diff options
author | jan.nijtmans <nijtmans@users.sourceforge.net> | 2018-10-28 19:26:49 (GMT) |
---|---|---|
committer | jan.nijtmans <nijtmans@users.sourceforge.net> | 2018-10-28 19:26:49 (GMT) |
commit | ccb97d88ffefe602e7eb5a9610bd356d66bc2f20 (patch) | |
tree | 417cc58c53b8773c08262c85128c46225c4d2ea0 /libtommath/bn_mp_toom_mul.c | |
parent | 3c4cc0a0013a0552c90518a995ae654571c18a5b (diff) | |
parent | fd548863930daf0e1bc3d01286f542844b1cc69c (diff) | |
download | tcl-ccb97d88ffefe602e7eb5a9610bd356d66bc2f20.zip tcl-ccb97d88ffefe602e7eb5a9610bd356d66bc2f20.tar.gz tcl-ccb97d88ffefe602e7eb5a9610bd356d66bc2f20.tar.bz2 |
Merge 8.7
Diffstat (limited to 'libtommath/bn_mp_toom_mul.c')
-rw-r--r-- | libtommath/bn_mp_toom_mul.c | 468 |
1 files changed, 233 insertions, 235 deletions
diff --git a/libtommath/bn_mp_toom_mul.c b/libtommath/bn_mp_toom_mul.c index 4731f8f..ff7df02 100644 --- a/libtommath/bn_mp_toom_mul.c +++ b/libtommath/bn_mp_toom_mul.c @@ -1,4 +1,4 @@ -#include <tommath_private.h> +#include "tommath_private.h" #ifdef BN_MP_TOOM_MUL_C /* LibTomMath, multiple-precision integer library -- Tom St Denis * @@ -11,8 +11,6 @@ * * The library is free for all purposes without any express * guarantee it works. - * - * Tom St Denis, tstdenis82@gmail.com, http://libtom.org */ /* multiplication using the Toom-Cook 3-way algorithm @@ -22,265 +20,265 @@ * 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) +int mp_toom_mul(const mp_int *a, const mp_int *b, mp_int *c) { - mp_int w0, w1, w2, w3, w4, tmp1, tmp2, a0, a1, a2, b0, b1, b2; - int res, B; + 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, - &b2, &tmp1, &tmp2, NULL)) != MP_OKAY) { - return res; - } + /* init temps */ + 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; + /* 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; - } + /* a = a2 * B**2 + a1 * B + a0 */ + if ((res = mp_mod_2d(a, DIGIT_BIT * B, &a0)) != MP_OKAY) { + goto LBL_ERR; + } - if ((res = mp_copy(a, &a1)) != MP_OKAY) { - goto ERR; - } - mp_rshd(&a1, B); - if ((res = mp_mod_2d(&a1, DIGIT_BIT * B, &a1)) != MP_OKAY) { - goto ERR; - } + if ((res = mp_copy(a, &a1)) != MP_OKAY) { + goto LBL_ERR; + } + mp_rshd(&a1, B); + if ((res = mp_mod_2d(&a1, DIGIT_BIT * B, &a1)) != MP_OKAY) { + goto LBL_ERR; + } - if ((res = mp_copy(a, &a2)) != MP_OKAY) { - goto ERR; - } - mp_rshd(&a2, B*2); + if ((res = mp_copy(a, &a2)) != MP_OKAY) { + goto LBL_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; - } + /* b = b2 * B**2 + b1 * B + b0 */ + if ((res = mp_mod_2d(b, DIGIT_BIT * B, &b0)) != MP_OKAY) { + goto LBL_ERR; + } - if ((res = mp_copy(b, &b1)) != MP_OKAY) { - goto ERR; - } - mp_rshd(&b1, B); - (void)mp_mod_2d(&b1, DIGIT_BIT * B, &b1); + if ((res = mp_copy(b, &b1)) != MP_OKAY) { + goto LBL_ERR; + } + mp_rshd(&b1, B); + (void)mp_mod_2d(&b1, DIGIT_BIT * B, &b1); - if ((res = mp_copy(b, &b2)) != MP_OKAY) { - goto ERR; - } - mp_rshd(&b2, B*2); + if ((res = mp_copy(b, &b2)) != MP_OKAY) { + goto LBL_ERR; + } + mp_rshd(&b2, B*2); - /* w0 = a0*b0 */ - if ((res = mp_mul(&a0, &b0, &w0)) != MP_OKAY) { - goto ERR; - } + /* w0 = a0*b0 */ + if ((res = mp_mul(&a0, &b0, &w0)) != MP_OKAY) { + goto LBL_ERR; + } - /* w4 = a2 * b2 */ - if ((res = mp_mul(&a2, &b2, &w4)) != MP_OKAY) { - goto ERR; - } + /* w4 = a2 * b2 */ + if ((res = mp_mul(&a2, &b2, &w4)) != MP_OKAY) { + goto LBL_ERR; + } - /* w1 = (a2 + 2(a1 + 2a0))(b2 + 2(b1 + 2b0)) */ - if ((res = mp_mul_2(&a0, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&tmp1, &a1, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_mul_2(&tmp1, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&tmp1, &a2, &tmp1)) != MP_OKAY) { - goto ERR; - } + /* w1 = (a2 + 2(a1 + 2a0))(b2 + 2(b1 + 2b0)) */ + if ((res = mp_mul_2(&a0, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&tmp1, &a1, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_mul_2(&tmp1, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&tmp1, &a2, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } - if ((res = mp_mul_2(&b0, &tmp2)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&tmp2, &b1, &tmp2)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_mul_2(&tmp2, &tmp2)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&tmp2, &b2, &tmp2)) != MP_OKAY) { - goto ERR; - } + if ((res = mp_mul_2(&b0, &tmp2)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&tmp2, &b1, &tmp2)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_mul_2(&tmp2, &tmp2)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&tmp2, &b2, &tmp2)) != MP_OKAY) { + goto LBL_ERR; + } - if ((res = mp_mul(&tmp1, &tmp2, &w1)) != MP_OKAY) { - goto ERR; - } + if ((res = mp_mul(&tmp1, &tmp2, &w1)) != MP_OKAY) { + goto LBL_ERR; + } - /* w3 = (a0 + 2(a1 + 2a2))(b0 + 2(b1 + 2b2)) */ - if ((res = mp_mul_2(&a2, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&tmp1, &a1, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_mul_2(&tmp1, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&tmp1, &a0, &tmp1)) != MP_OKAY) { - goto ERR; - } + /* w3 = (a0 + 2(a1 + 2a2))(b0 + 2(b1 + 2b2)) */ + if ((res = mp_mul_2(&a2, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&tmp1, &a1, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_mul_2(&tmp1, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&tmp1, &a0, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } - if ((res = mp_mul_2(&b2, &tmp2)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&tmp2, &b1, &tmp2)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_mul_2(&tmp2, &tmp2)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&tmp2, &b0, &tmp2)) != MP_OKAY) { - goto ERR; - } + if ((res = mp_mul_2(&b2, &tmp2)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&tmp2, &b1, &tmp2)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_mul_2(&tmp2, &tmp2)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&tmp2, &b0, &tmp2)) != MP_OKAY) { + goto LBL_ERR; + } - if ((res = mp_mul(&tmp1, &tmp2, &w3)) != MP_OKAY) { - goto ERR; - } + if ((res = mp_mul(&tmp1, &tmp2, &w3)) != MP_OKAY) { + goto LBL_ERR; + } - /* w2 = (a2 + a1 + a0)(b2 + b1 + b0) */ - if ((res = mp_add(&a2, &a1, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&tmp1, &a0, &tmp1)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&b2, &b1, &tmp2)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_add(&tmp2, &b0, &tmp2)) != MP_OKAY) { - goto ERR; - } - if ((res = mp_mul(&tmp1, &tmp2, &w2)) != MP_OKAY) { - goto ERR; - } + /* w2 = (a2 + a1 + a0)(b2 + b1 + b0) */ + if ((res = mp_add(&a2, &a1, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&tmp1, &a0, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&b2, &b1, &tmp2)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&tmp2, &b0, &tmp2)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_mul(&tmp1, &tmp2, &w2)) != MP_OKAY) { + goto LBL_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 + 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; - } + /* r1 - r4 */ + if ((res = mp_sub(&w1, &w4, &w1)) != MP_OKAY) { + goto LBL_ERR; + } + /* r3 - r0 */ + if ((res = mp_sub(&w3, &w0, &w3)) != MP_OKAY) { + goto LBL_ERR; + } + /* r1/2 */ + if ((res = mp_div_2(&w1, &w1)) != MP_OKAY) { + goto LBL_ERR; + } + /* r3/2 */ + if ((res = mp_div_2(&w3, &w3)) != MP_OKAY) { + goto LBL_ERR; + } + /* r2 - r0 - r4 */ + if ((res = mp_sub(&w2, &w0, &w2)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_sub(&w2, &w4, &w2)) != MP_OKAY) { + goto LBL_ERR; + } + /* r1 - r2 */ + if ((res = mp_sub(&w1, &w2, &w1)) != MP_OKAY) { + goto LBL_ERR; + } + /* r3 - r2 */ + if ((res = mp_sub(&w3, &w2, &w3)) != MP_OKAY) { + goto LBL_ERR; + } + /* r1 - 8r0 */ + if ((res = mp_mul_2d(&w0, 3, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_sub(&w1, &tmp1, &w1)) != MP_OKAY) { + goto LBL_ERR; + } + /* r3 - 8r4 */ + if ((res = mp_mul_2d(&w4, 3, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_sub(&w3, &tmp1, &w3)) != MP_OKAY) { + goto LBL_ERR; + } + /* 3r2 - r1 - r3 */ + if ((res = mp_mul_d(&w2, 3uL, &w2)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_sub(&w2, &w1, &w2)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_sub(&w2, &w3, &w2)) != MP_OKAY) { + goto LBL_ERR; + } + /* r1 - r2 */ + if ((res = mp_sub(&w1, &w2, &w1)) != MP_OKAY) { + goto LBL_ERR; + } + /* r3 - r2 */ + if ((res = mp_sub(&w3, &w2, &w3)) != MP_OKAY) { + goto LBL_ERR; + } + /* r1/3 */ + if ((res = mp_div_3(&w1, &w1, NULL)) != MP_OKAY) { + goto LBL_ERR; + } + /* r3/3 */ + if ((res = mp_div_3(&w3, &w3, NULL)) != MP_OKAY) { + goto LBL_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; - } + /* at this point shift W[n] by B*n */ + if ((res = mp_lshd(&w1, 1*B)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_lshd(&w2, 2*B)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_lshd(&w3, 3*B)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_lshd(&w4, 4*B)) != MP_OKAY) { + goto LBL_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; - } + if ((res = mp_add(&w0, &w1, c)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&w2, &w3, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&w4, &tmp1, &tmp1)) != MP_OKAY) { + goto LBL_ERR; + } + if ((res = mp_add(&tmp1, c, c)) != MP_OKAY) { + goto LBL_ERR; + } -ERR: - mp_clear_multi(&w0, &w1, &w2, &w3, &w4, - &a0, &a1, &a2, &b0, &b1, - &b2, &tmp1, &tmp2, NULL); - return res; +LBL_ERR: + mp_clear_multi(&w0, &w1, &w2, &w3, &w4, + &a0, &a1, &a2, &b0, &b1, + &b2, &tmp1, &tmp2, NULL); + return res; } #endif -/* $Source$ */ -/* $Revision$ */ -/* $Date$ */ +/* ref: $Format:%D$ */ +/* git commit: $Format:%H$ */ +/* commit time: $Format:%ai$ */ |