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

   1 /*
   2  * Copyright (C) 2004-2010 Michael Niedermayer <michaelni@gmx.at>
   3  * Copyright (C) 2008 David Conrad
   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 #include "libavutil/attributes.h"
  23 #include "dsputil.h"
  24 #include "dwt.h"
  25 #include "libavcodec/x86/dwt.h"
  26
  27 int ff_slice_buffer_init(slice_buffer *buf, int line_count,
  28                          int max_allocated_lines, int line_width,
  29                          IDWTELEM *base_buffer)
  30 {
  31     int i;
  32
  33     buf->base_buffer = base_buffer;
  34     buf->line_count  = line_count;
  35     buf->line_width  = line_width;
  36     buf->data_count  = max_allocated_lines;
  37     buf->line        = av_mallocz(sizeof(IDWTELEM *) * line_count);
  38     if (!buf->line)
  39         return AVERROR(ENOMEM);
  40     buf->data_stack  = av_malloc(sizeof(IDWTELEM *) * max_allocated_lines);
  41     if (!buf->data_stack) {
  42         av_freep(&buf->line);
  43         return AVERROR(ENOMEM);
  44     }
  45
  46     for (i = 0; i < max_allocated_lines; i++) {
  47         buf->data_stack[i] = av_malloc(sizeof(IDWTELEM) * line_width);
  48         if (!buf->data_stack[i]) {
  49             for (i--; i >=0; i--)
  50                 av_freep(&buf->data_stack[i]);
  51             av_freep(&buf->data_stack);
  52             av_freep(&buf->line);
  53             return AVERROR(ENOMEM);
  54         }
  55     }
  56
  57     buf->data_stack_top = max_allocated_lines - 1;
  58     return 0;
  59 }
  60
  61 IDWTELEM *ff_slice_buffer_load_line(slice_buffer *buf, int line)
  62 {
  63     IDWTELEM *buffer;
  64
  65     assert(buf->data_stack_top >= 0);
  66 //  assert(!buf->line[line]);
  67     if (buf->line[line])
  68         return buf->line[line];
  69
  70     buffer = buf->data_stack[buf->data_stack_top];
  71     buf->data_stack_top--;
  72     buf->line[line] = buffer;
  73
  74     return buffer;
  75 }
  76
  77 void ff_slice_buffer_release(slice_buffer *buf, int line)
  78 {
  79     IDWTELEM *buffer;
  80
  81     assert(line >= 0 && line < buf->line_count);
  82     assert(buf->line[line]);
  83
  84     buffer = buf->line[line];
  85     buf->data_stack_top++;
  86     buf->data_stack[buf->data_stack_top] = buffer;
  87     buf->line[line]                      = NULL;
  88 }
  89
  90 void ff_slice_buffer_flush(slice_buffer *buf)
  91 {
  92     int i;
  93     for (i = 0; i < buf->line_count; i++)
  94         if (buf->line[i])
  95             ff_slice_buffer_release(buf, i);
  96 }
  97
  98 void ff_slice_buffer_destroy(slice_buffer *buf)
  99 {
 100     int i;
 101     ff_slice_buffer_flush(buf);
 102
 103     for (i = buf->data_count - 1; i >= 0; i--)
 104         av_freep(&buf->data_stack[i]);
 105     av_freep(&buf->data_stack);
 106     av_freep(&buf->line);
 107 }
 108
 109 static inline int mirror(int v, int m)
 110 {
 111     while ((unsigned)v > (unsigned)m) {
 112         v = -v;
 113         if (v < 0)
 114             v += 2 * m;
 115     }
 116     return v;
 117 }
 118
 119 static av_always_inline void lift(DWTELEM *dst, DWTELEM *src, DWTELEM *ref,
 120                                   int dst_step, int src_step, int ref_step,
 121                                   int width, int mul, int add, int shift,
 122                                   int highpass, int inverse)
 123 {
 124     const int mirror_left  = !highpass;
 125     const int mirror_right = (width & 1) ^ highpass;
 126     const int w            = (width >> 1) - 1 + (highpass & width);
 127     int i;
 128
 129 #define LIFT(src, ref, inv) ((src) + ((inv) ? -(ref) : +(ref)))
 130     if (mirror_left) {
 131         dst[0] = LIFT(src[0], ((mul * 2 * ref[0] + add) >> shift), inverse);
 132         dst   += dst_step;
 133         src   += src_step;
 134     }
 135
 136     for (i = 0; i < w; i++)
 137         dst[i * dst_step] = LIFT(src[i * src_step],
 138                                  ((mul * (ref[i * ref_step] +
 139                                           ref[(i + 1) * ref_step]) +
 140                                    add) >> shift),
 141                                  inverse);
 142
 143     if (mirror_right)
 144         dst[w * dst_step] = LIFT(src[w * src_step],
 145                                  ((mul * 2 * ref[w * ref_step] + add) >> shift),
 146                                  inverse);
 147 }
 148
 149 static av_always_inline void inv_lift(IDWTELEM *dst, IDWTELEM *src, IDWTELEM *ref,
 150                                       int dst_step, int src_step, int ref_step,
 151                                       int width, int mul, int add, int shift,
 152                                       int highpass, int inverse)
 153 {
 154     const int mirror_left  = !highpass;
 155     const int mirror_right = (width & 1) ^ highpass;
 156     const int w            = (width >> 1) - 1 + (highpass & width);
 157     int i;
 158
 159 #define LIFT(src, ref, inv) ((src) + ((inv) ? -(ref) : +(ref)))
 160     if (mirror_left) {
 161         dst[0] = LIFT(src[0], ((mul * 2 * ref[0] + add) >> shift), inverse);
 162         dst   += dst_step;
 163         src   += src_step;
 164     }
 165
 166     for (i = 0; i < w; i++)
 167         dst[i * dst_step] = LIFT(src[i * src_step],
 168                                  ((mul * (ref[i * ref_step] +
 169                                           ref[(i + 1) * ref_step]) +
 170                                    add) >> shift),
 171                                  inverse);
 172
 173     if (mirror_right) {
 174         dst[w * dst_step] = LIFT(src[w * src_step],
 175                                  ((mul * 2 * ref[w * ref_step] + add) >> shift),
 176                                  inverse);
 177     }
 178 }
 179
 180 #ifndef liftS
 181 static av_always_inline void liftS(DWTELEM *dst, DWTELEM *src, DWTELEM *ref,
 182                                    int dst_step, int src_step, int ref_step,
 183                                    int width, int mul, int add, int shift,
 184                                    int highpass, int inverse)
 185 {
 186     const int mirror_left  = !highpass;
 187     const int mirror_right = (width & 1) ^ highpass;
 188     const int w            = (width >> 1) - 1 + (highpass & width);
 189     int i;
 190
 191     assert(shift == 4);
 192 #define LIFTS(src, ref, inv)                                            \
 193     ((inv) ? (src) + (((ref) + 4 * (src)) >> shift)                     \
 194            : -((-16 * (src) + (ref) + add /                             \
 195                 4 + 1 + (5 << 25)) / (5 * 4) - (1 << 23)))
 196     if (mirror_left) {
 197         dst[0] = LIFTS(src[0], mul * 2 * ref[0] + add, inverse);
 198         dst   += dst_step;
 199         src   += src_step;
 200     }
 201
 202     for (i = 0; i < w; i++)
 203         dst[i * dst_step] = LIFTS(src[i * src_step],
