| 1 |
/* |
| 2 |
This file is for implementing the Gao-Mateer FFT, see |
| 3 |
http://www.math.clemson.edu/~sgao/papers/GM10.pdf |
| 4 |
|
| 5 |
For the implementation strategy, see |
| 6 |
https://eprint.iacr.org/2017/793.pdf |
| 7 |
*/ |
| 8 |
// 20240805 djb: more use of cryptoint |
| 9 |
// 20240508 djb: include vec{128,256}_gf.h |
| 10 |
// 20221230 djb: split these arrays into separate .c files |
| 11 |
// 20221230 djb: rename consts array as fft_consts |
| 12 |
// 20221230 djb: rename s array as fft_scalars_2x |
| 13 |
// 20221230 djb: add linker lines |
| 14 |
|
| 15 |
// linker define fft |
| 16 |
// linker use vec128_mul_asm |
| 17 |
// linker use vec256_maa_asm |
| 18 |
// linker use transpose_64x256_sp_asm |
| 19 |
// linker use fft_scalars_2x fft_consts fft_powers |
| 20 |
|
| 21 |
#include "fft.h" |
| 22 |
#include "fft_scalars_2x.h" |
| 23 |
#include "fft_consts.h" |
| 24 |
#include "fft_powers.h" |
| 25 |
|
| 26 |
#include "transpose.h" |
| 27 |
#include "vec128_gf.h" |
| 28 |
#include "vec256_gf.h" |
| 29 |
|
| 30 |
#include <stdint.h> |
| 31 |
#include "crypto_int64.h" |
| 32 |
|
| 33 |
/* input: in, polynomial in bitsliced form */ |
| 34 |
/* output: in, result of applying the radix conversions on in */ |
| 35 |
static void radix_conversions(vec128 *in) |
| 36 |
{ |
| 37 |
int i, j, k; |
| 38 |
vec128 t; |
| 39 |
uint64_t v0, v1; |
| 40 |
|
| 41 |
const vec128 mask[5][2] = |
| 42 |
{ |
| 43 |
{vec128_set2x(0x8888888888888888, 0x8888888888888888), |
| 44 |
vec128_set2x(0x4444444444444444, 0x4444444444444444)}, |
| 45 |
{vec128_set2x(0xC0C0C0C0C0C0C0C0, 0xC0C0C0C0C0C0C0C0), |
| 46 |
vec128_set2x(0x3030303030303030, 0x3030303030303030)}, |
| 47 |
{vec128_set2x(0xF000F000F000F000, 0xF000F000F000F000), |
| 48 |
vec128_set2x(0x0F000F000F000F00, 0x0F000F000F000F00)}, |
| 49 |
{vec128_set2x(0xFF000000FF000000, 0xFF000000FF000000), |
| 50 |
vec128_set2x(0x00FF000000FF0000, 0x00FF000000FF0000)}, |
| 51 |
{vec128_set2x(0xFFFF000000000000, 0xFFFF000000000000), |
| 52 |
vec128_set2x(0x0000FFFF00000000, 0x0000FFFF00000000)} |
| 53 |
}; |
| 54 |
|
| 55 |
// |
| 56 |
|
| 57 |
for (j = 0; j <= 5; j++) |
| 58 |
{ |
| 59 |
for (i = 0; i < GFBITS; i++) |
| 60 |
{ |
| 61 |
v1 = vec128_extract(in[i], 1); |
| 62 |
v1 ^= v1 >> 32; |
| 63 |
v0 = vec128_extract(in[i], 0); |
| 64 |
v0 ^= v1 << 32; |
| 65 |
in[i] = vec128_set2x(v0, v1); |
| 66 |
} |
| 67 |
|
| 68 |
for (i = 0; i < GFBITS; i++) |
| 69 |
for (k = 4; k >= j; k--) |
