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 "libavutil/avassert.h"
  24 #include "libavutil/common.h"
  25 #include "dsputil.h"
  26 #include "dwt.h"
  27 #include "libavcodec/x86/dwt.h"
  28
  29 int ff_slice_buffer_init(slice_buffer *buf, int line_count,
  30                          int max_allocated_lines, int line_width,
  31                          IDWTELEM *base_buffer)
  32 {
  33     int i;
  34
  35     buf->base_buffer = base_buffer;
  36     buf->line_count  = line_count;
  37     buf->line_width  = line_width;
  38     buf->data_count  = max_allocated_lines;
  39     buf->line        = av_mallocz(sizeof(IDWTELEM *) * line_count);
  40     if (!buf->line)
  41         return AVERROR(ENOMEM);
  42     buf->data_stack  = av_malloc(sizeof(IDWTELEM *) * max_allocated_lines);
  43     if (!buf->data_stack) {
  44         av_freep(&buf->line);
  45         return AVERROR(ENOMEM);
  46     }
  47
  48     for (i = 0; i < max_allocated_lines; i++) {
  49         buf->data_stack[i] = av_malloc(sizeof(IDWTELEM) * line_width);
  50         if (!buf->data_stack[i]) {
  51             for (i--; i >=0; i--)
  52                 av_freep(&buf->data_stack[i]);
  53             av_freep(&buf->data_stack);
  54             av_freep(&buf->line);
  55             return AVERROR(ENOMEM);
  56         }
  57     }
  58
  59     buf->data_stack_top = max_allocated_lines - 1;
  60     return 0;
  61 }
  62
  63 IDWTELEM *ff_slice_buffer_load_line(slice_buffer *buf, int line)
  64 {
  65     IDWTELEM *buffer;
  66
  67     av_assert0(buf->data_stack_top >= 0);
  68 //  av_assert1(!buf->line[line]);
  69     if (buf->line[line])
  70         return buf->line[line];
  71
  72     buffer = buf->data_stack[buf->data_stack_top];
  73     buf->data_stack_top--;
  74     buf->line[line] = buffer;
  75
  76     return buffer;
  77 }
  78
  79 void ff_slice_buffer_release(slice_buffer *buf, int line)
  80 {
  81     IDWTELEM *buffer;
  82
  83     av_assert1(line >= 0 && line < buf->line_count);
  84     av_assert1(buf->line[line]);
  85
  86     buffer = buf->line[line];
  87     buf->data_stack_top++;
  88     buf->data_stack[buf->data_stack_top] = buffer;
  89     buf->line[line]                      = NULL;
  90 }
  91
  92 void ff_slice_buffer_flush(slice_buffer *buf)
  93 {
  94     int i;
  95     for (i = 0; i < buf->line_count; i++)
  96         if (buf->line[i])
  97             ff_slice_buffer_release(buf, i);
  98 }
  99
 100 void ff_slice_buffer_destroy(slice_buffer *buf)
 101 {
 102     int i;
 103     ff_slice_buffer_flush(buf);
 104
 105     for (i = buf->data_count - 1; i >= 0; i--)
 106         av_freep(&buf->data_stack[i]);
 107     av_freep(&buf->data_stack);
 108     av_freep(&buf->line);
 109 }
 110
 111 static inline int mirror(int v, int m)
 112 {
 113     while ((unsigned)v > (unsigned)m) {
 114         v = -v;
 115         if (v < 0)
 116             v += 2 * m;
 117     }
 118     return v;
 119 }
 120
 121 static av_always_inline void lift(DWTELEM *dst, DWTELEM *src, DWTELEM *ref,
 122                                   int dst_step, int src_step, int ref_step,
 123                                   int width, int mul, int add, int shift,
 124                                   int highpass, int inverse)
 125 {
 126     const int mirror_left  = !highpass;
 127     const int mirror_right = (width & 1) ^ highpass;
 128     const int w            = (width >> 1) - 1 + (highpass & width);
 129     int i;
 130
 131 #define LIFT(src, ref, inv) ((src) + ((inv) ? -(ref) : +(ref)))
 132     if (mirror_left) {
 133         dst[0] = LIFT(src[0], ((mul * 2 * ref[0] + add) >> shift), inverse);
 134         dst   += dst_step;
 135         src   += src_step;
 136     }
 137
 138     for (i = 0; i < w; i++)
 139         dst[i * dst_step] = LIFT(src[i * src_step],
 140                                  ((mul * (ref[i * ref_step] +
 141                                           ref[(i + 1) * ref_step]) +
 142                                    add) >> shift),
 143                                  inverse);
 144
 145     if (mirror_right)
 146         dst[w * dst_step] = LIFT(src[w * src_step],
 147                                  ((mul * 2 * ref[w * ref_step] + add) >> shift),
 148                                  inverse);
 149 }
 150
 151 static av_always_inline void liftS(DWTELEM *dst, DWTELEM *src, DWTELEM *ref,
 152                                    int dst_step, int src_step, int ref_step,
 153                                    int width, int mul, int add, int shift,
 154                                    int highpass, int inverse)
 155 {
 156     const int mirror_left  = !highpass;
 157     const int mirror_right = (width & 1) ^ highpass;
 158     const int w            = (width >> 1) - 1 + (highpass & width);
 159     int i;
 160
 161     av_assert1(shift == 4);
 162 #define LIFTS(src, ref, inv)                                            \
 163     ((inv) ? (src) + (((ref) + 4 * (src)) >> shift)                     \
 164            : -((-16 * (src) + (ref) + add /                             \
 165                 4 + 1 + (5 << 25)) / (5 * 4) - (1 << 23)))
 166     if (mirror_left) {
 167         dst[0] = LIFTS(src[0], mul * 2 * ref[0] + add, inverse);
 168         dst   += dst_step;
 169         src   += src_step;
 170     }
 171
 172     for (i = 0; i < w; i++)
 173         dst[i * dst_step] = LIFTS(src[i * src_step],
 174                                   mul * (ref[i * ref_step] +
 175                                          ref[(i + 1) * ref_step]) + add,
 176                                   inverse);
 177
 178     if (mirror_right)
 179         dst[w * dst_step] = LIFTS(src[w * src_step],
 180                                   mul * 2 * ref[w * ref_step] + add,
 181                                   inverse);
 182 }
 183
 184 static void horizontal_decompose53i(DWTELEM *b, DWTELEM *temp, int width)
 185 {
 186     const int width2 = width >> 1;
 187     int x;
 188     const int w2 = (width + 1) >> 1;
 189
 190     for (x = 0; x < width2; x++) {
 191         temp[x]      = b[2 * x];
 192         temp[x + w2] = b[2 * x + 1];
 193     }
 194     if (width & 1)
 195         temp[x] = b[2 * x];
 196     lift(b + w2, temp + w2, temp,   1, 1, 1, width, -1, 0, 1, 1, 0);
 197     lift(b,      temp,      b + w2, 1, 1, 1, width,  1, 2, 2, 0, 0);
 198 }
 199
 200 static void vertical_decompose53iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 201                                     int width)
 202 {
 203     int i;
