-rw-r--r-- 7996 libmceliece-20240726/crypto_kem/6688128/avx/fft.c raw
/*
This file is for implementing the Gao-Mateer FFT, see
http://www.math.clemson.edu/~sgao/papers/GM10.pdf
For the implementation strategy, see
https://eprint.iacr.org/2017/793.pdf
*/
// 20240508 djb: include vec{128,256}_gf.h
// 20221230 djb: split these arrays into separate .c files
// 20221230 djb: rename consts array as fft_consts
// 20221230 djb: rename s array as fft_scalars_2x
// 20221230 djb: add linker lines
// linker define fft
// linker use vec128_mul_asm
// linker use vec256_maa_asm
// linker use transpose_64x256_sp_asm
// linker use fft_scalars_2x fft_consts fft_powers
#include "fft.h"
#include "fft_scalars_2x.h"
#include "fft_consts.h"
#include "fft_powers.h"
#include "transpose.h"
#include "vec128_gf.h"
#include "vec256_gf.h"
#include <stdint.h>
/* input: in, polynomial in bitsliced form */
/* output: in, result of applying the radix conversions on in */
static void radix_conversions(vec128 *in)
{
int i, j, k;
vec128 t;
uint64_t v0, v1;
const vec128 mask[5][2] =
{
{vec128_set2x(0x8888888888888888, 0x8888888888888888),
vec128_set2x(0x4444444444444444, 0x4444444444444444)},
{vec128_set2x(0xC0C0C0C0C0C0C0C0, 0xC0C0C0C0C0C0C0C0),
vec128_set2x(0x3030303030303030, 0x3030303030303030)},
{vec128_set2x(0xF000F000F000F000, 0xF000F000F000F000),
vec128_set2x(0x0F000F000F000F00, 0x0F000F000F000F00)},
{vec128_set2x(0xFF000000FF000000, 0xFF000000FF000000),
vec128_set2x(0x00FF000000FF0000, 0x00FF000000FF0000)},
{vec128_set2x(0xFFFF000000000000, 0xFFFF000000000000),
vec128_set2x(0x0000FFFF00000000, 0x0000FFFF00000000)}
};
//
for (j = 0; j <= 5; j++)
{
for (i = 0; i < GFBITS; i++)
{
v1 = vec128_extract(in[i], 1);
v1 ^= v1 >> 32;
v0 = vec128_extract(in[i], 0);
v0 ^= v1 << 32;
in[i] = vec128_set2x(v0, v1);
}
for (i = 0; i < GFBITS; i++)
for (k = 4; k >= j; k--)
{
t = vec128_and(in[i], mask[k][0]);
t = vec128_srl_2x(t, 1 << k);
in[i] = vec128_xor(in[i], t);
t = vec128_and(in[i], mask[k][1]);
t = vec128_srl_2x(t, 1 << k);
in[i] = vec128_xor(in[i], t);
}
if (j < 5)
vec128_mul(in, in, fft_scalars_2x[j]); // scaling
}
}
/* input: in, result of applying the radix conversions to the input polynomial */
/* output: out, evaluation results (by applying the FFT butterflies) */
static void butterflies(vec256 out[][ GFBITS ], vec128 *in)
{
int i, j, k, s, b;
vec128 tmp[ GFBITS ];
vec256 tmp0[ GFBITS ];
vec256 tmp1[ GFBITS ];
vec128 t[ GFBITS ];
union {
vec128 v[8][ GFBITS+1 ];
vec256 V[8][ (GFBITS+1)/2 ];
