| 1 |
#include <tommath.h> |
| 2 |
#ifdef BN_MP_KARATSUBA_MUL_C |
| 3 |
/* LibTomMath, multiple-precision integer library -- Tom St Denis |
| 4 |
* |
| 5 |
* LibTomMath is a library that provides multiple-precision |
| 6 |
* integer arithmetic as well as number theoretic functionality. |
| 7 |
* |
| 8 |
* The library was designed directly after the MPI library by |
| 9 |
* Michael Fromberger but has been written from scratch with |
| 10 |
* additional optimizations in place. |
| 11 |
* |
| 12 |
* The library is free for all purposes without any express |
| 13 |
* guarantee it works. |
| 14 |
* |
| 15 |
* Tom St Denis, [email protected], http://libtom.org |
| 16 |
*/ |
| 17 |
|
| 18 |
/* c = |a| * |b| using Karatsuba Multiplication using |
| 19 |
* three half size multiplications |
| 20 |
* |
| 21 |
* Let B represent the radix [e.g. 2**DIGIT_BIT] and |
| 22 |
* let n represent half of the number of digits in |
| 23 |
* the min(a,b) |
| 24 |
* |
| 25 |
* a = a1 * B**n + a0 |
| 26 |
* b = b1 * B**n + b0 |
| 27 |
* |
| 28 |
* Then, a * b => |
| 29 |
a1b1 * B**2n + ((a1 + a0)(b1 + b0) - (a0b0 + a1b1)) * B + a0b0 |
| 30 |
* |
| 31 |
* Note that a1b1 and a0b0 are used twice and only need to be |
| 32 |
* computed once. So in total three half size (half # of |
| 33 |
* digit) multiplications are performed, a0b0, a1b1 and |
| 34 |
* (a1+b1)(a0+b0) |
| 35 |
* |
| 36 |
* Note that a multiplication of half the digits requires |
| 37 |
* 1/4th the number of single precision multiplications so in |
| 38 |
* total after one call 25% of the single precision multiplications |
| 39 |
* are saved. Note also that the call to mp_mul can end up back |
| 40 |
* in this function if the a0, a1, b0, or b1 are above the threshold. |
| 41 |
* This is known as divide-and-conquer and leads to the famous |
| 42 |
* O(N**lg(3)) or O(N**1.584) work which is asymptopically lower than |
| 43 |
* the standard O(N**2) that the baseline/comba methods use. |
| 44 |
* Generally though the overhead of this method doesn't pay off |
| 45 |
* until a certain size (N ~ 80) is reached. |
| 46 |
*/ |
| 47 |
int mp_karatsuba_mul (mp_int * a, mp_int * b, mp_int * c) |
| 48 |
{ |
| 49 |
mp_int x0, x1, y0, y1, t1, x0y0, x1y1; |
| 50 |
int B, err; |
| 51 |
|
| 52 |
/* default the return code to an error */ |
| 53 |
err = MP_MEM; |
| 54 |
|
| 55 |
/* min # of digits */ |
| 56 |
B = MIN (a->used, b->used); |
| 57 |
|
| 58 |
/* now divide in two */ |
| 59 |
B = B >> 1; |
| 60 |
|
| 61 |
/* init copy all the temps */ |
| 62 |
if (mp_init_size (&x0, B) != MP_OKAY) |
| 63 |
goto ERR; |
| 64 |
if (mp_init_size (&x1, a->used - B) != MP_OKAY) |
| 65 |
goto X0; |
| 66 |
if (mp_init_size (&y0, B) != MP_OKAY) |
| 67 |
goto X1; |
| 68 |
if (mp_init_size (&y1, b->used - B) != MP_OKAY) |
| 69 |
goto Y0; |
| 70 |
|
| 71 |
/* init temps */ |
| 72 |
if (mp_init_size (&t1, B * 2) != MP_OKAY) |
| 73 |
goto Y1; |
| 74 |
if (mp_init_size (&x0y0, B * 2) != MP_OKAY) |
| 75 |
goto T1; |
| 76 |