| 70 |
{ |
| 71 |
t = vec128_and(in[i], mask[k][0]); |
| 72 |
t = vec128_srl_2x(t, 1 << k); |
| 73 |
in[i] = vec128_xor(in[i], t); |
| 74 |
|
| 75 |
t = vec128_and(in[i], mask[k][1]); |
| 76 |
t = vec128_srl_2x(t, 1 << k); |
| 77 |
in[i] = vec128_xor(in[i], t); |
| 78 |
} |
| 79 |
|
| 80 |
if (j < 5) |
| 81 |
vec128_mul(in, in, fft_scalars_2x[j]); // scaling |
| 82 |
} |
| 83 |
} |
| 84 |
|
| 85 |
/* input: in, result of applying the radix conversions to the input polynomial */ |
| 86 |
/* output: out, evaluation results (by applying the FFT butterflies) */ |
| 87 |
static void butterflies(vec256 out[][ GFBITS ], vec128 *in) |
| 88 |
{ |
| 89 |
int i, j, k, s, b; |
| 90 |
|
| 91 |
vec128 tmp[ GFBITS ]; |
| 92 |
vec256 tmp0[ GFBITS ]; |
| 93 |
vec256 tmp1[ GFBITS ]; |
| 94 |
vec128 t[ GFBITS ]; |
| 95 |
|
| 96 |
union { |
| 97 |
vec128 v[8][ GFBITS+1 ]; |
| 98 |
vec256 V[8][ (GFBITS+1)/2 ]; |
| 99 |
} pre; |
| 100 |
|
| 101 |
union { |
| 102 |
vec128 v[64][ 2 ]; |
| 103 |
vec256 V[64]; |
| 104 |
} buf; |
| 105 |
|
| 106 |
uint64_t v0, v1; |
| 107 |
|
| 108 |
uint64_t consts_ptr = 2; |
| 109 |
|
| 110 |
const unsigned char reversal[64] = |
| 111 |
{ |
| 112 |
0, 32, 16, 48, 8, 40, 24, 56, |
| 113 |
4, 36, 20, 52, 12, 44, 28, 60, |
| 114 |
2, 34, 18, 50, 10, 42, 26, 58, |
| 115 |
6, 38, 22, 54, 14, 46, 30, 62, |
| 116 |
1, 33, 17, 49, 9, 41, 25, 57, |
| 117 |
5, 37, 21, 53, 13, 45, 29, 61, |
| 118 |
3, 35, 19, 51, 11, 43, 27, 59, |
| 119 |
7, 39, 23, 55, 15, 47, 31, 63 |
| 120 |
}; |
| 121 |
|
| 122 |
const uint16_t beta[8] = {2522, 7827, 7801, 8035, 6897, 8167, 3476, 0}; |
| 123 |
|
| 124 |
// boradcast |
| 125 |
|
| 126 |
for (j = 0; j < GFBITS; j++) |
| 127 |
t[j] = vec128_unpack_high(in[j], in[j]); |
| 128 |
|
| 129 |
for (i = 0; i < 8; i+=2) |
| 130 |
{ |
| 131 |
for (j = 0; j < GFBITS; j++) |
| 132 |
{ |
| 133 |
v0 = crypto_int64_bitmod_mask(beta[i+0], j); |
| 134 |
v1 = crypto_int64_bitmod_mask(beta[i+1], j); |
| 135 |
|
| 136 |
tmp[j] = vec128_set2x(v0, v1); |
| 137 |
} |
| 138 |
|
| 139 |
vec128_mul(tmp, t, tmp); |
| 140 |
|
| 141 |
for (j = 0; j < GFBITS; j++) |
| 142 |
{ |
| 143 |
pre.v[i+0][j] = vec128_unpack_low(tmp[j], tmp[j]); |
| 144 |
pre.v[i+1][j] = vec128_unpack_high(tmp[j], tmp[j]); |
| 145 |
} |
| 146 |
} |
| 147 |
|
| 148 |
for (i = 0; i < GFBITS; i+=2) |
| 149 |
{ |
| 150 |
if (i != GFBITS-1) |
| 151 |
buf.v[0][1] = vec128_unpack_low(in[i+1], in[i+1] ^ pre.v[6][i+1]); |
| 152 |