 204
 205     for (i = 0; i < width; i++)
 206         b1[i] -= (b0[i] + b2[i]) >> 1;
 207 }
 208
 209 static void vertical_decompose53iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 210                                     int width)
 211 {
 212     int i;
 213
 214     for (i = 0; i < width; i++)
 215         b1[i] += (b0[i] + b2[i] + 2) >> 2;
 216 }
 217
 218 static void spatial_decompose53i(DWTELEM *buffer, DWTELEM *temp,
 219                                  int width, int height, int stride)
 220 {
 221     int y;
 222     DWTELEM *b0 = buffer + mirror(-2 - 1, height - 1) * stride;
 223     DWTELEM *b1 = buffer + mirror(-2,     height - 1) * stride;
 224
 225     for (y = -2; y < height; y += 2) {
 226         DWTELEM *b2 = buffer + mirror(y + 1, height - 1) * stride;
 227         DWTELEM *b3 = buffer + mirror(y + 2, height - 1) * stride;
 228
 229         if (y + 1 < (unsigned)height)
 230             horizontal_decompose53i(b2, temp, width);
 231         if (y + 2 < (unsigned)height)
 232             horizontal_decompose53i(b3, temp, width);
 233
 234         if (y + 1 < (unsigned)height)
 235             vertical_decompose53iH0(b1, b2, b3, width);
 236         if (y + 0 < (unsigned)height)
 237             vertical_decompose53iL0(b0, b1, b2, width);
 238
 239         b0 = b2;
 240         b1 = b3;
 241     }
 242 }
 243
 244 static void horizontal_decompose97i(DWTELEM *b, DWTELEM *temp, int width)
 245 {
 246     const int w2 = (width + 1) >> 1;
 247
 248     lift(temp + w2, b + 1, b,         1, 2, 2, width, W_AM, W_AO, W_AS, 1, 1);
 249     liftS(temp,     b,     temp + w2, 1, 2, 1, width, W_BM, W_BO, W_BS, 0, 0);
 250     lift(b + w2, temp + w2, temp,     1, 1, 1, width, W_CM, W_CO, W_CS, 1, 0);
 251     lift(b,      temp,      b + w2,   1, 1, 1, width, W_DM, W_DO, W_DS, 0, 0);
 252 }
 253
 254 static void vertical_decompose97iH0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 255                                     int width)
 256 {
 257     int i;
 258
 259     for (i = 0; i < width; i++)
 260         b1[i] -= (W_AM * (b0[i] + b2[i]) + W_AO) >> W_AS;
 261 }
 262
 263 static void vertical_decompose97iH1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 264                                     int width)
 265 {
 266     int i;
 267
 268     for (i = 0; i < width; i++)
 269         b1[i] += (W_CM * (b0[i] + b2[i]) + W_CO) >> W_CS;
 270 }
 271
 272 static void vertical_decompose97iL0(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 273                                     int width)
 274 {
 275     int i;
 276
 277     for (i = 0; i < width; i++)
 278         b1[i] = (16 * 4 * b1[i] - 4 * (b0[i] + b2[i]) + W_BO * 5 + (5 << 27)) /
 279                 (5 * 16) - (1 << 23);
 280 }
 281
 282 static void vertical_decompose97iL1(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2,
 283                                     int width)
 284 {
 285     int i;
 286
 287     for (i = 0; i < width; i++)
 288         b1[i] += (W_DM * (b0[i] + b2[i]) + W_DO) >> W_DS;
 289 }
 290
 291 static void spatial_decompose97i(DWTELEM *buffer, DWTELEM *temp,
 292                                  int width, int height, int stride)
 293 {
 294     int y;
 295     DWTELEM *b0 = buffer + mirror(-4 - 1, height - 1) * stride;
 296     DWTELEM *b1 = buffer + mirror(-4,     height - 1) * stride;
 297     DWTELEM *b2 = buffer + mirror(-4 + 1, height - 1) * stride;
 298     DWTELEM *b3 = buffer + mirror(-4 + 2, height - 1) * stride;
 299
 300     for (y = -4; y < height; y += 2) {
 301         DWTELEM *b4 = buffer + mirror(y + 3, height - 1) * stride;
 302         DWTELEM *b5 = buffer + mirror(y + 4, height - 1) * stride;
 303
 304         if (y + 3 < (unsigned)height)
 305             horizontal_decompose97i(b4, temp, width);
 306         if (y + 4 < (unsigned)height)
 307             horizontal_decompose97i(b5, temp, width);
 308
 309         if (y + 3 < (unsigned)height)
 310             vertical_decompose97iH0(b3, b4, b5, width);
 311         if (y + 2 < (unsigned)height)
 312             vertical_decompose97iL0(b2, b3, b4, width);
 313         if (y + 1 < (unsigned)height)
 314             vertical_decompose97iH1(b1, b2, b3, width);
 315         if (y + 0 < (unsigned)height)
 316             vertical_decompose97iL1(b0, b1, b2, width);
 317
 318         b0 = b2;
 319         b1 = b3;
 320         b2 = b4;
 321         b3 = b5;
 322     }
 323 }
 324
 325 void ff_spatial_dwt(DWTELEM *buffer, DWTELEM *temp, int width, int height,
 326                     int stride, int type, int decomposition_count)
 327 {
 328     int level;
 329
 330     for (level = 0; level < decomposition_count; level++) {
 331         switch (type) {
 332         case DWT_97:
 333             spatial_decompose97i(buffer, temp,
 334                                  width >> level, height >> level,
 335                                  stride << level);
 336             break;
 337         case DWT_53:
 338             spatial_decompose53i(buffer, temp,
 339                                  width >> level, height >> level,
 340                                  stride << level);
 341             break;
 342         }
 343     }
 344 }
 345
 346 static void horizontal_compose53i(IDWTELEM *b, IDWTELEM *temp, int width)
 347 {
 348     const int width2 = width >> 1;
 349     const int w2     = (width + 1) >> 1;
 350     int x;
 351
 352     for (x = 0; x < width2; x++) {
 353         temp[2 * x]     = b[x];
 354         temp[2 * x + 1] = b[x + w2];
 355     }
 356     if (width & 1)
 357         temp[2 * x] = b[x];
 358
 359     b[0] = temp[0] - ((temp[1] + 1) >> 1);
 360     for (x = 2; x < width - 1; x += 2) {
 361         b[x]     = temp[x]     - ((temp[x - 1] + temp[x + 1] + 2) >> 2);
 362         b[x - 1] = temp[x - 1] + ((b[x - 2]    + b[x]        + 1) >> 1);
 363     }
 364     if (width & 1) {
 365         b[x]     = temp[x]     - ((temp[x - 1]     + 1) >> 1);
 366         b[x - 1] = temp[x - 1] + ((b[x - 2] + b[x] + 1) >> 1);
 367     } else
 368         b[x - 1] = temp[x - 1] + b[x - 2];
 369 }
 370
 371 static void vertical_compose53iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 372                                   int width)
 373 {
 374     int i;
 375
 376     for (i = 0; i < width; i++)
 377         b1[i] += (b0[i] + b2[i]) >> 1;
 378 }
 379
 380 static void vertical_compose53iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 381                                   int width)
 382 {
 383     int i;
 384
 385     for (i = 0; i < width; i++)
