git.sesse.net Git - ffmpeg/blob - libavcodec/me_cmp.c

   1 /*
   2  * DSP utils
   3  * Copyright (c) 2000, 2001 Fabrice Bellard
   4  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
   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 #include "libavutil/attributes.h"
  24 #include "libavutil/internal.h"
  25 #include "avcodec.h"
  26 #include "copy_block.h"
  27 #include "simple_idct.h"
  28 #include "me_cmp.h"
  29 #include "mpegvideo.h"
  30 #include "config.h"
  31
  32 uint32_t ff_square_tab[512] = { 0, };
  33
  34 static int sse4_c(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  35                   ptrdiff_t stride, int h)
  36 {
  37     int s = 0, i;
  38     uint32_t *sq = ff_square_tab + 256;
  39
  40     for (i = 0; i < h; i++) {
  41         s    += sq[pix1[0] - pix2[0]];
  42         s    += sq[pix1[1] - pix2[1]];
  43         s    += sq[pix1[2] - pix2[2]];
  44         s    += sq[pix1[3] - pix2[3]];
  45         pix1 += stride;
  46         pix2 += stride;
  47     }
  48     return s;
  49 }
  50
  51 static int sse8_c(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  52                   ptrdiff_t stride, int h)
  53 {
  54     int s = 0, i;
  55     uint32_t *sq = ff_square_tab + 256;
  56
  57     for (i = 0; i < h; i++) {
  58         s    += sq[pix1[0] - pix2[0]];
  59         s    += sq[pix1[1] - pix2[1]];
  60         s    += sq[pix1[2] - pix2[2]];
  61         s    += sq[pix1[3] - pix2[3]];
  62         s    += sq[pix1[4] - pix2[4]];
  63         s    += sq[pix1[5] - pix2[5]];
  64         s    += sq[pix1[6] - pix2[6]];
  65         s    += sq[pix1[7] - pix2[7]];
  66         pix1 += stride;
  67         pix2 += stride;
  68     }
  69     return s;
  70 }
  71
  72 static int sse16_c(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
  73                    ptrdiff_t stride, int h)
  74 {
  75     int s = 0, i;
  76     uint32_t *sq = ff_square_tab + 256;
  77
  78     for (i = 0; i < h; i++) {
  79         s += sq[pix1[0]  - pix2[0]];
  80         s += sq[pix1[1]  - pix2[1]];
  81         s += sq[pix1[2]  - pix2[2]];
  82         s += sq[pix1[3]  - pix2[3]];
  83         s += sq[pix1[4]  - pix2[4]];
  84         s += sq[pix1[5]  - pix2[5]];
  85         s += sq[pix1[6]  - pix2[6]];
  86         s += sq[pix1[7]  - pix2[7]];
  87         s += sq[pix1[8]  - pix2[8]];
  88         s += sq[pix1[9]  - pix2[9]];
  89         s += sq[pix1[10] - pix2[10]];
  90         s += sq[pix1[11] - pix2[11]];
  91         s += sq[pix1[12] - pix2[12]];
  92         s += sq[pix1[13] - pix2[13]];
  93         s += sq[pix1[14] - pix2[14]];
  94         s += sq[pix1[15] - pix2[15]];
  95
  96         pix1 += stride;
  97         pix2 += stride;
  98     }
  99     return s;
 100 }
 101
 102 static int sum_abs_dctelem_c(int16_t *block)
 103 {
 104     int sum = 0, i;
 105
 106     for (i = 0; i < 64; i++)
 107         sum += FFABS(block[i]);
 108     return sum;
 109 }
 110
 111 #define avg2(a, b) (((a) + (b) + 1) >> 1)
 112 #define avg4(a, b, c, d) (((a) + (b) + (c) + (d) + 2) >> 2)
 113
 114 static inline int pix_abs16_c(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
 115                               ptrdiff_t stride, int h)
 116 {
 117     int s = 0, i;
 118
 119     for (i = 0; i < h; i++) {
 120         s    += abs(pix1[0]  - pix2[0]);
 121         s    += abs(pix1[1]  - pix2[1]);
 122         s    += abs(pix1[2]  - pix2[2]);
 123         s    += abs(pix1[3]  - pix2[3]);
 124         s    += abs(pix1[4]  - pix2[4]);
 125         s    += abs(pix1[5]  - pix2[5]);
 126         s    += abs(pix1[6]  - pix2[6]);
 127         s    += abs(pix1[7]  - pix2[7]);
 128         s    += abs(pix1[8]  - pix2[8]);
 129         s    += abs(pix1[9]  - pix2[9]);
 130         s    += abs(pix1[10] - pix2[10]);
 131         s    += abs(pix1[11] - pix2[11]);
 132         s    += abs(pix1[12] - pix2[12]);
 133         s    += abs(pix1[13] - pix2[13]);
 134         s    += abs(pix1[14] - pix2[14]);
 135         s    += abs(pix1[15] - pix2[15]);
 136         pix1 += stride;
 137         pix2 += stride;
 138     }
 139     return s;
 140 }
 141
 142 static int pix_abs16_x2_c(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
 143                           ptrdiff_t stride, int h)
 144 {
 145     int s = 0, i;
 146
 147     for (i = 0; i < h; i++) {
 148         s    += abs(pix1[0]  - avg2(pix2[0],  pix2[1]));
 149         s    += abs(pix1[1]  - avg2(pix2[1],  pix2[2]));
 150         s    += abs(pix1[2]  - avg2(pix2[2],  pix2[3]));
 151         s    += abs(pix1[3]  - avg2(pix2[3],  pix2[4]));
 152         s    += abs(pix1[4]  - avg2(pix2[4],  pix2[5]));
 153         s    += abs(pix1[5]  - avg2(pix2[5],  pix2[6]));
 154         s    += abs(pix1[6]  - avg2(pix2[6],  pix2[7]));
 155         s    += abs(pix1[7]  - avg2(pix2[7],  pix2[8]));
 156         s    += abs(pix1[8]  - avg2(pix2[8],  pix2[9]));
