d36d25d1a6057526b438ba719ad07be85b9c1f7c
[ffmpeg.git] / 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 void ff_snow_horizontal_compose97i(IDWTELEM *b, IDWTELEM *temp, int width)
472 {
473     const int w2 = (width + 1) >> 1;
474     int x;
475
476     temp[0] = b[0] - ((3 * b[w2] + 2) >> 2);
477     for (x = 1; x < (width >> 1); x++) {
478         temp[2 * x]     = b[x] - ((3 * (b[x + w2 - 1] + b[x + w2]) + 4) >> 3);
479         temp[2 * x - 1] = b[x + w2 - 1] - temp[2 * x - 2] - temp[2 * x];
480     }
481     if (width & 1) {
482         temp[2 * x]     = b[x] - ((3 * b[x + w2 - 1] + 2) >> 2);
483         temp[2 * x - 1] = b[x + w2 - 1] - temp[2 * x - 2] - temp[2 * x];
484     } else
485         temp[2 * x - 1] = b[x + w2 - 1] - 2 * temp[2 * x - 2];
486
487     b[0] = temp[0] + ((2 * temp[0] + temp[1] + 4) >> 3);
488     for (x = 2; x < width - 1; x += 2) {
489         b[x]     = temp[x] + ((4 * temp[x] + temp[x - 1] + temp[x + 1] + 8) >> 4);
490         b[x - 1] = temp[x - 1] + ((3 * (b[x - 2] + b[x])) >> 1);
491     }
492     if (width & 1) {
493         b[x]     = temp[x] + ((2 * temp[x] + temp[x - 1] + 4) >> 3);
494         b[x - 1] = temp[x - 1] + ((3 * (b[x - 2] + b[x])) >> 1);
495     } else
496         b[x - 1] = temp[x - 1] + 3 * b[x - 2];
497 }
498
499 static void vertical_compose97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
500                                   int width)
501 {
502     int i;
503
504     for (i = 0; i < width; i++)
505         b1[i] += (W_AM * (b0[i] + b2[i]) + W_AO) >> W_AS;
506 }
507
508 static void vertical_compose97iH1(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_CM * (b0[i] + b2[i]) + W_CO) >> W_CS;
515 }
516
517 static void vertical_compose97iL0(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_BM * (b0[i] + b2[i]) + 4 * b1[i] + W_BO) >> W_BS;
524 }
525
526 static void vertical_compose97iL1(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_DM * (b0[i] + b2[i]) + W_DO) >> W_DS;
533 }
534
535 void ff_snow_vertical_compose97i(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
536                                  IDWTELEM *b3, IDWTELEM *b4, IDWTELEM *b5,
537                                  int width)
538 {
539     int i;
540
541     for (i = 0; i < width; i++) {
542         b4[i] -= (W_DM * (b3[i] + b5[i]) + W_DO) >> W_DS;
543         b3[i] -= (W_CM * (b2[i] + b4[i]) + W_CO) >> W_CS;
544         b2[i] += (W_BM * (b1[i] + b3[i]) + 4 * b2[i] + W_BO) >> W_BS;
545         b1[i] += (W_AM * (b0[i] + b2[i]) + W_AO) >> W_AS;
546     }
547 }
548
549 static void spatial_compose97i_buffered_init(DWTCompose *cs, slice_buffer *sb,
550                                              int height, int stride_line)
551 {
552     cs->b0 = slice_buffer_get_line(sb, mirror(-3 - 1, height - 1) * stride_line);
553     cs->b1 = slice_buffer_get_line(sb, mirror(-3,     height - 1) * stride_line);
554     cs->b2 = slice_buffer_get_line(sb, mirror(-3 + 1, height - 1) * stride_line);
555     cs->b3 = slice_buffer_get_line(sb, mirror(-3 + 2, height - 1) * stride_line);
556     cs->y  = -3;
557 }
558
559 static void spatial_compose97i_init(DWTCompose *cs, IDWTELEM *buffer, int height,
560                                     int stride)
561 {
562     cs->b0 = buffer + mirror(-3 - 1, height - 1) * stride;
563     cs->b1 = buffer + mirror(-3,     height - 1) * stride;
564     cs->b2 = buffer + mirror(-3 + 1, height - 1) * stride;
565     cs->b3 = buffer + mirror(-3 + 2, height - 1) * stride;
566     cs->y  = -3;
567 }
568
569 static void spatial_compose97i_dy_buffered(DWTContext *dsp, DWTCompose *cs,
570                                            slice_buffer * sb, IDWTELEM *temp,
571                                            int width, int height,
572                                            int stride_line)
573 {
574     int y = cs->y;
575
576     IDWTELEM *b0 = cs->b0;
577     IDWTELEM *b1 = cs->b1;
578     IDWTELEM *b2 = cs->b2;
579     IDWTELEM *b3 = cs->b3;
580     IDWTELEM *b4 = slice_buffer_get_line(sb,
581                                          mirror(y + 3, height - 1) *
582                                          stride_line);
583     IDWTELEM *b5 = slice_buffer_get_line(sb,
584                                          mirror(y + 4, height - 1) *