 386         b1[i] -= (b0[i] + b2[i] + 2) >> 2;
 387 }
 388
 389 static void spatial_compose53i_buffered_init(DWTCompose *cs, slice_buffer *sb,
 390                                              int height, int stride_line)
 391 {
 392     cs->b0 = slice_buffer_get_line(sb,
 393                                    mirror(-1 - 1, height - 1) * stride_line);
 394     cs->b1 = slice_buffer_get_line(sb, mirror(-1, height - 1) * stride_line);
 395     cs->y  = -1;
 396 }
 397
 398 static void spatial_compose53i_init(DWTCompose *cs, IDWTELEM *buffer,
 399                                     int height, int stride)
 400 {
 401     cs->b0 = buffer + mirror(-1 - 1, height - 1) * stride;
 402     cs->b1 = buffer + mirror(-1,     height - 1) * stride;
 403     cs->y  = -1;
 404 }
 405
 406 static void spatial_compose53i_dy_buffered(DWTCompose *cs, slice_buffer *sb,
 407                                            IDWTELEM *temp,
 408                                            int width, int height,
 409                                            int stride_line)
 410 {
 411     int y = cs->y;
 412
 413     IDWTELEM *b0 = cs->b0;
 414     IDWTELEM *b1 = cs->b1;
 415     IDWTELEM *b2 = slice_buffer_get_line(sb,
 416                                          mirror(y + 1, height - 1) *
 417                                          stride_line);
 418     IDWTELEM *b3 = slice_buffer_get_line(sb,
 419                                          mirror(y + 2, height - 1) *
 420                                          stride_line);
 421
 422     if (y + 1 < (unsigned)height && y < (unsigned)height) {
 423         int x;
 424
 425         for (x = 0; x < width; x++) {
 426             b2[x] -= (b1[x] + b3[x] + 2) >> 2;
 427             b1[x] += (b0[x] + b2[x])     >> 1;
 428         }
 429     } else {
 430         if (y + 1 < (unsigned)height)
 431             vertical_compose53iL0(b1, b2, b3, width);
 432         if (y + 0 < (unsigned)height)
 433             vertical_compose53iH0(b0, b1, b2, width);
 434     }
 435
 436     if (y - 1 < (unsigned)height)
 437         horizontal_compose53i(b0, temp, width);
 438     if (y + 0 < (unsigned)height)
 439         horizontal_compose53i(b1, temp, width);
 440
 441     cs->b0  = b2;
 442     cs->b1  = b3;
 443     cs->y  += 2;
 444 }
 445
 446 static void spatial_compose53i_dy(DWTCompose *cs, IDWTELEM *buffer,
 447                                   IDWTELEM *temp, int width, int height,
 448                                   int stride)
 449 {
 450     int y        = cs->y;
 451     IDWTELEM *b0 = cs->b0;
 452     IDWTELEM *b1 = cs->b1;
 453     IDWTELEM *b2 = buffer + mirror(y + 1, height - 1) * stride;
 454     IDWTELEM *b3 = buffer + mirror(y + 2, height - 1) * stride;
 455
 456     if (y + 1 < (unsigned)height)
 457         vertical_compose53iL0(b1, b2, b3, width);
 458     if (y + 0 < (unsigned)height)
 459         vertical_compose53iH0(b0, b1, b2, width);
 460
 461     if (y - 1 < (unsigned)height)
 462         horizontal_compose53i(b0, temp, width);
 463     if (y + 0 < (unsigned)height)
 464         horizontal_compose53i(b1, temp, width);
 465
 466     cs->b0  = b2;
 467     cs->b1  = b3;
 468     cs->y  += 2;
 469 }
 470
 471 static void av_unused spatial_compose53i(IDWTELEM *buffer, IDWTELEM *temp,
 472                                          int width, int height, int stride)
 473 {
 474     DWTCompose cs;
 475     spatial_compose53i_init(&cs, buffer, height, stride);
 476     while (cs.y <= height)
 477         spatial_compose53i_dy(&cs, buffer, temp, width, height, stride);
 478 }
 479
 480 void ff_snow_horizontal_compose97i(IDWTELEM *b, IDWTELEM *temp, int width)
 481 {
 482     const int w2 = (width + 1) >> 1;
 483     int x;
 484
 485     temp[0] = b[0] - ((3 * b[w2] + 2) >> 2);
 486     for (x = 1; x < (width >> 1); x++) {
 487         temp[2 * x]     = b[x] - ((3 * (b[x + w2 - 1] + b[x + w2]) + 4) >> 3);
 488         temp[2 * x - 1] = b[x + w2 - 1] - temp[2 * x - 2] - temp[2 * x];
 489     }
 490     if (width & 1) {
 491         temp[2 * x]     = b[x] - ((3 * b[x + w2 - 1] + 2) >> 2);
 492         temp[2 * x - 1] = b[x + w2 - 1] - temp[2 * x - 2] - temp[2 * x];
 493     } else
 494         temp[2 * x - 1] = b[x + w2 - 1] - 2 * temp[2 * x - 2];
 495
 496     b[0] = temp[0] + ((2 * temp[0] + temp[1] + 4) >> 3);
 497     for (x = 2; x < width - 1; x += 2) {
 498         b[x]     = temp[x] + ((4 * temp[x] + temp[x - 1] + temp[x + 1] + 8) >> 4);
 499         b[x - 1] = temp[x - 1] + ((3 * (b[x - 2] + b[x])) >> 1);
 500     }
 501     if (width & 1) {
 502         b[x]     = temp[x] + ((2 * temp[x] + temp[x - 1] + 4) >> 3);
 503         b[x - 1] = temp[x - 1] + ((3 * (b[x - 2] + b[x])) >> 1);
 504     } else
 505         b[x - 1] = temp[x - 1] + 3 * b[x - 2];
 506 }
 507
 508 static void vertical_compose97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 509                                   int width)
 510 {
 511     int i;
 512
 513     for (i = 0; i < width; i++)
 514         b1[i] += (W_AM * (b0[i] + b2[i]) + W_AO) >> W_AS;
 515 }
 516
 517 static void vertical_compose97iH1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 518                                   int width)
 519 {
 520     int i;
 521
 522     for (i = 0; i < width; i++)
 523         b1[i] -= (W_CM * (b0[i] + b2[i]) + W_CO) >> W_CS;
 524 }
 525
 526 static void vertical_compose97iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 527                                   int width)
 528 {
 529     int i;
 530
 531     for (i = 0; i < width; i++)
 532         b1[i] += (W_BM * (b0[i] + b2[i]) + 4 * b1[i] + W_BO) >> W_BS;
 533 }
 534
 535 static void vertical_compose97iL1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 536                                   int width)
 537 {
 538     int i;
 539
 540     for (i = 0; i < width; i++)
 541         b1[i] -= (W_DM * (b0[i] + b2[i]) + W_DO) >> W_DS;
 542 }
 543
 544 void ff_snow_vertical_compose97i(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
 545                                  IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5,
 546                                  int width)
 547 {
 548     int i;
 549
 550     for (i = 0; i < width; i++) {
 551         b4[i] -= (W_DM * (b3[i] + b5[i]) + W_DO) >> W_DS;
 552         b3[i] -= (W_CM * (b2[i] + b4[i]) + W_CO) >> W_CS;
 553         b2[i] += (W_BM * (b1[i] + b3[i]) + 4 * b2[i] + W_BO) >> W_BS;
 554         b1[i] += (W_AM * (b0[i] + b2[i]) + W_AO) >> W_AS;
 555     }
 556 }
 557