 157         s    += abs(pix1[9]  - avg2(pix2[9],  pix2[10]));
 158         s    += abs(pix1[10] - avg2(pix2[10], pix2[11]));
 159         s    += abs(pix1[11] - avg2(pix2[11], pix2[12]));
 160         s    += abs(pix1[12] - avg2(pix2[12], pix2[13]));
 161         s    += abs(pix1[13] - avg2(pix2[13], pix2[14]));
 162         s    += abs(pix1[14] - avg2(pix2[14], pix2[15]));
 163         s    += abs(pix1[15] - avg2(pix2[15], pix2[16]));
 164         pix1 += stride;
 165         pix2 += stride;
 166     }
 167     return s;
 168 }
 169
 170 static int pix_abs16_y2_c(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
 171                           ptrdiff_t stride, int h)
 172 {
 173     int s = 0, i;
 174     uint8_t *pix3 = pix2 + stride;
 175
 176     for (i = 0; i < h; i++) {
 177         s    += abs(pix1[0]  - avg2(pix2[0],  pix3[0]));
 178         s    += abs(pix1[1]  - avg2(pix2[1],  pix3[1]));
 179         s    += abs(pix1[2]  - avg2(pix2[2],  pix3[2]));
 180         s    += abs(pix1[3]  - avg2(pix2[3],  pix3[3]));
 181         s    += abs(pix1[4]  - avg2(pix2[4],  pix3[4]));
 182         s    += abs(pix1[5]  - avg2(pix2[5],  pix3[5]));
 183         s    += abs(pix1[6]  - avg2(pix2[6],  pix3[6]));
 184         s    += abs(pix1[7]  - avg2(pix2[7],  pix3[7]));
 185         s    += abs(pix1[8]  - avg2(pix2[8],  pix3[8]));
 186         s    += abs(pix1[9]  - avg2(pix2[9],  pix3[9]));
 187         s    += abs(pix1[10] - avg2(pix2[10], pix3[10]));
 188         s    += abs(pix1[11] - avg2(pix2[11], pix3[11]));
 189         s    += abs(pix1[12] - avg2(pix2[12], pix3[12]));
 190         s    += abs(pix1[13] - avg2(pix2[13], pix3[13]));
 191         s    += abs(pix1[14] - avg2(pix2[14], pix3[14]));
 192         s    += abs(pix1[15] - avg2(pix2[15], pix3[15]));
 193         pix1 += stride;
 194         pix2 += stride;
 195         pix3 += stride;
 196     }
 197     return s;
 198 }
 199
 200 static int pix_abs16_xy2_c(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
 201                            ptrdiff_t stride, int h)
 202 {
 203     int s = 0, i;
 204     uint8_t *pix3 = pix2 + stride;
 205
 206     for (i = 0; i < h; i++) {
 207         s    += abs(pix1[0]  - avg4(pix2[0],  pix2[1],  pix3[0],  pix3[1]));
 208         s    += abs(pix1[1]  - avg4(pix2[1],  pix2[2],  pix3[1],  pix3[2]));
 209         s    += abs(pix1[2]  - avg4(pix2[2],  pix2[3],  pix3[2],  pix3[3]));
 210         s    += abs(pix1[3]  - avg4(pix2[3],  pix2[4],  pix3[3],  pix3[4]));
 211         s    += abs(pix1[4]  - avg4(pix2[4],  pix2[5],  pix3[4],  pix3[5]));
 212         s    += abs(pix1[5]  - avg4(pix2[5],  pix2[6],  pix3[5],  pix3[6]));
 213         s    += abs(pix1[6]  - avg4(pix2[6],  pix2[7],  pix3[6],  pix3[7]));
 214         s    += abs(pix1[7]  - avg4(pix2[7],  pix2[8],  pix3[7],  pix3[8]));
 215         s    += abs(pix1[8]  - avg4(pix2[8],  pix2[9],  pix3[8],  pix3[9]));
 216         s    += abs(pix1[9]  - avg4(pix2[9],  pix2[10], pix3[9],  pix3[10]));
 217         s    += abs(pix1[10] - avg4(pix2[10], pix2[11], pix3[10], pix3[11]));
 218         s    += abs(pix1[11] - avg4(pix2[11], pix2[12], pix3[11], pix3[12]));
 219         s    += abs(pix1[12] - avg4(pix2[12], pix2[13], pix3[12], pix3[13]));
 220         s    += abs(pix1[13] - avg4(pix2[13], pix2[14], pix3[13], pix3[14]));
 221         s    += abs(pix1[14] - avg4(pix2[14], pix2[15], pix3[14], pix3[15]));
 222         s    += abs(pix1[15] - avg4(pix2[15], pix2[16], pix3[15], pix3[16]));
 223         pix1 += stride;
 224         pix2 += stride;
 225         pix3 += stride;
 226     }
 227     return s;
 228 }
 229
 230 static inline int pix_abs8_c(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
 231                              ptrdiff_t stride, int h)
 232 {
 233     int s = 0, i;
 234
 235     for (i = 0; i < h; i++) {
 236         s    += abs(pix1[0] - pix2[0]);
 237         s    += abs(pix1[1] - pix2[1]);
 238         s    += abs(pix1[2] - pix2[2]);
 239         s    += abs(pix1[3] - pix2[3]);
 240         s    += abs(pix1[4] - pix2[4]);
 241         s    += abs(pix1[5] - pix2[5]);
 242         s    += abs(pix1[6] - pix2[6]);
 243         s    += abs(pix1[7] - pix2[7]);
 244         pix1 += stride;
 245         pix2 += stride;
 246     }
 247     return s;
 248 }
 249
 250 static int pix_abs8_x2_c(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
 251                          ptrdiff_t stride, int h)
 252 {
 253     int s = 0, i;
 254
 255     for (i = 0; i < h; i++) {
 256         s    += abs(pix1[0] - avg2(pix2[0], pix2[1]));
 257         s    += abs(pix1[1] - avg2(pix2[1], pix2[2]));
 258         s    += abs(pix1[2] - avg2(pix2[2], pix2[3]));
 259         s    += abs(pix1[3] - avg2(pix2[3], pix2[4]));
 260         s    += abs(pix1[4] - avg2(pix2[4], pix2[5]));
 261         s    += abs(pix1[5] - avg2(pix2[5], pix2[6]));
 262         s    += abs(pix1[6] - avg2(pix2[6], pix2[7]));
 263         s    += abs(pix1[7] - avg2(pix2[7], pix2[8]));
