1 /* mpn_mu_bdiv_qr(qp,rp,np,nn,dp,dn,tp) -- Compute {np,nn} / {dp,dn} mod B^qn,
2 where qn = nn-dn, storing the result in {qp,qn}. Overlap allowed between Q
3 and N; all other overlap disallowed.
5 Contributed to the GNU project by Torbjorn Granlund.
7 THE FUNCTIONS IN THIS FILE ARE INTERNAL WITH MUTABLE INTERFACES. IT IS ONLY
8 SAFE TO REACH THEM THROUGH DOCUMENTED INTERFACES. IN FACT, IT IS ALMOST
9 GUARANTEED THAT THEY WILL CHANGE OR DISAPPEAR IN A FUTURE GMP RELEASE.
11 Copyright 2005, 2006, 2007, 2009, 2010 Free Software Foundation, Inc.
13 This file is part of the GNU MP Library.
15 The GNU MP Library is free software; you can redistribute it and/or modify
16 it under the terms of the GNU Lesser General Public License as published by
17 the Free Software Foundation; either version 3 of the License, or (at your
18 option) any later version.
20 The GNU MP Library is distributed in the hope that it will be useful, but
21 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
22 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public
23 License for more details.
25 You should have received a copy of the GNU Lesser General Public License
26 along with the GNU MP Library. If not, see http://www.gnu.org/licenses/. */
30 The idea of the algorithm used herein is to compute a smaller inverted value
31 than used in the standard Barrett algorithm, and thus save time in the
32 Newton iterations, and pay just a small price when using the inverted value
33 for developing quotient bits. This algorithm was presented at ICMS 2006.
48 scratch space as determined by mpn_mu_bdiv_qr_itch(nn,dn).
50 Write quotient to Q = {qp,nn-dn}.
52 FIXME: When iterating, perhaps do the small step before loop, not after.
53 FIXME: Try to avoid the scalar divisions when computing inverse size.
54 FIXME: Trim allocation for (qn > dn) case, 3*dn might be possible. In
55 particular, when dn==in, tp and rp could use the same space.
58 mpn_mu_bdiv_qr (mp_ptr qp,
60 mp_srcptr np, mp_size_t nn,
61 mp_srcptr dp, mp_size_t dn,
80 /* |_______________________| dividend
83 #define ip scratch /* in */
84 #define tp (scratch + in) /* dn+in or next_size(dn) or rest >= binvert_itch(in) */
85 #define scratch_out (scratch + in + tn)/* mulmod_bnm1_itch(next_size(dn)) */
87 /* Compute an inverse size that is a nice partition of the quotient. */
88 b = (qn - 1) / dn + 1; /* ceil(qn/dn), number of blocks */
89 in = (qn - 1) / b + 1; /* ceil(qn/b) = ceil(qn / ceil(qn/dn)) */
91 /* Some notes on allocation:
93 When in = dn, R dies when mpn_mullo returns, if in < dn the low in
94 limbs of R dies at that point. We could save memory by letting T live
95 just under R, and let the upper part of T expand into R. These changes
96 should reduce itch to perhaps 3dn.