 558 static void spatial_compose97i_buffered_init(DWTCompose *cs, slice_buffer *sb,
 559                                              int height, int stride_line)
 560 {
 561     cs->b0 = slice_buffer_get_line(sb, mirror(-3 - 1, height - 1) * stride_line);
 562     cs->b1 = slice_buffer_get_line(sb, mirror(-3,     height - 1) * stride_line);
 563     cs->b2 = slice_buffer_get_line(sb, mirror(-3 + 1, height - 1) * stride_line);
 564     cs->b3 = slice_buffer_get_line(sb, mirror(-3 + 2, height - 1) * stride_line);
 565     cs->y  = -3;
 566 }
 567
 568 static void spatial_compose97i_init(DWTCompose *cs, IDWTELEM *buffer, int height,
 569                                     int stride)
 570 {
 571     cs->b0 = buffer + mirror(-3 - 1, height - 1) * stride;
 572     cs->b1 = buffer + mirror(-3,     height - 1) * stride;
 573     cs->b2 = buffer + mirror(-3 + 1, height - 1) * stride;
 574     cs->b3 = buffer + mirror(-3 + 2, height - 1) * stride;
 575     cs->y  = -3;
 576 }
 577
 578 static void spatial_compose97i_dy_buffered(DWTContext *dsp, DWTCompose *cs,
 579                                            slice_buffer * sb, IDWTELEM *temp,
 580                                            int width, int height,
 581                                            int stride_line)
 582 {
 583     int y = cs->y;
 584
 585     IDWTELEM *b0 = cs->b0;
 586     IDWTELEM *b1 = cs->b1;
 587     IDWTELEM *b2 = cs->b2;
 588     IDWTELEM *b3 = cs->b3;
 589     IDWTELEM *b4 = slice_buffer_get_line(sb,
 590                                          mirror(y + 3, height - 1) *
 591                                          stride_line);
 592     IDWTELEM *b5 = slice_buffer_get_line(sb,
 593                                          mirror(y + 4, height - 1) *
 594                                          stride_line);
 595
 596     if (y > 0 && y + 4 < height) {
 597         dsp->vertical_compose97i(b0, b1, b2, b3, b4, b5, width);
 598     } else {
 599         if (y + 3 < (unsigned)height)
 600             vertical_compose97iL1(b3, b4, b5, width);
 601         if (y + 2 < (unsigned)height)
 602             vertical_compose97iH1(b2, b3, b4, width);
 603         if (y + 1 < (unsigned)height)
 604             vertical_compose97iL0(b1, b2, b3, width);
 605         if (y + 0 < (unsigned)height)
 606             vertical_compose97iH0(b0, b1, b2, width);
 607     }
 608
 609     if (y - 1 < (unsigned)height)
 610         dsp->horizontal_compose97i(b0, temp, width);
 611     if (y + 0 < (unsigned)height)
 612         dsp->horizontal_compose97i(b1, temp, width);
 613
 614     cs->b0  = b2;
 615     cs->b1  = b3;
 616     cs->b2  = b4;
 617     cs->b3  = b5;
 618     cs->y  += 2;
 619 }
 620
 621 static void spatial_compose97i_dy(DWTCompose *cs, IDWTELEM *buffer,
 622                                   IDWTELEM *temp, int width, int height,
 623                                   int stride)
 624 {
 625     int y        = cs->y;
 626     IDWTELEM *b0 = cs->b0;
 627     IDWTELEM *b1 = cs->b1;
 628     IDWTELEM *b2 = cs->b2;
 629     IDWTELEM *b3 = cs->b3;
 630     IDWTELEM *b4 = buffer + mirror(y + 3, height - 1) * stride;
 631     IDWTELEM *b5 = buffer + mirror(y + 4, height - 1) * stride;
 632
 633     if (y + 3 < (unsigned)height)
 634         vertical_compose97iL1(b3, b4, b5, width);
 635     if (y + 2 < (unsigned)height)
 636         vertical_compose97iH1(b2, b3, b4, width);
 637     if (y + 1 < (unsigned)height)
 638         vertical_compose97iL0(b1, b2, b3, width);
 639     if (y + 0 < (unsigned)height)
 640         vertical_compose97iH0(b0, b1, b2, width);
 641
 642     if (y - 1 < (unsigned)height)
 643         ff_snow_horizontal_compose97i(b0, temp, width);
 644     if (y + 0 < (unsigned)height)
 645         ff_snow_horizontal_compose97i(b1, temp, width);
 646
 647     cs->b0  = b2;
 648     cs->b1  = b3;
 649     cs->b2  = b4;
 650     cs->b3  = b5;
 651     cs->y  += 2;
 652 }
 653
 654 static void av_unused spatial_compose97i(IDWTELEM *buffer, IDWTELEM *temp,
 655                                          int width, int height, int stride)
 656 {
 657     DWTCompose cs;
 658     spatial_compose97i_init(&cs, buffer, height, stride);
 659     while (cs.y <= height)
 660         spatial_compose97i_dy(&cs, buffer, temp, width, height, stride);
 661 }
 662
 663 void ff_spatial_idwt_buffered_init(DWTCompose *cs, slice_buffer *sb, int width,
 664                                    int height, int stride_line, int type,
 665                                    int decomposition_count)
 666 {
 667     int level;
 668     for (level = decomposition_count - 1; level >= 0; level--) {
 669         switch (type) {
 670         case DWT_97:
 671             spatial_compose97i_buffered_init(cs + level, sb, height >> level,
 672                                              stride_line << level);
 673             break;
 674         case DWT_53:
 675             spatial_compose53i_buffered_init(cs + level, sb, height >> level,
 676                                              stride_line << level);
 677             break;
 678         }
 679     }
 680 }
 681
 682 void ff_spatial_idwt_buffered_slice(DWTContext *dsp, DWTCompose *cs,
 683                                     slice_buffer *slice_buf, IDWTELEM *temp,
 684                                     int width, int height, int stride_line,
 685                                     int type, int decomposition_count, int y)
 686 {
 687     const int support = type == 1 ? 3 : 5;
 688     int level;
 689     if (type == 2)
 690         return;
 691
 692     for (level = decomposition_count - 1; level >= 0; level--)
 693         while (cs[level].y <= FFMIN((y >> level) + support, height >> level)) {
 694             switch (type) {
 695             case DWT_97:
 696                 spatial_compose97i_dy_buffered(dsp, cs + level, slice_buf, temp,
 697                                                width >> level,
 698                                                height >> level,
 699                                                stride_line << level);
 700                 break;
 701             case DWT_53:
 702                 spatial_compose53i_dy_buffered(cs + level, slice_buf, temp,
 703                                                width >> level,
 704                                                height >> level,
 705                                                stride_line << level);
 706                 break;
 707             }
 708         }
 709 }
 710
 711 static void ff_spatial_idwt_init(DWTCompose *cs, IDWTELEM *buffer, int width,