 204                                   mul * (ref[i * ref_step] +
 205                                          ref[(i + 1) * ref_step]) + add,
 206                                   inverse);
 207
 208     if (mirror_right)
 209         dst[w * dst_step] = LIFTS(src[w * src_step],
 210                                   mul * 2 * ref[w * ref_step] + add,
 211                                   inverse);
 212 }
 213
 214 static av_always_inline void inv_liftS(IDWTELEM *dst, IDWTELEM *src,
 215                                        IDWTELEM *ref, int dst_step,
 216                                        int src_step, int ref_step,
 217                                        int width, int mul, int add, int shift,
 218                                        int highpass, int inverse)
 219 {
 220     const int mirror_left  = !highpass;
 221     const int mirror_right = (width & 1) ^ highpass;
 222     const int w            = (width >> 1) - 1 + (highpass & width);
 223     int i;
 224
 225     assert(shift == 4);
 226 #define LIFTS(src, ref, inv)                                            \
 227     ((inv) ? (src) + (((ref) + 4 * (src)) >> shift)                     \
 228            : -((-16 * (src) + (ref) + add /                             \
 229                 4 + 1 + (5 << 25)) / (5 * 4) - (1 << 23)))
 230     if (mirror_left) {
 231         dst[0] = LIFTS(src[0], mul * 2 * ref[0] + add, inverse);
 232         dst   += dst_step;
 233         src   += src_step;
 234     }
 235
 236     for (i = 0; i < w; i++)
 237         dst[i * dst_step] = LIFTS(src[i * src_step],
 238                                   mul * (ref[i * ref_step] +
 239                                          ref[(i + 1) * ref_step]) + add,
 240                                   inverse);
 241
 242     if (mirror_right)
 243         dst[w * dst_step] = LIFTS(src[w * src_step],
 244                                   mul * 2 * ref[w * ref_step] + add, inverse);
 245 }
 246 #endif /* ! liftS */
 247
 248 static void horizontal_decompose53i(DWTELEM *b, DWTELEM *temp, int width)
 249 {
 250     const int width2 = width >> 1;
 251     int x;
 252     const int w2 = (width + 1) >> 1;
 253
 254     for (x = 0; x < width2; x++) {
 255         temp[x]      = b[2 * x];
 256         temp[x + w2] = b[2 * x + 1];
 257     }
 258     if (width & 1)
 259         temp[x] = b[2 * x];
 260 #if 0
 261     {
 262         int A1, A2, A3, A4;
 263         A2             = temp[1];
 264         A4             = temp[0];
 265         A1             = temp[0 + width2];
 266         A1            -= (A2 + A4) >> 1;
 267         A4            += (A1 +  1) >> 1;
 268         b[0 + width2]  = A1;
 269         b[0]           = A4;
 270         for (x = 1; x + 1 < width2; x += 2) {
 271             A3                 = temp[x + width2];
 272             A4                 = temp[x + 1];
 273             A3                -= (A2 + A4)     >> 1;
 274             A2                += (A1 + A3 + 2) >> 2;
 275             b[x + width2]      = A3;
 276             b[x]               = A2;
 277
 278             A1                 = temp[x + 1 + width2];
 279             A2                 = temp[x + 2];
 280             A1                -= (A2 + A4)     >> 1;
 281             A4                += (A1 + A3 + 2) >> 2;
 282             b[x + 1 + width2]  = A1;
 283             b[x + 1]           = A4;
 284         }
 285         A3            = temp[width - 1];
 286         A3           -= A2;
 287         A2           += (A1 + A3 + 2) >> 2;
 288         b[width  - 1] = A3;
 289         b[width2 - 1] = A2;
 290     }
 291 #else
 292     lift(b + w2, temp + w2, temp,   1, 1, 1, width, -1, 0, 1, 1, 0);
 293     lift(b,      temp,      b + w2, 1, 1, 1, width,  1, 2, 2, 0, 0);
 294 #endif /* 0 */
 295 }
 296
 297 static void vertical_decompose53iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 298                                     int width)
 299 {
 300     int i;
 301
 302     for (i = 0; i < width; i++)
 303         b1[i] -= (b0[i] + b2[i]) >> 1;
 304 }
 305
 306 static void vertical_decompose53iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 307                                     int width)
 308 {
 309     int i;
 310
 311     for (i = 0; i < width; i++)
 312         b1[i] += (b0[i] + b2[i] + 2) >> 2;
 313 }
 314
 315 static void spatial_decompose53i(DWTELEM *buffer, DWTELEM *temp,
 316                                  int width, int height, int stride)
 317 {
 318     int y;
 319     DWTELEM *b0 = buffer + mirror(-2 - 1, height - 1) * stride;
 320     DWTELEM *b1 = buffer + mirror(-2,     height - 1) * stride;
 321
 322     for (y = -2; y < height; y += 2) {
 323         DWTELEM *b2 = buffer + mirror(y + 1, height - 1) * stride;
 324         DWTELEM *b3 = buffer + mirror(y + 2, height - 1) * stride;
 325
 326         if (y + 1 < (unsigned)height)
 327             horizontal_decompose53i(b2, temp, width);
 328         if (y + 2 < (unsigned)height)
 329             horizontal_decompose53i(b3, temp, width);
 330
 331         if (y + 1 < (unsigned)height)
 332             vertical_decompose53iH0(b1, b2, b3, width);
 333         if (y + 0 < (unsigned)height)
 334             vertical_decompose53iL0(b0, b1, b2, width);
 335
 336         b0 = b2;
 337         b1 = b3;
 338     }
 339 }
 340
 341 static void horizontal_decompose97i(DWTELEM *b, DWTELEM *temp, int width)
 342 {
 343     const int w2 = (width + 1) >> 1;
 344
 345     lift(temp + w2, b + 1, b,         1, 2, 2, width, W_AM, W_AO, W_AS, 1, 1);
 346     liftS(temp,     b,     temp + w2, 1, 2, 1, width, W_BM, W_BO, W_BS, 0, 0);
 347     lift(b + w2, temp + w2, temp,     1, 1, 1, width, W_CM, W_CO, W_CS, 1, 0);
 348     lift(b,      temp,      b + w2,   1, 1, 1, width, W_DM, W_DO, W_DS, 0, 0);
 349 }
 350
 351 static void vertical_decompose97iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 352                                     int width)
 353 {
 354     int i;
 355
 356     for (i = 0; i < width; i++)
 357         b1[i] -= (W_AM * (b0[i] + b2[i]) + W_AO) >> W_AS;
 358 }
 359
 360 static void vertical_decompose97iH1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 361                                     int width)
 362 {
 363     int i;
 364
 365     for (i = 0; i < width; i++)
 366         b1[i] += (W_CM * (b0[i] + b2[i]) + W_CO) >> W_CS;
 367 }
 368
 369 static void vertical_decompose97iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 370                                     int width)
 371 {
 372     int i;
 373
 374     for (i = 0; i < width; i++)
 375 #ifdef liftS
 376         b1[i] -= (W_BM * (b0[i] + b2[i]) + W_BO) >> W_BS;
 377 #else
 378         b1[i] = (16 * 4 * b1[i] - 4 * (b0[i] + b2[i]) + W_BO * 5 + (5 << 27)) /
 379                 (5 * 16) - (1 << 23);
 380 #endif
 381 }
 382
 383 static void vertical_decompose97iL1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 384                                     int width)