buf.v[0][0] = vec128_unpack_low(in[i+0], in[i+0] ^ pre.v[6][i+0]); |
| 153 |
|
| 154 |
#define XOR vec256_xor |
| 155 |
|
| 156 |
buf.V[1] = XOR(buf.V[0], pre.V[0][i/2]); buf.V[16] = XOR(buf.V[0], pre.V[4][i/2]); |
| 157 |
buf.V[3] = XOR(buf.V[1], pre.V[1][i/2]); buf.V[48] = XOR(buf.V[16], pre.V[5][i/2]); |
| 158 |
buf.V[49] = XOR(buf.V[48], pre.V[0][i/2]); |
| 159 |
buf.V[2] = XOR(buf.V[0], pre.V[1][i/2]); buf.V[51] = XOR(buf.V[49], pre.V[1][i/2]); |
| 160 |
buf.V[6] = XOR(buf.V[2], pre.V[2][i/2]); buf.V[50] = XOR(buf.V[51], pre.V[0][i/2]); |
| 161 |
buf.V[7] = XOR(buf.V[6], pre.V[0][i/2]); buf.V[54] = XOR(buf.V[50], pre.V[2][i/2]); |
| 162 |
buf.V[5] = XOR(buf.V[7], pre.V[1][i/2]); buf.V[55] = XOR(buf.V[54], pre.V[0][i/2]); |
| 163 |
buf.V[53] = XOR(buf.V[55], pre.V[1][i/2]); |
| 164 |
buf.V[4] = XOR(buf.V[0], pre.V[2][i/2]); buf.V[52] = XOR(buf.V[53], pre.V[0][i/2]); |
| 165 |
buf.V[12] = XOR(buf.V[4], pre.V[3][i/2]); buf.V[60] = XOR(buf.V[52], pre.V[3][i/2]); |
| 166 |
buf.V[13] = XOR(buf.V[12], pre.V[0][i/2]); buf.V[61] = XOR(buf.V[60], pre.V[0][i/2]); |
| 167 |
buf.V[15] = XOR(buf.V[13], pre.V[1][i/2]); buf.V[63] = XOR(buf.V[61], pre.V[1][i/2]); |
| 168 |
buf.V[14] = XOR(buf.V[15], pre.V[0][i/2]); buf.V[62] = XOR(buf.V[63], pre.V[0][i/2]); |
| 169 |
buf.V[10] = XOR(buf.V[14], pre.V[2][i/2]); buf.V[58] = XOR(buf.V[62], pre.V[2][i/2]); |
| 170 |
buf.V[11] = XOR(buf.V[10], pre.V[0][i/2]); buf.V[59] = XOR(buf.V[58], pre.V[0][i/2]); |
| 171 |
buf.V[9] = XOR(buf.V[11], pre.V[1][i/2]); buf.V[57] = XOR(buf.V[59], pre.V[1][i/2]); |
| 172 |
buf.V[56] = XOR(buf.V[57], pre.V[0][i/2]); |
| 173 |
buf.V[8] = XOR(buf.V[0], pre.V[3][i/2]); buf.V[40] = XOR(buf.V[56], pre.V[4][i/2]); |
| 174 |
buf.V[24] = XOR(buf.V[8], pre.V[4][i/2]); buf.V[41] = XOR(buf.V[40], pre.V[0][i/2]); |
| 175 |
buf.V[25] = XOR(buf.V[24], pre.V[0][i/2]); buf.V[43] = XOR(buf.V[41], pre.V[1][i/2]); |
| 176 |
buf.V[27] = XOR(buf.V[25], pre.V[1][i/2]); buf.V[42] = XOR(buf.V[43], pre.V[0][i/2]); |
| 177 |
buf.V[26] = XOR(buf.V[27], pre.V[0][i/2]); buf.V[46] = XOR(buf.V[42], pre.V[2][i/2]); |
| 178 |
buf.V[30] = XOR(buf.V[26], pre.V[2][i/2]); buf.V[47] = XOR(buf.V[46], pre.V[0][i/2]); |
| 179 |
buf.V[31] = XOR(buf.V[30], pre.V[0][i/2]); buf.V[45] = XOR(buf.V[47], pre.V[1][i/2]); |
| 180 |
buf.V[29] = XOR(buf.V[31], pre.V[1][i/2]); buf.V[44] = XOR(buf.V[45], pre.V[0][i/2]); |
| 181 |
buf.V[28] = XOR(buf.V[29], pre.V[0][i/2]); buf.V[36] = XOR(buf.V[44], pre.V[3][i/2]); |
| 182 |