 712                                  int height, int stride, int type,
 713                                  int decomposition_count)
 714 {
 715     int level;
 716     for (level = decomposition_count - 1; level >= 0; level--) {
 717         switch (type) {
 718         case DWT_97:
 719             spatial_compose97i_init(cs + level, buffer, height >> level,
 720                                     stride << level);
 721             break;
 722         case DWT_53:
 723             spatial_compose53i_init(cs + level, buffer, height >> level,
 724                                     stride << level);
 725             break;
 726         }
 727     }
 728 }
 729
 730 static void ff_spatial_idwt_slice(DWTCompose *cs, IDWTELEM *buffer,
 731                                   IDWTELEM *temp, int width, int height,
 732                                   int stride, int type,
 733                                   int decomposition_count, int y)
 734 {
 735     const int support = type == 1 ? 3 : 5;
 736     int level;
 737     if (type == 2)
 738         return;
 739
 740     for (level = decomposition_count - 1; level >= 0; level--)
 741         while (cs[level].y <= FFMIN((y >> level) + support, height >> level)) {
 742             switch (type) {
 743             case DWT_97:
 744                 spatial_compose97i_dy(cs + level, buffer, temp, width >> level,
 745                                       height >> level, stride << level);
 746                 break;
 747             case DWT_53:
 748                 spatial_compose53i_dy(cs + level, buffer, temp, width >> level,
 749                                       height >> level, stride << level);
 750                 break;
 751             }
 752         }
 753 }
 754
 755 void ff_spatial_idwt(IDWTELEM *buffer, IDWTELEM *temp, int width, int height,
 756                      int stride, int type, int decomposition_count)
 757 {
 758     DWTCompose cs[MAX_DECOMPOSITIONS];
 759     int y;
 760     ff_spatial_idwt_init(cs, buffer, width, height, stride, type,
 761                          decomposition_count);
 762     for (y = 0; y < height; y += 4)
 763         ff_spatial_idwt_slice(cs, buffer, temp, width, height, stride, type,
 764                               decomposition_count, y);
 765 }
 766
 767 static inline int w_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size,
 768                       int w, int h, int type)
 769 {
 770     int s, i, j;
 771     const int dec_count = w == 8 ? 3 : 4;
 772     int tmp[32 * 32], tmp2[32];
 773     int level, ori;
 774     static const int scale[2][2][4][4] = {
 775         {
 776             { // 9/7 8x8 dec=3
 777                 { 268, 239, 239, 213 },
 778                 { 0,   224, 224, 152 },
 779                 { 0,   135, 135, 110 },
 780             },
 781             { // 9/7 16x16 or 32x32 dec=4
 782                 { 344, 310, 310, 280 },
 783                 { 0,   320, 320, 228 },
 784                 { 0,   175, 175, 136 },
 785                 { 0,   129, 129, 102 },
 786             }
 787         },
 788         {
 789             { // 5/3 8x8 dec=3
 790                 { 275, 245, 245, 218 },
 791                 { 0,   230, 230, 156 },
 792                 { 0,   138, 138, 113 },
 793             },
 794             { // 5/3 16x16 or 32x32 dec=4
 795                 { 352, 317, 317, 286 },
 796                 { 0,   328, 328, 233 },
 797                 { 0,   180, 180, 140 },
 798                 { 0,   132, 132, 105 },
 799             }
 800         }
 801     };
 802
 803     for (i = 0; i < h; i++) {
 804         for (j = 0; j < w; j += 4) {
 805             tmp[32 * i + j + 0] = (pix1[j + 0] - pix2[j + 0]) << 4;
 806             tmp[32 * i + j + 1] = (pix1[j + 1] - pix2[j + 1]) << 4;
 807             tmp[32 * i + j + 2] = (pix1[j + 2] - pix2[j + 2]) << 4;
 808             tmp[32 * i + j + 3] = (pix1[j + 3] - pix2[j + 3]) << 4;
 809         }
 810         pix1 += line_size;
 811         pix2 += line_size;
 812     }
 813
 814     ff_spatial_dwt(tmp, tmp2, w, h, 32, type, dec_count);
 815
 816     s = 0;
 817     av_assert1(w == h);
 818     for (level = 0; level < dec_count; level++)
 819         for (ori = level ? 1 : 0; ori < 4; ori++) {
 820             int size   = w >> (dec_count - level);
 821             int sx     = (ori & 1) ? size : 0;
 822             int stride = 32 << (dec_count - level);
 823             int sy     = (ori & 2) ? stride >> 1 : 0;
 824
 825             for (i = 0; i < size; i++)
 826                 for (j = 0; j < size; j++) {
 827                     int v = tmp[sx + sy + i * stride + j] *
 828                             scale[type][dec_count - 3][level][ori];
 829                     s += FFABS(v);
 830                 }
 831         }
 832     av_assert1(s >= 0);
 833     return s >> 9;
 834 }
 835
 836 static int w53_8_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 837 {
 838     return w_c(v, pix1, pix2, line_size, 8, h, 1);
 839 }
 840
 841 static int w97_8_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 842 {
 843     return w_c(v, pix1, pix2, line_size, 8, h, 0);
 844 }
 845
 846 static int w53_16_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 847 {
 848     return w_c(v, pix1, pix2, line_size, 16, h, 1);
 849 }
 850
 851 static int w97_16_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 852 {
 853     return w_c(v, pix1, pix2, line_size, 16, h, 0);
 854 }
 855
 856 int ff_w53_32_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 857 {
 858     return w_c(v, pix1, pix2, line_size, 32, h, 1);
 859 }
 860
 861 int ff_w97_32_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
 862 {
 863     return w_c(v, pix1, pix2, line_size, 32, h, 0);
 864 }
 865
 866 void ff_dsputil_init_dwt(DSPContext *c)
 867 {
 868     c->w53[0] = w53_16_c;
 869     c->w53[1] = w53_8_c;
 870     c->w97[0] = w97_16_c;
 871     c->w97[1] = w97_8_c;
 872 }
 873
 874 void ff_dwt_init(DWTContext *c)
 875 {
 876     c->vertical_compose97i   = ff_snow_vertical_compose97i;
 877     c->horizontal_compose97i = ff_snow_horizontal_compose97i;
 878     c->inner_add_yblock      = ff_snow_inner_add_yblock;
 879
 880     if (HAVE_MMX)
 881         ff_dwt_init_x86(c);
 882 }
 883
 884
 885 static av_always_inline
 886 void interleave(IDWTELEM *dst, IDWTELEM *src0, IDWTELEM *src1, int w2, int add, int shift)
 887 {
 888     int i;
 889     for (i = 0; i < w2; i++) {
 890         dst[2*i  ] = (src0[i] + add) >> shift;
 891         dst[2*i+1] = (src1[i] + add) >> shift;
 892     }