 385 {
 386     int i;
 387
 388     for (i = 0; i < width; i++)
 389         b1[i] += (W_DM * (b0[i] + b2[i]) + W_DO) >> W_DS;
 390 }
 391
 392 static void spatial_decompose97i(DWTELEM *buffer, DWTELEM *temp,
 393                                  int width, int height, int stride)
 394 {
 395     int y;
 396     DWTELEM *b0 = buffer + mirror(-4 - 1, height - 1) * stride;
 397     DWTELEM *b1 = buffer + mirror(-4,     height - 1) * stride;
 398     DWTELEM *b2 = buffer + mirror(-4 + 1, height - 1) * stride;
 399     DWTELEM *b3 = buffer + mirror(-4 + 2, height - 1) * stride;
 400
 401     for (y = -4; y < height; y += 2) {
 402         DWTELEM *b4 = buffer + mirror(y + 3, height - 1) * stride;
 403         DWTELEM *b5 = buffer + mirror(y + 4, height - 1) * stride;
 404
 405         if (y + 3 < (unsigned)height)
 406             horizontal_decompose97i(b4, temp, width);
 407         if (y + 4 < (unsigned)height)
 408             horizontal_decompose97i(b5, temp, width);
 409
 410         if (y + 3 < (unsigned)height)
 411             vertical_decompose97iH0(b3, b4, b5, width);
 412         if (y + 2 < (unsigned)height)
 413             vertical_decompose97iL0(b2, b3, b4, width);
 414         if (y + 1 < (unsigned)height)
 415             vertical_decompose97iH1(b1, b2, b3, width);
 416         if (y + 0 < (unsigned)height)
 417             vertical_decompose97iL1(b0, b1, b2, width);
 418
 419         b0 = b2;
 420         b1 = b3;
 421         b2 = b4;
 422         b3 = b5;
 423     }
 424 }
 425
 426 void ff_spatial_dwt(DWTELEM *buffer, DWTELEM *temp, int width, int height,
 427                     int stride, int type, int decomposition_count)
 428 {
 429     int level;
 430
 431     for (level = 0; level < decomposition_count; level++) {
 432         switch (type) {
 433         case DWT_97:
 434             spatial_decompose97i(buffer, temp,
 435                                  width >> level, height >> level,
 436                                  stride << level);
 437             break;
 438         case DWT_53:
 439             spatial_decompose53i(buffer, temp,
 440                                  width >> level, height >> level,
 441                                  stride << level);
 442             break;
 443         }
 444     }
 445 }
 446
 447 static void horizontal_compose53i(IDWTELEM *b, IDWTELEM *temp, int width)
 448 {
 449     const int width2 = width >> 1;
 450     const int w2     = (width + 1) >> 1;
 451     int x;
 452
 453     for (x = 0; x < width2; x++) {
 454         temp[2 * x]     = b[x];
 455         temp[2 * x + 1] = b[x + w2];
 456     }
 457     if (width & 1)
 458         temp[2 * x] = b[x];
 459
 460     b[0] = temp[0] - ((temp[1] + 1) >> 1);
 461     for (x = 2; x < width - 1; x += 2) {
 462         b[x]     = temp[x]     - ((temp[x - 1] + temp[x + 1] + 2) >> 2);
 463         b[x - 1] = temp[x - 1] + ((b[x - 2]    + b[x]        + 1) >> 1);
 464     }
 465     if (width & 1) {
 466         b[x]     = temp[x]     - ((temp[x - 1]     + 1) >> 1);
 467         b[x - 1] = temp[x - 1] + ((b[x - 2] + b[x] + 1) >> 1);
 468     } else
 469         b[x - 1] = temp[x - 1] + b[x - 2];
 470 }
 471
 472 static void vertical_compose53iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 473                                   int width)
 474 {
 475     int i;
 476
 477     for (i = 0; i < width; i++)
 478         b1[i] += (b0[i] + b2[i]) >> 1;
 479 }
 480
 481 static void vertical_compose53iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 482                                   int width)
 483 {
 484     int i;
 485
 486     for (i = 0; i < width; i++)
 487         b1[i] -= (b0[i] + b2[i] + 2) >> 2;
 488 }
 489
 490 static void spatial_compose53i_buffered_init(DWTCompose *cs, slice_buffer *sb,
 491                                              int height, int stride_line)
 492 {
 493     cs->b0 = slice_buffer_get_line(sb,
 494                                    mirror(-1 - 1, height - 1) * stride_line);
 495     cs->b1 = slice_buffer_get_line(sb, mirror(-1, height - 1) * stride_line);
 496     cs->y  = -1;
 497 }
 498
 499 static void spatial_compose53i_init(DWTCompose *cs, IDWTELEM *buffer,
 500                                     int height, int stride)
 501 {
 502     cs->b0 = buffer + mirror(-1 - 1, height - 1) * stride;
 503     cs->b1 = buffer + mirror(-1,     height - 1) * stride;
 504     cs->y  = -1;
 505 }
 506
 507 static void spatial_compose53i_dy_buffered(DWTCompose *cs, slice_buffer *sb,
 508                                            IDWTELEM *temp,
 509                                            int width, int height,
 510                                            int stride_line)
 511 {
 512     int y = cs->y;
 513
 514     IDWTELEM *b0 = cs->b0;
 515     IDWTELEM *b1 = cs->b1;
 516     IDWTELEM *b2 = slice_buffer_get_line(sb,
 517                                          mirror(y + 1, height - 1) *
 518                                          stride_line);
 519     IDWTELEM *b3 = slice_buffer_get_line(sb,
 520                                          mirror(y + 2, height - 1) *
 521                                          stride_line);
 522
 523     if (y + 1 < (unsigned)height && y < (unsigned)height) {
 524         int x;
 525
 526         for (x = 0; x < width; x++) {
 527             b2[x] -= (b1[x] + b3[x] + 2) >> 2;
 528             b1[x] += (b0[x] + b2[x])     >> 1;
 529         }
 530     } else {
 531         if (y + 1 < (unsigned)height)
 532             vertical_compose53iL0(b1, b2, b3, width);
 533         if (y + 0 < (unsigned)height)
 534             vertical_compose53iH0(b0, b1, b2, width);
 535     }
 536
 537     if (y - 1 < (unsigned)height)
 538         horizontal_compose53i(b0, temp, width);
 539     if (y + 0 < (unsigned)height)
 540         horizontal_compose53i(b1, temp, width);
 541
 542     cs->b0  = b2;
 543     cs->b1  = b3;
 544     cs->y  += 2;
 545 }
 546
 547 static void spatial_compose53i_dy(DWTCompose *cs, IDWTELEM *buffer,
 548                                   IDWTELEM *temp, int width, int height,
 549                                   int stride)
 550 {
 551     int y        = cs->y;
 552     IDWTELEM *b0 = cs->b0;
 553     IDWTELEM *b1 = cs->b1;
 554     IDWTELEM *b2 = buffer + mirror(y + 1, height - 1) * stride;
 555     IDWTELEM *b3 = buffer + mirror(y + 2, height - 1) * stride;
 556
 557     if (y + 1 < (unsigned)height)
 558         vertical_compose53iL0(b1, b2, b3, width);
 559     if (y + 0 < (unsigned)height)
 560         vertical_compose53iH0(b0, b1, b2, width);
 561
 562     if (y - 1 < (unsigned)height)
 563         horizontal_compose53i(b0, temp, width);
 564     if (y + 0 < (unsigned)height)
 565         horizontal_compose53i(b1, temp, width);
 566
 567     cs->b0  = b2;
 568     cs->b1  = b3;
 569     cs->y  += 2;
 570 }
 571
 572 static void av_unused spatial_compose53i(IDWTELEM *buffer, IDWTELEM *temp,
