git.sesse.net Git - ffmpeg/blob - libavcodec/vc1dsp.c

   1 /*
   2  * VC-1 and WMV3 decoder - DSP functions
   3  * Copyright (c) 2006 Konstantin Shishkov
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * VC-1 and WMV3 decoder
  25  */
  26
  27 #include "libavutil/avassert.h"
  28 #include "libavutil/common.h"
  29 #include "libavutil/intreadwrite.h"
  30 #include "h264chroma.h"
  31 #include "qpeldsp.h"
  32 #include "rnd_avg.h"
  33 #include "vc1dsp.h"
  34 #include "startcode.h"
  35
  36 /* Apply overlap transform to horizontal edge */
  37 static void vc1_v_overlap_c(uint8_t *src, int stride)
  38 {
  39     int i;
  40     int a, b, c, d;
  41     int d1, d2;
  42     int rnd = 1;
  43     for (i = 0; i < 8; i++) {
  44         a  = src[-2 * stride];
  45         b  = src[-stride];
  46         c  = src[0];
  47         d  = src[stride];
  48         d1 = (a - d + 3 + rnd) >> 3;
  49         d2 = (a - d + b - c + 4 - rnd) >> 3;
  50
  51         src[-2 * stride] = a - d1;
  52         src[-stride]     = av_clip_uint8(b - d2);
  53         src[0]           = av_clip_uint8(c + d2);
  54         src[stride]      = d + d1;
  55         src++;
  56         rnd = !rnd;
  57     }
  58 }
  59
  60 /* Apply overlap transform to vertical edge */
  61 static void vc1_h_overlap_c(uint8_t *src, int stride)
  62 {
  63     int i;
  64     int a, b, c, d;
  65     int d1, d2;
  66     int rnd = 1;
  67     for (i = 0; i < 8; i++) {
  68         a  = src[-2];
  69         b  = src[-1];
  70         c  = src[0];
  71         d  = src[1];
  72         d1 = (a - d + 3 + rnd) >> 3;
  73         d2 = (a - d + b - c + 4 - rnd) >> 3;
  74
  75         src[-2] = a - d1;
  76         src[-1] = av_clip_uint8(b - d2);
  77         src[0]  = av_clip_uint8(c + d2);
  78         src[1]  = d + d1;
  79         src    += stride;
  80         rnd     = !rnd;
  81     }
  82 }
  83
  84 static void vc1_v_s_overlap_c(int16_t *top, int16_t *bottom)
  85 {
  86     int i;
  87     int a, b, c, d;
  88     int d1, d2;
  89     int rnd1 = 4, rnd2 = 3;
  90     for (i = 0; i < 8; i++) {
  91         a  = top[48];
  92         b  = top[56];
  93         c  = bottom[0];
  94         d  = bottom[8];
  95         d1 = a - d;
  96         d2 = a - d + b - c;
  97
  98         top[48]   = ((a * 8) - d1 + rnd1) >> 3;
  99         top[56]   = ((b * 8) - d2 + rnd2) >> 3;
 100         bottom[0] = ((c * 8) + d2 + rnd1) >> 3;
 101         bottom[8] = ((d * 8) + d1 + rnd2) >> 3;
 102
 103         bottom++;
 104         top++;
 105         rnd2 = 7 - rnd2;
 106         rnd1 = 7 - rnd1;
 107     }
 108 }
 109
 110 static void vc1_h_s_overlap_c(int16_t *left, int16_t *right, int left_stride, int right_stride, int flags)
 111 {
 112     int i;
 113     int a, b, c, d;
 114     int d1, d2;
 115     int rnd1 = flags & 2 ? 3 : 4;
 116     int rnd2 = 7 - rnd1;
 117     for (i = 0; i < 8; i++) {
 118         a  = left[6];
 119         b  = left[7];
 120         c  = right[0];
 121         d  = right[1];
 122         d1 = a - d;
 123         d2 = a - d + b - c;
 124
 125         left[6]  = ((a * 8) - d1 + rnd1) >> 3;
 126         left[7]  = ((b * 8) - d2 + rnd2) >> 3;
 127         right[0] = ((c * 8) + d2 + rnd1) >> 3;
 128         right[1] = ((d * 8) + d1 + rnd2) >> 3;
 129
 130         right += right_stride;
 131         left  += left_stride;
 132         if (flags & 1) {
 133             rnd2   = 7 - rnd2;
 134             rnd1   = 7 - rnd1;
 135         }
 136     }
 137 }
 138
 139 /**
 140  * VC-1 in-loop deblocking filter for one line
 141  * @param src source block type
 142  * @param stride block stride
 143  * @param pq block quantizer
 144  * @return whether other 3 pairs should be filtered or not
 145  * @see 8.6
 146  */
 147 static av_always_inline int vc1_filter_line(uint8_t *src, int stride, int pq)
 148 {
 149     int a0 = (2 * (src[-2 * stride] - src[1 * stride]) -
 150               5 * (src[-1 * stride] - src[0 * stride]) + 4) >> 3;
 151     int a0_sign = a0 >> 31;        /* Store sign */
 152
 153     a0 = (a0 ^ a0_sign) - a0_sign; /* a0 = FFABS(a0); */
 154     if (a0 < pq) {
 155         int a1 = FFABS((2 * (src[-4 * stride] - src[-1 * stride]) -
 156                         5 * (src[-3 * stride] - src[-2 * stride]) + 4) >> 3);
 157         int a2 = FFABS((2 * (src[ 0 * stride] - src[ 3 * stride]) -
 158                         5 * (src[ 1 * stride] - src[ 2 * stride]) + 4) >> 3);
 159         if (a1 < a0 || a2 < a0) {
 160             int clip      = src[-1 * stride] - src[0 * stride];
 161             int clip_sign = clip >> 31;
 162
 163             clip = ((clip ^ clip_sign) - clip_sign) >> 1;
 164             if (clip) {
 165                 int a3     = FFMIN(a1, a2);
 166                 int d      = 5 * (a3 - a0);
 167                 int d_sign = (d >> 31);
 168
 169                 d       = ((d ^ d_sign) - d_sign) >> 3;
 170                 d_sign ^= a0_sign;
 171
 172                 if (d_sign ^ clip_sign)
 173                     d = 0;
 174                 else {
 175                     d = FFMIN(d, clip);
 176                     d = (d ^ d_sign) - d_sign; /* Restore sign */
 177                     src[-1 * stride] = av_clip_uint8(src[-1 * stride] - d);
 178                     src[ 0 * stride] = av_clip_uint8(src[ 0 * stride] + d);
 179                 }
 180                 return 1;
 181             }
 182         }
 183     }
 184     return 0;
 185 }
 186
 187 /**
 188  * VC-1 in-loop deblocking filter
 189  * @param src source block type