 893 }
 894
 895 static void horizontal_compose_dirac53i(IDWTELEM *b, IDWTELEM *temp, int w)
 896 {
 897     const int w2 = w >> 1;
 898     int x;
 899
 900     temp[0] = COMPOSE_53iL0(b[w2], b[0], b[w2]);
 901     for (x = 1; x < w2; x++) {
 902         temp[x     ] = COMPOSE_53iL0     (b[x+w2-1], b[x     ], b[x+w2]);
 903         temp[x+w2-1] = COMPOSE_DIRAC53iH0(temp[x-1], b[x+w2-1], temp[x]);
 904     }
 905     temp[w-1] = COMPOSE_DIRAC53iH0(temp[w2-1], b[w-1], temp[w2-1]);
 906
 907     interleave(b, temp, temp+w2, w2, 1, 1);
 908 }
 909
 910 static void horizontal_compose_dd97i(IDWTELEM *b, IDWTELEM *tmp, int w)
 911 {
 912     const int w2 = w >> 1;
 913     int x;
 914
 915     tmp[0] = COMPOSE_53iL0(b[w2], b[0], b[w2]);
 916     for (x = 1; x < w2; x++)
 917         tmp[x] = COMPOSE_53iL0(b[x+w2-1], b[x], b[x+w2]);
 918
 919     // extend the edges
 920     tmp[-1]   = tmp[0];
 921     tmp[w2+1] = tmp[w2] = tmp[w2-1];
 922
 923     for (x = 0; x < w2; x++) {
 924         b[2*x  ] = (tmp[x] + 1)>>1;
 925         b[2*x+1] = (COMPOSE_DD97iH0(tmp[x-1], tmp[x], b[x+w2], tmp[x+1], tmp[x+2]) + 1)>>1;
 926     }
 927 }
 928
 929 static void horizontal_compose_dd137i(IDWTELEM *b, IDWTELEM *tmp, int w)
 930 {
 931     const int w2 = w >> 1;
 932     int x;
 933
 934     tmp[0] = COMPOSE_DD137iL0(b[w2], b[w2], b[0], b[w2  ], b[w2+1]);
 935     tmp[1] = COMPOSE_DD137iL0(b[w2], b[w2], b[1], b[w2+1], b[w2+2]);
 936     for (x = 2; x < w2-1; x++)
 937         tmp[x] = COMPOSE_DD137iL0(b[x+w2-2], b[x+w2-1], b[x], b[x+w2], b[x+w2+1]);
 938     tmp[w2-1] = COMPOSE_DD137iL0(b[w-3], b[w-2], b[w2-1], b[w-1], b[w-1]);
 939
 940     // extend the edges
 941     tmp[-1]   = tmp[0];
 942     tmp[w2+1] = tmp[w2] = tmp[w2-1];
 943
 944     for (x = 0; x < w2; x++) {
 945         b[2*x  ] = (tmp[x] + 1)>>1;
 946         b[2*x+1] = (COMPOSE_DD97iH0(tmp[x-1], tmp[x], b[x+w2], tmp[x+1], tmp[x+2]) + 1)>>1;
 947     }
 948 }
 949
 950 static av_always_inline
 951 void horizontal_compose_haari(IDWTELEM *b, IDWTELEM *temp, int w, int shift)
 952 {
 953     const int w2 = w >> 1;
 954     int x;
 955
 956     for (x = 0; x < w2; x++) {
 957         temp[x   ] = COMPOSE_HAARiL0(b[x   ], b[x+w2]);
 958         temp[x+w2] = COMPOSE_HAARiH0(b[x+w2], temp[x]);
 959     }
 960
 961     interleave(b, temp, temp+w2, w2, shift, shift);
 962 }
 963
 964 static void horizontal_compose_haar0i(IDWTELEM *b, IDWTELEM *temp, int w)
 965 {
 966     horizontal_compose_haari(b, temp, w, 0);
 967 }
 968
 969 static void horizontal_compose_haar1i(IDWTELEM *b, IDWTELEM *temp, int w)
 970 {
 971     horizontal_compose_haari(b, temp, w, 1);
 972 }
 973
 974 static void horizontal_compose_fidelityi(IDWTELEM *b, IDWTELEM *tmp, int w)
 975 {
 976     const int w2 = w >> 1;
 977     int i, x;
 978     IDWTELEM v[8];
 979
 980     for (x = 0; x < w2; x++) {
 981         for (i = 0; i < 8; i++)
 982             v[i] = b[av_clip(x-3+i, 0, w2-1)];
 983         tmp[x] = COMPOSE_FIDELITYiH0(v[0], v[1], v[2], v[3], b[x+w2], v[4], v[5], v[6], v[7]);
 984     }
 985
 986     for (x = 0; x < w2; x++) {
 987         for (i = 0; i < 8; i++)
 988             v[i] = tmp[av_clip(x-4+i, 0, w2-1)];
 989         tmp[x+w2] = COMPOSE_FIDELITYiL0(v[0], v[1], v[2], v[3], b[x], v[4], v[5], v[6], v[7]);
 990     }
 991
 992     interleave(b, tmp+w2, tmp, w2, 0, 0);
 993 }
 994
 995 static void horizontal_compose_daub97i(IDWTELEM *b, IDWTELEM *temp, int w)
 996 {
 997     const int w2 = w >> 1;
 998     int x, b0, b1, b2;
 999
1000     temp[0] = COMPOSE_DAUB97iL1(b[w2], b[0], b[w2]);
1001     for (x = 1; x < w2; x++) {
1002         temp[x     ] = COMPOSE_DAUB97iL1(b[x+w2-1], b[x     ], b[x+w2]);
1003         temp[x+w2-1] = COMPOSE_DAUB97iH1(temp[x-1], b[x+w2-1], temp[x]);
1004     }
1005     temp[w-1] = COMPOSE_DAUB97iH1(temp[w2-1], b[w-1], temp[w2-1]);
1006
1007     // second stage combined with interleave and shift
1008     b0 = b2 = COMPOSE_DAUB97iL0(temp[w2], temp[0], temp[w2]);
1009     b[0] = (b0 + 1) >> 1;
1010     for (x = 1; x < w2; x++) {
1011         b2 = COMPOSE_DAUB97iL0(temp[x+w2-1], temp[x     ], temp[x+w2]);
1012         b1 = COMPOSE_DAUB97iH0(          b0, temp[x+w2-1], b2        );
1013         b[2*x-1] = (b1 + 1) >> 1;
1014         b[2*x  ] = (b2 + 1) >> 1;
1015         b0 = b2;
1016     }
1017     b[w-1] = (COMPOSE_DAUB97iH0(b2, temp[w-1], b2) + 1) >> 1;
1018 }
1019
1020 static void vertical_compose_dirac53iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1021 {
1022     int i;
1023
1024     for(i=0; i<width; i++){
1025         b1[i] = COMPOSE_DIRAC53iH0(b0[i], b1[i], b2[i]);
1026     }
1027 }
1028
1029 static void vertical_compose_dd97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
1030                                   IDWTELEM *b3, IDWTELEM *b4, int width)
1031 {
1032     int i;
1033
1034     for(i=0; i<width; i++){
1035         b2[i] = COMPOSE_DD97iH0(b0[i], b1[i], b2[i], b3[i], b4[i]);
1036     }
1037 }
1038
1039 static void vertical_compose_dd137iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
1040                                       IDWTELEM *b3, IDWTELEM *b4, int width)
1041 {
1042     int i;
1043
1044     for(i=0; i<width; i++){
1045         b2[i] = COMPOSE_DD137iL0(b0[i], b1[i], b2[i], b3[i], b4[i]);
1046     }
1047 }
1048
1049 static void vertical_compose_haar(IDWTELEM *b0, IDWTELEM *b1, int width)
1050 {
1051     int i;
1052
1053     for (i = 0; i < width; i++) {
1054         b0[i] = COMPOSE_HAARiL0(b0[i], b1[i]);
1055         b1[i] = COMPOSE_HAARiH0(b1[i], b0[i]);
1056     }
1057 }
1058
1059 static void vertical_compose_fidelityiH0(IDWTELEM *dst, IDWTELEM *b[8], int width)
1060 {
1061     int i;
1062
1063     for(i=0; i<width; i++){
1064         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]);
1065     }
1066 }
1067
1068 static void vertical_compose_fidelityiL0(IDWTELEM *dst, IDWTELEM *b[8], int width)
1069 {
1070     int i;
1071
1072     for(i=0; i<width; i++){
1073         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]);
1074     }
1075 }
1076
1077 static void vertical_compose_daub97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1078 {
1079     int i;
1080
1081     for(i=0; i<width; i++){
1082         b1[i] = COMPOSE_DAUB97iH0(b0[i], b1[i], b2[i]);
1083     }
1084 }
1085
1086 static void vertical_compose_daub97iH1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1087 {
1088     int i;
1089
1090     for(i=0; i<width; i++){
1091         b1[i] = COMPOSE_DAUB97iH1(b0[i], b1[i], b2[i]);
1092     }
1093 }
1094