 264         pix1 += stride;
 265         pix2 += stride;
 266     }
 267     return s;
 268 }
 269
 270 static int pix_abs8_y2_c(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
 271                          ptrdiff_t stride, int h)
 272 {
 273     int s = 0, i;
 274     uint8_t *pix3 = pix2 + stride;
 275
 276     for (i = 0; i < h; i++) {
 277         s    += abs(pix1[0] - avg2(pix2[0], pix3[0]));
 278         s    += abs(pix1[1] - avg2(pix2[1], pix3[1]));
 279         s    += abs(pix1[2] - avg2(pix2[2], pix3[2]));
 280         s    += abs(pix1[3] - avg2(pix2[3], pix3[3]));
 281         s    += abs(pix1[4] - avg2(pix2[4], pix3[4]));
 282         s    += abs(pix1[5] - avg2(pix2[5], pix3[5]));
 283         s    += abs(pix1[6] - avg2(pix2[6], pix3[6]));
 284         s    += abs(pix1[7] - avg2(pix2[7], pix3[7]));
 285         pix1 += stride;
 286         pix2 += stride;
 287         pix3 += stride;
 288     }
 289     return s;
 290 }
 291
 292 static int pix_abs8_xy2_c(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
 293                           ptrdiff_t stride, int h)
 294 {
 295     int s = 0, i;
 296     uint8_t *pix3 = pix2 + stride;
 297
 298     for (i = 0; i < h; i++) {
 299         s    += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1]));
 300         s    += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2]));
 301         s    += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3]));
 302         s    += abs(pix1[3] - avg4(pix2[3], pix2[4], pix3[3], pix3[4]));
 303         s    += abs(pix1[4] - avg4(pix2[4], pix2[5], pix3[4], pix3[5]));
 304         s    += abs(pix1[5] - avg4(pix2[5], pix2[6], pix3[5], pix3[6]));
 305         s    += abs(pix1[6] - avg4(pix2[6], pix2[7], pix3[6], pix3[7]));
 306         s    += abs(pix1[7] - avg4(pix2[7], pix2[8], pix3[7], pix3[8]));
 307         pix1 += stride;
 308         pix2 += stride;
 309         pix3 += stride;
 310     }
 311     return s;
 312 }
 313
 314 static int nsse16_c(MpegEncContext *c, uint8_t *s1, uint8_t *s2,
 315                     ptrdiff_t stride, int h)
 316 {
 317     int score1 = 0, score2 = 0, x, y;
 318
 319     for (y = 0; y < h; y++) {
 320         for (x = 0; x < 16; x++)
 321             score1 += (s1[x] - s2[x]) * (s1[x] - s2[x]);
 322         if (y + 1 < h) {
 323             for (x = 0; x < 15; x++)
 324                 score2 += FFABS(s1[x]     - s1[x + stride] -
 325                                 s1[x + 1] + s1[x + stride + 1]) -
 326                           FFABS(s2[x]     - s2[x + stride] -
 327                                 s2[x + 1] + s2[x + stride + 1]);
 328         }
 329         s1 += stride;
 330         s2 += stride;
 331     }
 332
 333     if (c)
 334         return score1 + FFABS(score2) * c->avctx->nsse_weight;
 335     else
 336         return score1 + FFABS(score2) * 8;
 337 }
 338
 339 static int nsse8_c(MpegEncContext *c, uint8_t *s1, uint8_t *s2,
 340                    ptrdiff_t stride, int h)
 341 {
 342     int score1 = 0, score2 = 0, x, y;
 343
 344     for (y = 0; y < h; y++) {
 345         for (x = 0; x < 8; x++)
 346             score1 += (s1[x] - s2[x]) * (s1[x] - s2[x]);
 347         if (y + 1 < h) {
 348             for (x = 0; x < 7; x++)
 349                 score2 += FFABS(s1[x]     - s1[x + stride] -
 350                                 s1[x + 1] + s1[x + stride + 1]) -
 351                           FFABS(s2[x]     - s2[x + stride] -
 352                                 s2[x + 1] + s2[x + stride + 1]);
 353         }
 354         s1 += stride;
 355         s2 += stride;
 356     }
 357
 358     if (c)
 359         return score1 + FFABS(score2) * c->avctx->nsse_weight;
 360     else
 361         return score1 + FFABS(score2) * 8;
 362 }
 363
 364 static int zero_cmp(MpegEncContext *s, uint8_t *a, uint8_t *b,
 365                     ptrdiff_t stride, int h)
 366 {
 367     return 0;
 368 }
 369
 370 void ff_set_cmp(MECmpContext *c, me_cmp_func *cmp, int type)
 371 {
 372     int i;
 373
 374     memset(cmp, 0, sizeof(void *) * 6);
 375
 376     for (i = 0; i < 6; i++) {
 377         switch (type & 0xFF) {
 378         case FF_CMP_SAD:
 379             cmp[i] = c->sad[i];
 380             break;
 381         case FF_CMP_SATD:
 382             cmp[i] = c->hadamard8_diff[i];
 383             break;
 384         case FF_CMP_SSE:
 385             cmp[i] = c->sse[i];
 386             break;
 387         case FF_CMP_DCT:
 388             cmp[i] = c->dct_sad[i];
 389             break;
 390         case FF_CMP_DCT264:
 391             cmp[i] = c->dct264_sad[i];
 392             break;
 393         case FF_CMP_DCTMAX:
 394             cmp[i] = c->dct_max[i];
 395             break;
 396         case FF_CMP_PSNR:
 397             cmp[i] = c->quant_psnr[i];
 398             break;
 399         case FF_CMP_BIT:
 400             cmp[i] = c->bit[i];
 401             break;