} pre;
union {
vec128 v[64][ 2 ];
vec256 V[64];
} buf;
uint64_t v0, v1;
uint64_t consts_ptr = 2;
const unsigned char reversal[64] =
{
0, 32, 16, 48, 8, 40, 24, 56,
4, 36, 20, 52, 12, 44, 28, 60,
2, 34, 18, 50, 10, 42, 26, 58,
6, 38, 22, 54, 14, 46, 30, 62,
1, 33, 17, 49, 9, 41, 25, 57,
5, 37, 21, 53, 13, 45, 29, 61,
3, 35, 19, 51, 11, 43, 27, 59,
7, 39, 23, 55, 15, 47, 31, 63
};
const uint16_t beta[8] = {2522, 7827, 7801, 8035, 6897, 8167, 3476, 0};
// boradcast
for (j = 0; j < GFBITS; j++)
t[j] = vec128_unpack_high(in[j], in[j]);
for (i = 0; i < 8; i+=2)
{
for (j = 0; j < GFBITS; j++)
{
v0 = (beta[i+0] >> j) & 1; v0 = -v0;
v1 = (beta[i+1] >> j) & 1; v1 = -v1;
tmp[j] = vec128_set2x(v0, v1);
}
vec128_mul(tmp, t, tmp);
for (j = 0; j < GFBITS; j++)
{
pre.v[i+0][j] = vec128_unpack_low(tmp[j], tmp[j]);
pre.v[i+1][j] = vec128_unpack_high(tmp[j], tmp[j]);
}
}
for (i = 0; i < GFBITS; i+=2)
{
if (i != GFBITS-1)
buf.v[0][1] = vec128_unpack_low(in[i+1], in[i+1] ^ pre.v[6][i+1]);
buf.v[0][0] = vec128_unpack_low(in[i+0], in[i+0] ^ pre.v[6][i+0]);
#define xor vec256_xor
buf.V[1] = xor(buf.V[0], pre.V[0][i/2]); buf.V[16] = xor(buf.V[0], pre.V[4][i/2]);
buf.V[3] = xor(buf.V[1], pre.V[1][i/2]); buf.V[48] = xor(buf.V[16], pre.V[5][i/2]);
buf.V[49] = xor(buf.V[48], pre.V[0][i/2]);
buf.V[2] = xor(buf.V[0], pre.V[1][i/2]); buf.V[51] = xor(buf.V[49], pre.V[1][i/2]);
buf.V[6] = xor(buf.V[2], pre.V[2][i/2]); buf.V[50] = xor(buf.V[51], pre.V[0][i/2]);
buf.V[7] = xor(buf.V[6], pre.V[0][i/2]); buf.V[54] = xor(buf.V[50], pre.V[2][i/2]);
buf.V[5] = xor(buf.V[7], pre.V[1][i/2]); buf.V[55] = xor(buf.V[54], pre.V[0][i/2]);
buf.V[53] = xor(buf.V[55], pre.V[1][i/2]);
buf.V[4] = xor(buf.V[0], pre.V[2][i/2]); buf.V[52] = xor(buf.V[53], pre.V[0][i/2]);
buf.V[12] = xor(buf.V[4], pre.V[3][i/2]); buf.V[60] = xor(buf.V[52], pre.V[3][i/2]);
buf.V[13] = xor(buf.V[12], pre.V[0][i/2]); buf.V[61] = xor(buf.V[60], pre.V[0][i/2]);
buf.V[15] = xor(buf.V[13], pre.V[1][i/2]); buf.V[63] = xor(buf.V[61], pre.V[1][i/2]);
buf.V[14] = xor(buf.V[15], pre.V[0][i/2]); buf.V[62] = xor(buf.V[63], pre.V[0][i/2]);
buf.V[10] = xor(buf.V[14], pre.V[2][i/2]); buf.V[58] = xor(buf.V[62], pre.V[2][i/2]);
buf.V[11] = xor(buf.V[10], pre.V[0][i/2]); buf.V[59] = xor(buf.V[58], pre.V[0][i/2]);
buf.V[9] = xor(buf.V[11], pre.V[1][i/2]); buf.V[57] = xor(buf.V[59], pre.V[1][i/2]);
buf.V[56] = xor(buf.V[57], pre.V[0][i/2]);
buf.V[8] = xor(buf.V[0], pre.V[3][i/2]); buf.V[40] = xor(buf.V[56], pre.V[4][i/2]);