1095 static void vertical_compose_daub97iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1096 {
1097     int i;
1098
1099     for(i=0; i<width; i++){
1100         b1[i] = COMPOSE_DAUB97iL0(b0[i], b1[i], b2[i]);
1101     }
1102 }
1103
1104 static void vertical_compose_daub97iL1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1105 {
1106     int i;
1107
1108     for(i=0; i<width; i++){
1109         b1[i] = COMPOSE_DAUB97iL1(b0[i], b1[i], b2[i]);
1110     }
1111 }
1112
1113
1114 static void spatial_compose_dd97i_dy(DWTContext *d, int level, int width, int height, int stride)
1115 {
1116     vertical_compose_3tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1117     vertical_compose_5tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1118     DWTCompose *cs = d->cs + level;
1119
1120     int i, y = cs->y;
1121     IDWTELEM *b[8];
1122     for (i = 0; i < 6; i++)
1123         b[i] = cs->b[i];
1124     b[6] = d->buffer + av_clip(y+5, 0, height-2)*stride;
1125     b[7] = d->buffer + av_clip(y+6, 1, height-1)*stride;
1126
1127         if(y+5<(unsigned)height) vertical_compose_l0(      b[5], b[6], b[7],       width);
1128         if(y+1<(unsigned)height) vertical_compose_h0(b[0], b[2], b[3], b[4], b[6], width);
1129
1130         if(y-1<(unsigned)height) d->horizontal_compose(b[0], d->temp, width);
1131         if(y+0<(unsigned)height) d->horizontal_compose(b[1], d->temp, width);
1132
1133     for (i = 0; i < 6; i++)
1134         cs->b[i] = b[i+2];
1135     cs->y += 2;
1136 }
1137
1138 static void spatial_compose_dirac53i_dy(DWTContext *d, int level, int width, int height, int stride)
1139 {
1140     vertical_compose_3tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1141     vertical_compose_3tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1142     DWTCompose *cs = d->cs + level;
1143
1144     int y= cs->y;
1145     IDWTELEM *b[4] = { cs->b[0], cs->b[1] };
1146     b[2] = d->buffer + mirror(y+1, height-1)*stride;
1147     b[3] = d->buffer + mirror(y+2, height-1)*stride;
1148
1149         if(y+1<(unsigned)height) vertical_compose_l0(b[1], b[2], b[3], width);
1150         if(y+0<(unsigned)height) vertical_compose_h0(b[0], b[1], b[2], width);
1151
1152         if(y-1<(unsigned)height) d->horizontal_compose(b[0], d->temp, width);
1153         if(y+0<(unsigned)height) d->horizontal_compose(b[1], d->temp, width);
1154
1155     cs->b[0] = b[2];
1156     cs->b[1] = b[3];
1157     cs->y += 2;
1158 }
1159
1160
1161 static void spatial_compose_dd137i_dy(DWTContext *d, int level, int width, int height, int stride)
1162 {
1163     vertical_compose_5tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1164     vertical_compose_5tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1165     DWTCompose *cs = d->cs + level;
1166
1167     int i, y = cs->y;
1168     IDWTELEM *b[10];
1169     for (i = 0; i < 8; i++)
1170         b[i] = cs->b[i];
1171     b[8] = d->buffer + av_clip(y+7, 0, height-2)*stride;
1172     b[9] = d->buffer + av_clip(y+8, 1, height-1)*stride;
1173
1174         if(y+5<(unsigned)height) vertical_compose_l0(b[3], b[5], b[6], b[7], b[9], width);
1175         if(y+1<(unsigned)height) vertical_compose_h0(b[0], b[2], b[3], b[4], b[6], width);
1176
1177         if(y-1<(unsigned)height) d->horizontal_compose(b[0], d->temp, width);
1178         if(y+0<(unsigned)height) d->horizontal_compose(b[1], d->temp, width);
1179
1180     for (i = 0; i < 8; i++)
1181         cs->b[i] = b[i+2];
1182     cs->y += 2;
1183 }
1184
1185 // haar makes the assumption that height is even (always true for dirac)
1186 static void spatial_compose_haari_dy(DWTContext *d, int level, int width, int height, int stride)
1187 {
1188     vertical_compose_2tap vertical_compose = (void*)d->vertical_compose;
1189     int y = d->cs[level].y;
1190     IDWTELEM *b0 = d->buffer + (y-1)*stride;
1191     IDWTELEM *b1 = d->buffer + (y  )*stride;
1192
1193     vertical_compose(b0, b1, width);
1194     d->horizontal_compose(b0, d->temp, width);
1195     d->horizontal_compose(b1, d->temp, width);
1196
1197     d->cs[level].y += 2;
1198 }
1199
1200 // Don't do sliced idwt for fidelity; the 9 tap filter makes it a bit annoying
1201 // Fortunately, this filter isn't used in practice.
1202 static void spatial_compose_fidelity(DWTContext *d, int level, int width, int height, int stride)
1203 {
1204     vertical_compose_9tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1205     vertical_compose_9tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1206     int i, y;
1207     IDWTELEM *b[8];
1208
1209     for (y = 1; y < height; y += 2) {
1210         for (i = 0; i < 8; i++)
1211             b[i] = d->buffer + av_clip((y-7 + 2*i), 0, height-2)*stride;
1212         vertical_compose_h0(d->buffer + y*stride, b, width);
1213     }
1214
1215     for (y = 0; y < height; y += 2) {
1216         for (i = 0; i < 8; i++)
1217             b[i] = d->buffer + av_clip((y-7 + 2*i), 1, height-1)*stride;
1218         vertical_compose_l0(d->buffer + y*stride, b, width);
1219     }
1220
1221     for (y = 0; y < height; y++)
1222         d->horizontal_compose(d->buffer + y*stride, d->temp, width);
1223
1224     d->cs[level].y = height+1;
1225 }
1226
1227 static void spatial_compose_daub97i_dy(DWTContext *d, int level, int width, int height, int stride)
1228 {
1229     vertical_compose_3tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1230     vertical_compose_3tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1231     vertical_compose_3tap vertical_compose_l1 = (void*)d->vertical_compose_l1;
1232     vertical_compose_3tap vertical_compose_h1 = (void*)d->vertical_compose_h1;
1233     DWTCompose *cs = d->cs + level;
1234
1235     int i, y = cs->y;
1236     IDWTELEM *b[6];
1237     for (i = 0; i < 4; i++)
1238         b[i] = cs->b[i];
1239     b[4] = d->buffer + mirror(y+3, height-1)*stride;
1240     b[5] = d->buffer + mirror(y+4, height-1)*stride;
1241
1242         if(y+3<(unsigned)height) vertical_compose_l1(b[3], b[4], b[5], width);
1243         if(y+2<(unsigned)height) vertical_compose_h1(b[2], b[3], b[4], width);
1244         if(y+1<(unsigned)height) vertical_compose_l0(b[1], b[2], b[3], width);
1245         if(y+0<(unsigned)height) vertical_compose_h0(b[0], b[1], b[2], 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 < 4; i++)
1251         cs->b[i] = b[i+2];
1252     cs->y += 2;
1253 }
1254
1255
1256 static void spatial_compose97i_init2(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1257 {
1258     cs->b[0] = buffer + mirror(-3-1, height-1)*stride;
1259     cs->b[1] = buffer + mirror(-3  , height-1)*stride;