 402         case FF_CMP_RD:
 403             cmp[i] = c->rd[i];
 404             break;
 405         case FF_CMP_VSAD:
 406             cmp[i] = c->vsad[i];
 407             break;
 408         case FF_CMP_VSSE:
 409             cmp[i] = c->vsse[i];
 410             break;
 411         case FF_CMP_ZERO:
 412             cmp[i] = zero_cmp;
 413             break;
 414         case FF_CMP_NSSE:
 415             cmp[i] = c->nsse[i];
 416             break;
 417 #if CONFIG_DWT
 418         case FF_CMP_W53:
 419             cmp[i]= c->w53[i];
 420             break;
 421         case FF_CMP_W97:
 422             cmp[i]= c->w97[i];
 423             break;
 424 #endif
 425         default:
 426             av_log(NULL, AV_LOG_ERROR,
 427                    "internal error in cmp function selection\n");
 428         }
 429     }
 430 }
 431
 432 #define BUTTERFLY2(o1, o2, i1, i2)              \
 433     o1 = (i1) + (i2);                           \
 434     o2 = (i1) - (i2);
 435
 436 #define BUTTERFLY1(x, y)                        \
 437     {                                           \
 438         int a, b;                               \
 439         a = x;                                  \
 440         b = y;                                  \
 441         x = a + b;                              \
 442         y = a - b;                              \
 443     }
 444
 445 #define BUTTERFLYA(x, y) (FFABS((x) + (y)) + FFABS((x) - (y)))
 446
 447 static int hadamard8_diff8x8_c(MpegEncContext *s, uint8_t *dst,
 448                                uint8_t *src, ptrdiff_t stride, int h)
 449 {
 450     int i, temp[64], sum = 0;
 451
 452     av_assert2(h == 8);
 453
 454     for (i = 0; i < 8; i++) {
 455         // FIXME: try pointer walks
 456         BUTTERFLY2(temp[8 * i + 0], temp[8 * i + 1],
 457                    src[stride * i + 0] - dst[stride * i + 0],
 458                    src[stride * i + 1] - dst[stride * i + 1]);
 459         BUTTERFLY2(temp[8 * i + 2], temp[8 * i + 3],
 460                    src[stride * i + 2] - dst[stride * i + 2],
 461                    src[stride * i + 3] - dst[stride * i + 3]);
 462         BUTTERFLY2(temp[8 * i + 4], temp[8 * i + 5],
 463                    src[stride * i + 4] - dst[stride * i + 4],
 464                    src[stride * i + 5] - dst[stride * i + 5]);
 465         BUTTERFLY2(temp[8 * i + 6], temp[8 * i + 7],
 466                    src[stride * i + 6] - dst[stride * i + 6],
 467                    src[stride * i + 7] - dst[stride * i + 7]);
 468
 469         BUTTERFLY1(temp[8 * i + 0], temp[8 * i + 2]);
 470         BUTTERFLY1(temp[8 * i + 1], temp[8 * i + 3]);
 471         BUTTERFLY1(temp[8 * i + 4], temp[8 * i + 6]);
 472         BUTTERFLY1(temp[8 * i + 5], temp[8 * i + 7]);
 473
 474         BUTTERFLY1(temp[8 * i + 0], temp[8 * i + 4]);
 475         BUTTERFLY1(temp[8 * i + 1], temp[8 * i + 5]);
 476         BUTTERFLY1(temp[8 * i + 2], temp[8 * i + 6]);
 477         BUTTERFLY1(temp[8 * i + 3], temp[8 * i + 7]);
 478     }
 479
 480     for (i = 0; i < 8; i++) {
 481         BUTTERFLY1(temp[8 * 0 + i], temp[8 * 1 + i]);
 482         BUTTERFLY1(temp[8 * 2 + i], temp[8 * 3 + i]);
 483         BUTTERFLY1(temp[8 * 4 + i], temp[8 * 5 + i]);
 484         BUTTERFLY1(temp[8 * 6 + i], temp[8 * 7 + i]);
 485
 486         BUTTERFLY1(temp[8 * 0 + i], temp[8 * 2 + i]);
 487         BUTTERFLY1(temp[8 * 1 + i], temp[8 * 3 + i]);
 488         BUTTERFLY1(temp[8 * 4 + i], temp[8 * 6 + i]);
 489         BUTTERFLY1(temp[8 * 5 + i], temp[8 * 7 + i]);
 490
 491         sum += BUTTERFLYA(temp[8 * 0 + i], temp[8 * 4 + i]) +
 492                BUTTERFLYA(temp[8 * 1 + i], temp[8 * 5 + i]) +
 493                BUTTERFLYA(temp[8 * 2 + i], temp[8 * 6 + i]) +
 494                BUTTERFLYA(temp[8 * 3 + i], temp[8 * 7 + i]);
 495     }
 496     return sum;
 497 }
 498
 499 static int hadamard8_intra8x8_c(MpegEncContext *s, uint8_t *src,
 500                                 uint8_t *dummy, ptrdiff_t stride, int h)
 501 {
 502     int i, temp[64], sum = 0;
 503
 504     av_assert2(h == 8);
 505
 506     for (i = 0; i < 8; i++) {
 507         // FIXME: try pointer walks
 508         BUTTERFLY2(temp[8 * i + 0], temp[8 * i + 1],
 509                    src[stride * i + 0], src[stride * i + 1]);
 510         BUTTERFLY2(temp[8 * i + 2], temp[8 * i + 3],
 511                    src[stride * i + 2], src[stride * i + 3]);
 512         BUTTERFLY2(temp[8 * i + 4], temp[8 * i + 5],
 513                    src[stride * i + 4], src[stride * i + 5]);
 514         BUTTERFLY2(temp[8 * i + 6], temp[8 * i + 7],
 515                    src[stride * i + 6], src[stride * i + 7]);
 516
 517         BUTTERFLY1(temp[8 * i + 0], temp[8 * i + 2]);
 518         BUTTERFLY1(temp[8 * i + 1], temp[8 * i + 3]);
 519         BUTTERFLY1(temp[8 * i + 4], temp[8 * i + 6]);
 520         BUTTERFLY1(temp[8 * i + 5], temp[8 * i + 7]);
 521
