          expR[i] = mpz_normalize2 (R[i], R[i], expR[i], 1 - ql);
      /* the absolute error on R[i]*rr is still 2*i-1 ulps */
      expt = mpz_normalize2 (t, R[m-1], expR[m-1], 1 - ql);
      /* err(t) <= 2*m-1 ulps */
      /* computes t = 1 + r/(l+1) + ... + r^(m-1)*l!/(l+m-1)!
         using Horner's scheme */
      for (i = m-1 ; i-- != 0 ; )
        {
          mpz_fdiv_q_ui (t, t, l+i+1); /* err(t) += 1 ulp */
          mpz_add (t, t, R[i]);
        }
      /* now err(t) <= (3m-2) ulps */

      /* now multiplies t by r^l/l! and adds to s */
      mpz_mul (t, t, rr);
      expt += expr;
      expt = mpz_normalize2 (t, t, expt, *exps);
      /* err(t) <= (3m-1) + err_rr(l) <= (3m-2) + 2*l */
      MPFR_ASSERTD (expt == *exps);
      mpz_add (s, s, t); /* no error here */

      /* updates rr, the multiplication of the factors l+i could be done
         using binary splitting too, but it is not sure it would save much */
      mpz_mul (t, rr, R[m]); /* err(t) <= err(rr) + 2m-1 */
      expr += expR[m];
      mpz_set_ui (tmp, 1);
      for (i = 1 ; i <= m ; i++)
        mpz_mul_ui (tmp, tmp, l + i);
      mpz_fdiv_q (t, t, tmp); /* err(t) <= err(rr) + 2m */
      l += m;
      if (MPFR_UNLIKELY (mpz_sgn (t) == 0))
        break;
      expr += mpz_normalize (rr, t, ql); /* err_rr(l+1) <= err_rr(l) + 2m+1 */
      if (MPFR_UNLIKELY (mpz_sgn (rr) == 0))
        rrbit = 1;
      else
        MPFR_MPZ_SIZEINBASE2 (rrbit, rr);
      MPFR_MPZ_SIZEINBASE2 (sbit, s);
      ql = q - *exps - sbit + expr + rrbit;
      /* TODO: Wrong cast. I don't want what is right, but this is
         certainly wrong */
    }
  while ((size_t) expr + rrbit > (size_t) -q);

  for (i = 0 ; i <= m ; i++)
    mpz_clear (R[i]);
  MPFR_TMP_FREE(marker);
  mpz_clear (rr);
  mpz_clear (t);
  mpz_clear (tmp);

  return l * (l + 4);
}