 573                                          int width, int height, int stride)
 574 {
 575     DWTCompose cs;
 576     spatial_compose53i_init(&cs, buffer, height, stride);
 577     while (cs.y <= height)
 578         spatial_compose53i_dy(&cs, buffer, temp, width, height, stride);
 579 }
 580
 581 void ff_snow_horizontal_compose97i(IDWTELEM *b, IDWTELEM *temp, int width)
 582 {
 583     const int w2 = (width + 1) >> 1;
 584
 585 #if 0 //maybe more understadable but slower
 586     inv_lift(temp,     b,      b + w2, 2, 1, 1, width, W_DM, W_DO, W_DS, 0, 1);
 587     inv_lift(temp + 1, b + w2, temp,   2, 1, 2, width, W_CM, W_CO, W_CS, 1, 1);
 588
 589     inv_liftS(b,    temp,     temp + 1, 2, 2, 2, width, W_BM, W_BO, W_BS, 0, 1);
 590     inv_lift(b + 1, temp + 1, b,        2, 2, 2, width, W_AM, W_AO, W_AS, 1, 0);
 591 #else
 592     int x;
 593     temp[0] = b[0] - ((3 * b[w2] + 2) >> 2);
 594     for (x = 1; x < (width >> 1); x++) {
 595         temp[2 * x]     = b[x] - ((3 * (b[x + w2 - 1] + b[x + w2]) + 4) >> 3);
 596         temp[2 * x - 1] = b[x + w2 - 1] - temp[2 * x - 2] - temp[2 * x];
 597     }
 598     if (width & 1) {
 599         temp[2 * x]     = b[x] - ((3 * b[x + w2 - 1] + 2) >> 2);
 600         temp[2 * x - 1] = b[x + w2 - 1] - temp[2 * x - 2] - temp[2 * x];
 601     } else
 602         temp[2 * x - 1] = b[x + w2 - 1] - 2 * temp[2 * x - 2];
 603
 604     b[0] = temp[0] + ((2 * temp[0] + temp[1] + 4) >> 3);
 605     for (x = 2; x < width - 1; x += 2) {
 606         b[x]     = temp[x] + ((4 * temp[x] + temp[x - 1] + temp[x + 1] + 8) >> 4);
 607         b[x - 1] = temp[x - 1] + ((3 * (b[x - 2] + b[x])) >> 1);
 608     }
 609     if (width & 1) {
 610         b[x]     = temp[x] + ((2 * temp[x] + temp[x - 1] + 4) >> 3);
 611         b[x - 1] = temp[x - 1] + ((3 * (b[x - 2] + b[x])) >> 1);
 612     } else
 613         b[x - 1] = temp[x - 1] + 3 * b[x - 2];
 614 #endif
 615 }
 616
 617 static void vertical_compose97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 618                                   int width)
 619 {
 620     int i;
 621
 622     for (i = 0; i < width; i++)
 623         b1[i] += (W_AM * (b0[i] + b2[i]) + W_AO) >> W_AS;
 624 }
 625
 626 static void vertical_compose97iH1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 627                                   int width)
 628 {
 629     int i;
 630
 631     for (i = 0; i < width; i++)
 632         b1[i] -= (W_CM * (b0[i] + b2[i]) + W_CO) >> W_CS;
 633 }
 634
 635 static void vertical_compose97iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 636                                   int width)
 637 {
 638     int i;
 639
 640     for (i = 0; i < width; i++)
 641 #ifdef liftS
 642         b1[i] += (W_BM * (b0[i] + b2[i]) + W_BO) >> W_BS;
 643 #else
 644         b1[i] += (W_BM * (b0[i] + b2[i]) + 4 * b1[i] + W_BO) >> W_BS;
 645 #endif
 646 }
 647
 648 static void vertical_compose97iL1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 649                                   int width)
 650 {
 651     int i;
 652
 653     for (i = 0; i < width; i++)
 654         b1[i] -= (W_DM * (b0[i] + b2[i]) + W_DO) >> W_DS;
 655 }
 656
 657 void ff_snow_vertical_compose97i(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 658                                  IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5,
 659                                  int width)
 660 {
 661     int i;
 662
 663     for (i = 0; i < width; i++) {
 664         b4[i] -= (W_DM * (b3[i] + b5[i]) + W_DO) >> W_DS;
 665         b3[i] -= (W_CM * (b2[i] + b4[i]) + W_CO) >> W_CS;
 666 #ifdef liftS
 667         b2[i] += (W_BM * (b1[i] + b3[i]) + W_BO) >> W_BS;
 668 #else
 669         b2[i] += (W_BM * (b1[i] + b3[i]) + 4 * b2[i] + W_BO) >> W_BS;
 670 #endif
 671         b1[i] += (W_AM * (b0[i] + b2[i]) + W_AO) >> W_AS;
 672     }
 673 }
 674
 675 static void spatial_compose97i_buffered_init(DWTCompose *cs, slice_buffer *sb,
 676                                              int height, int stride_line)
 677 {
 678     cs->b0 = slice_buffer_get_line(sb, mirror(-3 - 1, height - 1) * stride_line);
 679     cs->b1 = slice_buffer_get_line(sb, mirror(-3,     height - 1) * stride_line);
 680     cs->b2 = slice_buffer_get_line(sb, mirror(-3 + 1, height - 1) * stride_line);
 681     cs->b3 = slice_buffer_get_line(sb, mirror(-3 + 2, height - 1) * stride_line);
 682     cs->y  = -3;
 683 }
 684
 685 static void spatial_compose97i_init(DWTCompose *cs, IDWTELEM *buffer, int height,
 686                                     int stride)
 687 {
 688     cs->b0 = buffer + mirror(-3 - 1, height - 1) * stride;
 689     cs->b1 = buffer + mirror(-3,     height - 1) * stride;
 690     cs->b2 = buffer + mirror(-3 + 1, height - 1) * stride;
 691     cs->b3 = buffer + mirror(-3 + 2, height - 1) * stride;
 692     cs->y  = -3;
 693 }
 694
 695 static void spatial_compose97i_dy_buffered(DWTContext *dsp, DWTCompose *cs,
 696                                            slice_buffer * sb, IDWTELEM *temp,
 697                                            int width, int height,
 698                                            int stride_line)
 699 {
 700     int y = cs->y;
 701
 702     IDWTELEM *b0 = cs->b0;
 703     IDWTELEM *b1 = cs->b1;
 704     IDWTELEM *b2 = cs->b2;
 705     IDWTELEM *b3 = cs->b3;
 706     IDWTELEM *b4 = slice_buffer_get_line(sb,
 707                                          mirror(y + 3, height - 1) *
 708                                          stride_line);
 709     IDWTELEM *b5 = slice_buffer_get_line(sb,
 710                                          mirror(y + 4, height - 1) *
 711                                          stride_line);
 712
 713     if (y > 0 && y + 4 < height) {
 714         dsp->vertical_compose97i(b0, b1, b2, b3, b4, b5, width);
 715     } else {
 716         if (y + 3 < (unsigned)height)
 717             vertical_compose97iL1(b3, b4, b5, width);
 718         if (y + 2 < (unsigned)height)
 719             vertical_compose97iH1(b2, b3, b4, width);
 720         if (y + 1 < (unsigned)height)
 721             vertical_compose97iL0(b1, b2, b3, width);
 722         if (y + 0 < (unsigned)height)
 723             vertical_compose97iH0(b0, b1, b2, width);
 724     }
 725
 726     if (y - 1 < (unsigned)height)
 727         dsp->horizontal_compose97i(b0, temp, width);
 728     if (y + 0 < (unsigned)height)
 729         dsp->horizontal_compose97i(b1, temp, width);
 730
 731     cs->b0  = b2;
 732     cs->b1  = b3;
 733     cs->b2  = b4;
 734     cs->b3  = b5;
 735     cs->y  += 2;
 736 }
 737
 738 static void spatial_compose97i_dy(DWTCompose *cs, IDWTELEM *buffer,
 739                                   IDWTELEM *temp, int width, int height,
 740                                   int stride)
 741 {
 742     int y        = cs->y;
 743     IDWTELEM *b0 = cs->b0;
 744     IDWTELEM *b1 = cs->b1;
 745     IDWTELEM *b2 = cs->b2;
 746     IDWTELEM *b3 = cs->b3;
 747     IDWTELEM *b4 = buffer + mirror(y + 3, height - 1) * stride;
