Mercurial > mplayer.hg
comparison libmpcodecs/pullup.c @ 10664:d47ca466c97b
pullup -- third generation inverse telecine engine. the backend
(pullup.[ch]) is not mplayer-specific and is designed to work well
with g2; vf_pullup.c is the g1 wrapper. see man page for details, and
keep in mind, this is a work in progress.
author | rfelker |
---|---|
date | Mon, 18 Aug 2003 15:24:08 +0000 |
parents | |
children | 67449e5936f3 |
comparison
equal
deleted
inserted
replaced
10663:711159267b2d | 10664:d47ca466c97b |
---|---|
1 | |
2 | |
3 #include <stdlib.h> | |
4 #include "pullup.h" | |
5 | |
6 | |
7 | |
8 | |
9 #ifdef HAVE_MMX | |
10 static int diff_y_mmx(unsigned char *a, unsigned char *b, int s) | |
11 { | |
12 int ret; | |
13 asm ( | |
14 "movl $4, %%ecx \n\t" | |
15 "pxor %%mm4, %%mm4 \n\t" | |
16 "pxor %%mm7, %%mm7 \n\t" | |
17 | |
18 ".balign 16 \n\t" | |
19 "1: \n\t" | |
20 | |
21 "movq (%%esi), %%mm0 \n\t" | |
22 "movq (%%esi), %%mm2 \n\t" | |
23 "addl %%eax, %%esi \n\t" | |
24 "movq (%%edi), %%mm1 \n\t" | |
25 "addl %%eax, %%edi \n\t" | |
26 "psubusb %%mm1, %%mm2 \n\t" | |
27 "psubusb %%mm0, %%mm1 \n\t" | |
28 "movq %%mm2, %%mm0 \n\t" | |
29 "movq %%mm1, %%mm3 \n\t" | |
30 "punpcklbw %%mm7, %%mm0 \n\t" | |
31 "punpcklbw %%mm7, %%mm1 \n\t" | |
32 "punpckhbw %%mm7, %%mm2 \n\t" | |
33 "punpckhbw %%mm7, %%mm3 \n\t" | |
34 "paddw %%mm0, %%mm4 \n\t" | |
35 "paddw %%mm1, %%mm4 \n\t" | |
36 "paddw %%mm2, %%mm4 \n\t" | |
37 "paddw %%mm3, %%mm4 \n\t" | |
38 | |
39 "decl %%ecx \n\t" | |
40 "jnz fb \n\t" | |
41 | |
42 "movq %%mm4, %%mm3 \n\t" | |
43 "punpcklwl %%mm7, %%mm4 \n\t" | |
44 "punpckhwl %%mm7, %%mm3 \n\t" | |
45 "paddl %%mm4, %%mm3 \n\t" | |
46 "movq %%mm3, %%mm2 \n\t" | |
47 "punpckllq %%mm7, %%mm3 \n\t" | |
48 "punpckhlq %%mm7, %%mm2 \n\t" | |
49 "paddl %%mm3, %%mm2 \n\t" | |
50 "movl %%mm2, %eax" | |
51 | |
52 "emms \n\t" | |
53 : "=a" (ret) | |
54 : "S" (a), "D" (b), "a" (s) | |
55 : | |
56 ); | |
57 return ret; | |
58 } | |
59 #endif | |
60 | |
61 #define ABS(a) (((a)^((a)>>31))-((a)>>31)) | |
62 | |
63 static int diff_y(unsigned char *a, unsigned char *b, int s) | |
64 { | |
65 int i, j, diff=0; | |
66 for (i=4; i; i--) { | |
67 for (j=0; j<8; j++) diff += ABS(a[j]-b[j]); | |
68 a+=s; b+=s; | |
69 } | |
70 return diff; | |
71 } | |
72 | |
73 static int licomb_y(unsigned char *a, unsigned char *b, int s) | |
74 { | |
75 int i, j, diff=0; | |
76 for (i=8; i; i--) { | |
77 for (j=0; j<8; j++) | |
78 diff += ABS((a[j]<<1) - b[j-s] - b[j]) | |
79 + ABS((b[j]<<1) - a[j] - a[j+s]); | |
80 a+=s; b+=s; | |
81 } | |
82 return diff; | |
83 } | |
84 | |
85 | |
86 | |
87 | |
88 | |
89 | |
90 | |
91 | |
92 | |
93 static void alloc_buffer(struct pullup_context *c, struct pullup_buffer *b) | |
94 { | |
95 int i; | |
96 if (b->planes) return; | |
97 b->planes = calloc(c->nplanes, sizeof(unsigned char *)); | |