 522         BUTTERFLY1(temp[8 * i + 0], temp[8 * i + 4]);
 523         BUTTERFLY1(temp[8 * i + 1], temp[8 * i + 5]);
 524         BUTTERFLY1(temp[8 * i + 2], temp[8 * i + 6]);
 525         BUTTERFLY1(temp[8 * i + 3], temp[8 * i + 7]);
 526     }
 527
 528     for (i = 0; i < 8; i++) {
 529         BUTTERFLY1(temp[8 * 0 + i], temp[8 * 1 + i]);
 530         BUTTERFLY1(temp[8 * 2 + i], temp[8 * 3 + i]);
 531         BUTTERFLY1(temp[8 * 4 + i], temp[8 * 5 + i]);
 532         BUTTERFLY1(temp[8 * 6 + i], temp[8 * 7 + i]);
 533
 534         BUTTERFLY1(temp[8 * 0 + i], temp[8 * 2 + i]);
 535         BUTTERFLY1(temp[8 * 1 + i], temp[8 * 3 + i]);
 536         BUTTERFLY1(temp[8 * 4 + i], temp[8 * 6 + i]);
 537         BUTTERFLY1(temp[8 * 5 + i], temp[8 * 7 + i]);
 538
 539         sum +=
 540             BUTTERFLYA(temp[8 * 0 + i], temp[8 * 4 + i])
 541             + BUTTERFLYA(temp[8 * 1 + i], temp[8 * 5 + i])
 542             + BUTTERFLYA(temp[8 * 2 + i], temp[8 * 6 + i])
 543             + BUTTERFLYA(temp[8 * 3 + i], temp[8 * 7 + i]);
 544     }
 545
 546     sum -= FFABS(temp[8 * 0] + temp[8 * 4]); // -mean
 547
 548     return sum;
 549 }
 550
 551 static int dct_sad8x8_c(MpegEncContext *s, uint8_t *src1,
 552                         uint8_t *src2, ptrdiff_t stride, int h)
 553 {
 554     LOCAL_ALIGNED_16(int16_t, temp, [64]);
 555
 556     av_assert2(h == 8);
 557
 558     s->pdsp.diff_pixels(temp, src1, src2, stride);
 559     s->fdsp.fdct(temp);
 560     return s->mecc.sum_abs_dctelem(temp);
 561 }
 562
 563 #if CONFIG_GPL
 564 #define DCT8_1D                                         \
 565     {                                                   \
 566         const int s07 = SRC(0) + SRC(7);                \
 567         const int s16 = SRC(1) + SRC(6);                \
 568         const int s25 = SRC(2) + SRC(5);                \
 569         const int s34 = SRC(3) + SRC(4);                \
 570         const int a0  = s07 + s34;                      \
 571         const int a1  = s16 + s25;                      \
 572         const int a2  = s07 - s34;                      \
 573         const int a3  = s16 - s25;                      \
 574         const int d07 = SRC(0) - SRC(7);                \
 575         const int d16 = SRC(1) - SRC(6);                \
 576         const int d25 = SRC(2) - SRC(5);                \
 577         const int d34 = SRC(3) - SRC(4);                \
 578         const int a4  = d16 + d25 + (d07 + (d07 >> 1)); \
 579         const int a5  = d07 - d34 - (d25 + (d25 >> 1)); \
 580         const int a6  = d07 + d34 - (d16 + (d16 >> 1)); \
 581         const int a7  = d16 - d25 + (d34 + (d34 >> 1)); \
 582         DST(0, a0 + a1);                                \
 583         DST(1, a4 + (a7 >> 2));                         \
 584         DST(2, a2 + (a3 >> 1));                         \
 585         DST(3, a5 + (a6 >> 2));                         \
 586         DST(4, a0 - a1);                                \
 587         DST(5, a6 - (a5 >> 2));                         \
 588         DST(6, (a2 >> 1) - a3);                         \
 589         DST(7, (a4 >> 2) - a7);                         \
 590     }
 591
 592 static int dct264_sad8x8_c(MpegEncContext *s, uint8_t *src1,
 593                            uint8_t *src2, ptrdiff_t stride, int h)
 594 {
 595     int16_t dct[8][8];
 596     int i, sum = 0;
 597
 598     s->pdsp.diff_pixels(dct[0], src1, src2, stride);
 599
 600 #define SRC(x) dct[i][x]
 601 #define DST(x, v) dct[i][x] = v
 602     for (i = 0; i < 8; i++)
 603         DCT8_1D
 604 #undef SRC
 605 #undef DST
 606
 607 #define SRC(x) dct[x][i]
 608 #define DST(x, v) sum += FFABS(v)
 609         for (i = 0; i < 8; i++)
 610             DCT8_1D
 611 #undef SRC
 612 #undef DST
 613             return sum;
 614 }
 615 #endif
 616
 617 static int dct_max8x8_c(MpegEncContext *s, uint8_t *src1,
 618                         uint8_t *src2, ptrdiff_t stride, int h)
 619 {
 620     LOCAL_ALIGNED_16(int16_t, temp, [64]);
 621     int sum = 0, i;
 622
 623     av_assert2(h == 8);
 624
 625     s->pdsp.diff_pixels(temp, src1, src2, stride);
 626     s->fdsp.fdct(temp);
 627
 628     for (i = 0; i < 64; i++)
 629         sum = FFMAX(sum, FFABS(temp[i]));
 630
 631     return sum;
 632 }
 633
 634 static int quant_psnr8x8_c(MpegEncContext *s, uint8_t *src1,
 635                            uint8_t *src2, ptrdiff_t stride, int h)
 636 {
 637     LOCAL_ALIGNED_16(int16_t, temp, [64 * 2]);
 638     int16_t *const bak = temp + 64;
 639     int sum = 0, i;
 640
 641     av_assert2(h == 8);
 642     s->mb_intra = 0;
 643
 644     s->pdsp.diff_pixels(temp, src1, src2, stride);
 645
 646     memcpy(bak, temp, 64 * sizeof(int16_t));
 647
 648     s->block_last_index[0 /* FIXME */] =
 649         s->fast_dct_quantize(s, temp, 0 /* FIXME */, s->qscale, &i);
 650     s->dct_unquantize_inter(s, temp, 0, s->qscale);
 651     ff_simple_idct_8(temp); // FIXME