 190  * @param step distance between horizontally adjacent elements
 191  * @param stride distance between vertically adjacent elements
 192  * @param len edge length to filter (4 or 8 pixels)
 193  * @param pq block quantizer
 194  * @see 8.6
 195  */
 196 static inline void vc1_loop_filter(uint8_t *src, int step, int stride,
 197                                    int len, int pq)
 198 {
 199     int i;
 200     int filt3;
 201
 202     for (i = 0; i < len; i += 4) {
 203         filt3 = vc1_filter_line(src + 2 * step, stride, pq);
 204         if (filt3) {
 205             vc1_filter_line(src + 0 * step, stride, pq);
 206             vc1_filter_line(src + 1 * step, stride, pq);
 207             vc1_filter_line(src + 3 * step, stride, pq);
 208         }
 209         src += step * 4;
 210     }
 211 }
 212
 213 static void vc1_v_loop_filter4_c(uint8_t *src, int stride, int pq)
 214 {
 215     vc1_loop_filter(src, 1, stride, 4, pq);
 216 }
 217
 218 static void vc1_h_loop_filter4_c(uint8_t *src, int stride, int pq)
 219 {
 220     vc1_loop_filter(src, stride, 1, 4, pq);
 221 }
 222
 223 static void vc1_v_loop_filter8_c(uint8_t *src, int stride, int pq)
 224 {
 225     vc1_loop_filter(src, 1, stride, 8, pq);
 226 }
 227
 228 static void vc1_h_loop_filter8_c(uint8_t *src, int stride, int pq)
 229 {
 230     vc1_loop_filter(src, stride, 1, 8, pq);
 231 }
 232
 233 static void vc1_v_loop_filter16_c(uint8_t *src, int stride, int pq)
 234 {
 235     vc1_loop_filter(src, 1, stride, 16, pq);
 236 }
 237
 238 static void vc1_h_loop_filter16_c(uint8_t *src, int stride, int pq)
 239 {
 240     vc1_loop_filter(src, stride, 1, 16, pq);
 241 }
 242
 243 /* Do inverse transform on 8x8 block */
 244 static void vc1_inv_trans_8x8_dc_c(uint8_t *dest, ptrdiff_t stride, int16_t *block)
 245 {
 246     int i;
 247     int dc = block[0];
 248
 249     dc = (3 * dc +  1) >> 1;
 250     dc = (3 * dc + 16) >> 5;
 251
 252     for (i = 0; i < 8; i++) {
 253         dest[0] = av_clip_uint8(dest[0] + dc);
 254         dest[1] = av_clip_uint8(dest[1] + dc);
 255         dest[2] = av_clip_uint8(dest[2] + dc);
 256         dest[3] = av_clip_uint8(dest[3] + dc);
 257         dest[4] = av_clip_uint8(dest[4] + dc);
 258         dest[5] = av_clip_uint8(dest[5] + dc);
 259         dest[6] = av_clip_uint8(dest[6] + dc);
 260         dest[7] = av_clip_uint8(dest[7] + dc);
 261         dest += stride;
 262     }
 263 }
 264
 265 static void vc1_inv_trans_8x8_c(int16_t block[64])
 266 {
 267     int i;
 268     register int t1, t2, t3, t4, t5, t6, t7, t8;
 269     int16_t *src, *dst, temp[64];
 270
 271     src = block;
 272     dst = temp;
 273     for (i = 0; i < 8; i++) {
 274         t1 = 12 * (src[ 0] + src[32]) + 4;
 275         t2 = 12 * (src[ 0] - src[32]) + 4;
 276         t3 = 16 * src[16] +  6 * src[48];
 277         t4 =  6 * src[16] - 16 * src[48];
 278
 279         t5 = t1 + t3;
 280         t6 = t2 + t4;
 281         t7 = t2 - t4;
 282         t8 = t1 - t3;
 283
 284         t1 = 16 * src[ 8] + 15 * src[24] +  9 * src[40] +  4 * src[56];
 285         t2 = 15 * src[ 8] -  4 * src[24] - 16 * src[40] -  9 * src[56];
 286         t3 =  9 * src[ 8] - 16 * src[24] +  4 * src[40] + 15 * src[56];
 287         t4 =  4 * src[ 8] -  9 * src[24] + 15 * src[40] - 16 * src[56];
 288
 289         dst[0] = (t5 + t1) >> 3;
 290         dst[1] = (t6 + t2) >> 3;
 291         dst[2] = (t7 + t3) >> 3;
 292         dst[3] = (t8 + t4) >> 3;
 293         dst[4] = (t8 - t4) >> 3;
 294         dst[5] = (t7 - t3) >> 3;
 295         dst[6] = (t6 - t2) >> 3;
 296         dst[7] = (t5 - t1) >> 3;
 297
 298         src += 1;
 299         dst += 8;
 300     }
 301
 302     src = temp;
 303     dst = block;
 304     for (i = 0; i < 8; i++) {
 305         t1 = 12 * (src[ 0] + src[32]) + 64;
 306         t2 = 12 * (src[ 0] - src[32]) + 64;
 307         t3 = 16 * src[16] +  6 * src[48];
 308         t4 =  6 * src[16] - 16 * src[48];
 309
 310         t5 = t1 + t3;
 311         t6 = t2 + t4;
 312         t7 = t2 - t4;
 313         t8 = t1 - t3;
 314
 315         t1 = 16 * src[ 8] + 15 * src[24] +  9 * src[40] +  4 * src[56];
 316         t2 = 15 * src[ 8] -  4 * src[24] - 16 * src[40] -  9 * src[56];
 317         t3 =  9 * src[ 8] - 16 * src[24] +  4 * src[40] + 15 * src[56];
 318         t4 =  4 * src[ 8] -  9 * src[24] + 15 * src[40] - 16 * src[56];
 319
 320         dst[ 0] = (t5 + t1) >> 7;
 321         dst[ 8] = (t6 + t2) >> 7;
 322         dst[16] = (t7 + t3) >> 7;
 323         dst[24] = (t8 + t4) >> 7;
 324         dst[32] = (t8 - t4 + 1) >> 7;
 325         dst[40] = (t7 - t3 + 1) >> 7;
 326         dst[48] = (t6 - t2 + 1) >> 7;
 327         dst[56] = (t5 - t1 + 1) >> 7;
 328
 329         src++;
 330         dst++;
 331     }
 332 }
 333
 334 /* Do inverse transform on 8x4 part of block */
 335 static void vc1_inv_trans_8x4_dc_c(uint8_t *dest, ptrdiff_t stride, int16_t *block)
 336 {
 337     int i;
 338     int dc = block[0];
 339
 340     dc =  (3 * dc +  1) >> 1;
 341     dc = (17 * dc + 64) >> 7;
 342
 343     for (i = 0; i < 4; i++) {
 344         dest[0] = av_clip_uint8(dest[0] + dc);
 345         dest[1] = av_clip_uint8(dest[1] + dc);
 346         dest[2] = av_clip_uint8(dest[2] + dc);
 347         dest[3] = av_clip_uint8(dest[3] + dc);
 348         dest[4] = av_clip_uint8(dest[4] + dc);
 349         dest[5] = av_clip_uint8(dest[5] + dc);
 350         dest[6] = av_clip_uint8(dest[6] + dc);
 351         dest[7] = av_clip_uint8(dest[7] + dc);
 352         dest += stride;
 353     }
