Coverage Report

Created: 2024-11-25 06:27

/src/gmp/mpn/toom22_mul.c
Line
Count
Source (jump to first uncovered line)
1
/* mpn_toom22_mul -- Multiply {ap,an} and {bp,bn} where an >= bn.  Or more
2
   accurately, bn <= an < 2bn.
3
4
   Contributed to the GNU project by Torbjorn Granlund.
5
6
   THE FUNCTION IN THIS FILE IS INTERNAL WITH A MUTABLE INTERFACE.  IT IS ONLY
7
   SAFE TO REACH IT THROUGH DOCUMENTED INTERFACES.  IN FACT, IT IS ALMOST
8
   GUARANTEED THAT IT WILL CHANGE OR DISAPPEAR IN A FUTURE GNU MP RELEASE.
9
10
Copyright 2006-2010, 2012, 2014, 2018, 2020 Free Software Foundation, Inc.
11
12
This file is part of the GNU MP Library.
13
14
The GNU MP Library is free software; you can redistribute it and/or modify
15
it under the terms of either:
16
17
  * the GNU Lesser General Public License as published by the Free
18
    Software Foundation; either version 3 of the License, or (at your
19
    option) any later version.
20
21
or
22
23
  * the GNU General Public License as published by the Free Software
24
    Foundation; either version 2 of the License, or (at your option) any
25
    later version.
26
27
or both in parallel, as here.
28
29
The GNU MP Library is distributed in the hope that it will be useful, but
30
WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
31
or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
32
for more details.
33
34
You should have received copies of the GNU General Public License and the
35
GNU Lesser General Public License along with the GNU MP Library.  If not,
36
see https://www.gnu.org/licenses/.  */
37
38
39
#include "gmp-impl.h"
40
41
/* Evaluate in: -1, 0, +inf
42
43
  <-s--><--n-->
44
   ____ ______
45
  |_a1_|___a0_|
46
   |b1_|___b0_|
47
   <-t-><--n-->
48
49
  v0  =  a0     * b0       #   A(0)*B(0)
50
  vm1 = (a0- a1)*(b0- b1)  #  A(-1)*B(-1)
51
  vinf=      a1 *     b1   # A(inf)*B(inf)
52
*/
53
54
#if TUNE_PROGRAM_BUILD || WANT_FAT_BINARY
55
#define MAYBE_mul_toom22   1
56
#else
57
#define MAYBE_mul_toom22            \
58
3.10M
  (MUL_TOOM33_THRESHOLD >= 2 * MUL_TOOM22_THRESHOLD)
59
#endif
60
61
#define TOOM22_MUL_N_REC(p, a, b, n, ws)        \
62
1.55M
  do {                 \
63
1.55M
    if (! MAYBE_mul_toom22            \
64
1.55M
  || BELOW_THRESHOLD (n, MUL_TOOM22_THRESHOLD))     \
65
1.55M
      mpn_mul_basecase (p, a, n, b, n);         \
66
1.55M
    else                \
67
1.55M
      mpn_toom22_mul (p, a, n, b, n, ws);       \
68
1.55M
  } while (0)
69
70
/* Normally, this calls mul_basecase or toom22_mul.  But when when the fraction
71
   MUL_TOOM33_THRESHOLD / MUL_TOOM22_THRESHOLD is large, an initially small
72
   relative unbalance will become a larger and larger relative unbalance with
73
   each recursion (the difference s-t will be invariant over recursive calls).
74
   Therefore, we need to call toom32_mul.  FIXME: Suppress depending on
75
   MUL_TOOM33_THRESHOLD / MUL_TOOM22_THRESHOLD and on MUL_TOOM22_THRESHOLD.  */
76
#define TOOM22_MUL_REC(p, a, an, b, bn, ws)       \
77
0
  do {                 \
78
0
    if (! MAYBE_mul_toom22            \
79
0
  || BELOW_THRESHOLD (bn, MUL_TOOM22_THRESHOLD))      \
80
0
      mpn_mul_basecase (p, a, an, b, bn);       \
81
0
    else if (4 * an < 5 * bn)           \
82
0
      mpn_toom22_mul (p, a, an, b, bn, ws);       \
83
0
    else                \
84
0
      mpn_toom32_mul (p, a, an, b, bn, ws);       \
85
0
  } while (0)
86
87
void
88
mpn_toom22_mul (mp_ptr pp,
89
    mp_srcptr ap, mp_size_t an,
90
    mp_srcptr bp, mp_size_t bn,
91
    mp_ptr scratch)
92
516k
{
93
516k
  const int __gmpn_cpuvec_initialized = 1;
94
516k
  mp_size_t n, s, t;