 652
 653     for (i = 0; i < 64; i++)
 654         sum += (temp[i] - bak[i]) * (temp[i] - bak[i]);
 655
 656     return sum;
 657 }
 658
 659 static int rd8x8_c(MpegEncContext *s, uint8_t *src1, uint8_t *src2,
 660                    ptrdiff_t stride, int h)
 661 {
 662     const uint8_t *scantable = s->intra_scantable.permutated;
 663     LOCAL_ALIGNED_16(int16_t, temp, [64]);
 664     LOCAL_ALIGNED_16(uint8_t, lsrc1, [64]);
 665     LOCAL_ALIGNED_16(uint8_t, lsrc2, [64]);
 666     int i, last, run, bits, level, distortion, start_i;
 667     const int esc_length = s->ac_esc_length;
 668     uint8_t *length, *last_length;
 669
 670     av_assert2(h == 8);
 671
 672     copy_block8(lsrc1, src1, 8, stride, 8);
 673     copy_block8(lsrc2, src2, 8, stride, 8);
 674
 675     s->pdsp.diff_pixels(temp, lsrc1, lsrc2, 8);
 676
 677     s->block_last_index[0 /* FIXME */] =
 678     last                               =
 679         s->fast_dct_quantize(s, temp, 0 /* FIXME */, s->qscale, &i);
 680
 681     bits = 0;
 682
 683     if (s->mb_intra) {
 684         start_i     = 1;
 685         length      = s->intra_ac_vlc_length;
 686         last_length = s->intra_ac_vlc_last_length;
 687         bits       += s->luma_dc_vlc_length[temp[0] + 256]; // FIXME: chroma
 688     } else {
 689         start_i     = 0;
 690         length      = s->inter_ac_vlc_length;
 691         last_length = s->inter_ac_vlc_last_length;
 692     }
 693
 694     if (last >= start_i) {
 695         run = 0;
 696         for (i = start_i; i < last; i++) {
 697             int j = scantable[i];
 698             level = temp[j];
 699
 700             if (level) {
 701                 level += 64;
 702                 if ((level & (~127)) == 0)
 703                     bits += length[UNI_AC_ENC_INDEX(run, level)];
 704                 else
 705                     bits += esc_length;
 706                 run = 0;
 707             } else
 708                 run++;
 709         }
 710         i = scantable[last];
 711
 712         level = temp[i] + 64;
 713
 714         av_assert2(level - 64);
 715
 716         if ((level & (~127)) == 0) {
 717             bits += last_length[UNI_AC_ENC_INDEX(run, level)];
 718         } else
 719             bits += esc_length;
 720     }
 721
 722     if (last >= 0) {
 723         if (s->mb_intra)
 724             s->dct_unquantize_intra(s, temp, 0, s->qscale);
 725         else
 726             s->dct_unquantize_inter(s, temp, 0, s->qscale);
 727     }
 728
 729     s->idsp.idct_add(lsrc2, 8, temp);
 730
 731     distortion = s->mecc.sse[1](NULL, lsrc2, lsrc1, 8, 8);
 732
 733     return distortion + ((bits * s->qscale * s->qscale * 109 + 64) >> 7);
 734 }
 735
 736 static int bit8x8_c(MpegEncContext *s, uint8_t *src1, uint8_t *src2,
 737                     ptrdiff_t stride, int h)
 738 {
 739     const uint8_t *scantable = s->intra_scantable.permutated;
 740     LOCAL_ALIGNED_16(int16_t, temp, [64]);
 741     int i, last, run, bits, level, start_i;
 742     const int esc_length = s->ac_esc_length;
 743     uint8_t *length, *last_length;
 744
 745     av_assert2(h == 8);
 746
 747     s->pdsp.diff_pixels(temp, src1, src2, stride);
 748
 749     s->block_last_index[0 /* FIXME */] =
 750     last                               =
 751         s->fast_dct_quantize(s, temp, 0 /* FIXME */, s->qscale, &i);
 752
 753     bits = 0;
 754
 755     if (s->mb_intra) {
 756         start_i     = 1;
 757         length      = s->intra_ac_vlc_length;
 758         last_length = s->intra_ac_vlc_last_length;
 759         bits       += s->luma_dc_vlc_length[temp[0] + 256]; // FIXME: chroma
 760     } else {
 761         start_i     = 0;
 762         length      = s->inter_ac_vlc_length;
 763         last_length = s->inter_ac_vlc_last_length;
 764     }
 765
 766     if (last >= start_i) {
 767         run = 0;
 768         for (i = start_i; i < last; i++) {
 769             int j = scantable[i];
 770             level = temp[j];
 771
 772             if (level) {
 773                 level += 64;
 774                 if ((level & (~127)) == 0)
 775                     bits += length[UNI_AC_ENC_INDEX(run, level)];
 776                 else
 777                     bits += esc_length;
 778                 run = 0;
 779             } else
 780                 run++;
 781         }
 782         i = scantable[last];
 783
 784         level = temp[i] + 64;
 785
 786         av_assert2(level - 64);
 787
 788         if ((level & (~127)) == 0)
 789             bits += last_length[UNI_AC_ENC_INDEX(run, level)];
 790         else
 791             bits += esc_length;
 792     }
 793
 794     return bits;
 795 }
 796
 797 #define VSAD_INTRA(size)                                                \
 798 static int vsad_intra ## size ## _c(MpegEncContext *c,                  \
 799                                     uint8_t *s, uint8_t *dummy,         \