if (mp_init_size (&x1y1, B * 2) != MP_OKAY) |
| 77 |
goto X0Y0; |
| 78 |
|
| 79 |
/* now shift the digits */ |
| 80 |
x0.used = y0.used = B; |
| 81 |
x1.used = a->used - B; |
| 82 |
y1.used = b->used - B; |
| 83 |
|
| 84 |
{ |
| 85 |
register int x; |
| 86 |
register mp_digit *tmpa, *tmpb, *tmpx, *tmpy; |
| 87 |
|
| 88 |
/* we copy the digits directly instead of using higher level functions |
| 89 |
* since we also need to shift the digits |
| 90 |
*/ |
| 91 |
tmpa = a->dp; |
| 92 |
tmpb = b->dp; |
| 93 |
|
| 94 |
tmpx = x0.dp; |
| 95 |
tmpy = y0.dp; |
| 96 |
for (x = 0; x < B; x++) { |
| 97 |
*tmpx++ = *tmpa++; |
| 98 |
*tmpy++ = *tmpb++; |
| 99 |
} |
| 100 |
|
| 101 |
tmpx = x1.dp; |
| 102 |
for (x = B; x < a->used; x++) { |
| 103 |
*tmpx++ = *tmpa++; |
| 104 |
} |
| 105 |
|
| 106 |
tmpy = y1.dp; |
| 107 |
for (x = B; x < b->used; x++) { |
| 108 |
*tmpy++ = *tmpb++; |
| 109 |
} |
| 110 |
} |
| 111 |
|
| 112 |
/* only need to clamp the lower words since by definition the |
| 113 |
* upper words x1/y1 must have a known number of digits |
| 114 |
*/ |
| 115 |
mp_clamp (&x0); |
| 116 |
mp_clamp (&y0); |
| 117 |
|
| 118 |
/* now calc the products x0y0 and x1y1 */ |
| 119 |
/* after this x0 is no longer required, free temp [x0==t2]! */ |
| 120 |
if (mp_mul (&x0, &y0, &x0y0) != MP_OKAY) |
| 121 |
goto X1Y1; /* x0y0 = x0*y0 */ |
| 122 |
if (mp_mul (&x1, &y1, &x1y1) != MP_OKAY) |
| 123 |
goto X1Y1; /* x1y1 = x1*y1 */ |
| 124 |
|
| 125 |
/* now calc x1+x0 and y1+y0 */ |
| 126 |
if (s_mp_add (&x1, &x0, &t1) != MP_OKAY) |
| 127 |
goto X1Y1; /* t1 = x1 - x0 */ |
| 128 |
if (s_mp_add (&y1, &y0, &x0) != MP_OKAY) |
| 129 |
goto X1Y1; /* t2 = y1 - y0 */ |
| 130 |
if (mp_mul (&t1, &x0, &t1) != MP_OKAY) |
| 131 |
goto X1Y1; /* t1 = (x1 + x0) * (y1 + y0) */ |
| 132 |
|
| 133 |
/* add x0y0 */ |
| 134 |
if (mp_add (&x0y0, &x1y1, &x0) != MP_OKAY) |
| 135 |
goto X1Y1; /* t2 = x0y0 + x1y1 */ |
| 136 |
if (s_mp_sub (&t1, &x0, &t1) != MP_OKAY) |
| 137 |
goto X1Y1; /* t1 = (x1+x0)*(y1+y0) - (x1y1 + x0y0) */ |
| 138 |
|
| 139 |
/* shift by B */ |
| 140 |
if (mp_lshd (&t1, B) != MP_OKAY) |
| 141 |
goto X1Y1; /* t1 = (x0y0 + x1y1 - (x1-x0)*(y1-y0))<<B */ |
| 142 |
if (mp_lshd (&x1y1, B * 2) != MP_OKAY) |
| 143 |
goto X1Y1; /* x1y1 = x1y1 << 2*B */ |
| 144 |
|
| 145 |
if (mp_add (&x0y0, &t1, &t1) != MP_OKAY) |
| 146 |
goto X1Y1; /* t1 = x0y0 + t1 */ |
| 147 |
if (mp_add (&t1, &x1y1, c) != MP_OKAY) |
| 148 |
goto X1Y1; /* t1 = x0y0 + t1 + x1y1 */ |
| 149 |
|
| 150 |
/* Algorithm succeeded set the return code to MP_OKAY */ |
| 151 |
err = MP_OKAY; |
| 152 |
|
| 153 |
X1Y1:mp_clear (&x1y1); |
| 154 |
X0Y0:mp_clear (&x0y0); |
| 155 |
T1:mp_clear (&t1); |
| 156 |
Y1:mp_clear (&y1); |
| 157 |
Y0:mp_clear (&y0); |
| 158 |
X1:mp_clear (&x1); |
| 159 |
X0:mp_clear (&x0); |
| 160 |
ERR: |
| 161 |
return err; |
| 162 |
} |
| 163 |
#endif |
| 164 |
|
| 165 |
/* $Source: /cvs/libtom/libtommath/bn_mp_karatsuba_mul.c,v $ */ |
| 166 |
/* $Revision: 1.6 $ */ |
| 167 |
/* $Date: 2006/12/28 01:25:13 $ */ |