| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* | ||
| 2 | * Xvid MPEG-4 IDCT | ||
| 3 | * | ||
| 4 | * Copyright (C) 2006-2011 Xvid Solutions GmbH | ||
| 5 | * | ||
| 6 | * This file is part of FFmpeg. | ||
| 7 | * | ||
| 8 | * FFmpeg is free software; you can redistribute it and/or | ||
| 9 | * modify it under the terms of the GNU Lesser General Public | ||
| 10 | * License as published by the Free Software Foundation; either | ||
| 11 | * version 2.1 of the License, or (at your option) any later version. | ||
| 12 | * | ||
| 13 | * FFmpeg is distributed in the hope that it will be useful, | ||
| 14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | ||
| 16 | * Lesser General Public License for more details. | ||
| 17 | * | ||
| 18 | * You should have received a copy of the GNU Lesser General Public | ||
| 19 | * License along with FFmpeg; if not, write to the Free Software | ||
| 20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | ||
| 21 | */ | ||
| 22 | |||
| 23 | /** | ||
| 24 | * @file | ||
| 25 | * Walken IDCT | ||
| 26 | * Alternative IDCT implementation for decoding compatibility. | ||
| 27 | * | ||
| 28 | * @author Skal | ||
| 29 | * @note This C version is not the original IDCT, but a modified one that | ||
| 30 | * yields the same error profile as the MMX/MMXEXT/SSE2 versions. | ||
| 31 | */ | ||
| 32 | |||
| 33 | #include "config.h" | ||
| 34 | #include "libavutil/attributes.h" | ||
| 35 | #include "idctdsp.h" | ||
| 36 | #include "xvididct.h" | ||
| 37 | |||
| 38 | #define ROW_SHIFT 11 | ||
| 39 | #define COL_SHIFT 6 | ||
| 40 | |||
| 41 | // #define FIX(x) (int)((x) * (1 << ROW_SHIFT)) | ||
| 42 | #define RND0 65536 // 1 << (COL_SHIFT + ROW_SHIFT - 1); | ||
| 43 | #define RND1 3597 // FIX (1.75683487303); | ||
| 44 | #define RND2 2260 // FIX (1.10355339059); | ||
| 45 | #define RND3 1203 // FIX (0.587788325588); | ||
| 46 | #define RND4 0 | ||
| 47 | #define RND5 120 // FIX (0.058658283817); | ||
| 48 | #define RND6 512 // FIX (0.25); | ||
| 49 | #define RND7 512 // FIX (0.25); | ||
| 50 | |||
| 51 | static const int TAB04[] = { 22725, 21407, 19266, 16384, 12873, 8867, 4520 }; | ||
| 52 | static const int TAB17[] = { 31521, 29692, 26722, 22725, 17855, 12299, 6270 }; | ||
| 53 | static const int TAB26[] = { 29692, 27969, 25172, 21407, 16819, 11585, 5906 }; | ||
| 54 | static const int TAB35[] = { 26722, 25172, 22654, 19266, 15137, 10426, 5315 }; | ||
| 55 | |||
| 56 | 480000 | static int idct_row(short *in, const int *const tab, int rnd) | |
| 57 | { | ||
| 58 | 480000 | const unsigned c1 = tab[0]; | |
| 59 | 480000 | const unsigned c2 = tab[1]; | |
| 60 | 480000 | const unsigned c3 = tab[2]; | |
| 61 | 480000 | const unsigned c4 = tab[3]; | |
| 62 | 480000 | const unsigned c5 = tab[4]; | |
| 63 | 480000 | const unsigned c6 = tab[5]; | |
| 64 | 480000 | const unsigned c7 = tab[6]; | |
| 65 | |||
| 66 | 480000 | const int right = in[5] | in[6] | in[7]; | |
| 67 | 480000 | const int left = in[1] | in[2] | in[3]; | |
| 68 |
2/2✓ Branch 0 taken 264353 times.
✓ Branch 1 taken 215647 times.
|
480000 | if (!(right | in[4])) { |
| 69 | 264353 | const int k = c4 * in[0] + rnd; | |
| 70 |
2/2✓ Branch 0 taken 24730 times.