 748     IDWTELEM *b5 = buffer + mirror(y + 4, height - 1) * stride;
 749
 750     if (y + 3 < (unsigned)height)
 751         vertical_compose97iL1(b3, b4, b5, width);
 752     if (y + 2 < (unsigned)height)
 753         vertical_compose97iH1(b2, b3, b4, width);
 754     if (y + 1 < (unsigned)height)
 755         vertical_compose97iL0(b1, b2, b3, width);
 756     if (y + 0 < (unsigned)height)
 757         vertical_compose97iH0(b0, b1, b2, width);
 758
 759     if (y - 1 < (unsigned)height)
 760         ff_snow_horizontal_compose97i(b0, temp, width);
 761     if (y + 0 < (unsigned)height)
 762         ff_snow_horizontal_compose97i(b1, temp, width);
 763
 764     cs->b0  = b2;
 765     cs->b1  = b3;
 766     cs->b2  = b4;
 767     cs->b3  = b5;
 768     cs->y  += 2;
 769 }
 770
 771 static void av_unused spatial_compose97i(IDWTELEM *buffer, IDWTELEM *temp,
 772                                          int width, int height, int stride)
 773 {
 774     DWTCompose cs;
 775     spatial_compose97i_init(&cs, buffer, height, stride);
 776     while (cs.y <= height)
 777         spatial_compose97i_dy(&cs, buffer, temp, width, height, stride);
 778 }
 779
 780 void ff_spatial_idwt_buffered_init(DWTCompose *cs, slice_buffer *sb, int width,
 781                                    int height, int stride_line, int type,
 782                                    int decomposition_count)
 783 {
 784     int level;
 785     for (level = decomposition_count - 1; level >= 0; level--) {
 786         switch (type) {
 787         case DWT_97:
 788             spatial_compose97i_buffered_init(cs + level, sb, height >> level,
 789                                              stride_line << level);
 790             break;
 791         case DWT_53:
 792             spatial_compose53i_buffered_init(cs + level, sb, height >> level,
 793                                              stride_line << level);
 794             break;
 795         }
 796     }
 797 }
 798
 799 void ff_spatial_idwt_buffered_slice(DWTContext *dsp, DWTCompose *cs,
 800                                     slice_buffer *slice_buf, IDWTELEM *temp,
 801                                     int width, int height, int stride_line,
 802                                     int type, int decomposition_count, int y)
 803 {
 804     const int support = type == 1 ? 3 : 5;
 805     int level;
 806     if (type == 2)
 807         return;
 808
 809     for (level = decomposition_count - 1; level >= 0; level--)
 810         while (cs[level].y <= FFMIN((y >> level) + support, height >> level)) {
 811             switch (type) {
 812             case DWT_97:
 813                 spatial_compose97i_dy_buffered(dsp, cs + level, slice_buf, temp,
 814                                                width >> level,
 815                                                height >> level,
 816                                                stride_line << level);
 817                 break;
 818             case DWT_53:
 819                 spatial_compose53i_dy_buffered(cs + level, slice_buf, temp,
 820                                                width >> level,
 821                                                height >> level,
 822                                                stride_line << level);
 823                 break;
 824             }
 825         }
 826 }
 827
 828 static void ff_spatial_idwt_init(DWTCompose *cs, IDWTELEM *buffer, int width,
 829                                  int height, int stride, int type,
 830                                  int decomposition_count)
 831 {
 832     int level;
 833     for (level = decomposition_count - 1; level >= 0; level--) {
 834         switch (type) {
 835         case DWT_97:
 836             spatial_compose97i_init(cs + level, buffer, height >> level,
 837                                     stride << level);
 838             break;
 839         case DWT_53:
 840             spatial_compose53i_init(cs + level, buffer, height >> level,
 841                                     stride << level);
 842             break;
 843         }
 844     }
 845 }
 846
 847 static void ff_spatial_idwt_slice(DWTCompose *cs, IDWTELEM *buffer,
 848                                   IDWTELEM *temp, int width, int height,
 849                                   int stride, int type,
 850                                   int decomposition_count, int y)
 851 {
 852     const int support = type == 1 ? 3 : 5;
 853     int level;
 854     if (type == 2)
 855         return;
 856
 857     for (level = decomposition_count - 1; level >= 0; level--)
 858         while (cs[level].y <= FFMIN((y >> level) + support, height >> level)) {
 859             switch (type) {
 860             case DWT_97:
 861                 spatial_compose97i_dy(cs + level, buffer, temp, width >> level,
 862                                       height >> level, stride << level);
 863                 break;
 864             case DWT_53:
 865                 spatial_compose53i_dy(cs + level, buffer, temp, width >> level,
 866                                       height >> level, stride << level);
 867                 break;
 868             }
 869         }
 870 }
 871
 872 void ff_spatial_idwt(IDWTELEM *buffer, IDWTELEM *temp, int width, int height,
 873                      int stride, int type, int decomposition_count)
 874 {
 875     DWTCompose cs[MAX_DECOMPOSITIONS];
 876     int y;
 877     ff_spatial_idwt_init(cs, buffer, width, height, stride, type,
 878                          decomposition_count);
 879     for (y = 0; y < height; y += 4)
 880         ff_spatial_idwt_slice(cs, buffer, temp, width, height, stride, type,
 881                               decomposition_count, y);
 882 }
 883
 884 static inline int w_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size,
 885                       int w, int h, int type)
 886 {
 887     int s, i, j;
 888     const int dec_count = w == 8 ? 3 : 4;
 889     int tmp[32 * 32], tmp2[32];
 890     int level, ori;
 891     static const int scale[2][2][4][4] = {
 892         {
 893             { // 9/7 8x8 dec=3
 894                 { 268, 239, 239, 213 },
 895                 { 0,   224, 224, 152 },
 896                 { 0,   135, 135, 110 },
 897             },
 898             { // 9/7 16x16 or 32x32 dec=4
 899                 { 344, 310, 310, 280 },
 900                 { 0,   320, 320, 228 },
 901                 { 0,   175, 175, 136 },
 902                 { 0,   129, 129, 102 },
 903             }
 904         },
 905         {
 906             { // 5/3 8x8 dec=3
 907                 { 275, 245, 245, 218 },
 908                 { 0,   230, 230, 156 },
 909                 { 0,   138, 138, 113 },
 910             },
 911             { // 5/3 16x16 or 32x32 dec=4
 912                 { 352, 317, 317, 286 },
 913                 { 0,   328, 328, 233 },
 914                 { 0,   180, 180, 140 },
 915                 { 0,   132, 132, 105 },
