/* * Copyright (c) 2002 by The XFree86 Project, Inc. * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included in * all copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE XFREE86 PROJECT BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. * * Except as contained in this notice, the name of the XFree86 Project shall * not be used in advertising or otherwise to promote the sale, use or other * dealings in this Software without prior written authorization from the * XFree86 Project. * * Author: Paulo César Pereira de Andrade */ /* $XFree86: xc/programs/xedit/lisp/mp/mp.c,v 1.2 2002/11/08 08:01:00 paulo Exp $ */ #include "mp.h" /* * TODO: * o Optimize squaring * o Write better division code and move from mpi.c to here * o Make multiplication code don't required memory to be zeroed * + The first step is easy, just multiply the low word, * then the high word, that may overlap with the result * of the first multiply (in case of carry), and then * just make sure carry is properly propagated in the * subsequent multiplications. * + Some code needs also to be rewritten because some * intermediate addition code in mp_mul, mp_karatsuba_mul, * and mp_toom_mul is assuming the memory is zeroed. */ /* * Prototypes */ /* out of memory handler */ static void mp_outmem(void); /* memory allocation fallback functions */ static void *_mp_malloc(size_t); static void *_mp_calloc(size_t, size_t); static void *_mp_realloc(void*, size_t); static void _mp_free(void*); /* * Initialization */ static mp_malloc_fun __mp_malloc = _mp_malloc; static mp_calloc_fun __mp_calloc = _mp_calloc; static mp_realloc_fun __mp_realloc = _mp_realloc; static mp_free_fun __mp_free = _mp_free; /* * Implementation */ static void mp_outmem(void) { fprintf(stderr, "out of memory in MP library.\n"); exit(1); } static void * _mp_malloc(size_t size) { return (malloc(size)); } void * mp_malloc(size_t size) { void *pointer = (*__mp_malloc)(size); if (pointer == NULL) mp_outmem(); return (pointer); } mp_malloc_fun mp_set_malloc(mp_malloc_fun fun) { mp_malloc_fun old = __mp_malloc; __mp_malloc = fun; return (old); } static void * _mp_calloc(size_t nmemb, size_t size) { return (calloc(nmemb, size)); } void * mp_calloc(size_t nmemb, size_t size) { void *pointer = (*__mp_calloc)(nmemb, size); if (pointer == NULL) mp_outmem(); return (pointer); } mp_calloc_fun mp_set_calloc(mp_calloc_fun fun) { mp_calloc_fun old = __mp_calloc; __mp_calloc = fun; return (old); } static void * _mp_realloc(void *old, size_t size) { return (realloc(old, size)); } void * mp_realloc(void *old, size_t size) { void *pointer = (*__mp_realloc)(old, size); if (pointer == NULL) mp_outmem(); return (pointer); } mp_realloc_fun mp_set_realloc(mp_realloc_fun fun) { mp_realloc_fun old = __mp_realloc; __mp_realloc = fun; return (old); } static void _mp_free(void *pointer) { free(pointer); } void mp_free(void *pointer) { (*__mp_free)(pointer); } mp_free_fun