1260     cs->b[2] = buffer + mirror(-3+1, height-1)*stride;
1261     cs->b[3] = buffer + mirror(-3+2, height-1)*stride;
1262     cs->y = -3;
1263 }
1264
1265 static void spatial_compose53i_init2(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1266 {
1267     cs->b[0] = buffer + mirror(-1-1, height-1)*stride;
1268     cs->b[1] = buffer + mirror(-1  , height-1)*stride;
1269     cs->y = -1;
1270 }
1271
1272 static void spatial_compose_dd97i_init(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1273 {
1274     cs->b[0] = buffer + av_clip(-5-1, 0, height-2)*stride;
1275     cs->b[1] = buffer + av_clip(-5  , 1, height-1)*stride;
1276     cs->b[2] = buffer + av_clip(-5+1, 0, height-2)*stride;
1277     cs->b[3] = buffer + av_clip(-5+2, 1, height-1)*stride;
1278     cs->b[4] = buffer + av_clip(-5+3, 0, height-2)*stride;
1279     cs->b[5] = buffer + av_clip(-5+4, 1, height-1)*stride;
1280     cs->y = -5;
1281 }
1282
1283 static void spatial_compose_dd137i_init(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1284 {
1285     cs->b[0] = buffer + av_clip(-5-1, 0, height-2)*stride;
1286     cs->b[1] = buffer + av_clip(-5  , 1, height-1)*stride;
1287     cs->b[2] = buffer + av_clip(-5+1, 0, height-2)*stride;
1288     cs->b[3] = buffer + av_clip(-5+2, 1, height-1)*stride;
1289     cs->b[4] = buffer + av_clip(-5+3, 0, height-2)*stride;
1290     cs->b[5] = buffer + av_clip(-5+4, 1, height-1)*stride;
1291     cs->b[6] = buffer + av_clip(-5+5, 0, height-2)*stride;
1292     cs->b[7] = buffer + av_clip(-5+6, 1, height-1)*stride;
1293     cs->y = -5;
1294 }
1295
1296 int ff_spatial_idwt_init2(DWTContext *d, IDWTELEM *buffer, int width, int height,
1297                           int stride, enum dwt_type type, int decomposition_count,
1298                           IDWTELEM *temp)
1299 {
1300     int level;
1301
1302     d->buffer = buffer;
1303     d->width = width;
1304     d->height = height;
1305     d->stride = stride;
1306     d->decomposition_count = decomposition_count;
1307     d->temp = temp + 8;
1308
1309     for(level=decomposition_count-1; level>=0; level--){
1310         int hl = height >> level;
1311         int stride_l = stride << level;
1312
1313         switch(type){
1314         case DWT_DIRAC_DD9_7:
1315             spatial_compose_dd97i_init(d->cs+level, buffer, hl, stride_l);
1316             break;
1317         case DWT_DIRAC_LEGALL5_3:
1318             spatial_compose53i_init2(d->cs+level, buffer, hl, stride_l);
1319             break;
1320         case DWT_DIRAC_DD13_7:
1321             spatial_compose_dd137i_init(d->cs+level, buffer, hl, stride_l);
1322             break;
1323         case DWT_DIRAC_HAAR0:
1324         case DWT_DIRAC_HAAR1:
1325             d->cs[level].y = 1;
1326             break;
1327         case DWT_DIRAC_DAUB9_7:
1328             spatial_compose97i_init2(d->cs+level, buffer, hl, stride_l);
1329             break;
1330         default:
1331             d->cs[level].y = 0;
1332             break;
1333         }
1334     }
1335
1336     switch (type) {
1337     case DWT_DIRAC_DD9_7:
1338         d->spatial_compose = spatial_compose_dd97i_dy;
1339         d->vertical_compose_l0 = (void*)vertical_compose53iL0;
1340         d->vertical_compose_h0 = (void*)vertical_compose_dd97iH0;
1341         d->horizontal_compose = horizontal_compose_dd97i;
1342         d->support = 7;
1343         break;
1344     case DWT_DIRAC_LEGALL5_3:
1345         d->spatial_compose = spatial_compose_dirac53i_dy;
1346         d->vertical_compose_l0 = (void*)vertical_compose53iL0;
1347         d->vertical_compose_h0 = (void*)vertical_compose_dirac53iH0;
1348         d->horizontal_compose = horizontal_compose_dirac53i;
1349         d->support = 3;
1350         break;
1351     case DWT_DIRAC_DD13_7:
1352         d->spatial_compose = spatial_compose_dd137i_dy;
1353         d->vertical_compose_l0 = (void*)vertical_compose_dd137iL0;
1354         d->vertical_compose_h0 = (void*)vertical_compose_dd97iH0;
1355         d->horizontal_compose = horizontal_compose_dd137i;
1356         d->support = 7;
1357         break;
1358     case DWT_DIRAC_HAAR0:
1359     case DWT_DIRAC_HAAR1:
1360         d->spatial_compose = spatial_compose_haari_dy;
1361         d->vertical_compose = (void*)vertical_compose_haar;
1362         if (type == DWT_DIRAC_HAAR0)
1363             d->horizontal_compose = horizontal_compose_haar0i;
1364         else
1365             d->horizontal_compose = horizontal_compose_haar1i;
1366         d->support = 1;
1367         break;
1368     case DWT_DIRAC_FIDELITY:
1369         d->spatial_compose = spatial_compose_fidelity;
1370         d->vertical_compose_l0 = (void*)vertical_compose_fidelityiL0;
1371         d->vertical_compose_h0 = (void*)vertical_compose_fidelityiH0;
1372         d->horizontal_compose = horizontal_compose_fidelityi;
1373         break;
1374     case DWT_DIRAC_DAUB9_7:
1375         d->spatial_compose = spatial_compose_daub97i_dy;
1376         d->vertical_compose_l0 = (void*)vertical_compose_daub97iL0;
1377         d->vertical_compose_h0 = (void*)vertical_compose_daub97iH0;
1378         d->vertical_compose_l1 = (void*)vertical_compose_daub97iL1;
1379         d->vertical_compose_h1 = (void*)vertical_compose_daub97iH1;
1380         d->horizontal_compose = horizontal_compose_daub97i;
1381         d->support = 5;
1382         break;
1383     default:
1384         av_log(NULL, AV_LOG_ERROR, "Unknown wavelet type %d\n", type);
1385         return -1;
1386     }
1387
1388     if (HAVE_MMX) ff_spatial_idwt_init_mmx(d, type);
1389
1390     return 0;
1391 }
1392
1393 void ff_spatial_idwt_slice2(DWTContext *d, int y)
1394 {
1395     int level, support = d->support;
1396
1397     for (level = d->decomposition_count-1; level >= 0; level--) {
1398         int wl = d->width  >> level;
1399         int hl = d->height >> level;
1400         int stride_l = d->stride << level;
1401
1402         while (d->cs[level].y <= FFMIN((y>>level)+support, hl))
1403             d->spatial_compose(d, level, wl, hl, stride_l);
1404     }
1405 }
1406
1407 int ff_spatial_idwt2(IDWTELEM *buffer, int width, int height, int stride,
1408                      enum dwt_type type, int decomposition_count, IDWTELEM *temp)
1409 {
1410     DWTContext d;
1411     int y;
1412
1413     if (ff_spatial_idwt_init2(&d, buffer, width, height, stride, type, decomposition_count, temp))
1414         return -1;
1415
1416     for (y = 0; y < d.height; y += 4)
1417         ff_spatial_idwt_slice2(&d, y);
1418
1419     return 0;
1420 }