 916             }
 917         }
 918     };
 919
 920     for (i = 0; i < h; i++) {
 921         for (j = 0; j < w; j += 4) {
 922             tmp[32 * i + j + 0] = (pix1[j + 0] - pix2[j + 0]) << 4;
 923             tmp[32 * i + j + 1] = (pix1[j + 1] - pix2[j + 1]) << 4;
 924             tmp[32 * i + j + 2] = (pix1[j + 2] - pix2[j + 2]) << 4;
 925             tmp[32 * i + j + 3] = (pix1[j + 3] - pix2[j + 3]) << 4;
 926         }
 927         pix1 += line_size;
 928         pix2 += line_size;
 929     }
 930
 931     ff_spatial_dwt(tmp, tmp2, w, h, 32, type, dec_count);
 932
 933     s = 0;
 934     assert(w == h);
 935     for (level = 0; level < dec_count; level++)
 936         for (ori = level ? 1 : 0; ori < 4; ori++) {
 937             int size   = w >> (dec_count - level);
 938             int sx     = (ori & 1) ? size : 0;
 939             int stride = 32 << (dec_count - level);
 940             int sy     = (ori & 2) ? stride >> 1 : 0;
 941
 942             for (i = 0; i < size; i++)
 943                 for (j = 0; j < size; j++) {
 944                     int v = tmp[sx + sy + i * stride + j] *
 945                             scale[type][dec_count - 3][level][ori];
 946                     s += FFABS(v);
 947                 }
 948         }
 949     assert(s >= 0);
 950     return s >> 9;
 951 }
 952
 953 static int w53_8_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 954 {
 955     return w_c(v, pix1, pix2, line_size, 8, h, 1);
 956 }
 957
 958 static int w97_8_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 959 {
 960     return w_c(v, pix1, pix2, line_size, 8, h, 0);
 961 }
 962
 963 static int w53_16_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 964 {
 965     return w_c(v, pix1, pix2, line_size, 16, h, 1);
 966 }
 967
 968 static int w97_16_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 969 {
 970     return w_c(v, pix1, pix2, line_size, 16, h, 0);
 971 }
 972
 973 int ff_w53_32_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 974 {
 975     return w_c(v, pix1, pix2, line_size, 32, h, 1);
 976 }
 977
 978 int ff_w97_32_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 979 {
 980     return w_c(v, pix1, pix2, line_size, 32, h, 0);
 981 }
 982
 983 void ff_dsputil_init_dwt(DSPContext *c)
 984 {
 985     c->w53[0] = w53_16_c;
 986     c->w53[1] = w53_8_c;
 987     c->w97[0] = w97_16_c;
 988     c->w97[1] = w97_8_c;
 989 }
 990
 991 void ff_dwt_init(DWTContext *c)
 992 {
 993     c->vertical_compose97i   = ff_snow_vertical_compose97i;
 994     c->horizontal_compose97i = ff_snow_horizontal_compose97i;
 995     c->inner_add_yblock      = ff_snow_inner_add_yblock;
 996
 997     if (HAVE_MMX)
 998         ff_dwt_init_x86(c);
 999 }
1000
1001
1002 static av_always_inline
1003 void interleave(IDWTELEM *dst, IDWTELEM *src0, IDWTELEM *src1, int w2, int add, int shift)
1004 {
1005     int i;
1006     for (i = 0; i < w2; i++) {
1007         dst[2*i  ] = (src0[i] + add) >> shift;
1008         dst[2*i+1] = (src1[i] + add) >> shift;
1009     }
1010 }
1011
1012 static void horizontal_compose_dirac53i(IDWTELEM *b, IDWTELEM *temp, int w)
1013 {
1014     const int w2 = w >> 1;
1015     int x;
1016
1017     temp[0] = COMPOSE_53iL0(b[w2], b[0], b[w2]);
1018     for (x = 1; x < w2; x++) {
1019         temp[x     ] = COMPOSE_53iL0     (b[x+w2-1], b[x     ], b[x+w2]);
1020         temp[x+w2-1] = COMPOSE_DIRAC53iH0(temp[x-1], b[x+w2-1], temp[x]);
1021     }
1022     temp[w-1] = COMPOSE_DIRAC53iH0(temp[w2-1], b[w-1], temp[w2-1]);
1023
1024     interleave(b, temp, temp+w2, w2, 1, 1);
1025 }
1026
1027 static void horizontal_compose_dd97i(IDWTELEM *b, IDWTELEM *tmp, int w)
1028 {
1029     const int w2 = w >> 1;
1030     int x;
1031
1032     tmp[0] = COMPOSE_53iL0(b[w2], b[0], b[w2]);
1033     for (x = 1; x < w2; x++)
1034         tmp[x] = COMPOSE_53iL0(b[x+w2-1], b[x], b[x+w2]);
1035
1036     // extend the edges
1037     tmp[-1]   = tmp[0];
1038     tmp[w2+1] = tmp[w2] = tmp[w2-1];
1039
1040     for (x = 0; x < w2; x++) {
1041         b[2*x  ] = (tmp[x] + 1)>>1;
1042         b[2*x+1] = (COMPOSE_DD97iH0(tmp[x-1], tmp[x], b[x+w2], tmp[x+1], tmp[x+2]) + 1)>>1;
1043     }
1044 }
1045
1046 static void horizontal_compose_dd137i(IDWTELEM *b, IDWTELEM *tmp, int w)
1047 {
1048     const int w2 = w >> 1;
1049     int x;
1050
1051     tmp[0] = COMPOSE_DD137iL0(b[w2], b[w2], b[0], b[w2  ], b[w2+1]);
1052     tmp[1] = COMPOSE_DD137iL0(b[w2], b[w2], b[1], b[w2+1], b[w2+2]);
1053     for (x = 2; x < w2-1; x++)
1054         tmp[x] = COMPOSE_DD137iL0(b[x+w2-2], b[x+w2-1], b[x], b[x+w2], b[x+w2+1]);
1055     tmp[w2-1] = COMPOSE_DD137iL0(b[w-3], b[w-2], b[w2-1], b[w-1], b[w-1]);
1056
1057     // extend the edges
1058     tmp[-1]   = tmp[0];
1059     tmp[w2+1] = tmp[w2] = tmp[w2-1];
1060
1061     for (x = 0; x < w2; x++) {
1062         b[2*x  ] = (tmp[x] + 1)>>1;
1063         b[2*x+1] = (COMPOSE_DD97iH0(tmp[x-1], tmp[x], b[x+w2], tmp[x+1], tmp[x+2]) + 1)>>1;
1064     }
1065 }
1066
1067 static av_always_inline
1068 void horizontal_compose_haari(IDWTELEM *b, IDWTELEM *temp, int w, int shift)
1069 {
1070     const int w2 = w >> 1;
1071     int x;
1072
1073     for (x = 0; x < w2; x++) {
1074         temp[x   ] = COMPOSE_HAARiL0(b[x   ], b[x+w2]);
1075         temp[x+w2] = COMPOSE_HAARiH0(b[x+w2], temp[x]);
1076     }
1077
1078     interleave(b, temp, temp+w2, w2, shift, shift);
1079 }
1080
1081 static void horizontal_compose_haar0i(IDWTELEM *b, IDWTELEM *temp, int w)
1082 {
1083     horizontal_compose_haari(b, temp, w, 0);
1084 }
1085
1086 static void horizontal_compose_haar1i(IDWTELEM *b, IDWTELEM *temp, int w)
1087 {
1088     horizontal_compose_haari(b, temp, w, 1);
1089 }
1090
1091 static void horizontal_compose_fidelityi(IDWTELEM *b, IDWTELEM *tmp, int w)
1092 {
1093     const int w2 = w >> 1;
1094     int i, x;
1095     IDWTELEM v[8];
1096
1097     for (x = 0; x < w2; x++) {
1098         for (i = 0; i < 8; i++)
1099             v[i] = b[av_clip(x-3+i, 0, w2-1)];
1100         tmp[x] = COMPOSE_FIDELITYiH0(v[0], v[1], v[2], v[3], b[x+w2], v[4], v[5], v[6], v[7]);
1101     }
1102
1103     for (x = 0; x < w2; x++) {
1104         for (i = 0; i < 8; i++)
1105             v[i] = tmp[av_clip(x-4+i, 0, w2-1)];
1106         tmp[x+w2] = COMPOSE_FIDELITYiL0(v[0], v[1], v[2], v[3], b[x], v[4], v[5], v[6], v[7]);
1107     }
1108
1109     interleave(b, tmp+w2, tmp, w2, 0, 0);
1110 }
1111
1112 static void horizontal_compose_daub97i(IDWTELEM *b, IDWTELEM *temp, int w)
1113 {
1114     const int w2 = w >> 1;
1115     int x, b0, b1, b2;
1116
1117     temp[0] = COMPOSE_DAUB97iL1(b[w2], b[0], b[w2]);
1118     for (x = 1; x < w2; x++) {
1119         temp[x     ] = COMPOSE_DAUB97iL1(b[x+w2-1], b[x     ], b[x+w2]);
1120         temp[x+w2-1] = COMPOSE_DAUB97iH1(temp[x-1], b[x+w2-1], temp[x]);
1121     }
1122     temp[w-1] = COMPOSE_DAUB97iH1(temp[w2-1], b[w-1], temp[w2-1]);
1123
1124     // second stage combined with interleave and shift