✓ Branch 1 taken 239623 times.
|
264353 | if (left) { |
| 71 | 24730 | const unsigned a0 = k + c2 * in[2]; | |
| 72 | 24730 | const unsigned a1 = k + c6 * in[2]; | |
| 73 | 24730 | const unsigned a2 = k - c6 * in[2]; | |
| 74 | 24730 | const unsigned a3 = k - c2 * in[2]; | |
| 75 | |||
| 76 | 24730 | const int b0 = c1 * in[1] + c3 * in[3]; | |
| 77 | 24730 | const int b1 = c3 * in[1] - c7 * in[3]; | |
| 78 | 24730 | const int b2 = c5 * in[1] - c1 * in[3]; | |
| 79 | 24730 | const int b3 = c7 * in[1] - c5 * in[3]; | |
| 80 | |||
| 81 | 24730 | in[0] = (int)(a0 + b0) >> ROW_SHIFT; | |
| 82 | 24730 | in[1] = (int)(a1 + b1) >> ROW_SHIFT; | |
| 83 | 24730 | in[2] = (int)(a2 + b2) >> ROW_SHIFT; | |
| 84 | 24730 | in[3] = (int)(a3 + b3) >> ROW_SHIFT; | |
| 85 | 24730 | in[4] = (int)(a3 - b3) >> ROW_SHIFT; | |
| 86 | 24730 | in[5] = (int)(a2 - b2) >> ROW_SHIFT; | |
| 87 | 24730 | in[6] = (int)(a1 - b1) >> ROW_SHIFT; | |
| 88 | 24730 | in[7] = (int)(a0 - b0) >> ROW_SHIFT; | |
| 89 | } else { | ||
| 90 | 239623 | const int a0 = k >> ROW_SHIFT; | |
| 91 |
2/2✓ Branch 0 taken 97939 times.
✓ Branch 1 taken 141684 times.
|
239623 | if (a0) { |
| 92 | 97939 | in[0] = | |
| 93 | 97939 | in[1] = | |
| 94 | 97939 | in[2] = | |
| 95 | 97939 | in[3] = | |
| 96 | 97939 | in[4] = | |
| 97 | 97939 | in[5] = | |
| 98 | 97939 | in[6] = | |
| 99 | 97939 | in[7] = a0; | |
| 100 | } else | ||
| 101 | 141684 | return 0; | |
| 102 | } | ||
| 103 |
2/2✓ Branch 0 taken 7429 times.
✓ Branch 1 taken 208218 times.
|
215647 | } else if (!(left | right)) { |
| 104 | 7429 | const int a0 = (int)(rnd + c4 * (in[0] + in[4])) >> ROW_SHIFT; | |
| 105 | 7429 | const int a1 = (int)(rnd + c4 * (in[0] - in[4])) >> ROW_SHIFT; | |
| 106 | |||
| 107 | 7429 | in[0] = a0; | |
| 108 | 7429 | in[3] = a0; | |
| 109 | 7429 | in[4] = a0; | |
| 110 | 7429 | in[7] = a0; | |
| 111 | 7429 | in[1] = a1; | |
| 112 | 7429 | in[2] = a1; | |
| 113 | 7429 | in[5] = a1; | |
| 114 | 7429 | in[6] = a1; | |
| 115 | } else { | ||
| 116 | 208218 | const unsigned int k = c4 * in[0] + rnd; | |
| 117 | 208218 | const unsigned int a0 = k + c2 * in[2] + c4 * in[4] + c6 * in[6]; | |
| 118 | 208218 | const unsigned int a1 = k + c6 * in[2] - c4 * in[4] - c2 * in[6]; | |
| 119 | 208218 | const unsigned int a2 = k - c6 * in[2] - c4 * in[4] + c2 * in[6]; | |
| 120 | 208218 | const unsigned int a3 = k - c2 * in[2] + c4 * in[4] - c6 * in[6]; | |
| 121 | |||