585                                          stride_line);
586
587     if (y > 0 && y + 4 < height) {
588         dsp->vertical_compose97i(b0, b1, b2, b3, b4, b5, width);
589     } else {
590         if (y + 3 < (unsigned)height)
591             vertical_compose97iL1(b3, b4, b5, width);
592         if (y + 2 < (unsigned)height)
593             vertical_compose97iH1(b2, b3, b4, width);
594         if (y + 1 < (unsigned)height)
595             vertical_compose97iL0(b1, b2, b3, width);
596         if (y + 0 < (unsigned)height)
597             vertical_compose97iH0(b0, b1, b2, width);
598     }
599
600     if (y - 1 < (unsigned)height)
601         dsp->horizontal_compose97i(b0, temp, width);
602     if (y + 0 < (unsigned)height)
603         dsp->horizontal_compose97i(b1, temp, width);
604
605     cs->b0  = b2;
606     cs->b1  = b3;
607     cs->b2  = b4;
608     cs->b3  = b5;
609     cs->y  += 2;
610 }
611
612 static void spatial_compose97i_dy(DWTCompose *cs, IDWTELEM *buffer,
613                                   IDWTELEM *temp, int width, int height,
614                                   int stride)
615 {
616     int y        = cs->y;
617     IDWTELEM *b0 = cs->b0;
618     IDWTELEM *b1 = cs->b1;
619     IDWTELEM *b2 = cs->b2;
620     IDWTELEM *b3 = cs->b3;
621     IDWTELEM *b4 = buffer + mirror(y + 3, height - 1) * stride;
622     IDWTELEM *b5 = buffer + mirror(y + 4, height - 1) * stride;
623
624     if (y + 3 < (unsigned)height)
625         vertical_compose97iL1(b3, b4, b5, width);
626     if (y + 2 < (unsigned)height)
627         vertical_compose97iH1(b2, b3, b4, width);
628     if (y + 1 < (unsigned)height)
629         vertical_compose97iL0(b1, b2, b3, width);
630     if (y + 0 < (unsigned)height)
631         vertical_compose97iH0(b0, b1, b2, width);
632
633     if (y - 1 < (unsigned)height)
634         ff_snow_horizontal_compose97i(b0, temp, width);
635     if (y + 0 < (unsigned)height)
636         ff_snow_horizontal_compose97i(b1, temp, width);
637
638     cs->b0  = b2;
639     cs->b1  = b3;
640     cs->b2  = b4;
641     cs->b3  = b5;
642     cs->y  += 2;
643 }
644
645 void ff_spatial_idwt_buffered_init(DWTCompose *cs, slice_buffer *sb, int width,
646                                    int height, int stride_line, int type,
647                                    int decomposition_count)
648 {
649     int level;
650     for (level = decomposition_count - 1; level >= 0; level--) {
651         switch (type) {
652         case DWT_97:
653             spatial_compose97i_buffered_init(cs + level, sb, height >> level,
654                                              stride_line << level);
655             break;
656         case DWT_53:
657             spatial_compose53i_buffered_init(cs + level, sb, height >> level,
658                                              stride_line << level);
659             break;
660         }
661     }
662 }
663
664 void ff_spatial_idwt_buffered_slice(DWTContext *dsp, DWTCompose *cs,
665                                     slice_buffer *slice_buf, IDWTELEM *temp,
666                                     int width, int height, int stride_line,
667                                     int type, int decomposition_count, int y)
668 {
669     const int support = type == 1 ? 3 : 5;
670     int level;
671     if (type == 2)
672         return;
673
674     for (level = decomposition_count - 1; level >= 0; level--)
675         while (cs[level].y <= FFMIN((y >> level) + support, height >> level)) {
676             switch (type) {
677             case DWT_97:
678                 spatial_compose97i_dy_buffered(dsp, cs + level, slice_buf, temp,
679                                                width >> level,
680                                                height >> level,
681                                                stride_line << level);
682                 break;
683             case DWT_53:
684                 spatial_compose53i_dy_buffered(cs + level, slice_buf, temp,
685                                                width >> level,
686                                                height >> level,
687                                                stride_line << level);
688                 break;
689             }
690         }
691 }
692