98 for (i = 0; i < c->nplanes; i++) { | |
99 b->planes[i] = malloc(c->h[i]*c->stride[i]); | |
100 /* Deal with idiotic 128=0 for chroma: */ | |
101 memset(b->planes[i], c->background[i], c->h[i]*c->stride[i]); | |
102 } | |
103 } | |
104 | |
105 struct pullup_buffer *pullup_lock_buffer(struct pullup_buffer *b, int parity) | |
106 { | |
107 if (parity+1 & 1) b->lock[0]++; | |
108 if (parity+1 & 2) b->lock[1]++; | |
109 return b; | |
110 } | |
111 | |
112 void pullup_release_buffer(struct pullup_buffer *b, int parity) | |
113 { | |
114 if (parity+1 & 1) b->lock[0]--; | |
115 if (parity+1 & 2) b->lock[1]--; | |
116 } | |
117 | |
118 struct pullup_buffer *pullup_get_buffer(struct pullup_context *c, int parity) | |
119 { | |
120 int i; | |
121 | |
122 /* Try first to get the sister buffer for the previous field */ | |
123 if (parity < 2 && c->last && parity != c->last->parity | |
124 && !c->last->buffer->lock[parity]) { | |
125 alloc_buffer(c, c->last->buffer); | |
126 return pullup_lock_buffer(c->last->buffer, parity); | |
127 } | |
128 | |
129 /* Prefer a buffer with both fields open */ | |
130 for (i = 0; i < c->nbuffers; i++) { | |
131 if (c->buffers[i].lock[0]) continue; | |
132 if (c->buffers[i].lock[1]) continue; | |
133 alloc_buffer(c, &c->buffers[i]); | |
134 return pullup_lock_buffer(&c->buffers[i], parity); | |
135 } | |
136 | |
137 if (parity == 2) return 0; | |
138 | |
139 /* Search for any half-free buffer */ | |
140 for (i = 0; i < c->nbuffers; i++) { | |
141 if (parity+1 & 1 && c->buffers[i].lock[0]) continue; | |
142 if (parity+1 & 2 && c->buffers[i].lock[1]) continue; | |
143 alloc_buffer(c, &c->buffers[i]); | |
144 return pullup_lock_buffer(&c->buffers[i], parity); | |
145 } | |
146 | |
147 return 0; | |
148 } | |
149 | |
150 | |
151 | |
152 | |
153 | |
154 | |
155 static void compute_metric(struct pullup_context *c, | |
156 struct pullup_field *fa, int pa, | |
157 struct pullup_field *fb, int pb, | |
158 int (*func)(unsigned char *, unsigned char *, int), int *dest) | |
159 { | |
160 unsigned char *a, *b; | |
161 int x, y; | |
162 int xstep = c->bpp[0]; | |
163 int ystep = c->stride[0]<<3; | |
164 int s = c->stride[0]<<1; /* field stride */ | |
165 int w = c->metric_w*xstep; | |
166 | |
167 if (!fa->buffer || !fb->buffer) return; | |
168 | |
169 /* Shortcut for duplicate fields (e.g. from RFF flag) */ | |
170 if (fa->buffer == fb->buffer && pa == pb) { | |
171 memset(dest, 0, c->metric_len * sizeof(int)); | |
172 return; | |
173 } | |
174 | |
175 a = fa->buffer->planes[0] + pa * c->stride[0] + c->metric_offset; | |
176 b = fb->buffer->planes[0] + pb * c->stride[0] + c->metric_offset; | |
177 | |
178 for (y = c->metric_h; y; y--) { | |
179 for (x = 0; x < w; x += xstep) { | |
180 *dest++ = func(a + x, b + x, s); | |
181 } | |
182 a += ystep; b += ystep; | |
183 } | |
184 } | |
185 | |
186 | |