| 122 | 208218 | const unsigned int b0 = c1 * in[1] + c3 * in[3] + c5 * in[5] + c7 * in[7]; | |
| 123 | 208218 | const unsigned int b1 = c3 * in[1] - c7 * in[3] - c1 * in[5] - c5 * in[7]; | |
| 124 | 208218 | const unsigned int b2 = c5 * in[1] - c1 * in[3] + c7 * in[5] + c3 * in[7]; | |
| 125 | 208218 | const unsigned int b3 = c7 * in[1] - c5 * in[3] + c3 * in[5] - c1 * in[7]; | |
| 126 | |||
| 127 | 208218 | in[0] = (int)(a0 + b0) >> ROW_SHIFT; | |
| 128 | 208218 | in[1] = (int)(a1 + b1) >> ROW_SHIFT; | |
| 129 | 208218 | in[2] = (int)(a2 + b2) >> ROW_SHIFT; | |
| 130 | 208218 | in[3] = (int)(a3 + b3) >> ROW_SHIFT; | |
| 131 | 208218 | in[4] = (int)(a3 - b3) >> ROW_SHIFT; | |
| 132 | 208218 | in[5] = (int)(a2 - b2) >> ROW_SHIFT; | |
| 133 | 208218 | in[6] = (int)(a1 - b1) >> ROW_SHIFT; | |
| 134 | 208218 | in[7] = (int)(a0 - b0) >> ROW_SHIFT; | |
| 135 | } | ||
| 136 | 338316 | return 1; | |
| 137 | } | ||
| 138 | |||
| 139 | #define TAN1 0x32EC | ||
| 140 | #define TAN2 0x6A0A | ||
| 141 | #define TAN3 0xAB0E | ||
| 142 | #define SQRT2 0x5A82 | ||
| 143 | |||
| 144 | #define MULT(c, x, n) ((unsigned)((int)((c) * (unsigned)(x)) >> (n))) | ||
| 145 | // 12b version => #define MULT(c,x, n) ((((c) >> 3) * (x)) >> ((n) - 3)) | ||
| 146 | // 12b zero-testing version: | ||
| 147 | |||
| 148 | #define BUTTERFLY(a, b, tmp) \ | ||
| 149 | (tmp) = (a) + (b); \ | ||
| 150 | (b) = (a) - (b); \ | ||
| 151 | (a) = (tmp) | ||
| 152 | |||
| 153 | #define LOAD_BUTTERFLY(m1, m2, a, b, tmp, s) \ | ||
| 154 | (m1) = (s)[(a)] + (s)[(b)]; \ | ||
| 155 | (m2) = (s)[(a)] - (s)[(b)] | ||
| 156 | |||
| 157 | 384504 | static void idct_col_8(short *const in) | |
| 158 | { | ||
| 159 | int mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7, spill; | ||
| 160 | |||
| 161 | // odd | ||
| 162 | |||
| 163 | 384504 | mm4 = (int) in[7 * 8]; | |
| 164 | 384504 | mm5 = (int) in[5 * 8]; | |
| 165 | 384504 | mm6 = (int) in[3 * 8]; | |
| 166 | 384504 | mm7 = (int) in[1 * 8]; | |
| 167 | |||
| 168 | 384504 | mm0 = MULT(TAN1, mm4, 16) + mm7; | |
| 169 | 384504 | mm1 = MULT(TAN1, mm7, 16) - mm4; | |
| 170 | 384504 | mm2 = MULT(TAN3, mm5, 16) + mm6; | |
| 171 | 384504 | mm3 = MULT(TAN3, mm6, 16) - mm5; | |
| 172 | |||
| 173 | 384504 | mm7 = mm0 + mm2; | |
| 174 | 384504 | mm4 = mm1 - mm3; | |
| 175 | 384504 | mm0 = mm0 - mm2; | |
| 176 | 384504 | mm1 = mm1 + mm3; | |
| 177 | 384504 | mm6 = mm0 + mm1; | |
| 178 | 384504 | mm5 = mm0 - mm1; | |
| 179 | 384504 | mm5 = 2 * MULT(SQRT2, mm5, 16); // 2*sqrt2 | |