693 static void ff_spatial_idwt_init(DWTCompose *cs, IDWTELEM *buffer, int width,
694                                  int height, int stride, int type,
695                                  int decomposition_count)
696 {
697     int level;
698     for (level = decomposition_count - 1; level >= 0; level--) {
699         switch (type) {
700         case DWT_97:
701             spatial_compose97i_init(cs + level, buffer, height >> level,
702                                     stride << level);
703             break;
704         case DWT_53:
705             spatial_compose53i_init(cs + level, buffer, height >> level,
706                                     stride << level);
707             break;
708         }
709     }
710 }
711
712 static void ff_spatial_idwt_slice(DWTCompose *cs, IDWTELEM *buffer,
713                                   IDWTELEM *temp, int width, int height,
714                                   int stride, int type,
715                                   int decomposition_count, int y)
716 {
717     const int support = type == 1 ? 3 : 5;
718     int level;
719     if (type == 2)
720         return;
721
722     for (level = decomposition_count - 1; level >= 0; level--)
723         while (cs[level].y <= FFMIN((y >> level) + support, height >> level)) {
724             switch (type) {
725             case DWT_97:
726                 spatial_compose97i_dy(cs + level, buffer, temp, width >> level,
727                                       height >> level, stride << level);
728                 break;
729             case DWT_53:
730                 spatial_compose53i_dy(cs + level, buffer, temp, width >> level,
731                                       height >> level, stride << level);
732                 break;
733             }
734         }
735 }
736
737 void ff_spatial_idwt(IDWTELEM *buffer, IDWTELEM *temp, int width, int height,
738                      int stride, int type, int decomposition_count)
739 {
740     DWTCompose cs[MAX_DECOMPOSITIONS];
741     int y;
742     ff_spatial_idwt_init(cs, buffer, width, height, stride, type,
743                          decomposition_count);
744     for (y = 0; y < height; y += 4)
745         ff_spatial_idwt_slice(cs, buffer, temp, width, height, stride, type,
746                               decomposition_count, y);
747 }
748
749 static inline int w_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size,
750                       int w, int h, int type)
751 {
752     int s, i, j;
753     const int dec_count = w == 8 ? 3 : 4;
754     int tmp[32 * 32], tmp2[32];
755     int level, ori;
756     static const int scale[2][2][4][4] = {
757         {
758             { // 9/7 8x8 dec=3
759                 { 268, 239, 239, 213 },
760                 { 0,   224, 224, 152 },
761                 { 0,   135, 135, 110 },
762             },
763             { // 9/7 16x16 or 32x32 dec=4
764                 { 344, 310, 310, 280 },
765                 { 0,   320, 320, 228 },
766                 { 0,   175, 175, 136 },
767                 { 0,   129, 129, 102 },
768             }
769         },
770         {
771             { // 5/3 8x8 dec=3
772                 { 275, 245, 245, 218 },
773                 { 0,   230, 230, 156 },
774                 { 0,   138, 138, 113 },
775             },
776             { // 5/3 16x16 or 32x32 dec=4
777                 { 352, 317, 317, 286 },
778                 { 0,   328, 328, 233 },
779                 { 0,   180, 180, 140 },
780                 { 0,   132, 132, 105 },
781             }
782         }
783     };
784
785     for (i = 0; i < h; i++) {
786         for (j = 0; j < w; j += 4) {
787             tmp[32 * i + j + 0] = (pix1[j + 0] - pix2[j + 0]) << 4;
788             tmp[32 * i + j + 1] = (pix1[j + 1] - pix2[j + 1]) << 4;
789             tmp[32 * i + j + 2] = (pix1[j + 2] - pix2[j + 2]) << 4;
790             tmp[32 * i + j + 3] = (pix1[j + 3] - pix2[j + 3]) << 4;
791         }
792         pix1 += line_size;
793         pix2 += line_size;
794     }
795
796     ff_spatial_dwt(tmp, tmp2, w, h, 32, type, dec_count);
797
798     s = 0;
799     av_assert1(w == h);
800     for (level = 0; level < dec_count; level++)
801         for (ori = level ? 1 : 0; ori < 4; ori++) {
802             int size   = w >> (dec_count - level);