1125     b0 = b2 = COMPOSE_DAUB97iL0(temp[w2], temp[0], temp[w2]);
1126     b[0] = (b0 + 1) >> 1;
1127     for (x = 1; x < w2; x++) {
1128         b2 = COMPOSE_DAUB97iL0(temp[x+w2-1], temp[x     ], temp[x+w2]);
1129         b1 = COMPOSE_DAUB97iH0(          b0, temp[x+w2-1], b2        );
1130         b[2*x-1] = (b1 + 1) >> 1;
1131         b[2*x  ] = (b2 + 1) >> 1;
1132         b0 = b2;
1133     }
1134     b[w-1] = (COMPOSE_DAUB97iH0(b2, temp[w-1], b2) + 1) >> 1;
1135 }
1136
1137 static void vertical_compose_dirac53iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1138 {
1139     int i;
1140
1141     for(i=0; i<width; i++){
1142         b1[i] = COMPOSE_DIRAC53iH0(b0[i], b1[i], b2[i]);
1143     }
1144 }
1145
1146 static void vertical_compose_dd97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
1147                                   IDWTELEM *b3, IDWTELEM *b4, int width)
1148 {
1149     int i;
1150
1151     for(i=0; i<width; i++){
1152         b2[i] = COMPOSE_DD97iH0(b0[i], b1[i], b2[i], b3[i], b4[i]);
1153     }
1154 }
1155
1156 static void vertical_compose_dd137iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
1157                                       IDWTELEM *b3, IDWTELEM *b4, int width)
1158 {
1159     int i;
1160
1161     for(i=0; i<width; i++){
1162         b2[i] = COMPOSE_DD137iL0(b0[i], b1[i], b2[i], b3[i], b4[i]);
1163     }
1164 }
1165
1166 static void vertical_compose_haar(IDWTELEM *b0, IDWTELEM *b1, int width)
1167 {
1168     int i;
1169
1170     for (i = 0; i < width; i++) {
1171         b0[i] = COMPOSE_HAARiL0(b0[i], b1[i]);
1172         b1[i] = COMPOSE_HAARiH0(b1[i], b0[i]);
1173     }
1174 }
1175
1176 static void vertical_compose_fidelityiH0(IDWTELEM *dst, IDWTELEM *b[8], int width)
1177 {
1178     int i;
1179
1180     for(i=0; i<width; i++){
1181         dst[i] = COMPOSE_FIDELITYiH0(b[0][i], b[1][i], b[2][i], b[3][i], dst[i], b[4][i], b[5][i], b[6][i], b[7][i]);
1182     }
1183 }
1184
1185 static void vertical_compose_fidelityiL0(IDWTELEM *dst, IDWTELEM *b[8], int width)
1186 {
1187     int i;
1188
1189     for(i=0; i<width; i++){
1190         dst[i] = COMPOSE_FIDELITYiL0(b[0][i], b[1][i], b[2][i], b[3][i], dst[i], b[4][i], b[5][i], b[6][i], b[7][i]);
1191     }
1192 }
1193
1194 static void vertical_compose_daub97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1195 {
1196     int i;
1197
1198     for(i=0; i<width; i++){
1199         b1[i] = COMPOSE_DAUB97iH0(b0[i], b1[i], b2[i]);
1200     }
1201 }
1202
1203 static void vertical_compose_daub97iH1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1204 {
1205     int i;
1206
1207     for(i=0; i<width; i++){
1208         b1[i] = COMPOSE_DAUB97iH1(b0[i], b1[i], b2[i]);
1209     }
1210 }
1211
1212 static void vertical_compose_daub97iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1213 {
1214     int i;
1215
1216     for(i=0; i<width; i++){
1217         b1[i] = COMPOSE_DAUB97iL0(b0[i], b1[i], b2[i]);
1218     }
1219 }
1220
1221 static void vertical_compose_daub97iL1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1222 {
1223     int i;
1224
1225     for(i=0; i<width; i++){
1226         b1[i] = COMPOSE_DAUB97iL1(b0[i], b1[i], b2[i]);
1227     }
1228 }
1229
1230
1231 static void spatial_compose_dd97i_dy(DWTContext *d, int level, int width, int height, int stride)
1232 {
1233     vertical_compose_3tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1234     vertical_compose_5tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1235     DWTCompose *cs = d->cs + level;
1236
1237     int i, y = cs->y;
1238     IDWTELEM *b[8];
1239     for (i = 0; i < 6; i++)
1240         b[i] = cs->b[i];
1241     b[6] = d->buffer + av_clip(y+5, 0, height-2)*stride;
1242     b[7] = d->buffer + av_clip(y+6, 1, height-1)*stride;
1243
1244         if(y+5<(unsigned)height) vertical_compose_l0(      b[5], b[6], b[7],       width);
1245         if(y+1<(unsigned)height) vertical_compose_h0(b[0], b[2], b[3], b[4], b[6], width);
1246
1247         if(y-1<(unsigned)height) d->horizontal_compose(b[0], d->temp, width);
1248         if(y+0<(unsigned)height) d->horizontal_compose(b[1], d->temp, width);
1249
1250     for (i = 0; i < 6; i++)
1251         cs->b[i] = b[i+2];
1252     cs->y += 2;
1253 }
1254
1255 static void spatial_compose_dirac53i_dy(DWTContext *d, int level, int width, int height, int stride)
1256 {
1257     vertical_compose_3tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1258     vertical_compose_3tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1259     DWTCompose *cs = d->cs + level;
1260
1261     int y= cs->y;
1262     IDWTELEM *b[4] = { cs->b[0], cs->b[1] };
1263     b[2] = d->buffer + mirror(y+1, height-1)*stride;
1264     b[3] = d->buffer + mirror(y+2, height-1)*stride;
1265
1266         if(y+1<(unsigned)height) vertical_compose_l0(b[1], b[2], b[3], width);
1267         if(y+0<(unsigned)height) vertical_compose_h0(b[0], b[1], b[2], width);
1268
1269         if(y-1<(unsigned)height) d->horizontal_compose(b[0], d->temp, width);
1270         if(y+0<(unsigned)height) d->horizontal_compose(b[1], d->temp, width);
1271
1272     cs->b[0] = b[2];
1273     cs->b[1] = b[3];
1274     cs->y += 2;
1275 }
1276
1277
1278 static void spatial_compose_dd137i_dy(DWTContext *d, int level, int width, int height, int stride)
1279 {
1280     vertical_compose_5tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1281     vertical_compose_5tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1282     DWTCompose *cs = d->cs + level;
1283
1284     int i, y = cs->y;
1285     IDWTELEM *b[10];
1286     for (i = 0; i < 8; i++)
1287         b[i] = cs->b[i];
1288     b[8] = d->buffer + av_clip(y+7, 0, height-2)*stride;
1289     b[9] = d->buffer + av_clip(y+8, 1, height-1)*stride;
1290
1291         if(y+5<(unsigned)height) vertical_compose_l0(b[3], b[5], b[6], b[7], b[9], width);
1292         if(y+1<(unsigned)height) vertical_compose_h0(b[0], b[2], b[3], b[4], b[6], width);
1293
1294         if(y-1<(unsigned)height) d->horizontal_compose(b[0], d->temp, width);
1295         if(y+0<(unsigned)height) d->horizontal_compose(b[1], d->temp, width);
1296
1297     for (i = 0; i < 8; i++)
1298         cs->b[i] = b[i+2];
1299     cs->y += 2;
1300 }
1301
1302 // haar makes the assumption that height is even (always true for dirac)
1303 static void spatial_compose_haari_dy(DWTContext *d, int level, int width, int height, int stride)
1304 {
1305     vertical_compose_2tap vertical_compose = (void*)d->vertical_compose;
1306     int y = d->cs[level].y;
1307     IDWTELEM *b0 = d->buffer + (y-1)*stride;
1308     IDWTELEM *b1 = d->buffer + (y  )*stride;
1309
1310     vertical_compose(b0, b1, width);
1311     d->horizontal_compose(b0, d->temp, width);
1312     d->horizontal_compose(b1, d->temp, width);
1313
1314     d->cs[level].y += 2;
1315 }
1316
1317 // Don't do sliced idwt for fidelity; the 9 tap filter makes it a bit annoying
1318 // Fortunately, this filter isn't used in practice.