mp_set_free(mp_free_fun fun) { mp_free_fun old = __mp_free; __mp_free = fun; return (old); } long mp_add(BNS *rop, BNS *op1, BNS *op2, BNI len1, BNI len2) { BNI value; /* intermediate result */ BNS carry; /* carry flag */ long size; /* result size */ if (len1 < len2) MP_SWAP(op1, op2, len1, len2); /* unroll start of loop */ value = op1[0] + op2[0]; rop[0] = value; carry = value >> BNSBITS; /* add op1 and op2 */ for (size = 1; size < len2; size++) { value = op1[size] + op2[size] + carry; rop[size] = value; carry = value >> BNSBITS; } if (rop != op1) { for (; size < len1; size++) { value = op1[size] + carry; rop[size] = value; carry = value >> BNSBITS; } } else { /* if rop == op1, than just adjust carry */ for (; carry && size < len1; size++) { value = op1[size] + carry; rop[size] = value; carry = value >> BNSBITS; } size = len1; } if (carry) rop[size++] = carry; return (size); } long mp_sub(BNS *rop, BNS *op1, BNS *op2, BNI len1, BNI len2) { long svalue; /* intermediate result */ BNS carry; /* carry flag */ long size; /* result size */ /* special case */ if (op1 == op2) { rop[0] = 0; return (1); } /* unroll start of loop */ svalue = op1[0] - op2[0]; rop[0] = svalue; carry = svalue < 0; /* subtracts op2 from op1 */ for (size = 1; size < len2; size++) { svalue = (long)(op1[size]) - op2[size] - carry; rop[size] = svalue; carry = svalue < 0; } if (rop != op1) { for (; size < len1; size++) { svalue = op1[size] - carry; rop[size] = svalue; carry = svalue < 0; } } else { /* if rop == op1, than just adjust carry */ for (; carry && size < len1; size++) { svalue = op1[size] - carry; rop[size] = svalue; carry = svalue < 0; } size = len1; } /* calculate result size */ while (size > 1 && rop[size - 1] == 0) --size; return (size); } long mp_lshift(BNS *rop, BNS *op, BNI len, long shift) { long i, size; BNI words, bits; /* how many word and bit shifts */ words = shift / BNSBITS; bits = shift % BNSBITS; size = len + words; if (bits) { BNS hi, lo; BNI carry; int adj; for (i = 1, carry = CARRY >> 1; carry; i++, carry >>= 1) if (op[len - 1] & carry) break; adj = (bits + (BNSBITS - i)) / BNSBITS; size += adj; lo = hi = op[0]; rop[words] = lo << bits; for (i = 1; i < len; i++) { hi = op[i]; rop[words + i] = hi << bits | (lo >> (BNSBITS - bits)); lo = hi; } if (adj) rop[size - 1] = hi >> (BNSBITS - bits); } else memmove(rop + size - len, op, sizeof(BNS) * len); if (words) memset(rop, '\0', sizeof(BNS) * words); return (size); } long mp_rshift(BNS *rop, BNS *op, BNI len, long shift) { int adj = 0; long i, size; BNI words, bits; /* how many word and bit shifts */ words = shift / BNSBITS; bits = shift % BNSBITS; size = len - words; if (bits) { BNS hi, lo; BNI carry; for (i = 0, carry = CARRY >> 1; carry; i++, carry >>= 1) if (op[len - 1] & carry) break; adj = (bits + i) / BNSBITS; if (size - adj == 0) { rop[0] = 0; return (1); } hi = lo = op[words + size - 1]; rop[size - 1] = hi >> bits; for (i = size - 2; i >= 0; i--) { lo = op[words + i]; rop[i] = (lo >> bits) | (hi << (BNSBITS - bits)); hi = lo; } if (adj) rop[0] |= lo << (BNSBITS - bits); } else memmove(rop, op + len - size, size * sizeof(BNS)); return (size - adj); } /* rop must be a pointer to len1 + len2 elements * rop cannot be either op1 or op2 * rop must be all zeros */ long mp_base_mul(BNS *rop, BNS *op1, BNS *op2, BNI len1, BNI len2) { long i, j; /* counters */ BNI value; /* intermediate result */ BNS carry; /* carry value */ long size = len1 + len2; /* simple optimization: first pass does not need to deference rop[i+j] */ if (op1[0]) { value = (BNI)(op1[0]) * op2[0]; rop[0] = value; carry = (BNS)(value >> BNSBITS); for (j = 1; j < len2; j++) { value = (BNI)(op1[0]) * op2[j] + carry; rop[j] = value; carry = (BNS)(value >> BNSBITS); } rop[j] = carry; } /* do the multiplication */ for (i = 1; i < len1; i++) { if (op1[i]) { /* unrool loop initialization */ value = (BNI)(op1[i]) * op2[0] + rop[i]; rop[i] = value; carry = (BNS)(value >> BNSBITS); /* multiply */ for (j = 1; j < len2; j++) { value = (BNI)(op1[i]) * op2[j] + rop[i + j] + carry; rop[i + j] = value; carry = (BNS)(value >> BNSBITS); } rop[i + j] = carry; } } if (size > 1 && rop[size - 1] == 0) --size; return (size); } /* Karatsuba method * t + ((a0 + a1) (b0 + b1) - t - u) x + ux² * where t = a0b0 and u = a1b1 * * Karatsuba method reduces the number of multiplications. Example: * Square a 40 length number * instead of a plain 40*40 = 1600 multiplies/adds, it does: * 20*20+20*20+20*20 = 1200 * but since it is recursive, every 20*20=400 is reduced to * 10*10+10*10+10*10=300 * and so on. * The multiplication by x and x² is a just a shift, as it is a * power of two, and is implemented below by just writting at the * correct offset */ long mp_karatsuba_mul(BNS *rop, BNS *op1, BNS *op2, BNI len1, BNI len2) { BNI x; /* shift count */ BNI la0, la1, lb0, lb1; /* length of a0, a1, b0, and b1 */ BNS *t; /* temporary memory for t product */ BNS *u; /* temporary memory for u product */ BNS *r; /* pointer to rop */ long xlen, tlen, ulen; /* calculate value of x, that is 2^(BNSBITS*x) */ if (len1 >= len2) x = (len1 + 1) >> 1; else x = (len2 + 1) >> 1; /* calculate length of operands */ la0 = x; la1 = len1 - x; lb0 = x; lb1 = len2 - x; /* allocate buffer for t and (a0 + a1) */ tlen = la0 + lb0; t = mp_malloc(sizeof(BNS) * tlen); /* allocate buffer for u and (b0 + b1) */ if (la1 + lb1 < lb0 + lb1 + 1) ulen = lb0 + lb1 + 1; else ulen = la1 + lb1; u = mp_malloc(sizeof(BNS) * ulen); /* calculate a0 + a1, store result in t */ tlen = mp_add(t, op1, op1 + x, la0, la1); /* calculate b0 + b1, store result in u */ ulen = mp_add(u, op2, op2 + x, lb0, lb1); /* store (a0 + a1) * (b0 + b1) in rop */ r = rop + x; /* multiplied by 2^(BNSBITS*x) */ xlen = mp_mul(r, t, u, tlen, ulen); /* must zero t and u memory, this is required for mp_mul */ /* calculate t = a0 * b0 */ tlen = la0 + lb0; memset(t, '\0', sizeof(BNS) * tlen); tlen = mp_mul(t, op1, op2, la0, lb0); /* calculate u = a1 * b1 */ ulen = la1 + lb1; memset(u, '\0', sizeof(BNS) * ulen); ulen = mp_mul(u, op1 + x, op2 + x, la1, lb1); /* subtract t from partial result */ xlen = mp_sub(r, r, t, xlen, tlen); /* subtract u form partial result */ xlen = mp_sub(r, r, u, xlen, ulen); /* add ux^2 to partial result */ r = rop + (x << 1); /* multiplied by x^2 = 2^(BNSBITS*x*2) */ xlen = len1 + len2; xlen = mp_add(r, r, u, xlen, ulen); /* now add t to final result */ xlen = mp_add(rop, rop, t, xlen, tlen); mp_free(t); mp_free(u); if (xlen > 1 && rop[xlen - 1] == 0) --xlen; return (xlen); } /* Toom method (partially based on GMP documentation) * Evaluation at k = [ 0 1/2 1 2 oo ] * U(x) = (U2k + U1)k + U0 * V(x) = (V2k + V1)k + V0 * W(x) = U(x)V(x) * * Sample: * 123 * 456 * * EVALUATION: * U(0) = (1*0+2)*0+3 => 3 * U(1) = 1+(2+3*2)*2 => 17 * U(2) = 1+2+3 => 6 * U(3) = (1*2+2)*2+3 => 11 * U(4) = 1+(2+3*0)*0 => 1 * * V(0) = (4*0+5)*0+6 => 6 * V(1) = 4+(5+6*2)*2 => 38 * V(2) = 4+5+6 => 15 * V(3) = (4*2+5)*2+6 => 32 * V(4) = 4+(5+6*0)*0 => 4 * * U = [ 3 17 6 11 1 ] * V = [ 6 38 15 32 4 ] * W = [ 18 646 90 352 4 ] * * After that, we have: * a = 18 (w0 already known) * b = 16w0 + 8w1 + 4w2 + 2w3 + w4 * c = w0 + w1 + w2 + w3 + w4 * d = w0 + 2w1 + 4w2 + 8w3 + 16w4 * e = 4 (w4 already known) * * INTERPOLATION: * b = b -16a - e (354) * c = c - a - e (68) * d = d - a - 16e (270) * * w = (b + d) - 8c = (10w1+8w2+10w3) - (8w1+8w2+8w3) = 2w1+2w3 * w = 2c - w (56) * b = b/2 = 4w1+w+w3 * b = b-c = 4w1+w+w3 - w1+w2+w3 = 3w1+w2 * c = w/2 (w2 = 28) * b = b-c = 3w1+c - c = 3w1 * b = b/3 (w1 = 27) * d = d/2 * d = d-b-w = b+w+4w3 - b-w = 4w3 * d = d/4 (w3 = 13) * * RESULT: * w4*10^4 + w3*10³ + w2*10² + w1*10 + w0 * 40000 + 13000 + 2800 + 270 + 18 * 10 is the base where the calculation was done * * This sample uses small numbers, so it does not show the * advantage of the method. But for example (in base 10), when squaring * 123456789012345678901234567890 * The normal method would do 30*30=900 multiplications * Karatsuba method would do 15*15*3=675 multiplications * Toom method would do 10*10*5=500 multiplications * Toom method has a larger overhead if compared with Karatsuba method, * due to evaluation and interpolation, so it should be used for larger * numbers, so that the computation time of evaluation/interpolation * would be smaller than the time spent using other methods. * * Note that Karatsuba method can be seen as a special case of * Toom method, i.e: * U1U0 * V1V0 * with k = [ 0 1 oo ] * U = [ U0 U1+U0 U1 ] * V = [ V0 V1+V0 V1 ] * W = [ U0*V0 (U1+U0)*(V1+V0) (U1+V1) ] * * w0 = U0*V0 * w = (U1+U0)*(V1+V0) * w2 = (U1*V1) * * w1 = w - w0 - w2 * w2x² + w1x + w0 * * See Knuth's Seminumerical Algorithms for a sample implemention * using 4 stacks and k = [ 0 1 2 3 ... ], based on the size of the * input. */ long mp_toom_mul(BNS *rop, BNS *op1, BNS *op2, BNI len1, BNI len2) { long size, xsize, i; BNI value; /* used in division */ BNS carry; BNI x; /* shift count */ BNI l1, l2; BNI al, bl, cl, dl, el, Ul[3], Vl[3]; BNS *a, *b, *c, *d, *e, *U[3], *V[3]; /* x is the base i.e. 2^(BNSBITS*x) */ x = (len1 + len2 + 4) / 6; l1 = len1 - (x << 1); /* length of remaining piece of op1 */ l2 = len2 - (x << 1); /* length of remaining piece of op2 */ /* allocate memory for storing U and V */ U[0] = mp_malloc(sizeof(BNS) * (x + 2)); V[0] = mp_malloc(sizeof(BNS) * (x + 2)); U[1] = mp_malloc(sizeof(BNS) * (x + 1)); V[1] = mp_malloc(sizeof(BNS) * (x + 1)); U[2] = mp_malloc(sizeof(BNS) * (x + 2)); V[2] = mp_malloc(sizeof(BNS) * (x + 2)); /* EVALUATE U AND V */ /* Numbers are in the format U2x²+U1x+U0 and V2x²+V1x+V0 */ /* U[0] = U2+U1*2+U0*4 */ /* store U1*2 in U[1], this value is used twice */ Ul[1] = mp_lshift(U[1], op1 + x, x, 1); /* store U0*4 in U[0] */ Ul[0] = mp_lshift(U[0], op1, x, 2); /* add U1*2 to U[0] */ Ul[0] = mp_add(U[0], U[0], U[1], Ul[0], Ul[1]); /* add U2 to U[0] */ Ul[0] = mp_add(U[0], U[0], op1 + x + x, Ul[0], l1); /* U[2] = U2*4+U1*2+U0 */ /* store U2*4 in U[2] */ Ul[2] = mp_lshift(U[2], op1 + x + x, l1, 2); /* add U1*2 to U[2] */ Ul[2] = mp_add(U[2], U[2], U[1], Ul[2], Ul[1]); /* add U0 to U[2] */ Ul[2] = mp_add(U[2], U[2], op1, Ul[2], x); /* U[1] = U2+U1+U0 */ Ul[1] = mp_add(U[1], op1, op1 + x, x, x); Ul[1] = mp_add(U[1], U[1], op1 + x + x, Ul[1], l1); /* Evaluate V[x], same code as U[x] */ Vl[1] = mp_lshift(V[1], op2 + x, x, 1); Vl[0] = mp_lshift(V[0], op2, x, 2); Vl[0] = mp_add(V[0], V[0], V[1], Vl[0], Vl[1]); Vl[0] = mp_add(V[0], V[0], op2 + x + x, Vl[0], l2); Vl[2] = mp_lshift(V[2], op2 + x + x, l2, 2); Vl[2] = mp_add(V[2], V[2], V[1], Vl[2], Vl[1]); Vl[2] = mp_add(V[2], V[2], op2, Vl[2], x); Vl[1] = mp_add(V[1], op2, op2 + x, x, x); Vl[1] = mp_add(V[1], V[1], op2 + x + x, Vl[1], l2); /* MULTIPLY U[] AND V[] */ /* calculate (U2+U1*2+U0*4) * (V2+V1*2+V0*4) */ b = mp_calloc(1, sizeof(BNS) * (Ul[0] * Vl[0])); bl = mp_mul(b, U[0], V[0], Ul[0], Vl[0]); mp_free(U[0]); mp_free(V[0]); /* calculate (U2+U1+U0) * (V2+V1+V0) */ c = mp_calloc(1, sizeof(BNS) * (Ul[1] * Vl[1])); cl = mp_mul(c, U[1], V[1], Ul[1], Vl[1]); mp_free(U[1]); mp_free(V[1]); /* calculate (U2*4+U1*2+U0) * (V2*4+V1*2+V0) */ d = mp_calloc(1, sizeof(BNS) * (Ul[2] * Vl[2])); dl = mp_mul(d, U[2], V[2], Ul[2], Vl[2]); mp_free(U[2]); mp_free(V[2]); /* calculate U0 * V0 */ a = mp_calloc(1, sizeof(BNS) * (x + x)); al = mp_mul(a, op1, op2, x, x); /* calculate U2 * V2 */ e = mp_calloc(1, sizeof(BNS) * (l1 + l2)); el = mp_mul(e, op1 + x + x, op2 + x + x, l1, l2); /* INTERPOLATE COEFFICIENTS */ /* b = b - 16a - e */ size = mp_lshift(rop, a, al, 4); bl = mp_sub(b, b, rop, bl, size); bl = mp_sub(b, b, e, bl, el); /* c = c - a - e*/ cl = mp_sub(c, c, a, cl, al); cl = mp_sub(c, c, e, cl, el); /* d = d - a - 16e */ dl = mp_sub(d, d, a, dl, al); size = mp_lshift(rop, e, el, 