| 180 | 384504 | mm6 = 2 * MULT(SQRT2, mm6, 16); // Watch out: precision loss but done to match | |
| 181 | // the pmulhw used in MMX/MMXEXT/SSE2 versions | ||
| 182 | |||
| 183 | // even | ||
| 184 | |||
| 185 | 384504 | mm1 = (int) in[2 * 8]; | |
| 186 | 384504 | mm2 = (int) in[6 * 8]; | |
| 187 | 384504 | mm3 = MULT(TAN2, mm2, 16) + mm1; | |
| 188 | 384504 | mm2 = MULT(TAN2, mm1, 16) - mm2; | |
| 189 | |||
| 190 | 384504 | LOAD_BUTTERFLY(mm0, mm1, 0 * 8, 4 * 8, spill, in); | |
| 191 | |||
| 192 | 384504 | BUTTERFLY(mm0, mm3, spill); | |
| 193 | 384504 | BUTTERFLY(mm0, mm7, spill); | |
| 194 | 384504 | in[8 * 0] = (int16_t) (mm0 >> COL_SHIFT); | |
| 195 | 384504 | in[8 * 7] = (int16_t) (mm7 >> COL_SHIFT); | |
| 196 | 384504 | BUTTERFLY(mm3, mm4, mm0); | |
| 197 | 384504 | in[8 * 3] = (int16_t) (mm3 >> COL_SHIFT); | |
| 198 | 384504 | in[8 * 4] = (int16_t) (mm4 >> COL_SHIFT); | |
| 199 | |||
| 200 | 384504 | BUTTERFLY(mm1, mm2, mm0); | |
| 201 | 384504 | BUTTERFLY(mm1, mm6, mm0); | |
| 202 | 384504 | in[8 * 1] = (int16_t) (mm1 >> COL_SHIFT); | |
| 203 | 384504 | in[8 * 6] = (int16_t) (mm6 >> COL_SHIFT); | |
| 204 | 384504 | BUTTERFLY(mm2, mm5, mm0); | |
| 205 | 384504 | in[8 * 2] = (int16_t) (mm2 >> COL_SHIFT); | |
| 206 | 384504 | in[8 * 5] = (int16_t) (mm5 >> COL_SHIFT); | |
| 207 | 384504 | } | |
| 208 | |||
| 209 | 5896 | static void idct_col_4(short *const in) | |
| 210 | { | ||
| 211 | int mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7, spill; | ||
| 212 | |||
| 213 | // odd | ||
| 214 | |||
| 215 | 5896 | mm0 = (int) in[1 * 8]; | |
| 216 | 5896 | mm2 = (int) in[3 * 8]; | |
| 217 | |||
| 218 | 5896 | mm1 = MULT(TAN1, mm0, 16); | |
| 219 | 5896 | mm3 = MULT(TAN3, mm2, 16); | |
| 220 | |||
| 221 | 5896 | mm7 = mm0 + mm2; | |
| 222 | 5896 | mm4 = mm1 - mm3; | |
| 223 | 5896 | mm0 = mm0 - mm2; | |
| 224 | 5896 | mm1 = mm1 + mm3; | |
| 225 | 5896 | mm6 = mm0 + mm1; | |
| 226 | 5896 | mm5 = mm0 - mm1; | |
| 227 | 5896 | mm6 = 2 * MULT(SQRT2, mm6, 16); // 2*sqrt2 | |
| 228 | 5896 | mm5 = 2 * MULT(SQRT2, mm5, 16); | |
| 229 | |||
| 230 | // even | ||
| 231 | |||
| 232 | 5896 | mm0 = mm1 = (int) in[0 * 8]; | |
| 233 | 5896 | mm3 = (int) in[2 * 8]; | |
| 234 | 5896 | mm2 = MULT(TAN2, mm3, 16); | |
| 235 | |||
| 236 | 5896 | BUTTERFLY(mm0, mm3, spill); | |
| 237 | 5896 | BUTTERFLY(mm0, mm7, spill); | |
| 238 | 5896 | in[8 * 0] = (int16_t) (mm0 >> COL_SHIFT); | |
| 239 | 5896 | in[8 * 7] = (int16_t) (mm7 >> COL_SHIFT); | |