1319 static void spatial_compose_fidelity(DWTContext *d, int level, int width, int height, int stride)
1320 {
1321     vertical_compose_9tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1322     vertical_compose_9tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1323     int i, y;
1324     IDWTELEM *b[8];
1325
1326     for (y = 1; y < height; y += 2) {
1327         for (i = 0; i < 8; i++)
1328             b[i] = d->buffer + av_clip((y-7 + 2*i), 0, height-2)*stride;
1329         vertical_compose_h0(d->buffer + y*stride, b, width);
1330     }
1331
1332     for (y = 0; y < height; y += 2) {
1333         for (i = 0; i < 8; i++)
1334             b[i] = d->buffer + av_clip((y-7 + 2*i), 1, height-1)*stride;
1335         vertical_compose_l0(d->buffer + y*stride, b, width);
1336     }
1337
1338     for (y = 0; y < height; y++)
1339         d->horizontal_compose(d->buffer + y*stride, d->temp, width);
1340
1341     d->cs[level].y = height+1;
1342 }
1343
1344 static void spatial_compose_daub97i_dy(DWTContext *d, int level, int width, int height, int stride)
1345 {
1346     vertical_compose_3tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1347     vertical_compose_3tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1348     vertical_compose_3tap vertical_compose_l1 = (void*)d->vertical_compose_l1;
1349     vertical_compose_3tap vertical_compose_h1 = (void*)d->vertical_compose_h1;
1350     DWTCompose *cs = d->cs + level;
1351
1352     int i, y = cs->y;
1353     IDWTELEM *b[6];
1354     for (i = 0; i < 4; i++)
1355         b[i] = cs->b[i];
1356     b[4] = d->buffer + mirror(y+3, height-1)*stride;
1357     b[5] = d->buffer + mirror(y+4, height-1)*stride;
1358
1359         if(y+3<(unsigned)height) vertical_compose_l1(b[3], b[4], b[5], width);
1360         if(y+2<(unsigned)height) vertical_compose_h1(b[2], b[3], b[4], width);
1361         if(y+1<(unsigned)height) vertical_compose_l0(b[1], b[2], b[3], width);
1362         if(y+0<(unsigned)height) vertical_compose_h0(b[0], b[1], b[2], width);
1363
1364         if(y-1<(unsigned)height) d->horizontal_compose(b[0], d->temp, width);
1365         if(y+0<(unsigned)height) d->horizontal_compose(b[1], d->temp, width);
1366
1367     for (i = 0; i < 4; i++)
1368         cs->b[i] = b[i+2];
1369     cs->y += 2;
1370 }
1371
1372
1373 static void spatial_compose97i_init2(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1374 {
1375     cs->b[0] = buffer + mirror(-3-1, height-1)*stride;
1376     cs->b[1] = buffer + mirror(-3  , height-1)*stride;
1377     cs->b[2] = buffer + mirror(-3+1, height-1)*stride;
1378     cs->b[3] = buffer + mirror(-3+2, height-1)*stride;
1379     cs->y = -3;
1380 }
1381
1382 static void spatial_compose53i_init2(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1383 {
1384     cs->b[0] = buffer + mirror(-1-1, height-1)*stride;
1385     cs->b[1] = buffer + mirror(-1  , height-1)*stride;
1386     cs->y = -1;
1387 }
1388
1389 static void spatial_compose_dd97i_init(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1390 {
1391     cs->b[0] = buffer + av_clip(-5-1, 0, height-2)*stride;
1392     cs->b[1] = buffer + av_clip(-5  , 1, height-1)*stride;
1393     cs->b[2] = buffer + av_clip(-5+1, 0, height-2)*stride;
1394     cs->b[3] = buffer + av_clip(-5+2, 1, height-1)*stride;
1395     cs->b[4] = buffer + av_clip(-5+3, 0, height-2)*stride;
1396     cs->b[5] = buffer + av_clip(-5+4, 1, height-1)*stride;
1397     cs->y = -5;
1398 }
1399
1400 static void spatial_compose_dd137i_init(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1401 {
1402     cs->b[0] = buffer + av_clip(-5-1, 0, height-2)*stride;
1403     cs->b[1] = buffer + av_clip(-5  , 1, height-1)*stride;
1404     cs->b[2] = buffer + av_clip(-5+1, 0, height-2)*stride;
1405     cs->b[3] = buffer + av_clip(-5+2, 1, height-1)*stride;
1406     cs->b[4] = buffer + av_clip(-5+3, 0, height-2)*stride;
1407     cs->b[5] = buffer + av_clip(-5+4, 1, height-1)*stride;
1408     cs->b[6] = buffer + av_clip(-5+5, 0, height-2)*stride;
1409     cs->b[7] = buffer + av_clip(-5+6, 1, height-1)*stride;
1410     cs->y = -5;
1411 }
1412
1413 int ff_spatial_idwt_init2(DWTContext *d, IDWTELEM *buffer, int width, int height,
1414                           int stride, enum dwt_type type, int decomposition_count,
1415                           IDWTELEM *temp)
1416 {
1417     int level;
1418
1419     d->buffer = buffer;
1420     d->width = width;
1421     d->height = height;
1422     d->stride = stride;
1423     d->decomposition_count = decomposition_count;
1424     d->temp = temp + 8;
1425
1426     for(level=decomposition_count-1; level>=0; level--){
1427         int hl = height >> level;
1428         int stride_l = stride << level;
1429
1430         switch(type){
1431         case DWT_DIRAC_DD9_7:
1432             spatial_compose_dd97i_init(d->cs+level, buffer, hl, stride_l);
1433             break;
1434         case DWT_DIRAC_LEGALL5_3:
1435             spatial_compose53i_init2(d->cs+level, buffer, hl, stride_l);
1436             break;
1437         case DWT_DIRAC_DD13_7:
1438             spatial_compose_dd137i_init(d->cs+level, buffer, hl, stride_l);
1439             break;
1440         case DWT_DIRAC_HAAR0:
1441         case DWT_DIRAC_HAAR1:
1442             d->cs[level].y = 1;
1443             break;
1444         case DWT_DIRAC_DAUB9_7:
1445             spatial_compose97i_init2(d->cs+level, buffer, hl, stride_l);
1446             break;
1447         default:
1448             d->cs[level].y = 0;
1449             break;
1450         }
1451     }
1452
1453     switch (type) {
1454     case DWT_DIRAC_DD9_7:
1455         d->spatial_compose = spatial_compose_dd97i_dy;
1456         d->vertical_compose_l0 = (void*)vertical_compose53iL0;
1457         d->vertical_compose_h0 = (void*)vertical_compose_dd97iH0;
1458         d->horizontal_compose = horizontal_compose_dd97i;
1459         d->support = 7;
1460         break;
1461     case DWT_DIRAC_LEGALL5_3:
1462         d->spatial_compose = spatial_compose_dirac53i_dy;
1463         d->vertical_compose_l0 = (void*)vertical_compose53iL0;
1464         d->vertical_compose_h0 = (void*)vertical_compose_dirac53iH0;
1465         d->horizontal_compose = horizontal_compose_dirac53i;
1466         d->support = 3;
1467         break;
1468     case DWT_DIRAC_DD13_7:
1469         d->spatial_compose = spatial_compose_dd137i_dy;
1470         d->vertical_compose_l0 = (void*)vertical_compose_dd137iL0;
1471         d->vertical_compose_h0 = (void*)vertical_compose_dd97iH0;
1472         d->horizontal_compose = horizontal_compose_dd137i;
1473         d->support = 7;
1474         break;
1475     case DWT_DIRAC_HAAR0:
1476     case DWT_DIRAC_HAAR1:
1477         d->spatial_compose = spatial_compose_haari_dy;
1478         d->vertical_compose = (void*)vertical_compose_haar;
1479         if (type == DWT_DIRAC_HAAR0)
1480             d->horizontal_compose = horizontal_compose_haar0i;
1481         else
1482             d->horizontal_compose = horizontal_compose_haar1i;
1483         d->support = 1;
1484         break;
1485     case DWT_DIRAC_FIDELITY:
1486         d->spatial_compose = spatial_compose_fidelity;
1487         d->vertical_compose_l0 = (void*)vertical_compose_fidelityiL0;
1488         d->vertical_compose_h0 = (void*)vertical_compose_fidelityiH0;
1489         d->horizontal_compose = horizontal_compose_fidelityi;
1490         break;
1491     case DWT_DIRAC_DAUB9_7:
1492         d->spatial_compose = spatial_compose_daub97i_dy;
1493         d->vertical_compose_l0 = (void*)vertical_compose_daub97iL0;
1494         d->vertical_compose_h0 = (void*)vertical_compose_daub97iH0;
1495         d->vertical_compose_l1 = (void*)vertical_compose_daub97iL1;
1496         d->vertical_compose_h1 = (void*)vertical_compose_daub97iH1;
1497         d->horizontal_compose = horizontal_compose_daub97i;
1498         d->support = 5;
1499         break;
1500     default:
1501         av_log(NULL, AV_LOG_ERROR, "Unknown wavelet type %d\n", type);
1502         return -1;
1503     }
1504
1505     if (HAVE_MMX) ff_spatial_idwt_init_mmx(d, type);
1506
1507     return 0;
1508 }
1509
1510 void ff_spatial_idwt_slice2(DWTContext *d, int y)
1511 {
1512     int level, support = d->support;
1513
1514     for (level = d->decomposition_count-1; level >= 0; level--) {
1515         int wl = d->width  >> level;
1516         int hl = d->height >> level;
1517         int stride_l = d->stride << level;
1518
1519         while (d->cs[level].y <= FFMIN((y>>level)+support, hl))
1520             d->spatial_compose(d, level, wl, hl, stride_l);
1521     }
1522 }
1523
1524 int ff_spatial_idwt2(IDWTELEM *buffer, int width, int height, int stride,
1525                      enum dwt_type type, int decomposition_count, IDWTELEM *temp)
1526 {
1527     DWTContext d;
1528     int y;
1529
1530     if (ff_spatial_idwt_init2(&d, buffer, width, height, stride, type, decomposition_count, temp))
1531         return -1;
1532
1533     for (y = 0; y < d.height; y += 4)
1534         ff_spatial_idwt_slice2(&d, y);
1535
1536     return 0;
1537 }