187 | |
188 | |
189 | |
190 static void alloc_metrics(struct pullup_context *c, struct pullup_field *f) | |
191 { | |
192 f->diffs = calloc(c->metric_len, sizeof(int)); | |
193 f->licomb = calloc(c->metric_len, sizeof(int)); | |
194 /* add more metrics here as needed */ | |
195 } | |
196 | |
197 static struct pullup_field *make_field_queue(struct pullup_context *c, int len) | |
198 { | |
199 struct pullup_field *head, *f; | |
200 f = head = calloc(1, sizeof(struct pullup_field)); | |
201 alloc_metrics(c, f); | |
202 for (; len > 0; len--) { | |
203 f->next = calloc(1, sizeof(struct pullup_field)); | |
204 f->next->prev = f; | |
205 f = f->next; | |
206 alloc_metrics(c, f); | |
207 } | |
208 f->next = head; | |
209 head->prev = f; | |
210 return head; | |
211 } | |
212 | |
213 static void check_field_queue(struct pullup_context *c) | |
214 { | |
215 if (c->head->next == c->first) { | |
216 struct pullup_field *f = calloc(1, sizeof(struct pullup_field)); | |
217 alloc_metrics(c, f); | |
218 f->prev = c->head; | |
219 f->next = c->first; | |
220 c->head->next = f; | |
221 c->first->prev = f; | |
222 } | |
223 } | |
224 | |
225 int pullup_submit_field(struct pullup_context *c, struct pullup_buffer *b, int parity) | |
226 { | |
227 struct pullup_field *f; | |
228 | |
229 /* Grow the circular list if needed */ | |
230 check_field_queue(c); | |
231 | |
232 /* Cannot have two fields of same parity in a row; drop the new one */ | |
233 if (c->last && c->last->parity == parity) return 0; | |
234 | |
235 f = c->head; | |
236 f->parity = parity; | |
237 f->buffer = pullup_lock_buffer(b, parity); | |
238 f->flags = 0; | |
239 f->breaks = 0; | |
240 f->affinity = 0; | |
241 | |
242 compute_metric(c, f, parity, f->prev->prev, parity, c->diff, f->diffs); | |
243 compute_metric(c, parity?f->prev:f, 0, parity?f:f->prev, 1, c->licomb, f->licomb); | |
244 | |
245 /* Advance the circular list */ | |
246 if (!c->first) c->first = c->head; | |
247 c->last = c->head; | |
248 c->head = c->head->next; | |
249 } | |
250 | |
251 void pullup_flush_fields(struct pullup_context *c) | |
252 { | |
253 struct pullup_field *f; | |
254 | |
255 for (f = c->first; f && f != c->head; f = f->next) { | |
256 pullup_release_buffer(f->buffer, f->parity); | |
257 f->buffer = 0; | |
258 } | |
259 c->first = c->last = 0; | |
260 } | |
261 | |
262 | |
263 | |
264 | |
265 | |
266 | |
267 | |
268 | |
269 #define F_HAVE_BREAKS 1 | |
270 #define F_HAVE_AFFINITY 2 | |
271 | |
272 | |
273 #define BREAK_LEFT 1 | |
274 #define BREAK_RIGHT 2 | |
275 | |
276 | |
277 | |
278 | |
279 static int queue_length(struct pullup_field *begin, struct pullup_field *end) | |
280 { | |
281 int count = 1; | |
282 struct pullup_field *f; | |
283 | |
284 if (!begin || !end) return 0; | |
285 for (f = begin; f != end; f = f->next) count++; | |
286 return count; | |
287 } | |
288 | |
