diff options
author | jan.nijtmans <nijtmans@users.sourceforge.net> | 2016-11-16 15:22:26 (GMT) |
---|---|---|
committer | jan.nijtmans <nijtmans@users.sourceforge.net> | 2016-11-16 15:22:26 (GMT) |
commit | 68111aa5bf7fc228dcfda8beb9de265734925b56 (patch) | |
tree | fed7fb157cbaef79f43a45b27379a0fd2d64e6ea /libtommath/bn_mp_toom_mul.c | |
parent | 29606e4a7b43adb9f923fb5781d3b9a93d9ba1c8 (diff) | |
parent | 2adcff3e5ba6e09366ef4208ab81768803ba15bd (diff) | |
download | tcl-68111aa5bf7fc228dcfda8beb9de265734925b56.zip tcl-68111aa5bf7fc228dcfda8beb9de265734925b56.tar.gz tcl-68111aa5bf7fc228dcfda8beb9de265734925b56.tar.bz2 |
(experimental) Upgrade to libtommath 1.0 (actually by merging all changes between libtommath 0.42.0 and 1.0). Still to be tested thourougly, before doing anything with it.
Diffstat (limited to 'libtommath/bn_mp_toom_mul.c')
-rw-r--r-- | libtommath/bn_mp_toom_mul.c | 276 |
1 files changed, 141 insertions, 135 deletions
diff --git a/libtommath/bn_mp_toom_mul.c b/libtommath/bn_mp_toom_mul.c index 9daefbd..4731f8f 100644 --- a/libtommath/bn_mp_toom_mul.c +++ b/libtommath/bn_mp_toom_mul.c @@ -1,4 +1,4 @@ -#include <tommath.h> +#include <tommath_private.h> #ifdef BN_MP_TOOM_MUL_C /* LibTomMath, multiple-precision integer library -- Tom St Denis * @@ -12,31 +12,31 @@ * The library is free for all purposes without any express * guarantee it works. * - * Tom St Denis, tomstdenis@gmail.com, http://math.libtomcrypt.com + * Tom St Denis, tstdenis82@gmail.com, http://libtom.org */ -/* multiplication using the Toom-Cook 3-way algorithm +/* multiplication using the Toom-Cook 3-way algorithm * - * Much more complicated than Karatsuba but has a lower - * asymptotic running time of O(N**1.464). This algorithm is - * only particularly useful on VERY large inputs + * Much more complicated than Karatsuba but has a lower + * asymptotic running time of O(N**1.464). This algorithm is + * only particularly useful on VERY large inputs * (we're talking 1000s of digits here...). */ int mp_toom_mul(mp_int *a, mp_int *b, mp_int *c) { mp_int w0, w1, w2, w3, w4, tmp1, tmp2, a0, a1, a2, b0, b1, b2; int res, B; - + /* init temps */ - if ((res = mp_init_multi(&w0, &w1, &w2, &w3, &w4, - &a0, &a1, &a2, &b0, &b1, + if ((res = mp_init_multi(&w0, &w1, &w2, &w3, &w4, + &a0, &a1, &a2, &b0, &b1, &b2, &tmp1, &tmp2, NULL)) != MP_OKAY) { return res; } - + /* B */ B = MIN(a->used, b->used) / 3; - + /* a = a2 * B**2 + a1 * B + a0 */ if ((res = mp_mod_2d(a, DIGIT_BIT * B, &a0)) != MP_OKAY) { goto ERR; @@ -46,13 +46,15 @@ int mp_toom_mul(mp_int *a, mp_int *b, mp_int *c) goto ERR; } mp_rshd(&a1, B); - mp_mod_2d(&a1, DIGIT_BIT * B, &a1); + if ((res = mp_mod_2d(&a1, DIGIT_BIT * B, &a1)) != MP_OKAY) { + goto ERR; + } if ((res = mp_copy(a, &a2)) != MP_OKAY) { goto ERR; } mp_rshd(&a2, B*2); - + /* b = b2 * B**2 + b1 * B + b0 */ if ((res = mp_mod_2d(b, DIGIT_BIT * B, &b0)) != MP_OKAY) { goto ERR; @@ -62,23 +64,23 @@ int mp_toom_mul(mp_int *a, mp_int *b, mp_int *c) goto ERR; } mp_rshd(&b1, B); - mp_mod_2d(&b1, DIGIT_BIT * B, &b1); + (void)mp_mod_2d(&b1, DIGIT_BIT * B, &b1); if ((res = mp_copy(b, &b2)) != MP_OKAY) { goto ERR; } mp_rshd(&b2, B*2); - + /* w0 = a0*b0 */ if ((res = mp_mul(&a0, &b0, &w0)) != MP_OKAY) { goto ERR; } - + /* w4 = a2 * b2 */ if ((res = mp_mul(&a2, &b2, &w4)) != MP_OKAY) { goto ERR; } - + /* w1 = (a2 + 2(a1 + 2a0))(b2 + 2(b1 + 2b0)) */ if ((res = mp_mul_2(&a0, &tmp1)) != MP_OKAY) { goto ERR; @@ -92,7 +94,7 @@ int mp_toom_mul(mp_int *a, mp_int *b, mp_int *c) if ((res = mp_add(&tmp1, &a2, &tmp1)) != MP_OKAY) { goto ERR; } - + if ((res = mp_mul_2(&b0, &tmp2)) != MP_OKAY) { goto ERR; } @@ -105,11 +107,11 @@ int mp_toom_mul(mp_int *a, mp_int *b, mp_int *c) if ((res = mp_add(&tmp2, &b2, &tmp2)) != MP_OKAY) { goto ERR; } - + if ((res = mp_mul(&tmp1, &tmp2, &w1)) != MP_OKAY) { goto ERR; } - + /* w3 = (a0 + 2(a1 + 2a2))(b0 + 2(b1 + 2b2)) */ if ((res = mp_mul_2(&a2, &tmp1)) != MP_OKAY) { goto ERR; @@ -123,7 +125,7 @@ int mp_toom_mul(mp_int *a, mp_int *b, mp_int *c) if ((res = mp_add(&tmp1, &a0, &tmp1)) != MP_OKAY) { goto ERR; } - + if ((res = mp_mul_2(&b2, &tmp2)) != MP_OKAY) { goto ERR; } @@ -136,11 +138,11 @@ int mp_toom_mul(mp_int *a, mp_int *b, mp_int *c) if ((res = mp_add(&tmp2, &b0, &tmp2)) != MP_OKAY) { goto ERR; } - + if ((res = mp_mul(&tmp1, &tmp2, &w3)) != MP_OKAY) { goto ERR; } - + /* w2 = (a2 + a1 + a0)(b2 + b1 + b0) */ if ((res = mp_add(&a2, &a1, &tmp1)) != MP_OKAY) { @@ -158,123 +160,127 @@ int mp_toom_mul(mp_int *a, mp_int *b, mp_int *c) if ((res = mp_mul(&tmp1, &tmp2, &w2)) != MP_OKAY) { goto ERR; } - - /* now solve the matrix - + + /* now solve the matrix + 0 0 0 0 1 1 2 4 8 16 1 1 1 1 1 16 8 4 2 1 1 0 0 0 0 - - using 12 subtractions, 4 shifts, - 2 small divisions and 1 small multiplication + + using 12 subtractions, 4 shifts, + 2 small divisions and 1 small multiplication */ - - /* r1 - r4 */ - if ((res = mp_sub(&w1, &w4, &w1)) != MP_OKAY) { - goto ERR; - } - /* r3 - r0 */ - if ((res = mp_sub(&w3, &w0, &w3)) != MP_OKAY) { - goto ERR; - } - /* r1/2 */ - if ((res = mp_div_2(&w1, &w1)) != MP_OKAY) { - goto ERR; - } - /* r3/2 */ - if ((res = mp_div_2(&w3, &w3)) != MP_OKAY) { - goto ERR; - } - /* r2 - r0 - r4 */ - if ((res = mp_sub(&w2, &w0, &w2)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_sub(&w2, &w4, &w2)) != MP_OKAY) { - goto ERR; - } - /* r1 - r2 */ - if ((res = mp_sub(&w1, &w2, &w1)) != MP_OKAY) { - goto ERR; - } - /* r3 - r2 */ - if ((res = mp_sub(&w3, &w2, &w3)) != MP_OKAY) { - goto ERR; - } - /* r1 - 8r0 */ - if ((res = mp_mul_2d(&w0, 3, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_sub(&w1, &tmp1, &w1)) != MP_OKAY) { - goto ERR; - } - /* r3 - 8r4 */ - if ((res = mp_mul_2d(&w4, 3, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_sub(&w3, &tmp1, &w3)) != MP_OKAY) { - goto ERR; - } - /* 3r2 - r1 - r3 */ - if ((res = mp_mul_d(&w2, 3, &w2)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_sub(&w2, &w1, &w2)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_sub(&w2, &w3, &w2)) != MP_OKAY) { - goto ERR; - } - /* r1 - r2 */ - if ((res = mp_sub(&w1, &w2, &w1)) != MP_OKAY) { - goto ERR; - } - /* r3 - r2 */ - if ((res = mp_sub(&w3, &w2, &w3)) != MP_OKAY) { - goto