 354 }
 355
 356 static void vc1_inv_trans_8x4_c(uint8_t *dest, ptrdiff_t stride, int16_t *block)
 357 {
 358     int i;
 359     register int t1, t2, t3, t4, t5, t6, t7, t8;
 360     int16_t *src, *dst;
 361
 362     src = block;
 363     dst = block;
 364
 365     for (i = 0; i < 4; i++) {
 366         t1 = 12 * (src[0] + src[4]) + 4;
 367         t2 = 12 * (src[0] - src[4]) + 4;
 368         t3 = 16 * src[2] +  6 * src[6];
 369         t4 =  6 * src[2] - 16 * src[6];
 370
 371         t5 = t1 + t3;
 372         t6 = t2 + t4;
 373         t7 = t2 - t4;
 374         t8 = t1 - t3;
 375
 376         t1 = 16 * src[1] + 15 * src[3] +  9 * src[5] +  4 * src[7];
 377         t2 = 15 * src[1] -  4 * src[3] - 16 * src[5] -  9 * src[7];
 378         t3 =  9 * src[1] - 16 * src[3] +  4 * src[5] + 15 * src[7];
 379         t4 =  4 * src[1] -  9 * src[3] + 15 * src[5] - 16 * src[7];
 380
 381         dst[0] = (t5 + t1) >> 3;
 382         dst[1] = (t6 + t2) >> 3;
 383         dst[2] = (t7 + t3) >> 3;
 384         dst[3] = (t8 + t4) >> 3;
 385         dst[4] = (t8 - t4) >> 3;
 386         dst[5] = (t7 - t3) >> 3;
 387         dst[6] = (t6 - t2) >> 3;
 388         dst[7] = (t5 - t1) >> 3;
 389
 390         src += 8;
 391         dst += 8;
 392     }
 393
 394     src = block;
 395     for (i = 0; i < 8; i++) {
 396         t1 = 17 * (src[ 0] + src[16]) + 64;
 397         t2 = 17 * (src[ 0] - src[16]) + 64;
 398         t3 = 22 * src[ 8] + 10 * src[24];
 399         t4 = 22 * src[24] - 10 * src[ 8];
 400
 401         dest[0 * stride] = av_clip_uint8(dest[0 * stride] + ((t1 + t3) >> 7));
 402         dest[1 * stride] = av_clip_uint8(dest[1 * stride] + ((t2 - t4) >> 7));
 403         dest[2 * stride] = av_clip_uint8(dest[2 * stride] + ((t2 + t4) >> 7));
 404         dest[3 * stride] = av_clip_uint8(dest[3 * stride] + ((t1 - t3) >> 7));
 405
 406         src++;
 407         dest++;
 408     }
 409 }
 410
 411 /* Do inverse transform on 4x8 parts of block */
 412 static void vc1_inv_trans_4x8_dc_c(uint8_t *dest, ptrdiff_t stride, int16_t *block)
 413 {
 414     int i;
 415     int dc = block[0];
 416
 417     dc = (17 * dc +  4) >> 3;
 418     dc = (12 * dc + 64) >> 7;
 419
 420     for (i = 0; i < 8; i++) {
 421         dest[0] = av_clip_uint8(dest[0] + dc);
 422         dest[1] = av_clip_uint8(dest[1] + dc);
 423         dest[2] = av_clip_uint8(dest[2] + dc);
 424         dest[3] = av_clip_uint8(dest[3] + dc);
 425         dest += stride;
 426     }
 427 }
 428
 429 static void vc1_inv_trans_4x8_c(uint8_t *dest, ptrdiff_t stride, int16_t *block)
 430 {
 431     int i;
 432     register int t1, t2, t3, t4, t5, t6, t7, t8;
 433     int16_t *src, *dst;
 434
 435     src = block;
 436     dst = block;
 437
 438     for (i = 0; i < 8; i++) {
 439         t1 = 17 * (src[0] + src[2]) + 4;
 440         t2 = 17 * (src[0] - src[2]) + 4;
 441         t3 = 22 * src[1] + 10 * src[3];
 442         t4 = 22 * src[3] - 10 * src[1];
 443
 444         dst[0] = (t1 + t3) >> 3;
 445         dst[1] = (t2 - t4) >> 3;
 446         dst[2] = (t2 + t4) >> 3;
 447         dst[3] = (t1 - t3) >> 3;
 448
 449         src += 8;
 450         dst += 8;
 451     }
 452
 453     src = block;
 454     for (i = 0; i < 4; i++) {
 455         t1 = 12 * (src[ 0] + src[32]) + 64;
 456         t2 = 12 * (src[ 0] - src[32]) + 64;
 457         t3 = 16 * src[16] +  6 * src[48];
 458         t4 =  6 * src[16] - 16 * src[48];
 459
 460         t5 = t1 + t3;
 461         t6 = t2 + t4;
 462         t7 = t2 - t4;
 463         t8 = t1 - t3;
 464
 465         t1 = 16 * src[ 8] + 15 * src[24] +  9 * src[40] +  4 * src[56];
 466         t2 = 15 * src[ 8] -  4 * src[24] - 16 * src[40] -  9 * src[56];
 467         t3 =  9 * src[ 8] - 16 * src[24] +  4 * src[40] + 15 * src[56];
 468         t4 =  4 * src[ 8] -  9 * src[24] + 15 * src[40] - 16 * src[56];
 469
 470         dest[0 * stride] = av_clip_uint8(dest[0 * stride] + ((t5 + t1)     >> 7));
 471         dest[1 * stride] = av_clip_uint8(dest[1 * stride] + ((t6 + t2)     >> 7));
 472         dest[2 * stride] = av_clip_uint8(dest[2 * stride] + ((t7 + t3)     >> 7));
 473         dest[3 * stride] = av_clip_uint8(dest[3 * stride] + ((t8 + t4)     >> 7));
 474         dest[4 * stride] = av_clip_uint8(dest[4 * stride] + ((t8 - t4 + 1) >> 7));
 475         dest[5 * stride] = av_clip_uint8(dest[5 * stride] + ((t7 - t3 + 1) >> 7));
 476         dest[6 * stride] = av_clip_uint8(dest[6 * stride] + ((t6 - t2 + 1) >> 7));
 477         dest[7 * stride] = av_clip_uint8(dest[7 * stride] + ((t5 - t1 + 1) >> 7));
 478
 479         src++;
 480         dest++;
 481     }
 482 }
 483
 484 /* Do inverse transform on 4x4 part of block */
 485 static void vc1_inv_trans_4x4_dc_c(uint8_t *dest, ptrdiff_t stride, int16_t *block)
 486 {
 487     int i;
 488     int dc = block[0];
 489
 490     dc = (17 * dc +  4) >> 3;
 491     dc = (17 * dc + 64) >> 7;
 492
 493     for (i = 0; i < 4; i++) {
 494         dest[0] = av_clip_uint8(dest[0] + dc);
 495         dest[1] = av_clip_uint8(dest[1] + dc);
 496         dest[2] = av_clip_uint8(dest[2] + dc);
 497         dest[3] = av_clip_uint8(dest[3] + dc);
 498         dest += stride;
 499     }
 500 }
 501
 502 static void vc1_inv_trans_4x4_c(uint8_t *dest, ptrdiff_t stride, int16_t *block)
 503 {
 504     int i;
 505     register int t1, t2, t3, t4;
 506     int16_t *src, *dst;
 507
 508     src = block;
 509     dst = block;
 510     for (i = 0; i < 4; i++) {
 511         t1 = 17 * (src[0] + src[2]) + 4;
 512         t2 = 17 * (src[0] - src[2]) + 4;
