-rw-r--r-- 4730 libmceliece-20240726/crypto_kem/348864/vec/bm.c raw
/*
  This file is for implementating the inversion-free Berlekamp-Massey algorithm
  see https://ieeexplore.ieee.org/document/87857

  For the implementation strategy, see
  https://eprint.iacr.org/2017/793.pdf
*/
// 20240503 djb: use crypto_*_mask functions
// 20221230 djb: add linker lines

// linker define bm
// linker use vec_mul
// linker use gf_inv

#include "bm.h"

#include "util.h"
#include "gf.h"
#include "crypto_uint64.h"

#include <stdint.h>
#include <assert.h>

static inline void vec_cmov(vec * out, vec * in, uint16_t mask)
{
	int i;

	vec m0, m1;

	m0 = vec_set1_16b(mask);
	m1 = ~m0;

	for (i = 0; i < GFBITS; i++)
	{
		out[i] = (in[i] & m0) | (out[i] & m1);
		out[i] = (in[i] & m0) | (out[i] & m1);
	}
}

static inline void interleave(vec *in, int idx0, int idx1, vec *mask, int b)
{
	int s = 1 << b;

	vec x, y;

	x = (in[idx0] & mask[0]) | ((in[idx1] & mask[0]) << s);
	y = ((in[idx0] & mask[1]) >> s) | (in[idx1] & mask[1]);

	in[idx0] = x;
	in[idx1] = y;
}

/* input: in, field elements in bitsliced form */
/* output: out, field elements in non-bitsliced form */
static inline void get_coefs(gf *out, vec *in)
{
	int i, k;

	vec mask[4][2];
	vec buf[16];
	
	for (i =  0; i < GFBITS; i++) buf[i] = in[i];
	for (i = GFBITS; i < 16; i++) buf[i] = 0;

	mask[0][0] = vec_set1_16b(0x5555);
	mask[0][1] = vec_set1_16b(0xAAAA);
	mask[1][0] = vec_set1_16b(0x3333);
	mask[1][1] = vec_set1_16b(0xCCCC);
	mask[2][0] = vec_set1_16b(0x0F0F);
	mask[2][1] = vec_set1_16b(0xF0F0);
	mask[3][0] = vec_set1_16b(0x00FF);
	mask[3][1] = vec_set1_16b(0xFF00);

	interleave(buf,  0,  8, mask[3], 3);
	interleave(buf,  1,  9, mask[3], 3);
	interleave(buf,  2, 10, mask[3], 3);
	interleave(buf,  3, 11, mask[3], 3);
	interleave(buf,  4, 12, mask[3], 3);
	interleave(buf,  5, 13, mask[3], 3);
	interleave(buf,  6, 14, mask[3], 3);
	interleave(buf,  7, 15, mask[3], 3);

	interleave(buf,  0,  4, mask[2], 2);
	interleave(buf,  1,  5, mask[2], 2);
	interleave(buf,  2,  6, mask[2], 2);
	interleave(buf,  3,  7, mask[2], 2);
	interleave(buf,  8, 12, mask[2], 2);
	interleave(buf,  9, 13, mask[2], 2);
	interleave(buf, 10, 14, mask[2], 2);
	interleave(buf, 11, 15, mask[2], 2);

	interleave(buf,  0,  2, mask[1], 1);
	interleave(buf,  1,  3, mask[1], 1);
	interleave(buf,  4,  6, mask[1], 1);
	interleave(buf,  5,  7, mask[1], 1);
	interleave(buf,  8, 10, mask[1], 1);
	interleave(buf,  9, 11, mask[1], 1);
	interleave(buf, 12, 14, mask[1], 1);
	interleave(buf, 13, 15, mask[1], 1);

	interleave(buf,  0,  1, mask[0], 0);
	interleave(buf,  2,  3, mask[0], 0);
	interleave(buf,  4,  5, mask[0], 0);
	interleave(buf,  6,  7, mask[0], 0);
	interleave(buf,  8,  9, mask[0], 0);
	interleave(buf, 10, 11, mask[0], 0);
	interleave(buf, 12, 13, mask[0], 0);
	interleave(buf, 14, 15, mask[0], 0);

	for (i = 0; i < 16; i++)
	for (k = 0; k <  4; k++)
		out[ k*16 + i ] = (buf[i] >> (k*16)) & GFMASK;
}

static inline gf vec_reduce(vec *in)
{
	int i;
	vec tmp;
	gf ret = 0;

	for (i = GFBITS-1; i >= 0; i--)
	{
		tmp = in[i];

		tmp ^= tmp >> 32;
		tmp ^= tmp >> 16;
		tmp ^= tmp >> 8;
		tmp ^= tmp >> 4;
		tmp ^= tmp >> 2;
		tmp ^= tmp >> 1;
	
		ret <<= 1;
		ret |= tmp & 1;
	}

	return ret;
}

static void update(vec *in, const gf e)
{
	int i;
	vec tmp;

	for (i = 0; i < GFBITS; i++)
	{
		tmp = (e >> i) & 1;

		in[i] = (in[i] >> 1) | (tmp << 63);
	}
}

/* input: in, sequence of field elements */
/* output: out, minimal polynomial of in */
void bm(vec *out, vec in[][ GFBITS ])
{
	uint16_t i;
	uint16_t N, L;

	vec prod[ GFBITS ];
	vec in_tmp[ GFBITS ];

	vec d_vec[ GFBITS ];
	vec b_vec[ GFBITS ];
	vec B[ GFBITS ], C[ GFBITS ];
	vec B_tmp[ GFBITS ], C_tmp[ GFBITS ];

	vec mask, t;

	gf d, b, c0=1;

	gf coefs[SYS_T * 2];

	// init

	get_coefs(&coefs[  0], in[0]);
	get_coefs(&coefs[ 64], in[1]);

	C[0] = 0;
	B[0] = 1; 
	B[0] <<= 63;

	for (i = 1; i < GFBITS; i++)
		B[i] = C[i] = 0;

	b = 1;
	L = 0;

	//

	for (i = 0; i < GFBITS; i++)
		in_tmp[i] = 0;

	for (N = 0; N < SYS_T * 2; N++)
	{
		// computing d

		vec_mul(prod, in_tmp, C);

		update(in_tmp, coefs[N]);

		d = vec_reduce(prod);

		t = gf_mul2(c0, coefs[N], b);
		d ^= t & 0xFFFFFFFF;

		// 3 cases

		mask = crypto_uint64_nonzero_mask(d) & crypto_uint64_leq_mask(L*2, N);

		for (i = 0; i < GFBITS; i++) 
		{
			d_vec[i] = vec_setbits((d >> i) & 1);
			b_vec[i] = vec_setbits((b >> i) & 1);
		}
		
		vec_mul(B_tmp, d_vec, B);
		vec_mul(C_tmp, b_vec, C);

		vec_cmov(B, C, mask);
		update(B, mask & c0);

		for (i = 0; i < GFBITS; i++) 
			C[i] = B_tmp[i] ^ C_tmp[i];

		c0 = t >> 32; 
		b = (d & mask) | (b & ~mask);
		L = ((N+1-L) & mask) | (L & ~mask);

	}

	c0 = gf_inv(c0);

	for (i = 0; i < GFBITS; i++) 
		out[i] = vec_setbits((c0 >> i) & 1);

	vec_mul(out, out, C);
}