289 static int find_first_break(struct pullup_field *f, int max) | |
290 { | |
291 int i; | |
292 for (i = 0; i < max; i++) { | |
293 if (f->breaks & BREAK_RIGHT || f->next->breaks & BREAK_LEFT) | |
294 return i+1; | |
295 f = f->next; | |
296 } | |
297 return 0; | |
298 } | |
299 | |
300 static void compute_breaks(struct pullup_context *c, struct pullup_field *f0) | |
301 { | |
302 int i; | |
303 struct pullup_field *f1 = f0->next; | |
304 struct pullup_field *f2 = f1->next; | |
305 struct pullup_field *f3 = f2->next; | |
306 int l, max_l=0, max_r=0; | |
307 | |
308 if (f0->flags & F_HAVE_BREAKS) return; | |
309 f0->flags |= F_HAVE_BREAKS; | |
310 | |
311 /* Special case when fields are 100% identical */ | |
312 if (f0->buffer == f2->buffer && f1->buffer != f3->buffer) { | |
313 f0->breaks |= BREAK_LEFT; | |
314 f2->breaks |= BREAK_RIGHT; | |
315 return; | |
316 } | |
317 | |
318 for (i = 0; i < c->metric_len; i++) { | |
319 l = f2->diffs[i] - f3->diffs[i]; | |
320 if (l > max_l) max_l = l; | |
321 if (-l > max_r) max_r = -l; | |
322 } | |
323 /* Don't get tripped up when differences are mostly quant error */ | |
324 if (max_l + max_r < 64) return; | |
325 if (max_l > 4*max_r) f1->breaks |= BREAK_LEFT; | |
326 if (max_r > 4*max_l) f2->breaks |= BREAK_RIGHT; | |
327 //printf("max_l=%d max_r=%d\n", max_l, max_r); | |
328 } | |
329 | |
330 static void compute_affinity(struct pullup_context *c, struct pullup_field *f) | |
331 { | |
332 int i; | |
333 int max_l=0, max_r=0, l; | |
334 if (f->flags & F_HAVE_AFFINITY) return; | |
335 f->flags |= F_HAVE_AFFINITY; | |
336 for (i = 0; i < c->metric_len; i++) { | |
337 l = f->licomb[i] - f->next->licomb[i]; | |
338 if (l > max_l) max_l = l; | |
339 if (-l > max_r) max_r = -l; | |
340 } | |
341 if (max_l + max_r < 64) return; | |
342 if (max_r > 3*max_l) f->affinity = -1; | |
343 else if (max_l > 3*max_r) f->affinity = 1; | |
344 } | |
345 | |
346 static void foo(struct pullup_context *c) | |
347 { | |
348 struct pullup_field *f = c->first; | |
349 int i, n = queue_length(f, c->last); | |
350 for (i = 0; i < n; i++) { | |
351 if (i < n-3) compute_breaks(c, f); | |
352 compute_affinity(c, f); | |
353 f = f->next; | |
354 } | |
355 } | |
356 | |
357 static int decide_frame_length(struct pullup_context *c) | |
358 { | |
359 int n; | |
360 struct pullup_field *f0 = c->first; | |
361 struct pullup_field *f1 = f0->next; | |
362 struct pullup_field *f2 = f1->next; | |
363 struct pullup_field *f3 = f2->next; | |
364 struct pullup_field *f4 = f3->next; | |
365 struct pullup_field *f5 = f4->next; | |
366 | |
367 if (queue_length(c->first, c->last) < 6) return 0; | |
368 foo(c); | |
369 | |
370 n = find_first_break(f0, 3); | |
371 | |
372 switch (n) { | |
373 case 1: | |
374 return 1; | |
375 case 2: | |
376 if (f0->affinity == -1 || f1->affinity == 1) return 1; | |