 513         t3 = 22 * src[1] + 10 * src[3];
 514         t4 = 22 * src[3] - 10 * src[1];
 515
 516         dst[0] = (t1 + t3) >> 3;
 517         dst[1] = (t2 - t4) >> 3;
 518         dst[2] = (t2 + t4) >> 3;
 519         dst[3] = (t1 - t3) >> 3;
 520
 521         src += 8;
 522         dst += 8;
 523     }
 524
 525     src = block;
 526     for (i = 0; i < 4; i++) {
 527         t1 = 17 * (src[0] + src[16]) + 64;
 528         t2 = 17 * (src[0] - src[16]) + 64;
 529         t3 = 22 * src[8] + 10 * src[24];
 530         t4 = 22 * src[24] - 10 * src[8];
 531
 532         dest[0 * stride] = av_clip_uint8(dest[0 * stride] + ((t1 + t3) >> 7));
 533         dest[1 * stride] = av_clip_uint8(dest[1 * stride] + ((t2 - t4) >> 7));
 534         dest[2 * stride] = av_clip_uint8(dest[2 * stride] + ((t2 + t4) >> 7));
 535         dest[3 * stride] = av_clip_uint8(dest[3 * stride] + ((t1 - t3) >> 7));
 536
 537         src++;
 538         dest++;
 539     }
 540 }
 541
 542 /* motion compensation functions */
 543
 544 /* Filter in case of 2 filters */
 545 #define VC1_MSPEL_FILTER_16B(DIR, TYPE)                                       \
 546 static av_always_inline int vc1_mspel_ ## DIR ## _filter_16bits(const TYPE *src, \
 547                                                                 int stride,   \
 548                                                                 int mode)     \
 549 {                                                                             \
 550     switch(mode) {                                                            \
 551     case 0: /* no shift - should not occur */                                 \
 552         return 0;                                                             \
 553     case 1: /* 1/4 shift */                                                   \
 554         return -4 * src[-stride] + 53 * src[0] +                              \
 555                18 * src[stride]  -  3 * src[stride * 2];                      \
 556     case 2: /* 1/2 shift */                                                   \
 557         return -1 * src[-stride] +  9 * src[0] +                              \
 558                 9 * src[stride]  -  1 * src[stride * 2];                      \
 559     case 3: /* 3/4 shift */                                                   \
 560         return -3 * src[-stride] + 18 * src[0] +                              \
 561                53 * src[stride]  -  4 * src[stride * 2];                      \
 562     }                                                                         \
 563     return 0; /* should not occur */                                          \
 564 }
 565
 566 VC1_MSPEL_FILTER_16B(ver, uint8_t)
 567 VC1_MSPEL_FILTER_16B(hor, int16_t)
 568
 569 /* Filter used to interpolate fractional pel values */
 570 static av_always_inline int vc1_mspel_filter(const uint8_t *src, int stride,
 571                                              int mode, int r)
 572 {
 573     switch (mode) {
 574     case 0: // no shift
 575         return src[0];
 576     case 1: // 1/4 shift
 577         return (-4 * src[-stride] + 53 * src[0] +
 578                 18 * src[stride]  -  3 * src[stride * 2] + 32 - r) >> 6;
 579     case 2: // 1/2 shift
 580         return (-1 * src[-stride] +  9 * src[0] +
 581                  9 * src[stride]  -  1 * src[stride * 2] + 8 - r) >> 4;
 582     case 3: // 3/4 shift
 583         return (-3 * src[-stride] + 18 * src[0] +
 584                 53 * src[stride]  -  4 * src[stride * 2] + 32 - r) >> 6;
 585     }
 586     return 0; // should not occur
 587 }
 588
 589 /* Function used to do motion compensation with bicubic interpolation */
 590 #define VC1_MSPEL_MC(OP, OP4, OPNAME)                                         \
 591 static av_always_inline void OPNAME ## vc1_mspel_mc(uint8_t *dst,             \
 592                                                     const uint8_t *src,       \
 593                                                     ptrdiff_t stride,         \
 594                                                     int hmode,                \
 595                                                     int vmode,                \
 596                                                     int rnd)                  \
 597 {                                                                             \
 598     int i, j;                                                                 \
 599                                                                               \
 600     if (vmode) { /* Horizontal filter to apply */                             \
 601         int r;                                                                \
 602                                                                               \
 603         if (hmode) { /* Vertical filter to apply, output to tmp */            \
 604             static const int shift_value[] = { 0, 5, 1, 5 };                  \
 605             int shift = (shift_value[hmode] + shift_value[vmode]) >> 1;       \
 606             int16_t tmp[11 * 8], *tptr = tmp;                                 \
 607                                                                               \
 608             r = (1 << (shift - 1)) + rnd - 1;                                 \
 609                                                                               \
 610             src -= 1;                                                         \
 611             for (j = 0; j < 8; j++) {                                         \