803             int sx     = (ori & 1) ? size : 0;
804             int stride = 32 << (dec_count - level);
805             int sy     = (ori & 2) ? stride >> 1 : 0;
806
807             for (i = 0; i < size; i++)
808                 for (j = 0; j < size; j++) {
809                     int v = tmp[sx + sy + i * stride + j] *
810                             scale[type][dec_count - 3][level][ori];
811                     s += FFABS(v);
812                 }
813         }
814     av_assert1(s >= 0);
815     return s >> 9;
816 }
817
818 static int w53_8_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
819 {
820     return w_c(v, pix1, pix2, line_size, 8, h, 1);
821 }
822
823 static int w97_8_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
824 {
825     return w_c(v, pix1, pix2, line_size, 8, h, 0);
826 }
827
828 static int w53_16_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
829 {
830     return w_c(v, pix1, pix2, line_size, 16, h, 1);
831 }
832
833 static int w97_16_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
834 {
835     return w_c(v, pix1, pix2, line_size, 16, h, 0);
836 }
837
838 int ff_w53_32_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
839 {
840     return w_c(v, pix1, pix2, line_size, 32, h, 1);
841 }
842
843 int ff_w97_32_c(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
844 {
845     return w_c(v, pix1, pix2, line_size, 32, h, 0);
846 }
847
848 void ff_dsputil_init_dwt(DSPContext *c)
849 {
850     c->w53[0] = w53_16_c;
851     c->w53[1] = w53_8_c;
852     c->w97[0] = w97_16_c;
853     c->w97[1] = w97_8_c;
854 }
855
856 void ff_dwt_init(DWTContext *c)
857 {
858     c->vertical_compose97i   = ff_snow_vertical_compose97i;
859     c->horizontal_compose97i = ff_snow_horizontal_compose97i;
860     c->inner_add_yblock      = ff_snow_inner_add_yblock;
861
862     if (HAVE_MMX)
863         ff_dwt_init_x86(c);
864 }
865
866
867 static av_always_inline
868 void interleave(IDWTELEM *dst, IDWTELEM *src0, IDWTELEM *src1, int w2, int add, int shift)
869 {
870     int i;
871     for (i = 0; i < w2; i++) {
872         dst[2*i  ] = (src0[i] + add) >> shift;
873         dst[2*i+1] = (src1[i] + add) >> shift;
874     }
875 }
876
877 static void horizontal_compose_dirac53i(IDWTELEM *b, IDWTELEM *temp, int w)
878 {
879     const int w2 = w >> 1;
880     int x;
881
882     temp[0] = COMPOSE_53iL0(b[w2], b[0], b[w2]);
883     for (x = 1; x < w2; x++) {
884         temp[x     ] = COMPOSE_53iL0     (b[x+w2-1], b[x     ], b[x+w2]);
885         temp[x+w2-1] = COMPOSE_DIRAC53iH0(temp[x-1], b[x+w2-1], temp[x]);
886     }
887     temp[w-1] = COMPOSE_DIRAC53iH0(temp[w2-1], b[w-1], temp[w2-1]);
888
889     interleave(b, temp, temp+w2, w2, 1, 1);
890 }
891
892 static void horizontal_compose_dd97i(IDWTELEM *b, IDWTELEM *tmp, int w)
893 {
894     const int w2 = w >> 1;
895     int x;
896
897     tmp[0] = COMPOSE_53iL0(b[w2], b[0], b[w2]);
898     for (x = 1; x < w2; x++)
899         tmp[x] = COMPOSE_53iL0(b[x+w2-1], b[x], b[x+w2]);
900
901     // extend the edges
902     tmp[-1]   = tmp[0];
903     tmp[w2+1] = tmp[w2] = tmp[w2-1];
904
905     for (x = 0; x < w2; x++) {
906         b[2*x  ] = (tmp[x] + 1)>>1;
907         b[2*x+1] = (COMPOSE_DD97iH0(tmp[x-1], tmp[x], b[x+w2], tmp[x+1], tmp[x+2]) + 1)>>1;
908     }
909 }
910
911 static void horizontal_compose_dd137i(IDWTELEM *b, IDWTELEM *tmp, int w)
912 {
913     const int w2 = w >> 1;
914     int x;
915
916     tmp[0] = COMPOSE_DD137iL0(b[w2], b[w2], b[0], b[w2  ], b[w2+1]);
917     tmp[1] = COMPOSE_DD137iL0(b[w2], b[w2], b[1], b[w2+1], b[w2+2]);
918     for (x = 2; x < w2-1; x++)
919         tmp[x] = COMPOSE_DD137iL0(b[x+w2-2], b[x+w2-1], b[x], b[x+w2], b[x+w2+1]);
920     tmp[w2-1] = COMPOSE_DD137iL0(b[w-3], b[w-2], b[w2-1], b[w-1], b[w-1]);
921
922     // extend the edges
923     tmp[-1]   = tmp[0];
924     tmp[w2+1] = tmp[w2] = tmp[w2-1];
925
926     for (x = 0; x < w2; x++) {
927         b[2*x  ] = (tmp[x] + 1)>>1;
928         b[2*x+1] = (COMPOSE_DD97iH0(tmp[x-1], tmp[x], b[x+w2], tmp[x+1], tmp[x+2]) + 1)>>1;
929     }
930 }
931
932 static av_always_inline
933 void horizontal_compose_haari(IDWTELEM *b, IDWTELEM *temp, int w, int shift)
934 {
935     const int w2 = w >> 1;
936     int x;
937
938     for (x = 0; x < w2; x++) {