ERR; - } - /* r1/3 */ - if ((res = mp_div_3(&w1, &w1, NULL)) != MP_OKAY) { - goto ERR; - } - /* r3/3 */ - if ((res = mp_div_3(&w3, &w3, NULL)) != MP_OKAY) { - goto ERR; - } - - /* at this point shift W[n] by B*n */ - if ((res = mp_lshd(&w1, 1*B)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_lshd(&w2, 2*B)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_lshd(&w3, 3*B)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_lshd(&w4, 4*B)) != MP_OKAY) { - goto ERR; - } - - if ((res = mp_add(&w0, &w1, c)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&w2, &w3, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&w4, &tmp1, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&tmp1, c, c)) != MP_OKAY) { - goto ERR; - } - + + /* r1 - r4 */ + if ((res = mp_sub(&w1, &w4, &w1)) != MP_OKAY) { + goto ERR; + } + /* r3 - r0 */ + if ((res = mp_sub(&w3, &w0, &w3)) != MP_OKAY) { + goto ERR; + } + /* r1/2 */ + if ((res = mp_div_2(&w1, &w1)) != MP_OKAY) { + goto ERR; + } + /* r3/2 */ + if ((res = mp_div_2(&w3, &w3)) != MP_OKAY) { + goto ERR; + } + /* r2 - r0 - r4 */ + if ((res = mp_sub(&w2, &w0, &w2)) != MP_OKAY) { + goto ERR; + } + if ((res = mp_sub(&w2, &w4, &w2)) != MP_OKAY) { + goto ERR; + } + /* r1 - r2 */ + if ((res = mp_sub(&w1, &w2, &w1)) != MP_OKAY) { + goto ERR; + } + /* r3 - r2 */ + if ((res = mp_sub(&w3, &w2, &w3)) != MP_OKAY) { + goto ERR; + } + /* r1 - 8r0 */ + if ((res = mp_mul_2d(&w0, 3, &tmp1)) != MP_OKAY) { + goto ERR; + } + if ((res = mp_sub(&w1, &tmp1, &w1)) != MP_OKAY) { + goto ERR; + } + /* r3 - 8r4 */ + if ((res = mp_mul_2d(&w4, 3, &tmp1)) != MP_OKAY) { + goto ERR; + } + if ((res = mp_sub(&w3, &tmp1, &w3)) != MP_OKAY) { + goto ERR; + } + /* 3r2 - r1 - r3 */ + if ((res = mp_mul_d(&w2, 3, &w2)) != MP_OKAY) { + goto ERR; + } + if ((res = mp_sub(&w2, &w1, &w2)) != MP_OKAY) { + goto ERR; + } + if ((res = mp_sub(&w2, &w3, &w2)) != MP_OKAY) { + goto ERR; + } + /* r1 - r2 */ + if ((res = mp_sub(&w1, &w2, &w1)) != MP_OKAY) { + goto ERR; + } + /* r3 - r2 */ + if ((res = mp_sub(&w3, &w2, &w3)) != MP_OKAY) { + goto ERR; + } + /* r1/3 */ + if ((res = mp_div_3(&w1, &w1, NULL)) != MP_OKAY) { + goto ERR; + } + /* r3/3 */ + if ((res = mp_div_3(&w3, &w3, NULL)) != MP_OKAY) { + goto ERR; + } + + /* at this point shift W[n] by B*n */ + if ((res = mp_lshd(&w1, 1*B)) != MP_OKAY) { + goto ERR; + } + if ((res = mp_lshd(&w2, 2*B)) != MP_OKAY) { + goto ERR; + } + if ((res = mp_lshd(&w3, 3*B)) != MP_OKAY) { + goto ERR; + } + if ((res = mp_lshd(&w4, 4*B)) != MP_OKAY) { + goto ERR; + } + + if ((res = mp_add(&w0, &w1, c)) != MP_OKAY) { + goto ERR; + } + if ((res = mp_add(&w2, &w3, &tmp1)) != MP_OKAY) { + goto ERR; + } + if ((res = mp_add(&w4, &tmp1, &tmp1)) != MP_OKAY) { + goto ERR; + } + if ((res = mp_add(&tmp1, c, c)) != MP_OKAY) { + goto ERR; + } + ERR: - mp_clear_multi(&w0, &w1, &w2, &w3, &w4, - &a0, &a1, &a2, &b0, &b1, - &b2, &tmp1, &tmp2, NULL); - return res; -} - + mp_clear_multi(&w0, &w1, &w2, &w3, &w4, + &a0, &a1, &a2, &b0, &b1, + &b2, &tmp1, &tmp2, NULL); + return res; +} + #endif + +/* $Source$ */ +/* $Revision$ */ +/* $Date$ */ |