 612                 for (i = 0; i < 11; i++)                                      \
 613                     tptr[i] = (vc1_mspel_ver_filter_16bits(src + i, stride, vmode) + r) >> shift; \
 614                 src  += stride;                                               \
 615                 tptr += 11;                                                   \
 616             }                                                                 \
 617                                                                               \
 618             r    = 64 - rnd;                                                  \
 619             tptr = tmp + 1;                                                   \
 620             for (j = 0; j < 8; j++) {                                         \
 621                 for (i = 0; i < 8; i++)                                       \
 622                     OP(dst[i], (vc1_mspel_hor_filter_16bits(tptr + i, 1, hmode) + r) >> 7); \
 623                 dst  += stride;                                               \
 624                 tptr += 11;                                                   \
 625             }                                                                 \
 626                                                                               \
 627             return;                                                           \
 628         } else { /* No horizontal filter, output 8 lines to dst */            \
 629             r = 1 - rnd;                                                      \
 630                                                                               \
 631             for (j = 0; j < 8; j++) {                                         \
 632                 for (i = 0; i < 8; i++)                                       \
 633                     OP(dst[i], vc1_mspel_filter(src + i, stride, vmode, r));  \
 634                 src += stride;                                                \
 635                 dst += stride;                                                \
 636             }                                                                 \
 637             return;                                                           \
 638         }                                                                     \
 639     }                                                                         \
 640                                                                               \
 641     /* Horizontal mode with no vertical mode */                               \
 642     for (j = 0; j < 8; j++) {                                                 \
 643         for (i = 0; i < 8; i++)                                               \
 644             OP(dst[i], vc1_mspel_filter(src + i, 1, hmode, rnd));             \
 645         dst += stride;                                                        \
 646         src += stride;                                                        \
 647     }                                                                         \
 648 }\
 649 static av_always_inline void OPNAME ## vc1_mspel_mc_16(uint8_t *dst,          \
 650                                                        const uint8_t *src,    \
 651                                                        ptrdiff_t stride,      \
 652                                                        int hmode,             \
 653                                                        int vmode,             \
 654                                                        int rnd)               \
 655 {                                                                             \
 656     int i, j;                                                                 \
 657                                                                               \
 658     if (vmode) { /* Horizontal filter to apply */                             \
 659         int r;                                                                \
 660                                                                               \
 661         if (hmode) { /* Vertical filter to apply, output to tmp */            \
 662             static const int shift_value[] = { 0, 5, 1, 5 };                  \
 663             int shift = (shift_value[hmode] + shift_value[vmode]) >> 1;       \
 664             int16_t tmp[19 * 16], *tptr = tmp;                                \
 665                                                                               \
 666             r = (1 << (shift - 1)) + rnd - 1;                                 \
 667                                                                               \
 668             src -= 1;                                                         \
 669             for (j = 0; j < 16; j++) {                                        \
 670                 for (i = 0; i < 19; i++)                                      \
 671                     tptr[i] = (vc1_mspel_ver_filter_16bits(src + i, stride, vmode) + r) >> shift; \
 672                 src  += stride;                                               \
 673                 tptr += 19;                                                   \
 674             }                                                                 \
 675                                                                               \
 676             r    = 64 - rnd;                                                  \