939         temp[x   ] = COMPOSE_HAARiL0(b[x   ], b[x+w2]);
940         temp[x+w2] = COMPOSE_HAARiH0(b[x+w2], temp[x]);
941     }
942
943     interleave(b, temp, temp+w2, w2, shift, shift);
944 }
945
946 static void horizontal_compose_haar0i(IDWTELEM *b, IDWTELEM *temp, int w)
947 {
948     horizontal_compose_haari(b, temp, w, 0);
949 }
950
951 static void horizontal_compose_haar1i(IDWTELEM *b, IDWTELEM *temp, int w)
952 {
953     horizontal_compose_haari(b, temp, w, 1);
954 }
955
956 static void horizontal_compose_fidelityi(IDWTELEM *b, IDWTELEM *tmp, int w)
957 {
958     const int w2 = w >> 1;
959     int i, x;
960     IDWTELEM v[8];
961
962     for (x = 0; x < w2; x++) {
963         for (i = 0; i < 8; i++)
964             v[i] = b[av_clip(x-3+i, 0, w2-1)];
965         tmp[x] = COMPOSE_FIDELITYiH0(v[0], v[1], v[2], v[3], b[x+w2], v[4], v[5], v[6], v[7]);
966     }
967
968     for (x = 0; x < w2; x++) {
969         for (i = 0; i < 8; i++)
970             v[i] = tmp[av_clip(x-4+i, 0, w2-1)];
971         tmp[x+w2] = COMPOSE_FIDELITYiL0(v[0], v[1], v[2], v[3], b[x], v[4], v[5], v[6], v[7]);
972     }
973
974     interleave(b, tmp+w2, tmp, w2, 0, 0);
975 }
976
977 static void horizontal_compose_daub97i(IDWTELEM *b, IDWTELEM *temp, int w)
978 {
979     const int w2 = w >> 1;
980     int x, b0, b1, b2;
981
982     temp[0] = COMPOSE_DAUB97iL1(b[w2], b[0], b[w2]);
983     for (x = 1; x < w2; x++) {
984         temp[x     ] = COMPOSE_DAUB97iL1(b[x+w2-1], b[x     ], b[x+w2]);
985         temp[x+w2-1] = COMPOSE_DAUB97iH1(temp[x-1], b[x+w2-1], temp[x]);
986     }
987     temp[w-1] = COMPOSE_DAUB97iH1(temp[w2-1], b[w-1], temp[w2-1]);
988
989     // second stage combined with interleave and shift
990     b0 = b2 = COMPOSE_DAUB97iL0(temp[w2], temp[0], temp[w2]);
991     b[0] = (b0 + 1) >> 1;
992     for (x = 1; x < w2; x++) {
993         b2 = COMPOSE_DAUB97iL0(temp[x+w2-1], temp[x     ], temp[x+w2]);
994         b1 = COMPOSE_DAUB97iH0(          b0, temp[x+w2-1], b2        );
995         b[2*x-1] = (b1 + 1) >> 1;
996         b[2*x  ] = (b2 + 1) >> 1;
997         b0 = b2;
998     }
999     b[w-1] = (COMPOSE_DAUB97iH0(b2, temp[w-1], b2) + 1) >> 1;
1000 }
1001
1002 static void vertical_compose_dirac53iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1003 {
1004     int i;
1005
1006     for(i=0; i<width; i++){
1007         b1[i] = COMPOSE_DIRAC53iH0(b0[i], b1[i], b2[i]);
1008     }
1009 }
1010
1011 static void vertical_compose_dd97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
1012                                   IDWTELEM *b3, IDWTELEM *b4, int width)
1013 {
1014     int i;
1015
1016     for(i=0; i<width; i++){
1017         b2[i] = COMPOSE_DD97iH0(b0[i], b1[i], b2[i], b3[i], b4[i]);
1018     }
1019 }
1020
1021 static void vertical_compose_dd137iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2,
1022                                       IDWTELEM *b3, IDWTELEM *b4, int width)
1023 {
1024     int i;
1025
1026     for(i=0; i<width; i++){
1027         b2[i] = COMPOSE_DD137iL0(b0[i], b1[i], b2[i], b3[i], b4[i]);
1028     }
1029 }
1030
1031 static void vertical_compose_haar(IDWTELEM *b0, IDWTELEM *b1, int width)
1032 {
1033     int i;
1034
1035     for (i = 0; i < width; i++) {
1036         b0[i] = COMPOSE_HAARiL0(b0[i], b1[i]);
1037         b1[i] = COMPOSE_HAARiH0(b1[i], b0[i]);
1038     }
1039 }
1040
1041 static void vertical_compose_fidelityiH0(IDWTELEM *dst, IDWTELEM *b[8], int width)
1042 {
1043     int i;
1044
1045     for(i=0; i<width; i++){
1046         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]);
1047     }
1048 }
1049
1050 static void vertical_compose_fidelityiL0(IDWTELEM *dst, IDWTELEM *b[8], int width)
1051 {
1052     int i;
1053
1054     for(i=0; i<width; i++){
1055         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]);
1056     }
1057 }
1058
1059 static void vertical_compose_daub97iH0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1060 {
1061     int i;
1062
1063     for(i=0; i<width; i++){
1064         b1[i] = COMPOSE_DAUB97iH0(b0[i], b1[i], b2[i]);
1065     }
1066 }
1067
1068 static void vertical_compose_daub97iH1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1069 {
1070     int i;
1071
1072     for(i=0; i<width; i++){
1073         b1[i] = COMPOSE_DAUB97iH1(b0[i], b1[i], b2[i]);