95
516k
  int vm1_neg;
96
516k
  mp_limb_t cy, cy2;
97
516k
  mp_ptr asm1;
98
516k
  mp_ptr bsm1;
99
100
1.20M
#define a0  ap
101
861k
#define a1  (ap + n)
102
1.03M
#define b0  bp
103
861k
#define b1  (bp + n)
104
105
516k
  s = an >> 1;
106
516k
  n = an - s;
107
516k
  t = bn - n;
108
109
516k
  ASSERT (an >= bn);
110
111
516k
  ASSERT (0 < s && s <= n && (n - s) == (an & 1));
112
516k
  ASSERT (0 < t && t <= s);
113
114
516k
  asm1 = pp;
115
516k
  bsm1 = pp + n;
116
117
516k
  vm1_neg = 0;
118
119
  /* Compute asm1.  */
120
516k
  if ((an & 1) == 0) /* s == n */
121
344k
    {
122
344k
      if (mpn_cmp (a0, a1, n) < 0)
123
163k
  {
124
163k
    mpn_sub_n (asm1, a1, a0, n);
125
163k
    vm1_neg = 1;
126
163k
  }
127
181k
      else
128
181k
  {
129
181k
    mpn_sub_n (asm1, a0, a1, n);
130
181k
  }
131
344k
    }
132
172k
  else /* n - s == 1 */
133
172k
    {
134
172k
      if (a0[s] == 0 && mpn_cmp (a0, a1, s) < 0)
135
0
  {
136
0
    mpn_sub_n (asm1, a1, a0, s);
137
0
    asm1[s] = 0;
138
0
    vm1_neg = 1;
139
0
  }
140
172k
      else
141
172k
  {
142
172k
    asm1[s] = a0[s] - mpn_sub_n (asm1, a0, a1, s);
143
172k
  }
144
172k
    }
145
146
  /* Compute bsm1.  */
147
516k
  if (t == n)
148
344k
    {
149
344k
      if (mpn_cmp (b0, b1, n) < 0)
150
229k
  {
151
229k
    mpn_sub_n (bsm1, b1, b0, n);
152
229k
    vm1_neg ^= 1;
153
229k
  }
154
115k
      else
155
115k
  {
156
115k
    mpn_sub_n (bsm1, b0, b1, n);
157
115k
  }
158
344k
    }
159
172k
  else
160
172k
    {
161
172k
      if (mpn_zero_p (b0 + t, n - t) && mpn_cmp (b0, b1, t) < 0)
162
0
  {
163
0
    mpn_sub_n (bsm1, b1, b0, t);
164
0
    MPN_ZERO (bsm1 + t, n - t);
165
0
    vm1_neg ^= 1;
166
0
  }
167
172k
      else
168
172k
  {
169
172k
    mpn_sub (bsm1, b0, n, b1, t);
170
172k
  }
171
172k
    }
172
173
1.03M
#define v0  pp        /* 2n */
174
1.03M
#define vinf  (pp + 2 * n)      /* s+t */
175
516k
#define vm1 scratch        /* 2n */
176
516k
#define scratch_out scratch + 2 * n
177
178
  /* vm1, 2n limbs */
179
516k
  TOOM22_MUL_N_REC (vm1, asm1, bsm1, n, scratch_out);
180
181
516k
  if (s > t)  TOOM22_MUL_REC (vinf, a1, s, b1, t, scratch_out);
182
516k
  else        TOOM22_MUL_N_REC (vinf, a1, b1, s, scratch_out);
183
184
  /* v0, 2n limbs */
185
516k
  TOOM22_MUL_N_REC (v0, ap, bp, n, scratch_out);
186
187
  /* H(v0) + L(vinf) */
188
516k
  cy = mpn_add_n (pp + 2 * n, v0 + n, vinf, n);
189
190
  /* L(v0) + (H(v0) + L(vinf)) */
191
516k
  cy2 = cy + mpn_add_n (pp + n, pp + 2 * n, v0, n);
192
193
  /* (H(v0) + L(vinf)) + H(vinf) */
194
516k
  cy += mpn_add (pp + 2 * n, pp + 2 * n, n, vinf + n, s + t - n);
195
196
516k
  if (vm1_neg)
197
163k
    cy += mpn_add_n (pp + n, pp + n, vm1, 2 * n);
198
353k
  else {
199
353k
    cy -= mpn_sub_n (pp + n, pp + n, vm1, 2 * n);
200
353k
    if (UNLIKELY (cy + 1 == 0)) { /* cy is negative */
201
      /* The total contribution of v0+vinf-vm1 can not be negative. */
202
#if WANT_ASSERT
203
      /* The borrow in cy stops the propagation of the carry cy2, */
204
      ASSERT (cy2 == 1);
205
      cy += mpn_add_1 (pp + 2 * n, pp + 2 * n, n, cy2);
206
      ASSERT (cy == 0);
207
#else
208
      /* we simply fill the area with zeros. */
209
21
      MPN_FILL (pp + 2 * n, n, 0);
210
      /* ASSERT (s + t == n || mpn_zero_p (pp + 3 * n, s + t - n)); */
211
21
#endif
212
21
      return;
213
21
    }
214
353k
  }
215
216
516k
  ASSERT (cy  <= 2);
217
516k
  ASSERT (cy2 <= 2);
218
219
516k
  MPN_INCR_U (pp + 2 * n, s + t, cy2);
220
  /* if s+t==n, cy is zero, but we should not access pp[3*n] at all. */
221
516k
  MPN_INCR_U (pp + 3 * n, s + t - n, cy);
222
516k
}