 677             tptr = tmp + 1;                                                   \
 678             for (j = 0; j < 16; j++) {                                        \
 679                 for (i = 0; i < 16; i++)                                      \
 680                     OP(dst[i], (vc1_mspel_hor_filter_16bits(tptr + i, 1, hmode) + r) >> 7); \
 681                 dst  += stride;                                               \
 682                 tptr += 19;                                                   \
 683             }                                                                 \
 684                                                                               \
 685             return;                                                           \
 686         } else { /* No horizontal filter, output 8 lines to dst */            \
 687             r = 1 - rnd;                                                      \
 688                                                                               \
 689             for (j = 0; j < 16; j++) {                                        \
 690                 for (i = 0; i < 16; i++)                                      \
 691                     OP(dst[i], vc1_mspel_filter(src + i, stride, vmode, r));  \
 692                 src += stride;                                                \
 693                 dst += stride;                                                \
 694             }                                                                 \
 695             return;                                                           \
 696         }                                                                     \
 697     }                                                                         \
 698                                                                               \
 699     /* Horizontal mode with no vertical mode */                               \
 700     for (j = 0; j < 16; j++) {                                                \
 701         for (i = 0; i < 16; i++)                                              \
 702             OP(dst[i], vc1_mspel_filter(src + i, 1, hmode, rnd));             \
 703         dst += stride;                                                        \
 704         src += stride;                                                        \
 705     }                                                                         \
 706 }\
 707 static void OPNAME ## pixels8x8_c(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int rnd){\
 708     int i;\
 709     for(i=0; i<8; i++){\
 710         OP4(*(uint32_t*)(block  ), AV_RN32(pixels  ));\
 711         OP4(*(uint32_t*)(block+4), AV_RN32(pixels+4));\
 712         pixels+=line_size;\
 713         block +=line_size;\
 714     }\
 715 }\
 716 static void OPNAME ## pixels16x16_c(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int rnd){\
 717     int i;\
 718     for(i=0; i<16; i++){\
 719         OP4(*(uint32_t*)(block   ), AV_RN32(pixels   ));\
 720         OP4(*(uint32_t*)(block+ 4), AV_RN32(pixels+ 4));\
 721         OP4(*(uint32_t*)(block+ 8), AV_RN32(pixels+ 8));\
 722         OP4(*(uint32_t*)(block+12), AV_RN32(pixels+12));\
 723         pixels+=line_size;\
 724         block +=line_size;\
 725     }\
 726 }
 727
 728 #define op_put(a, b) (a) = av_clip_uint8(b)
 729 #define op_avg(a, b) (a) = ((a) + av_clip_uint8(b) + 1) >> 1
 730 #define op4_avg(a, b) (a) = rnd_avg32(a, b)
 731 #define op4_put(a, b) (a) = (b)
 732
 733 VC1_MSPEL_MC(op_put, op4_put, put_)
 734 VC1_MSPEL_MC(op_avg, op4_avg, avg_)
 735
 736 /* pixel functions - really are entry points to vc1_mspel_mc */
 737
 738 #define PUT_VC1_MSPEL(a, b)                                                   \
 739 static void put_vc1_mspel_mc ## a ## b ## _c(uint8_t *dst,                    \
 740                                              const uint8_t *src,              \
 741                                              ptrdiff_t stride, int rnd)       \
 742 {                                                                             \
 743     put_vc1_mspel_mc(dst, src, stride, a, b, rnd);                            \
 744 }                                                                             \
 745 static void avg_vc1_mspel_mc ## a ## b ## _c(uint8_t *dst,                    \
 746                                              const uint8_t *src,              \
 747                                              ptrdiff_t stride, int rnd)       \
 748 {                                                                             \
 749     avg_vc1_mspel_mc(dst, src, stride, a, b, rnd);                            \
 750 }                                                                             \
 751 static void put_vc1_mspel_mc ## a ## b ## _16_c(uint8_t *dst,                 \
 752                                                 const uint8_t *src,           \
 753                                                 ptrdiff_t stride, int rnd)    \
 754 {                                                                             \
 755     put_vc1_mspel_mc_16(dst, src, stride, a, b, rnd);                         \
 756 }                                                                             \
 757 static void avg_vc1_mspel_mc ## a ## b ## _16_c(uint8_t *dst,                 \
 758                                                 const uint8_t *src,           \
 759                                                 ptrdiff_t stride, int rnd)    \