1074     }
1075 }
1076
1077 static void vertical_compose_daub97iL0(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1078 {
1079     int i;
1080
1081     for(i=0; i<width; i++){
1082         b1[i] = COMPOSE_DAUB97iL0(b0[i], b1[i], b2[i]);
1083     }
1084 }
1085
1086 static void vertical_compose_daub97iL1(IDWTELEM *b0, IDWTELEM *b1, IDWTELEM *b2, int width)
1087 {
1088     int i;
1089
1090     for(i=0; i<width; i++){
1091         b1[i] = COMPOSE_DAUB97iL1(b0[i], b1[i], b2[i]);
1092     }
1093 }
1094
1095
1096 static void spatial_compose_dd97i_dy(DWTContext *d, int level, int width, int height, int stride)
1097 {
1098     vertical_compose_3tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1099     vertical_compose_5tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1100     DWTCompose *cs = d->cs + level;
1101
1102     int i, y = cs->y;
1103     IDWTELEM *b[8];
1104     for (i = 0; i < 6; i++)
1105         b[i] = cs->b[i];
1106     b[6] = d->buffer + av_clip(y+5, 0, height-2)*stride;
1107     b[7] = d->buffer + av_clip(y+6, 1, height-1)*stride;
1108
1109         if(y+5<(unsigned)height) vertical_compose_l0(      b[5], b[6], b[7],       width);
1110         if(y+1<(unsigned)height) vertical_compose_h0(b[0], b[2], b[3], b[4], b[6], width);
1111
1112         if(y-1<(unsigned)height) d->horizontal_compose(b[0], d->temp, width);
1113         if(y+0<(unsigned)height) d->horizontal_compose(b[1], d->temp, width);
1114
1115     for (i = 0; i < 6; i++)
1116         cs->b[i] = b[i+2];
1117     cs->y += 2;
1118 }
1119
1120 static void spatial_compose_dirac53i_dy(DWTContext *d, int level, int width, int height, int stride)
1121 {
1122     vertical_compose_3tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1123     vertical_compose_3tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1124     DWTCompose *cs = d->cs + level;
1125
1126     int y= cs->y;
1127     IDWTELEM *b[4] = { cs->b[0], cs->b[1] };
1128     b[2] = d->buffer + mirror(y+1, height-1)*stride;
1129     b[3] = d->buffer + mirror(y+2, height-1)*stride;
1130
1131         if(y+1<(unsigned)height) vertical_compose_l0(b[1], b[2], b[3], width);
1132         if(y+0<(unsigned)height) vertical_compose_h0(b[0], b[1], b[2], width);
1133
1134         if(y-1<(unsigned)height) d->horizontal_compose(b[0], d->temp, width);
1135         if(y+0<(unsigned)height) d->horizontal_compose(b[1], d->temp, width);
1136
1137     cs->b[0] = b[2];
1138     cs->b[1] = b[3];
1139     cs->y += 2;
1140 }
1141
1142
1143 static void spatial_compose_dd137i_dy(DWTContext *d, int level, int width, int height, int stride)
1144 {
1145     vertical_compose_5tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1146     vertical_compose_5tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1147     DWTCompose *cs = d->cs + level;
1148
1149     int i, y = cs->y;
1150     IDWTELEM *b[10];
1151     for (i = 0; i < 8; i++)
1152         b[i] = cs->b[i];
1153     b[8] = d->buffer + av_clip(y+7, 0, height-2)*stride;
1154     b[9] = d->buffer + av_clip(y+8, 1, height-1)*stride;
1155
1156         if(y+5<(unsigned)height) vertical_compose_l0(b[3], b[5], b[6], b[7], b[9], width);
1157         if(y+1<(unsigned)height) vertical_compose_h0(b[0], b[2], b[3], b[4], b[6], width);
1158
1159         if(y-1<(unsigned)height) d->horizontal_compose(b[0], d->temp, width);
1160         if(y+0<(unsigned)height) d->horizontal_compose(b[1], d->temp, width);
1161
1162     for (i = 0; i < 8; i++)
1163         cs->b[i] = b[i+2];
1164     cs->y += 2;
1165 }
1166
1167 // haar makes the assumption that height is even (always true for dirac)
1168 static void spatial_compose_haari_dy(DWTContext *d, int level, int width, int height, int stride)
1169 {
1170     vertical_compose_2tap vertical_compose = (void*)d->vertical_compose;
1171     int y = d->cs[level].y;
1172     IDWTELEM *b0 = d->buffer + (y-1)*stride;
1173     IDWTELEM *b1 = d->buffer + (y  )*stride;
1174
1175     vertical_compose(b0, b1, width);
1176     d->horizontal_compose(b0, d->temp, width);
1177     d->horizontal_compose(b1, d->temp, width);
1178
1179     d->cs[level].y += 2;
1180 }
1181
1182 // Don't do sliced idwt for fidelity; the 9 tap filter makes it a bit annoying
1183 // Fortunately, this filter isn't used in practice.