4); dl = mp_sub(d, d, rop, dl, size); /* w = (b + d) - 8c */ size = mp_add(rop, b, d, bl, dl); xsize = mp_lshift(rop + size, c, cl, 3); /* rop has enough storage */ size = mp_sub(rop, rop, rop + size, size, xsize); /* w = 2c - w*/ xsize = mp_lshift(rop + size, c, cl, 1); size = mp_sub(rop, rop + size, rop, xsize, size); /* b = b/2 */ bl = mp_rshift(b, b, bl, 1); /* b = b - c */ bl = mp_sub(b, b, c, bl, cl); /* c = w / 2 */ cl = mp_rshift(c, rop, size, 1); /* b = b - c */ bl = mp_sub(b, b, c, bl, cl); /* b = b/3 */ /* maybe the most expensive calculation */ i = bl - 1; value = b[i]; b[i] = value / 3; for (--i; i >= 0; i--) { carry = value % 3; value = ((BNI)carry << BNSBITS) + b[i]; b[i] = (BNS)(value / 3); } /* d = d/2 */ dl = mp_rshift(d, d, dl, 1); /* d = d - b - w */ dl = mp_sub(d, d, b, dl, bl); dl = mp_sub(d, d, rop, dl, size); /* d = d/4 */ dl = mp_rshift(d, d, dl, 2); /* STORE RESULT IN ROP */ /* first clear memory used as temporary variable w and 8c */ memset(rop, '\0', sizeof(BNS) * (len1 + len2)); i = x * 4; xsize = (len1 + len2) - i; size = mp_add(rop + i, rop + i, e, xsize, el) + i; i = x * 3; xsize = size - i; size = mp_add(rop + i, rop + i, d, xsize, dl) + i; i = x * 2; xsize = size - i; size = mp_add(rop + i, rop + i, c, xsize, cl) + i; i = x; xsize = size - i; size = mp_add(rop + i, rop + i, b, xsize, bl) + i; size = mp_add(rop, rop, a, size, al); mp_free(e); mp_free(d); mp_free(c); mp_free(b); mp_free(a); if (size > 1 && rop[size - 1] == 0) --size; return (size); } long mp_mul(BNS *rop, BNS *op1, BNS *op2, BNI len1, BNI len2) { if (len1 < len2) MP_SWAP(op1, op2, len1, len2); if (len1 < KARATSUBA || len2 < KARATSUBA) return (mp_base_mul(rop, op1, op2, len1, len2)); else if (len1 < TOOM && len2 < TOOM && len2 > ((len1 + 1) >> 1)) return (mp_karatsuba_mul(rop, op1, op2, len1, len2)); else if (len1 >= TOOM && len2 >= TOOM && (len2 + 2) / 3 == (len1 + 2) / 3) return (mp_toom_mul(rop, op1, op2, len1, len2)); else { long xsize, psize, isize; BNS *ptr; /* adjust index pointer and estimated size of result */ isize = 0; xsize = len1 + len2; mp_mul(rop, op1, op2, len2, len2); /* adjust pointers */ len1 -= len2; op1 += len2; /* allocate buffer for intermediate multiplications */ if (len1 > len2) ptr = mp_calloc(1, sizeof(BNS) * (len2 + len2)); else ptr = mp_calloc(1, sizeof(BNS) * (len1 + len2)); /* loop multiplying len2 size operands at a time */ while (len1 >= len2) { isize += len2; psize = mp_mul(ptr, op1, op2, len2, len2); mp_add(rop + isize, rop + isize, ptr, xsize - isize, psize); len1 -= len2; op1 += len2; /* multiplication routines require zeroed memory */ memset(ptr, '\0', sizeof(BNS) * (MIN(len1, len2) + len2)); } /* len1 was not a multiple of len2 */ if (len1) { isize += len2; psize = mp_mul(ptr, op2, op1, len2, len1); mp_add(rop + isize, rop + isize, ptr, xsize, psize); } /* adjust result size */ if (rop[xsize - 1] == 0) --xsize; mp_free(ptr); return (xsize); } }