 760 {                                                                             \
 761     avg_vc1_mspel_mc_16(dst, src, stride, a, b, rnd);                         \
 762 }
 763
 764 PUT_VC1_MSPEL(1, 0)
 765 PUT_VC1_MSPEL(2, 0)
 766 PUT_VC1_MSPEL(3, 0)
 767
 768 PUT_VC1_MSPEL(0, 1)
 769 PUT_VC1_MSPEL(1, 1)
 770 PUT_VC1_MSPEL(2, 1)
 771 PUT_VC1_MSPEL(3, 1)
 772
 773 PUT_VC1_MSPEL(0, 2)
 774 PUT_VC1_MSPEL(1, 2)
 775 PUT_VC1_MSPEL(2, 2)
 776 PUT_VC1_MSPEL(3, 2)
 777
 778 PUT_VC1_MSPEL(0, 3)
 779 PUT_VC1_MSPEL(1, 3)
 780 PUT_VC1_MSPEL(2, 3)
 781 PUT_VC1_MSPEL(3, 3)
 782
 783 #define chroma_mc(a) \
 784     ((A * src[a] + B * src[a + 1] + \
 785       C * src[stride + a] + D * src[stride + a + 1] + 32 - 4) >> 6)
 786 static void put_no_rnd_vc1_chroma_mc8_c(uint8_t *dst /* align 8 */,
 787                                         uint8_t *src /* align 1 */,
 788                                         ptrdiff_t stride, int h, int x, int y)
 789 {
 790     const int A = (8 - x) * (8 - y);
 791     const int B =     (x) * (8 - y);
 792     const int C = (8 - x) *     (y);
 793     const int D =     (x) *     (y);
 794     int i;
 795
 796     av_assert2(x < 8 && y < 8 && x >= 0 && y >= 0);
 797
 798     for (i = 0; i < h; i++) {
 799         dst[0] = chroma_mc(0);
 800         dst[1] = chroma_mc(1);
 801         dst[2] = chroma_mc(2);
 802         dst[3] = chroma_mc(3);
 803         dst[4] = chroma_mc(4);
 804         dst[5] = chroma_mc(5);
 805         dst[6] = chroma_mc(6);
 806         dst[7] = chroma_mc(7);
 807         dst += stride;
 808         src += stride;
 809     }
 810 }
 811
 812 static void put_no_rnd_vc1_chroma_mc4_c(uint8_t *dst, uint8_t *src,
 813                                         ptrdiff_t stride, int h, int x, int y)
 814 {
 815     const int A = (8 - x) * (8 - y);
 816     const int B =     (x) * (8 - y);
 817     const int C = (8 - x) *     (y);
 818     const int D =     (x) *     (y);
 819     int i;
 820
 821     av_assert2(x < 8 && y < 8 && x >= 0 && y >= 0);
 822
 823     for (i = 0; i < h; i++) {
 824         dst[0] = chroma_mc(0);
 825         dst[1] = chroma_mc(1);
 826         dst[2] = chroma_mc(2);
 827         dst[3] = chroma_mc(3);
 828         dst += stride;
 829         src += stride;
 830     }
 831 }
 832
 833 #define avg2(a, b) (((a) + (b) + 1) >> 1)
 834 static void avg_no_rnd_vc1_chroma_mc8_c(uint8_t *dst /* align 8 */,
 835                                         uint8_t *src /* align 1 */,
 836                                         ptrdiff_t stride, int h, int x, int y)
 837 {
 838     const int A = (8 - x) * (8 - y);
 839     const int B =     (x) * (8 - y);
 840     const int C = (8 - x) *     (y);
 841     const int D =     (x) *     (y);
 842     int i;
 843
 844     av_assert2(x < 8 && y < 8 && x >= 0 && y >= 0);
 845
 846     for (i = 0; i < h; i++) {
 847         dst[0] = avg2(dst[0], chroma_mc(0));
 848         dst[1] = avg2(dst[1], chroma_mc(1));
 849         dst[2] = avg2(dst[2], chroma_mc(2));
 850         dst[3] = avg2(dst[3], chroma_mc(3));
 851         dst[4] = avg2(dst[4], chroma_mc(4));
 852         dst[5] = avg2(dst[5], chroma_mc(5));
 853         dst[6] = avg2(dst[6], chroma_mc(6));
 854         dst[7] = avg2(dst[7], chroma_mc(7));
 855         dst += stride;
 856         src += stride;
 857     }
 858 }
 859
 860 static void avg_no_rnd_vc1_chroma_mc4_c(uint8_t *dst /* align 8 */,
 861                                         uint8_t *src /* align 1 */,
 862                                         ptrdiff_t stride, int h, int x, int y)
 863 {
 864     const int A = (8 - x) * (8 - y);
 865     const int B = (    x) * (8 - y);
 866     const int C = (8 - x) * (    y);
 867     const int D = (    x) * (    y);
 868     int i;
 869
 870     av_assert2(x < 8 && y < 8 && x >= 0 && y >= 0);
 871
 872     for (i = 0; i < h; i++) {
 873         dst[0] = avg2(dst[0], chroma_mc(0));
 874         dst[1] = avg2(dst[1], chroma_mc(1));
 875         dst[2] = avg2(dst[2], chroma_mc(2));
 876         dst[3] = avg2(dst[3], chroma_mc(3));
 877         dst += stride;
 878         src += stride;
 879     }
 880 }
 881
 882 #if CONFIG_WMV3IMAGE_DECODER || CONFIG_VC1IMAGE_DECODER
 883
 884 static void sprite_h_c(uint8_t *dst, const uint8_t *src, int offset,
 885                        int advance, int count)
 886 {
 887     while (count--) {
 888         int a = src[(offset >> 16)];
 889         int b = src[(offset >> 16) + 1];
 890         *dst++  = a + ((b - a) * (offset & 0xFFFF) >> 16);
 891         offset += advance;
 892     }
 893 }
 894
 895 static av_always_inline void sprite_v_template(uint8_t *dst,
 896                                                const uint8_t *src1a,
 897                                                const uint8_t *src1b,
 898                                                int offset1,
 899                                                int two_sprites,
 900                                                const uint8_t *src2a,
 901                                                const uint8_t *src2b,
 902                                                int offset2,
 903                                                int alpha, int scaled,
 904                                                int width)
 905 {
 906     int a1, b1, a2, b2;
 907     while (width--) {
 908         a1 = *src1a++;
 909         if (scaled) {
 910             b1 = *src1b++;
 911             a1 = a1 + ((b1 - a1) * offset1 >> 16);
 912         }
 913         if (two_sprites) {