| 240 | 5896 | BUTTERFLY(mm3, mm4, mm0); | |
| 241 | 5896 | in[8 * 3] = (int16_t) (mm3 >> COL_SHIFT); | |
| 242 | 5896 | in[8 * 4] = (int16_t) (mm4 >> COL_SHIFT); | |
| 243 | |||
| 244 | 5896 | BUTTERFLY(mm1, mm2, mm0); | |
| 245 | 5896 | BUTTERFLY(mm1, mm6, mm0); | |
| 246 | 5896 | in[8 * 1] = (int16_t) (mm1 >> COL_SHIFT); | |
| 247 | 5896 | in[8 * 6] = (int16_t) (mm6 >> COL_SHIFT); | |
| 248 | 5896 | BUTTERFLY(mm2, mm5, mm0); | |
| 249 | 5896 | in[8 * 2] = (int16_t) (mm2 >> COL_SHIFT); | |
| 250 | 5896 | in[8 * 5] = (int16_t) (mm5 >> COL_SHIFT); | |
| 251 | 5896 | } | |
| 252 | |||
| 253 | 89600 | static void idct_col_3(short *const in) | |
| 254 | { | ||
| 255 | int mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7, spill; | ||
| 256 | |||
| 257 | // odd | ||
| 258 | |||
| 259 | 89600 | mm7 = (int) in[1 * 8]; | |
| 260 | 89600 | mm4 = MULT(TAN1, mm7, 16); | |
| 261 | |||
| 262 | 89600 | mm6 = mm7 + mm4; | |
| 263 | 89600 | mm5 = mm7 - mm4; | |
| 264 | 89600 | mm6 = 2 * MULT(SQRT2, mm6, 16); // 2*sqrt2 | |
| 265 | 89600 | mm5 = 2 * MULT(SQRT2, mm5, 16); | |
| 266 | |||
| 267 | // even | ||
| 268 | |||
| 269 | 89600 | mm0 = mm1 = (int) in[0 * 8]; | |
| 270 | 89600 | mm3 = (int) in[2 * 8]; | |
| 271 | 89600 | mm2 = MULT(TAN2, mm3, 16); | |
| 272 | |||
| 273 | 89600 | BUTTERFLY(mm0, mm3, spill); | |
| 274 | 89600 | BUTTERFLY(mm0, mm7, spill); | |
| 275 | 89600 | in[8 * 0] = (int16_t) (mm0 >> COL_SHIFT); | |
| 276 | 89600 | in[8 * 7] = (int16_t) (mm7 >> COL_SHIFT); | |
| 277 | 89600 | BUTTERFLY(mm3, mm4, mm0); | |
| 278 | 89600 | in[8 * 3] = (int16_t) (mm3 >> COL_SHIFT); | |
| 279 | 89600 | in[8 * 4] = (int16_t) (mm4 >> COL_SHIFT); | |
| 280 | |||
| 281 | 89600 | BUTTERFLY(mm1, mm2, mm0); | |
| 282 | 89600 | BUTTERFLY(mm1, mm6, mm0); | |
| 283 | 89600 | in[8 * 1] = (int16_t) (mm1 >> COL_SHIFT); | |
| 284 | 89600 | in[8 * 6] = (int16_t) (mm6 >> COL_SHIFT); | |
| 285 | 89600 | BUTTERFLY(mm2, mm5, mm0); | |
| 286 | 89600 | in[8 * 2] = (int16_t) (mm2 >> COL_SHIFT); | |
| 287 | 89600 | in[8 * 5] = (int16_t) (mm5 >> COL_SHIFT); | |
| 288 | 89600 | } | |
| 289 | |||
| 290 | 60000 | void ff_xvid_idct(int16_t *const in) | |
| 291 | { | ||
| 292 | 60000 | int i, rows = 0x07; | |
| 293 | |||
| 294 | 60000 | idct_row(in + 0 * 8, TAB04, RND0); | |
| 295 | 60000 | idct_row(in + 1 * 8, TAB17, RND1); | |
| 296 | 60000 | idct_row(in + 2 * 8, TAB26, RND2); | |
| 297 |
2/2✓ Branch 1 taken 29558 times.