 800                                     ptrdiff_t stride, int h)            \
 801 {                                                                       \
 802     int score = 0, x, y;                                                \
 803                                                                         \
 804     for (y = 1; y < h; y++) {                                           \
 805         for (x = 0; x < size; x += 4) {                                 \
 806             score += FFABS(s[x]     - s[x + stride])     +              \
 807                      FFABS(s[x + 1] - s[x + stride + 1]) +              \
 808                      FFABS(s[x + 2] - s[x + 2 + stride]) +              \
 809                      FFABS(s[x + 3] - s[x + 3 + stride]);               \
 810         }                                                               \
 811         s += stride;                                                    \
 812     }                                                                   \
 813                                                                         \
 814     return score;                                                       \
 815 }
 816 VSAD_INTRA(8)
 817 VSAD_INTRA(16)
 818
 819 #define VSAD(size)                                                             \
 820 static int vsad ## size ## _c(MpegEncContext *c,                               \
 821                               uint8_t *s1, uint8_t *s2,                        \
 822                               ptrdiff_t stride, int h)                               \
 823 {                                                                              \
 824     int score = 0, x, y;                                                       \
 825                                                                                \
 826     for (y = 1; y < h; y++) {                                                  \
 827         for (x = 0; x < size; x++)                                             \
 828             score += FFABS(s1[x] - s2[x] - s1[x + stride] + s2[x + stride]);   \
 829         s1 += stride;                                                          \
 830         s2 += stride;                                                          \
 831     }                                                                          \
 832                                                                                \
 833     return score;                                                              \
 834 }
 835 VSAD(8)
 836 VSAD(16)
 837
 838 #define SQ(a) ((a) * (a))
 839 #define VSSE_INTRA(size)                                                \
 840 static int vsse_intra ## size ## _c(MpegEncContext *c,                  \
 841                                     uint8_t *s, uint8_t *dummy,         \
 842                                     ptrdiff_t stride, int h)            \
 843 {                                                                       \
 844     int score = 0, x, y;                                                \
 845                                                                         \
 846     for (y = 1; y < h; y++) {                                           \
 847         for (x = 0; x < size; x += 4) {                                 \
 848             score += SQ(s[x]     - s[x + stride]) +                     \
 849                      SQ(s[x + 1] - s[x + stride + 1]) +                 \
 850                      SQ(s[x + 2] - s[x + stride + 2]) +                 \
 851                      SQ(s[x + 3] - s[x + stride + 3]);                  \
 852         }                                                               \
 853         s += stride;                                                    \
 854     }                                                                   \
 855                                                                         \
 856     return score;                                                       \
 857 }
 858 VSSE_INTRA(8)
 859 VSSE_INTRA(16)
 860
 861 #define VSSE(size)                                                             \
 862 static int vsse ## size ## _c(MpegEncContext *c, uint8_t *s1, uint8_t *s2,     \
 863                               ptrdiff_t stride, int h)                         \
 864 {                                                                              \
 865     int score = 0, x, y;                                                       \
 866                                                                                \
 867     for (y = 1; y < h; y++) {                                                  \
 868         for (x = 0; x < size; x++)                                             \
 869             score += SQ(s1[x] - s2[x] - s1[x + stride] + s2[x + stride]);      \
 870         s1 += stride;                                                          \