 914             a2 = *src2a++;
 915             if (scaled > 1) {
 916                 b2 = *src2b++;
 917                 a2 = a2 + ((b2 - a2) * offset2 >> 16);
 918             }
 919             a1 = a1 + ((a2 - a1) * alpha >> 16);
 920         }
 921         *dst++ = a1;
 922     }
 923 }
 924
 925 static void sprite_v_single_c(uint8_t *dst, const uint8_t *src1a,
 926                               const uint8_t *src1b,
 927                               int offset, int width)
 928 {
 929     sprite_v_template(dst, src1a, src1b, offset, 0, NULL, NULL, 0, 0, 1, width);
 930 }
 931
 932 static void sprite_v_double_noscale_c(uint8_t *dst, const uint8_t *src1a,
 933                                       const uint8_t *src2a,
 934                                       int alpha, int width)
 935 {
 936     sprite_v_template(dst, src1a, NULL, 0, 1, src2a, NULL, 0, alpha, 0, width);
 937 }
 938
 939 static void sprite_v_double_onescale_c(uint8_t *dst,
 940                                        const uint8_t *src1a,
 941                                        const uint8_t *src1b,
 942                                        int offset1,
 943                                        const uint8_t *src2a,
 944                                        int alpha, int width)
 945 {
 946     sprite_v_template(dst, src1a, src1b, offset1, 1, src2a, NULL, 0, alpha, 1,
 947                       width);
 948 }
 949
 950 static void sprite_v_double_twoscale_c(uint8_t *dst,
 951                                        const uint8_t *src1a,
 952                                        const uint8_t *src1b,
 953                                        int offset1,
 954                                        const uint8_t *src2a,
 955                                        const uint8_t *src2b,
 956                                        int offset2,
 957                                        int alpha,
 958                                        int width)
 959 {
 960     sprite_v_template(dst, src1a, src1b, offset1, 1, src2a, src2b, offset2,
 961                       alpha, 2, width);
 962 }
 963
 964 #endif /* CONFIG_WMV3IMAGE_DECODER || CONFIG_VC1IMAGE_DECODER */
 965 #define FN_ASSIGN(X, Y) \
 966     dsp->put_vc1_mspel_pixels_tab[1][X+4*Y] = put_vc1_mspel_mc##X##Y##_c; \
 967     dsp->put_vc1_mspel_pixels_tab[0][X+4*Y] = put_vc1_mspel_mc##X##Y##_16_c; \
 968     dsp->avg_vc1_mspel_pixels_tab[1][X+4*Y] = avg_vc1_mspel_mc##X##Y##_c; \
 969     dsp->avg_vc1_mspel_pixels_tab[0][X+4*Y] = avg_vc1_mspel_mc##X##Y##_16_c
 970
 971 av_cold void ff_vc1dsp_init(VC1DSPContext *dsp)
 972 {
 973     dsp->vc1_inv_trans_8x8    = vc1_inv_trans_8x8_c;
 974     dsp->vc1_inv_trans_4x8    = vc1_inv_trans_4x8_c;
 975     dsp->vc1_inv_trans_8x4    = vc1_inv_trans_8x4_c;
 976     dsp->vc1_inv_trans_4x4    = vc1_inv_trans_4x4_c;
 977     dsp->vc1_inv_trans_8x8_dc = vc1_inv_trans_8x8_dc_c;
 978     dsp->vc1_inv_trans_4x8_dc = vc1_inv_trans_4x8_dc_c;
 979     dsp->vc1_inv_trans_8x4_dc = vc1_inv_trans_8x4_dc_c;
 980     dsp->vc1_inv_trans_4x4_dc = vc1_inv_trans_4x4_dc_c;
 981
 982     dsp->vc1_h_overlap        = vc1_h_overlap_c;
 983     dsp->vc1_v_overlap        = vc1_v_overlap_c;
 984     dsp->vc1_h_s_overlap      = vc1_h_s_overlap_c;
 985     dsp->vc1_v_s_overlap      = vc1_v_s_overlap_c;
 986
 987     dsp->vc1_v_loop_filter4   = vc1_v_loop_filter4_c;
 988     dsp->vc1_h_loop_filter4   = vc1_h_loop_filter4_c;
 989     dsp->vc1_v_loop_filter8   = vc1_v_loop_filter8_c;
 990     dsp->vc1_h_loop_filter8   = vc1_h_loop_filter8_c;
 991     dsp->vc1_v_loop_filter16  = vc1_v_loop_filter16_c;
 992     dsp->vc1_h_loop_filter16  = vc1_h_loop_filter16_c;
 993
 994     dsp->put_vc1_mspel_pixels_tab[0][0] = put_pixels16x16_c;
 995     dsp->avg_vc1_mspel_pixels_tab[0][0] = avg_pixels16x16_c;
 996     dsp->put_vc1_mspel_pixels_tab[1][0] = put_pixels8x8_c;
 997     dsp->avg_vc1_mspel_pixels_tab[1][0] = avg_pixels8x8_c;
 998     FN_ASSIGN(0, 1);
 999     FN_ASSIGN(0, 2);
1000     FN_ASSIGN(0, 3);
1001
1002     FN_ASSIGN(1, 0);
1003     FN_ASSIGN(1, 1);
1004     FN_ASSIGN(1, 2);
1005     FN_ASSIGN(1, 3);
1006
1007     FN_ASSIGN(2, 0);
1008     FN_ASSIGN(2, 1);
1009     FN_ASSIGN(2, 2);
1010     FN_ASSIGN(2, 3);
1011
1012     FN_ASSIGN(3, 0);
1013     FN_ASSIGN(3, 1);
1014     FN_ASSIGN(3, 2);
1015     FN_ASSIGN(3, 3);
1016
1017     dsp->put_no_rnd_vc1_chroma_pixels_tab[0] = put_no_rnd_vc1_chroma_mc8_c;
1018     dsp->avg_no_rnd_vc1_chroma_pixels_tab[0] = avg_no_rnd_vc1_chroma_mc8_c;
1019     dsp->put_no_rnd_vc1_chroma_pixels_tab[1] = put_no_rnd_vc1_chroma_mc4_c;
1020     dsp->avg_no_rnd_vc1_chroma_pixels_tab[1] = avg_no_rnd_vc1_chroma_mc4_c;
1021
1022 #if CONFIG_WMV3IMAGE_DECODER || CONFIG_VC1IMAGE_DECODER
1023     dsp->sprite_h                 = sprite_h_c;
1024     dsp->sprite_v_single          = sprite_v_single_c;
1025     dsp->sprite_v_double_noscale  = sprite_v_double_noscale_c;
1026     dsp->sprite_v_double_onescale = sprite_v_double_onescale_c;
1027     dsp->sprite_v_double_twoscale = sprite_v_double_twoscale_c;
1028 #endif /* CONFIG_WMV3IMAGE_DECODER || CONFIG_VC1IMAGE_DECODER */
1029
1030     dsp->startcode_find_candidate = ff_startcode_find_candidate_c;
1031
1032     if (ARCH_AARCH64)
1033         ff_vc1dsp_init_aarch64(dsp);
1034     if (ARCH_ARM)
1035         ff_vc1dsp_init_arm(dsp);
1036     if (ARCH_PPC)
1037         ff_vc1dsp_init_ppc(dsp);
1038     if (ARCH_X86)
1039         ff_vc1dsp_init_x86(dsp);
1040     if (ARCH_MIPS)
1041         ff_vc1dsp_init_mips(dsp);
1042 }