buf.V[24] = xor(buf.V[8], pre.V[4][i/2]); buf.V[41] = xor(buf.V[40], pre.V[0][i/2]);
buf.V[25] = xor(buf.V[24], pre.V[0][i/2]); buf.V[43] = xor(buf.V[41], pre.V[1][i/2]);
buf.V[27] = xor(buf.V[25], pre.V[1][i/2]); buf.V[42] = xor(buf.V[43], pre.V[0][i/2]);
buf.V[26] = xor(buf.V[27], pre.V[0][i/2]); buf.V[46] = xor(buf.V[42], pre.V[2][i/2]);
buf.V[30] = xor(buf.V[26], pre.V[2][i/2]); buf.V[47] = xor(buf.V[46], pre.V[0][i/2]);
buf.V[31] = xor(buf.V[30], pre.V[0][i/2]); buf.V[45] = xor(buf.V[47], pre.V[1][i/2]);
buf.V[29] = xor(buf.V[31], pre.V[1][i/2]); buf.V[44] = xor(buf.V[45], pre.V[0][i/2]);
buf.V[28] = xor(buf.V[29], pre.V[0][i/2]); buf.V[36] = xor(buf.V[44], pre.V[3][i/2]);
buf.V[20] = xor(buf.V[28], pre.V[3][i/2]); buf.V[37] = xor(buf.V[36], pre.V[0][i/2]);
buf.V[21] = xor(buf.V[20], pre.V[0][i/2]); buf.V[39] = xor(buf.V[37], pre.V[1][i/2]);
buf.V[23] = xor(buf.V[21], pre.V[1][i/2]); buf.V[38] = xor(buf.V[39], pre.V[0][i/2]);
buf.V[22] = xor(buf.V[23], pre.V[0][i/2]); buf.V[34] = xor(buf.V[38], pre.V[2][i/2]);
buf.V[18] = xor(buf.V[22], pre.V[2][i/2]); buf.V[35] = xor(buf.V[34], pre.V[0][i/2]);
buf.V[19] = xor(buf.V[18], pre.V[0][i/2]); buf.V[33] = xor(buf.V[35], pre.V[1][i/2]);
buf.V[17] = xor(buf.V[19], pre.V[1][i/2]); buf.V[32] = xor(buf.V[33], pre.V[0][i/2]);
#undef xor
// transpose
transpose_64x256_sp(buf.V);
for (j = 0; j < 32; j++)
{
if (i != GFBITS-1)
out[j][i+1] = vec256_unpack_high(buf.V[ reversal[2*j+0] ], buf.V[ reversal[2*j+1] ]);
out[j][i+0] = vec256_unpack_low (buf.V[ reversal[2*j+0] ], buf.V[ reversal[2*j+1] ]);
}
}
// butterflies
for (k = 0; k < 32; k+=2)
{
for (b = 0; b < GFBITS; b++) tmp0[b] = vec256_unpack_low (out[k][b], out[k+1][b]);
for (b = 0; b < GFBITS; b++) tmp1[b] = vec256_unpack_high (out[k][b], out[k+1][b]);
vec256_maa_asm(tmp0, tmp1, fft_consts[1]);
for (b = 0; b < GFBITS; b++) out[k][b] = vec256_unpack_low (tmp0[b], tmp1[b]);
for (b = 0; b < GFBITS; b++) out[k+1][b] = vec256_unpack_high (tmp0[b], tmp1[b]);
}
for (i = 0; i <= 4; i++)
{
s = 1 << i;
for (j = 0; j < 32; j += 2*s)
for (k = j; k < j+s; k++)
{
vec256_maa_asm(out[k], out[k+s], fft_consts[ consts_ptr + (k-j) ]);
}
consts_ptr += (1 << i);
}
// adding the part contributed by x^128
for (i = 0; i < 32; i++)
for (b = 0; b < GFBITS; b++)
out[i][b] = vec256_xor(out[i][b], fft_powers[i][b]);
}
/* input: in, polynomial in bitsliced form */
/* output: out, bitsliced results of evaluating in all the field elements */
void fft(vec256 out[][GFBITS], vec128 *in)
{
radix_conversions(in);
butterflies(out, in);
}