1184 static void spatial_compose_fidelity(DWTContext *d, int level, int width, int height, int stride)
1185 {
1186     vertical_compose_9tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1187     vertical_compose_9tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1188     int i, y;
1189     IDWTELEM *b[8];
1190
1191     for (y = 1; y < height; y += 2) {
1192         for (i = 0; i < 8; i++)
1193             b[i] = d->buffer + av_clip((y-7 + 2*i), 0, height-2)*stride;
1194         vertical_compose_h0(d->buffer + y*stride, b, width);
1195     }
1196
1197     for (y = 0; y < height; y += 2) {
1198         for (i = 0; i < 8; i++)
1199             b[i] = d->buffer + av_clip((y-7 + 2*i), 1, height-1)*stride;
1200         vertical_compose_l0(d->buffer + y*stride, b, width);
1201     }
1202
1203     for (y = 0; y < height; y++)
1204         d->horizontal_compose(d->buffer + y*stride, d->temp, width);
1205
1206     d->cs[level].y = height+1;
1207 }
1208
1209 static void spatial_compose_daub97i_dy(DWTContext *d, int level, int width, int height, int stride)
1210 {
1211     vertical_compose_3tap vertical_compose_l0 = (void*)d->vertical_compose_l0;
1212     vertical_compose_3tap vertical_compose_h0 = (void*)d->vertical_compose_h0;
1213     vertical_compose_3tap vertical_compose_l1 = (void*)d->vertical_compose_l1;
1214     vertical_compose_3tap vertical_compose_h1 = (void*)d->vertical_compose_h1;
1215     DWTCompose *cs = d->cs + level;
1216
1217     int i, y = cs->y;
1218     IDWTELEM *b[6];
1219     for (i = 0; i < 4; i++)
1220         b[i] = cs->b[i];
1221     b[4] = d->buffer + mirror(y+3, height-1)*stride;
1222     b[5] = d->buffer + mirror(y+4, height-1)*stride;
1223
1224         if(y+3<(unsigned)height) vertical_compose_l1(b[3], b[4], b[5], width);
1225         if(y+2<(unsigned)height) vertical_compose_h1(b[2], b[3], b[4], width);
1226         if(y+1<(unsigned)height) vertical_compose_l0(b[1], b[2], b[3], width);
1227         if(y+0<(unsigned)height) vertical_compose_h0(b[0], b[1], b[2], width);
1228
1229         if(y-1<(unsigned)height) d->horizontal_compose(b[0], d->temp, width);
1230         if(y+0<(unsigned)height) d->horizontal_compose(b[1], d->temp, width);
1231
1232     for (i = 0; i < 4; i++)
1233         cs->b[i] = b[i+2];
1234     cs->y += 2;
1235 }
1236
1237
1238 static void spatial_compose97i_init2(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1239 {
1240     cs->b[0] = buffer + mirror(-3-1, height-1)*stride;
1241     cs->b[1] = buffer + mirror(-3  , height-1)*stride;
1242     cs->b[2] = buffer + mirror(-3+1, height-1)*stride;
1243     cs->b[3] = buffer + mirror(-3+2, height-1)*stride;
1244     cs->y = -3;
1245 }
1246
1247 static void spatial_compose53i_init2(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1248 {
1249     cs->b[0] = buffer + mirror(-1-1, height-1)*stride;
1250     cs->b[1] = buffer + mirror(-1  , height-1)*stride;
1251     cs->y = -1;
1252 }
1253
1254 static void spatial_compose_dd97i_init(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1255 {
1256     cs->b[0] = buffer + av_clip(-5-1, 0, height-2)*stride;
1257     cs->b[1] = buffer + av_clip(-5  , 1, height-1)*stride;
1258     cs->b[2] = buffer + av_clip(-5+1, 0, height-2)*stride;
1259     cs->b[3] = buffer + av_clip(-5+2, 1, height-1)*stride;
1260     cs->b[4] = buffer + av_clip(-5+3, 0, height-2)*stride;
1261     cs->b[5] = buffer + av_clip(-5+4, 1, height-1)*stride;
1262     cs->y = -5;
1263 }
1264
1265 static void spatial_compose_dd137i_init(DWTCompose *cs, IDWTELEM *buffer, int height, int stride)
1266 {
1267     cs->b[0] = buffer + av_clip(-5-1, 0, height-2)*stride;
1268     cs->b[1] = buffer + av_clip(-5  , 1, height-1)*stride;
1269     cs->b[2] = buffer + av_clip(-5+1, 0, height-2)*stride;
1270     cs->b[3] = buffer + av_clip(-5+2, 1, height-1)*stride;
1271     cs->b[4] = buffer + av_clip(-5+3, 0, height-2)*stride;
1272     cs->b[5] = buffer + av_clip(-5+4, 1, height-1)*stride;
1273     cs->b[6] = buffer + av_clip(-5+5, 0, height-2)*stride;
1274     cs->b[7] = buffer + av_clip(-5+6, 1, height-1)*stride;
1275     cs->y = -5;
1276 }
1277
1278 int ff_spatial_idwt_init2(DWTContext *d, IDWTELEM *buffer, int width, int height,
1279                           int stride, enum dwt_type type, int decomposition_count,
1280                           IDWTELEM *temp)
1281 {
1282     int level;
1283
1284     d->buffer = buffer;
1285     d->width = width;
1286     d->height = height;
1287     d->stride = stride;
1288     d->decomposition_count = decomposition_count;
