1 /* mpn_mu_divappr_q, mpn_preinv_mu_divappr_q.
3 Compute Q = floor(N / D) + e. N is nn limbs, D is dn limbs and must be
4 normalized, and Q must be nn-dn limbs, 0 <= e <= 4. The requirement that Q
5 is nn-dn limbs (and not nn-dn+1 limbs) was put in place in order to allow us
6 to let N be unmodified during the operation.
8 Contributed to the GNU project by Torbjorn Granlund.
10 THE FUNCTIONS IN THIS FILE ARE INTERNAL WITH MUTABLE INTERFACES. IT IS ONLY
11 SAFE TO REACH THEM THROUGH DOCUMENTED INTERFACES. IN FACT, IT IS ALMOST
12 GUARANTEED THAT THEY WILL CHANGE OR DISAPPEAR IN A FUTURE GMP RELEASE.
14 Copyright 2005, 2006, 2007, 2009, 2010 Free Software Foundation, Inc.
16 This file is part of the GNU MP Library.
18 The GNU MP Library is free software; you can redistribute it and/or modify
19 it under the terms of the GNU Lesser General Public License as published by
20 the Free Software Foundation; either version 3 of the License, or (at your
21 option) any later version.
23 The GNU MP Library is distributed in the hope that it will be useful, but
24 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
25 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public
26 License for more details.
28 You should have received a copy of the GNU Lesser General Public License
29 along with the GNU MP Library. If not, see http://www.gnu.org/licenses/. */
33 The idea of the algorithm used herein is to compute a smaller inverted value
34 than used in the standard Barrett algorithm, and thus save time in the
35 Newton iterations, and pay just a small price when using the inverted value
36 for developing quotient bits. This algorithm was presented at ICMS 2006.
39 /* CAUTION: This code and the code in mu_div_qr.c should be edited in sync.
43 * The itch/scratch scheme isn't perhaps such a good idea as it once seemed,
44 demonstrated by the fact that the mpn_invertappr function's scratch needs
45 mean that we need to keep a large allocation long after it is needed.
46 Things are worse as mpn_mul_fft does not accept any scratch parameter,
47 which means we'll have a large memory hole while in mpn_mul_fft. In
48 general, a peak scratch need in the beginning of a function isn't
49 well-handled by the itch/scratch scheme.
59 #include <stdlib.h> /* for NULL */
65 mpn_mu_divappr_q (mp_ptr qp,
80 /* If Q is smaller than D, truncate operands. */
89 /* Compute the inverse size. */
90 in = mpn_mu_divappr_q_choose_in (qn, dn, 0);
94 /* This alternative inverse computation method gets slightly more accurate
95 results. FIXMEs: (1) Temp allocation needs not analysed (2) itch function
96 not adapted (3) mpn_invertappr scratch needs not met. */
98 tp = scratch + in + 1;
100 /* compute an approximate inverse on (in+1) limbs */
103 MPN_COPY (tp + 1, dp, in);
105 mpn_invertappr (ip, tp, in + 1, NULL);
106 MPN_COPY_INCR (ip, ip + 1, in);
110 cy = mpn_add_1 (tp, dp + dn - (in + 1), in + 1, 1);
111 if (UNLIKELY (cy != 0))
115 mpn_invertappr (ip, tp, in + 1, NULL);
116 MPN_COPY_INCR (ip, ip + 1, in);
120 /* This older inverse computation method gets slightly worse results than the
125 /* Compute inverse of D to in+1 limbs, then round to 'in' limbs. Ideally the
126 inversion function should do this automatically. */
130 MPN_COPY (tp + in + 2, dp, in);
131 mpn_invertappr (tp, tp + in + 1, in + 1, NULL);
135 mpn_invertappr (tp, dp + dn - (in + 1), in + 1, NULL);
137 cy = mpn_sub_1 (tp, tp, in + 1, GMP_NUMB_HIGHBIT);
138 if (UNLIKELY (cy != 0))
139 MPN_ZERO (tp + 1, in);
140 MPN_COPY (ip, tp + 1, in);
143 qh = mpn_preinv_mu_divappr_q (qp, np, nn, dp, dn, ip, in, scratch + in);
149 mpn_preinv_mu_divappr_q (mp_ptr qp,