99 mpn_binvert (ip, dp, in, tp);
101 MPN_COPY (rp, np, dn);
107 mpn_mullo_n (qp, rp, ip, in);
109 if (BELOW_THRESHOLD (in, MUL_TO_MULMOD_BNM1_FOR_2NXN_THRESHOLD))
110 mpn_mul (tp, dp, dn, qp, in); /* mulhi, need tp[dn+in-1...in] */
113 tn = mpn_mulmod_bnm1_next_size (dn);
114 mpn_mulmod_bnm1 (tp, tn, dp, dn, qp, in, scratch_out);
115 wn = dn + in - tn; /* number of wrapped limbs */
118 c0 = mpn_sub_n (tp + tn, tp, rp, wn);
119 mpn_decr_u (tp + wn, c0);
128 /* Subtract tp[dn-1...in] from partial remainder. */
129 cy += mpn_sub_n (rp, rp + in, tp + in, dn - in);
132 mpn_incr_u (tp + dn, 1);
136 /* Subtract tp[dn+in-1...dn] from dividend. */
137 cy = mpn_sub_nc (rp + dn - in, np, tp + dn, in, cy);
141 /* Generate last qn limbs. */
142 mpn_mullo_n (qp, rp, ip, qn);
144 if (BELOW_THRESHOLD (qn, MUL_TO_MULMOD_BNM1_FOR_2NXN_THRESHOLD))
145 mpn_mul (tp, dp, dn, qp, qn); /* mulhi, need tp[qn+in-1...in] */
148 tn = mpn_mulmod_bnm1_next_size (dn);
149 mpn_mulmod_bnm1 (tp, tn, dp, dn, qp, qn, scratch_out);
150 wn = dn + qn - tn; /* number of wrapped limbs */
153 c0 = mpn_sub_n (tp + tn, tp, rp, wn);
154 mpn_decr_u (tp + wn, c0);
160 cy += mpn_sub_n (rp, rp + qn, tp + qn, dn - qn);
163 mpn_incr_u (tp + dn, 1);
167 return mpn_sub_nc (rp + dn - qn, np, tp + dn, qn, cy);
175 /* |_______________________| dividend
176 |________________| divisor */
178 #define ip scratch /* in */
179 #define tp (scratch + in) /* dn+in or next_size(dn) or rest >= binvert_itch(in) */
180 #define scratch_out (scratch + in + tn)/* mulmod_bnm1_itch(next_size(dn)) */
182 /* Compute half-sized inverse. */
185 mpn_binvert (ip, dp, in, tp);
187 mpn_mullo_n (qp, np, ip, in); /* low `in' quotient limbs */
189 if (BELOW_THRESHOLD (in, MUL_TO_MULMOD_BNM1_FOR_2NXN_THRESHOLD))
190 mpn_mul (tp, dp, dn, qp, in); /* mulhigh */
193 tn = mpn_mulmod_bnm1_next_size (dn);
194 mpn_mulmod_bnm1 (tp, tn, dp, dn, qp, in, scratch_out);
195 wn = dn + in - tn; /* number of wrapped limbs */
198 c0 = mpn_sub_n (tp + tn, tp, np, wn);
199 mpn_decr_u (tp + wn, c0);
206 cy = mpn_sub_n (rp, np + in, tp + in, dn);
207 mpn_mullo_n (qp, rp, ip, qn); /* high qn quotient limbs */
209 if (BELOW_THRESHOLD (qn, MUL_TO_MULMOD_BNM1_FOR_2NXN_THRESHOLD))
210 mpn_mul (tp, dp, dn, qp, qn); /* mulhigh */
213 tn = mpn_mulmod_bnm1_next_size (dn);
214 mpn_mulmod_bnm1 (tp, tn, dp, dn, qp, qn, scratch_out);
215 wn = dn + qn - tn; /* number of wrapped limbs */
218 c0 = mpn_sub_n (tp + tn, tp, rp, wn);
219 mpn_decr_u (tp + wn, c0);
223 cy += mpn_sub_n (rp, rp + qn, tp + qn, dn - qn);
226 mpn_incr_u (tp + dn, 1);
229 return mpn_sub_nc (rp + dn - qn, np + dn + in, tp + dn, qn, cy);
238 mpn_mu_bdiv_qr_itch (mp_size_t nn, mp_size_t dn)
240 mp_size_t qn, in, tn, itch_binvert, itch_out, itches;
247 b = (qn - 1) / dn + 1; /* ceil(qn/dn), number of blocks */
248 in = (qn - 1) / b + 1; /* ceil(qn/b) = ceil(qn / ceil(qn/dn)) */
249 if (BELOW_THRESHOLD (in, MUL_TO_MULMOD_BNM1_FOR_2NXN_THRESHOLD))
256 tn = mpn_mulmod_bnm1_next_size (dn);
257 itch_out = mpn_mulmod_bnm1_itch (tn, dn, in);
259 itch_binvert = mpn_binvert_itch (in);
260 itches = tn + itch_out;
261 return in + MAX (itches, itch_binvert);
266 if (BELOW_THRESHOLD (in, MUL_TO_MULMOD_BNM1_FOR_2NXN_THRESHOLD))
273 tn = mpn_mulmod_bnm1_next_size (dn);
274 itch_out = mpn_mulmod_bnm1_itch (tn, dn, in);
277 itch_binvert = mpn_binvert_itch (in);
278 itches = tn + itch_out;
279 return in + MAX (itches, itch_binvert);