377 else return 2; | |
378 case 3: | |
379 if (f1->affinity == -1 && f2->affinity != -1) return 2; | |
380 else if (f1->affinity == 1 && f0->affinity != 1) return 1; | |
381 else return 3; | |
382 default: | |
383 if (f0->affinity == -1 && f1->affinity != -1) return 1; | |
384 else if (f1->affinity == 1 && f2->affinity == -1) return 1; | |
385 else return 2; | |
386 } | |
387 } | |
388 | |
389 | |
390 static void print_aff_and_breaks(struct pullup_context *c, struct pullup_field *f) | |
391 { | |
392 int i; | |
393 int max_l, max_r, l; | |
394 struct pullup_field *f0 = f; | |
395 const char aff_l[] = "+..", aff_r[] = "..+"; | |
396 printf("\naffinity: "); | |
397 for (i = 0; i < 6; i++) { | |
398 printf("%c%d%c", aff_l[1+f->affinity], i, aff_r[1+f->affinity]); | |
399 f = f->next; | |
400 } | |
401 f = f0; | |
402 printf("\nbreaks: "); | |
403 for (i=0; i<6; i++) { | |
404 printf("%c%d%c", f->breaks & BREAK_LEFT ? '|' : '.', i, f->breaks & BREAK_RIGHT ? '|' : '.'); | |
405 f = f->next; | |
406 } | |
407 printf("\n"); | |
408 } | |
409 | |
410 | |
411 | |
412 | |
413 | |
414 struct pullup_frame *pullup_get_frame(struct pullup_context *c) | |
415 { | |
416 int i; | |
417 struct pullup_frame *fr = c->frame; | |
418 int n = decide_frame_length(c); | |
419 | |
420 if (!n) return 0; | |
421 if (fr->lock) return 0; | |
422 | |
423 print_aff_and_breaks(c, c->first); | |
424 printf("duration: %d \n", n); | |
425 | |
426 fr->lock++; | |
427 fr->length = n; | |
428 fr->parity = c->first->parity; | |
429 fr->buffer = 0; | |
430 for (i = 0; i < n; i++) { | |
431 /* We cheat and steal the buffer without release+relock */ | |
432 fr->fields[i] = c->first->buffer; | |
433 c->first->buffer = 0; | |
434 c->first = c->first->next; | |
435 } | |
436 /* Export the entire frame as one buffer, if possible! */ | |
437 if (n == 2 && fr->fields[0] == fr->fields[1]) { | |
438 fr->buffer = fr->fields[0]; | |
439 pullup_lock_buffer(fr->buffer, 2); | |
440 return fr; | |
441 } | |
442 /* (loop is in case we ever support frames longer than 3 fields) */ | |
443 for (i = 1; i < n-1; i++) { | |
444 if (fr->fields[i] == fr->fields[i-1] | |
445 || fr->fields[i] == fr->fields[i+1]) { | |
446 fr->buffer = fr->fields[i]; | |
447 pullup_lock_buffer(fr->buffer, 2); | |
448 break; | |
449 } | |
450 } | |
451 return fr; | |
452 } | |
453 | |
454 static void copy_field(struct pullup_context *c, struct pullup_buffer *dest, | |
455 struct pullup_buffer *src, int parity) | |
456 { | |
457 int i, j; | |
458 unsigned char *d, *s; | |
459 for (i = 0; i < c->nplanes; i++) { | |
460 s = src->planes[i] + parity*c->stride[i]; | |
461 d = dest->planes[i] + parity*c->stride[i]; | |
462 for (j = c->h[i]>>1; j; j--) { | |
463 memcpy(d, s, c->stride[i]); | |
464 s += c->stride[i]<<1; | |
465 d += c->stride[i]<<1; | |
466 } | |
467 } | |
468 } | |
469 | |