buf.V[20] = XOR(buf.V[28], pre.V[3][i/2]); buf.V[37] = XOR(buf.V[36], pre.V[0][i/2]); |
| 183 |
buf.V[21] = XOR(buf.V[20], pre.V[0][i/2]); buf.V[39] = XOR(buf.V[37], pre.V[1][i/2]); |
| 184 |
buf.V[23] = XOR(buf.V[21], pre.V[1][i/2]); buf.V[38] = XOR(buf.V[39], pre.V[0][i/2]); |
| 185 |
buf.V[22] = XOR(buf.V[23], pre.V[0][i/2]); buf.V[34] = XOR(buf.V[38], pre.V[2][i/2]); |
| 186 |
buf.V[18] = XOR(buf.V[22], pre.V[2][i/2]); buf.V[35] = XOR(buf.V[34], pre.V[0][i/2]); |
| 187 |
buf.V[19] = XOR(buf.V[18], pre.V[0][i/2]); buf.V[33] = XOR(buf.V[35], pre.V[1][i/2]); |
| 188 |
buf.V[17] = XOR(buf.V[19], pre.V[1][i/2]); buf.V[32] = XOR(buf.V[33], pre.V[0][i/2]); |
| 189 |
|
| 190 |
#undef XOR |
| 191 |
|
| 192 |
// transpose |
| 193 |
|
| 194 |
transpose_64x256_sp(buf.V); |
| 195 |
|
| 196 |
for (j = 0; j < 32; j++) |
| 197 |
{ |
| 198 |
if (i != GFBITS-1) |
| 199 |
out[j][i+1] = vec256_unpack_high(buf.V[ reversal[2*j+0] ], buf.V[ reversal[2*j+1] ]); |
| 200 |
out[j][i+0] = vec256_unpack_low (buf.V[ reversal[2*j+0] ], buf.V[ reversal[2*j+1] ]); |
| 201 |
} |
| 202 |
} |
| 203 |
|
| 204 |
// butterflies |
| 205 |
|
| 206 |
for (k = 0; k < 32; k+=2) |
| 207 |
{ |
| 208 |
for (b = 0; b < GFBITS; b++) tmp0[b] = vec256_unpack_low (out[k][b], out[k+1][b]); |
| 209 |
for (b = 0; b < GFBITS; b++) tmp1[b] = vec256_unpack_high (out[k][b], out[k+1][b]); |
| 210 |
|
| 211 |
vec256_maa_asm(tmp0, tmp1, fft_consts[1]); |
| 212 |
|
| 213 |
for (b = 0; b < GFBITS; b++) out[k][b] = vec256_unpack_low (tmp0[b], tmp1[b]); |
| 214 |
for (b = 0; b < GFBITS; b++) out[k+1][b] = vec256_unpack_high (tmp0[b], tmp1[b]); |
| 215 |
} |
| 216 |
|
| 217 |
for (i = 0; i <= 4; i++) |
| 218 |
{ |
| 219 |
s = 1 << i; |
| 220 |
|
| 221 |
for (j = 0; j < 32; j += 2*s) |
| 222 |
for (k = j; k < j+s; k++) |
| 223 |
{ |
| 224 |
vec256_maa_asm(out[k], out[k+s], fft_consts[ consts_ptr + (k-j) ]); |
| 225 |
} |
| 226 |
|
| 227 |
consts_ptr += (1 << i); |
| 228 |
} |
| 229 |
|
| 230 |
// adding the part contributed by x^128 |
| 231 |
|
| 232 |
for (i = 0; i < 32; i++) |
| 233 |
for (b = 0; b < GFBITS; b++) |
| 234 |
out[i][b] = vec256_xor(out[i][b], fft_powers[i][b]); |
| 235 |
} |
| 236 |
|
| 237 |
/* input: in, polynomial in bitsliced form */ |
| 238 |
/* output: out, bitsliced results of evaluating in all the field elements */ |
| 239 |
void fft(vec256 out[][GFBITS], vec128 *in) |
| 240 |
{ |
| 241 |
radix_conversions(in); |
| 242 |
butterflies(out, in); |
| 243 |
} |
| 244 |
|