159 mp_limb_t cy, cx, qh;
164 #define tp (scratch + dn)
165 #define scratch_out (scratch + dn + tn)
172 qh = mpn_cmp (np, dp, dn) >= 0;
174 mpn_sub_n (rp, np, dp, dn);
176 MPN_COPY (rp, np, dn);
179 return qh; /* Degenerate use. Should we allow this? */
191 /* Compute the next block of quotient limbs by multiplying the inverse I
192 by the upper part of the partial remainder R. */
193 mpn_mul_n (tp, rp + dn - in, ip, in); /* mulhi */
194 cy = mpn_add_n (qp, tp + in, rp + dn - in, in); /* I's msb implicit */
195 ASSERT_ALWAYS (cy == 0);
201 /* Compute the product of the quotient block and the divisor D, to be
202 subtracted from the partial remainder combined with new limbs from the
203 dividend N. We only really need the low dn limbs. */
205 if (BELOW_THRESHOLD (in, MUL_TO_MULMOD_BNM1_FOR_2NXN_THRESHOLD))
206 mpn_mul (tp, dp, dn, qp, in); /* dn+in limbs, high 'in' cancels */
209 tn = mpn_mulmod_bnm1_next_size (dn + 1);
210 mpn_mulmod_bnm1 (tp, tn, dp, dn, qp, in, scratch_out);
211 wn = dn + in - tn; /* number of wrapped limbs */
214 cy = mpn_sub_n (tp, tp, rp + dn - wn, wn);
215 cy = mpn_sub_1 (tp + wn, tp + wn, tn - wn, cy);
216 cx = mpn_cmp (rp + dn - in, tp + dn, tn - dn) < 0;
217 ASSERT_ALWAYS (cx >= cy);
218 mpn_incr_u (tp, cx - cy);
222 r = rp[dn - in] - tp[dn];
224 /* Subtract the product from the partial remainder combined with new
225 limbs from the dividend N, generating a new partial remainder R. */
228 cy = mpn_sub_n (tp, np, tp, in); /* get next 'in' limbs from N */
229 cy = mpn_sub_nc (tp + in, rp, tp + in, dn - in, cy);
230 MPN_COPY (rp, tp, dn); /* FIXME: try to avoid this */
234 cy = mpn_sub_n (rp, np, tp, in); /* get next 'in' limbs from N */
237 STAT (int i; int err = 0;
238 static int errarr[5]; static int err_rec; static int tot);
240 /* Check the remainder R and adjust the quotient as needed. */
244 /* We loop 0 times with about 69% probability, 1 time with about 31%
245 probability, 2 times with about 0.6% probability, if inverse is
246 computed as recommended. */
248 cy = mpn_sub_n (rp, rp, dp, dn);
252 if (mpn_cmp (rp, dp, dn) >= 0)
254 /* This is executed with about 76% probability. */
256 cy = mpn_sub_n (rp, rp, dp, dn);
265 if (tot % 0x10000 == 0)
267 for (i = 0; i <= err_rec; i++)
268 printf (" %d(%.1f%%)", errarr[i], 100.0*errarr[i]/tot);
274 /* FIXME: We should perhaps be somewhat more elegant in our rounding of the
275 quotient. For now, just make sure the returned quotient is >= the real
276 quotient; add 3 with saturating arithmetic. */
278 cy += mpn_add_1 (qp, qp, qn, 3);
283 /* Return a quotient of just 1-bits, with qh set. */
285 for (i = 0; i < qn; i++)
286 qp[i] = GMP_NUMB_MAX;
290 /* Propagate carry into qh. */
298 /* In case k=0 (automatic choice), we distinguish 3 cases:
299 (a) dn < qn: in = ceil(qn / ceil(qn/dn))
300 (b) dn/3 < qn <= dn: in = ceil(qn / 2)
301 (c) qn < dn/3: in = qn
302 In all cases we have in <= dn.
305 mpn_mu_divappr_q_choose_in (mp_size_t qn, mp_size_t dn, int k)
314 /* Compute an inverse size that is a nice partition of the quotient. */
315 b = (qn - 1) / dn + 1; /* ceil(qn/dn), number of blocks */
316 in = (qn - 1) / b + 1; /* ceil(qn/b) = ceil(qn / ceil(qn/dn)) */
318 else if (3 * qn > dn)
320 in = (qn - 1) / 2 + 1; /* b = 2 */
324 in = (qn - 1) / 1 + 1; /* b = 1 */
331 in = (xn - 1) / k + 1;
338 mpn_mu_divappr_q_itch (mp_size_t nn, mp_size_t dn, int mua_k)
340 mp_size_t qn, in, itch_local, itch_out;
347 in = mpn_mu_divappr_q_choose_in (qn, dn, mua_k);
349 itch_local = mpn_mulmod_bnm1_next_size (dn + 1);
350 itch_out = mpn_mulmod_bnm1_itch (itch_local, dn, in);
351 return in + dn + itch_local + itch_out;