470 void pullup_pack_frame(struct pullup_context *c, struct pullup_frame *fr) | |
471 { | |
472 int i; | |
473 int par = fr->parity; | |
474 if (fr->buffer) return; | |
475 if (fr->length < 2) return; /* FIXME: deal with this */ | |
476 for (i = 0; i < fr->length; i++) | |
477 { | |
478 if (fr->fields[i]->lock[par ^ (i&1) ^ 1]) continue; | |
479 fr->buffer = fr->fields[i]; | |
480 pullup_lock_buffer(fr->buffer, 2); | |
481 copy_field(c, fr->buffer, fr->fields[i+(i>0?-1:1)], par^(i&1)^1); | |
482 return; | |
483 } | |
484 fr->buffer = pullup_get_buffer(c, 2); | |
485 copy_field(c, fr->buffer, fr->fields[0], par); | |
486 copy_field(c, fr->buffer, fr->fields[1], par^1); | |
487 } | |
488 | |
489 void pullup_release_frame(struct pullup_frame *fr) | |
490 { | |
491 int i; | |
492 for (i = 0; i < fr->length; i++) | |
493 pullup_release_buffer(fr->fields[i], fr->parity ^ (i&1)); | |
494 if (fr->buffer) pullup_release_buffer(fr->buffer, 2); | |
495 fr->lock--; | |
496 } | |
497 | |
498 | |
499 | |
500 | |
501 | |
502 | |
503 struct pullup_context *pullup_alloc_context() | |
504 { | |
505 struct pullup_context *c; | |
506 | |
507 c = calloc(1, sizeof(struct pullup_context)); | |
508 | |
509 return c; | |
510 } | |
511 | |
512 void pullup_preinit_context(struct pullup_context *c) | |
513 { | |
514 c->bpp = calloc(c->nplanes, sizeof(int)); | |
515 c->w = calloc(c->nplanes, sizeof(int)); | |
516 c->h = calloc(c->nplanes, sizeof(int)); | |
517 c->stride = calloc(c->nplanes, sizeof(int)); | |
518 c->background = calloc(c->nplanes, sizeof(int)); | |
519 } | |
520 | |
521 void pullup_init_context(struct pullup_context *c) | |
522 { | |
523 if (c->nbuffers < 10) c->nbuffers = 10; | |
524 c->buffers = calloc(c->nbuffers, sizeof (struct pullup_buffer)); | |
525 | |
526 c->metric_w = (c->w[0] - (c->junk_left + c->junk_right << 3)) >> 3; | |
527 c->metric_h = (c->h[0] - (c->junk_top + c->junk_bottom << 1)) >> 3; | |
528 c->metric_offset = c->junk_left*c->bpp[0] + (c->junk_top<<1)*c->stride[0]; | |
529 c->metric_len = c->metric_w * c->metric_h; | |
530 | |
531 c->head = make_field_queue(c, 8); | |
532 | |
533 c->frame = calloc(1, sizeof (struct pullup_frame)); | |
534 c->frame->fields = calloc(3, sizeof (struct pullup_buffer *)); | |
535 | |
536 switch(c->format) { | |
537 case PULLUP_FMT_Y: | |
538 c->diff = diff_y; | |
539 c->licomb = licomb_y; | |
540 #ifdef HAVE_MMX | |
541 if (c->cpu & PULLUP_CPU_MMX) c->diff = diff_y_mmx; | |
542 #endif | |
543 break; | |
544 #if 0 | |
545 case PULLUP_FMT_YUY2: | |
546 c->diff = diff_yuy2; | |
547 break; | |
548 case PULLUP_FMT_RGB32: | |
549 c->diff = diff_rgb32; | |
550 break; | |
551 #endif | |
552 } | |
553 } | |
554 | |
555 void pullup_free_context(struct pullup_context *c) | |
556 { | |
557 /* FIXME: free! */ | |
558 } | |
559 | |
560 | |
561 | |
562 | |
563 | |
564 | |
565 | |
566 | |
567 |