 871         s2 += stride;                                                          \
 872     }                                                                          \
 873                                                                                \
 874     return score;                                                              \
 875 }
 876 VSSE(8)
 877 VSSE(16)
 878
 879 #define WRAPPER8_16_SQ(name8, name16)                                   \
 880 static int name16(MpegEncContext *s, uint8_t *dst, uint8_t *src,        \
 881                   ptrdiff_t stride, int h)                              \
 882 {                                                                       \
 883     int score = 0;                                                      \
 884                                                                         \
 885     score += name8(s, dst, src, stride, 8);                             \
 886     score += name8(s, dst + 8, src + 8, stride, 8);                     \
 887     if (h == 16) {                                                      \
 888         dst   += 8 * stride;                                            \
 889         src   += 8 * stride;                                            \
 890         score += name8(s, dst, src, stride, 8);                         \
 891         score += name8(s, dst + 8, src + 8, stride, 8);                 \
 892     }                                                                   \
 893     return score;                                                       \
 894 }
 895
 896 WRAPPER8_16_SQ(hadamard8_diff8x8_c, hadamard8_diff16_c)
 897 WRAPPER8_16_SQ(hadamard8_intra8x8_c, hadamard8_intra16_c)
 898 WRAPPER8_16_SQ(dct_sad8x8_c, dct_sad16_c)
 899 #if CONFIG_GPL
 900 WRAPPER8_16_SQ(dct264_sad8x8_c, dct264_sad16_c)
 901 #endif
 902 WRAPPER8_16_SQ(dct_max8x8_c, dct_max16_c)
 903 WRAPPER8_16_SQ(quant_psnr8x8_c, quant_psnr16_c)
 904 WRAPPER8_16_SQ(rd8x8_c, rd16_c)
 905 WRAPPER8_16_SQ(bit8x8_c, bit16_c)
 906
 907 av_cold void ff_me_cmp_init_static(void)
 908 {
 909     int i;
 910
 911     for (i = 0; i < 512; i++)
 912         ff_square_tab[i] = (i - 256) * (i - 256);
 913 }
 914
 915 int ff_check_alignment(void)
 916 {
 917     static int did_fail = 0;
 918     LOCAL_ALIGNED_16(int, aligned, [4]);
 919
 920     if ((intptr_t)aligned & 15) {
 921         if (!did_fail) {
 922 #if HAVE_MMX || HAVE_ALTIVEC
 923             av_log(NULL, AV_LOG_ERROR,
 924                 "Compiler did not align stack variables. Libavcodec has been miscompiled\n"
 925                 "and may be very slow or crash. This is not a bug in libavcodec,\n"
 926                 "but in the compiler. You may try recompiling using gcc >= 4.2.\n"
 927                 "Do not report crashes to FFmpeg developers.\n");
 928 #endif
 929             did_fail=1;
 930         }
 931         return -1;
 932     }
 933     return 0;
 934 }
 935
 936 av_cold void ff_me_cmp_init(MECmpContext *c, AVCodecContext *avctx)
 937 {
 938     ff_check_alignment();
 939
 940     c->sum_abs_dctelem = sum_abs_dctelem_c;
 941
 942     /* TODO [0] 16  [1] 8 */
 943     c->pix_abs[0][0] = pix_abs16_c;
 944     c->pix_abs[0][1] = pix_abs16_x2_c;
 945     c->pix_abs[0][2] = pix_abs16_y2_c;
 946     c->pix_abs[0][3] = pix_abs16_xy2_c;
 947     c->pix_abs[1][0] = pix_abs8_c;
 948     c->pix_abs[1][1] = pix_abs8_x2_c;
 949     c->pix_abs[1][2] = pix_abs8_y2_c;
 950     c->pix_abs[1][3] = pix_abs8_xy2_c;
 951
 952 #define SET_CMP_FUNC(name)                      \
 953     c->name[0] = name ## 16_c;                  \
 954     c->name[1] = name ## 8x8_c;
 955
 956     SET_CMP_FUNC(hadamard8_diff)
 957     c->hadamard8_diff[4] = hadamard8_intra16_c;
 958     c->hadamard8_diff[5] = hadamard8_intra8x8_c;
 959     SET_CMP_FUNC(dct_sad)
 960     SET_CMP_FUNC(dct_max)
 961 #if CONFIG_GPL
 962     SET_CMP_FUNC(dct264_sad)
 963 #endif
 964     c->sad[0] = pix_abs16_c;
 965     c->sad[1] = pix_abs8_c;
 966     c->sse[0] = sse16_c;
 967     c->sse[1] = sse8_c;
 968     c->sse[2] = sse4_c;
 969     SET_CMP_FUNC(quant_psnr)
 970     SET_CMP_FUNC(rd)
 971     SET_CMP_FUNC(bit)
 972     c->vsad[0] = vsad16_c;
 973     c->vsad[1] = vsad8_c;
 974     c->vsad[4] = vsad_intra16_c;
 975     c->vsad[5] = vsad_intra8_c;
 976     c->vsse[0] = vsse16_c;
 977     c->vsse[1] = vsse8_c;
 978     c->vsse[4] = vsse_intra16_c;
 979     c->vsse[5] = vsse_intra8_c;
 980     c->nsse[0] = nsse16_c;
 981     c->nsse[1] = nsse8_c;
 982 #if CONFIG_SNOW_DECODER || CONFIG_SNOW_ENCODER
 983     ff_dsputil_init_dwt(c);
 984 #endif
 985
 986     if (ARCH_ALPHA)
 987         ff_me_cmp_init_alpha(c, avctx);
 988     if (ARCH_ARM)
 989         ff_me_cmp_init_arm(c, avctx);
 990     if (ARCH_PPC)
 991         ff_me_cmp_init_ppc(c, avctx);
 992     if (ARCH_X86)
 993         ff_me_cmp_init_x86(c, avctx);
 994     if (ARCH_MIPS)
 995         ff_me_cmp_init_mips(c, avctx);
 996 }