1289     d->temp = temp + 8;
1290
1291     for(level=decomposition_count-1; level>=0; level--){
1292         int hl = height >> level;
1293         int stride_l = stride << level;
1294
1295         switch(type){
1296         case DWT_DIRAC_DD9_7:
1297             spatial_compose_dd97i_init(d->cs+level, buffer, hl, stride_l);
1298             break;
1299         case DWT_DIRAC_LEGALL5_3:
1300             spatial_compose53i_init2(d->cs+level, buffer, hl, stride_l);
1301             break;
1302         case DWT_DIRAC_DD13_7:
1303             spatial_compose_dd137i_init(d->cs+level, buffer, hl, stride_l);
1304             break;
1305         case DWT_DIRAC_HAAR0:
1306         case DWT_DIRAC_HAAR1:
1307             d->cs[level].y = 1;
1308             break;
1309         case DWT_DIRAC_DAUB9_7:
1310             spatial_compose97i_init2(d->cs+level, buffer, hl, stride_l);
1311             break;
1312         default:
1313             d->cs[level].y = 0;
1314             break;
1315         }
1316     }
1317
1318     switch (type) {
1319     case DWT_DIRAC_DD9_7:
1320         d->spatial_compose = spatial_compose_dd97i_dy;
1321         d->vertical_compose_l0 = (void*)vertical_compose53iL0;
1322         d->vertical_compose_h0 = (void*)vertical_compose_dd97iH0;
1323         d->horizontal_compose = horizontal_compose_dd97i;
1324         d->support = 7;
1325         break;
1326     case DWT_DIRAC_LEGALL5_3:
1327         d->spatial_compose = spatial_compose_dirac53i_dy;
1328         d->vertical_compose_l0 = (void*)vertical_compose53iL0;
1329         d->vertical_compose_h0 = (void*)vertical_compose_dirac53iH0;
1330         d->horizontal_compose = horizontal_compose_dirac53i;
1331         d->support = 3;
1332         break;
1333     case DWT_DIRAC_DD13_7:
1334         d->spatial_compose = spatial_compose_dd137i_dy;
1335         d->vertical_compose_l0 = (void*)vertical_compose_dd137iL0;
1336         d->vertical_compose_h0 = (void*)vertical_compose_dd97iH0;
1337         d->horizontal_compose = horizontal_compose_dd137i;
1338         d->support = 7;
1339         break;
1340     case DWT_DIRAC_HAAR0:
1341     case DWT_DIRAC_HAAR1:
1342         d->spatial_compose = spatial_compose_haari_dy;
1343         d->vertical_compose = (void*)vertical_compose_haar;
1344         if (type == DWT_DIRAC_HAAR0)
1345             d->horizontal_compose = horizontal_compose_haar0i;
1346         else
1347             d->horizontal_compose = horizontal_compose_haar1i;
1348         d->support = 1;
1349         break;
1350     case DWT_DIRAC_FIDELITY:
1351         d->spatial_compose = spatial_compose_fidelity;
1352         d->vertical_compose_l0 = (void*)vertical_compose_fidelityiL0;
1353         d->vertical_compose_h0 = (void*)vertical_compose_fidelityiH0;
1354         d->horizontal_compose = horizontal_compose_fidelityi;
1355         break;
1356     case DWT_DIRAC_DAUB9_7:
1357         d->spatial_compose = spatial_compose_daub97i_dy;
1358         d->vertical_compose_l0 = (void*)vertical_compose_daub97iL0;
1359         d->vertical_compose_h0 = (void*)vertical_compose_daub97iH0;
1360         d->vertical_compose_l1 = (void*)vertical_compose_daub97iL1;
1361         d->vertical_compose_h1 = (void*)vertical_compose_daub97iH1;
1362         d->horizontal_compose = horizontal_compose_daub97i;
1363         d->support = 5;
1364         break;
1365     default:
1366         av_log(NULL, AV_LOG_ERROR, "Unknown wavelet type %d\n", type);
1367         return -1;
1368     }
1369
1370     if (HAVE_MMX) ff_spatial_idwt_init_mmx(d, type);
1371
1372     return 0;
1373 }
1374
1375 void ff_spatial_idwt_slice2(DWTContext *d, int y)
1376 {
1377     int level, support = d->support;
1378
1379     for (level = d->decomposition_count-1; level >= 0; level--) {
1380         int wl = d->width  >> level;
1381         int hl = d->height >> level;
1382         int stride_l = d->stride << level;
1383
1384         while (d->cs[level].y <= FFMIN((y>>level)+support, hl))
1385             d->spatial_compose(d, level, wl, hl, stride_l);
1386     }
1387 }
1388
1389 int ff_spatial_idwt2(IDWTELEM *buffer, int width, int height, int stride,
1390                      enum dwt_type type, int decomposition_count, IDWTELEM *temp)
1391 {
1392     DWTContext d;
1393     int y;
1394
1395     if (ff_spatial_idwt_init2(&d, buffer, width, height, stride, type, decomposition_count, temp))
1396         return -1;
1397
1398     for (y = 0; y < d.height; y += 4)
1399         ff_spatial_idwt_slice2(&d, y);
1400
1401     return 0;
1402 }