✓ Branch 2 taken 30442 times.
|
60000 | if (idct_row(in + 3 * 8, TAB35, RND3)) |
| 298 | 29558 | rows |= 0x08; | |
| 299 |
2/2✓ Branch 1 taken 29661 times.
✓ Branch 2 taken 30339 times.
|
60000 | if (idct_row(in + 4 * 8, TAB04, RND4)) |
| 300 | 29661 | rows |= 0x10; | |
| 301 |
2/2✓ Branch 1 taken 29720 times.
✓ Branch 2 taken 30280 times.
|
60000 | if (idct_row(in + 5 * 8, TAB35, RND5)) |
| 302 | 29720 | rows |= 0x20; | |
| 303 |
2/2✓ Branch 1 taken 29731 times.
✓ Branch 2 taken 30269 times.
|
60000 | if (idct_row(in + 6 * 8, TAB26, RND6)) |
| 304 | 29731 | rows |= 0x40; | |
| 305 |
2/2✓ Branch 1 taken 39653 times.
✓ Branch 2 taken 20347 times.
|
60000 | if (idct_row(in + 7 * 8, TAB17, RND7)) |
| 306 | 39653 | rows |= 0x80; | |
| 307 | |||
| 308 |
2/2✓ Branch 0 taken 48063 times.
✓ Branch 1 taken 11937 times.
|
60000 | if (rows & 0xF0) { |
| 309 |
2/2✓ Branch 0 taken 384504 times.
✓ Branch 1 taken 48063 times.
|
432567 | for (i = 0; i < 8; i++) |
| 310 | 384504 | idct_col_8(in + i); | |
| 311 |
2/2✓ Branch 0 taken 737 times.
✓ Branch 1 taken 11200 times.
|
11937 | } else if (rows & 0x08) { |
| 312 |
2/2✓ Branch 0 taken 5896 times.
✓ Branch 1 taken 737 times.
|
6633 | for (i = 0; i < 8; i++) |
| 313 | 5896 | idct_col_4(in + i); | |
| 314 | } else { | ||
| 315 |
2/2✓ Branch 0 taken 89600 times.
✓ Branch 1 taken 11200 times.
|
100800 | for (i = 0; i < 8; i++) |
| 316 | 89600 | idct_col_3(in + i); | |
| 317 | } | ||
| 318 | 60000 | } | |
| 319 | |||
| 320 | ✗ | static void xvid_idct_put(uint8_t *dest, ptrdiff_t line_size, int16_t *block) | |
| 321 | { | ||
| 322 | ✗ | ff_xvid_idct(block); | |
| 323 | ✗ | ff_put_pixels_clamped_c(block, dest, line_size); | |
| 324 | ✗ | } | |
| 325 | |||
| 326 | ✗ | static void xvid_idct_add(uint8_t *dest, ptrdiff_t line_size, int16_t *block) | |
| 327 | { | ||
| 328 | ✗ | ff_xvid_idct(block); | |
| 329 | ✗ | ff_add_pixels_clamped_c(block, dest, line_size); | |
| 330 | ✗ | } | |
| 331 | |||
| 332 | 4 | av_cold void ff_xvid_idct_init(IDCTDSPContext *c) | |
| 333 | { | ||
| 334 | 4 | c->idct_put = xvid_idct_put; | |
| 335 | 4 | c->idct_add = xvid_idct_add; | |
| 336 | 4 | c->idct = ff_xvid_idct; | |
| 337 | 4 | c->perm_type = FF_IDCT_PERM_NONE; | |
| 338 | |||
| 339 | #if ARCH_X86 && HAVE_X86ASM | ||
| 340 | 4 | ff_xvid_idct_init_x86(c); | |
| 341 | #elif ARCH_MIPS | ||
| 342 | ff_xvid_idct_init_mips(c); | ||
| 343 | #endif | ||
| 344 | 4 | } | |
| 345 |