1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
|
/* mpn_mu_div_q.
Contributed to the GNU project by Torbjorn Granlund and Marco Bodrato.
THE FUNCTIONS IN THIS FILE ARE INTERNAL WITH MUTABLE INTERFACES. IT IS ONLY
SAFE TO REACH THEM THROUGH DOCUMENTED INTERFACES. IN FACT, IT IS ALMOST
GUARANTEED THAT THEY WILL CHANGE OR DISAPPEAR IN A FUTURE GMP RELEASE.
Copyright 2005-2007, 2009, 2010, 2013 Free Software Foundation, Inc.
This file is part of the GNU MP Library.
The GNU MP 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 MP 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 MP Library. If not, see https://www.gnu.org/licenses/. */
/*
The idea of the algorithm used herein is to compute a smaller inverted value
than used in the standard Barrett algorithm, and thus save time in the
Newton iterations, and pay just a small price when using the inverted value
for developing quotient bits. This algorithm was presented at ICMS 2006.
*/
/*
Things to work on:
1. This is a rudimentary implementation of mpn_mu_div_q. The algorithm is
probably close to optimal, except when mpn_mu_divappr_q fails.
2. We used to fall back to mpn_mu_div_qr when we detect a possible
mpn_mu_divappr_q rounding problem, now we multiply and compare.
Unfortunately, since mpn_mu_divappr_q does not return the partial
remainder, this also doesn't become optimal. A mpn_mu_divappr_qr could
solve that.
3. The allocations done here should be made from the scratch area, which
then would need to be amended.
*/
#include <stdlib.h> /* for NULL */
#include "gmp.h"
#include "gmp-impl.h"
mp_limb_t
mpn_mu_div_q (mp_ptr qp,
mp_srcptr np, mp_size_t nn,
mp_srcptr dp, mp_size_t dn,
mp_ptr scratch)
{
mp_ptr tp, rp;
mp_size_t qn;
mp_limb_t cy, qh;
TMP_DECL;
TMP_MARK;
qn = nn - dn;
tp = TMP_BALLOC_LIMBS (qn + 1);
if (qn >= dn) /* nn >= 2*dn + 1 */
{
/* |_______________________| dividend
|________| divisor */
rp = TMP_BALLOC_LIMBS (nn + 1);
MPN_COPY (rp + 1, np, nn);
rp[0] = 0;
qh = mpn_cmp (rp + 1 + nn - dn, dp, dn) >= 0;
if (qh != 0)
mpn_sub_n (rp + 1 + nn - dn, rp + 1 + nn - dn, dp, dn);
cy = mpn_mu_divappr_q (tp, rp, nn + 1, dp, dn, scratch);
if (UNLIKELY (cy != 0))
{
/* Since the partial remainder fed to mpn_preinv_mu_divappr_q was
canonically reduced, replace the returned value of B^(qn-dn)+eps
by the largest possible value. */
mp_size_t i;
for (i = 0; i < qn + 1; i++)
tp[i] = GMP_NUMB_MAX;
}
/* The max error of mpn_mu_divappr_q is +4. If the low quotient limb is
greater than the max error, we cannot trust the quotient. */
if (tp[0] > 4)
{
MPN_COPY (qp, tp + 1, qn);
}
else
{
mp_limb_t cy;
mp_ptr pp;
pp = rp;
mpn_mul (pp, tp + 1, qn, dp, dn);
cy = (qh != 0) ? mpn_add_n (pp + qn, pp + qn, dp, dn) : 0;
if (cy || mpn_cmp (pp, np, nn) > 0) /* At most is wrong by one, no cycle. */
qh -= mpn_sub_1 (qp, tp + 1, qn, 1);
else /* Same as above */
MPN_COPY (qp, tp + 1, qn);
}
}
else
{
/* |_______________________| dividend
|________________| divisor */
/* FIXME: When nn = 2dn-1, qn becomes dn-1, and the numerator size passed
here becomes 2dn, i.e., more than nn. This shouldn't hurt, since only
the most significant dn-1 limbs will actually be read, but it is not
pretty. */
qh = mpn_mu_divappr_q (tp, np + nn - (2 * qn + 2), 2 * qn + 2,
dp + dn - (qn + 1), qn + 1, scratch);
/* The max error of mpn_mu_divappr_q is +4, but we get an additional
error from the divisor truncation. */
if (tp[0] > 6)
{
MPN_COPY (qp, tp + 1, qn);
}
else
{
mp_limb_t cy;
/* FIXME: a shorter product should be enough; we may use already
allocated space... */
rp = TMP_BALLOC_LIMBS (nn);
mpn_mul (rp, dp, dn, tp + 1, qn);
cy = (qh != 0) ? mpn_add_n (rp + qn, rp + qn, dp, dn) : 0;
if (cy || mpn_cmp (rp, np, nn) > 0) /* At most is wrong by one, no cycle. */
qh -= mpn_sub_1 (qp, tp + 1, qn, 1);
else /* Same as above */
MPN_COPY (qp, tp + 1, qn);
}
}
TMP_FREE;
return qh;
}
mp_size_t
mpn_mu_div_q_itch (mp_size_t nn, mp_size_t dn, int mua_k)
{
mp_size_t qn;
qn = nn - dn;
if (qn >= dn)
{
return mpn_mu_divappr_q_itch (nn + 1, dn, mua_k);
}
else
{
return mpn_mu_divappr_q_itch (2 * qn + 2, qn + 1, mua_k);
}
}
|
