?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Add support for VDPAU video out, including hardware decoding.
[mplayer/glamo.git] / libmpcodecs / vf_filmdint.c
blob477530e5f15ed67a6dfbef9015d03ccf637069bf
1 #include <stdio.h>
2 #include <stdlib.h>
3 #include <string.h>
4 #include <sys/time.h>
5
6 #include "config.h"
7 #include "mp_msg.h"
8 #include "cpudetect.h"
9
10 #include "img_format.h"
11 #include "mp_image.h"
12 #include "vf.h"
13 #include "cmmx.h"
14
15 #include "libvo/fastmemcpy.h"
16
17 #define NUM_STORED 4
18
19 enum pu_field_type_t {
20 PU_1ST_OF_3,
21 PU_2ND_OF_3,
22 PU_3RD_OF_3,
23 PU_1ST_OF_2,
24 PU_2ND_OF_2,
25 PU_INTERLACED
26 };
27
28 struct metrics {
29 /* This struct maps to a packed word 64-bit MMX register */
30 unsigned short int even;
31 unsigned short int odd;
32 unsigned short int noise;
33 unsigned short int temp;
34 } __attribute__ ((aligned (8)));
35
36 struct frame_stats {
37 struct metrics tiny, low, high, bigger, twox, max;
38 struct { unsigned int even, odd, noise, temp; } sad;
39 unsigned short interlaced_high;
40 unsigned short interlaced_low;
41 unsigned short num_blocks;
42 };
43
44 struct vf_priv_s {
45 unsigned long inframes;
46 unsigned long outframes;
47 enum pu_field_type_t prev_type;
48 unsigned swapped, chroma_swapped;
49 unsigned luma_only;
50 unsigned verbose;
51 unsigned fast;
52 unsigned long w, h, cw, ch, stride, chroma_stride, nplanes;
53 unsigned long sad_thres;
54 unsigned long dint_thres;
55 unsigned char *memory_allocated;
56 unsigned char *planes[2*NUM_STORED][4];
57 unsigned char **old_planes;
58 unsigned long static_idx;
59 unsigned long temp_idx;
60 unsigned long crop_x, crop_y, crop_cx, crop_cy;
61 unsigned long export_count, merge_count;
62 unsigned long num_breaks;
63 unsigned long num_copies;
64 long in_inc, out_dec, iosync;
65 long num_fields;
66 long prev_fields;
67 long notout;
68 long mmx2;
69 unsigned small_bytes[2];
70 unsigned mmx_temp[2];
71 struct frame_stats stats[2];
72 struct metrics thres;
73 char chflag;
74 double diff_time, merge_time, decode_time, vo_time, filter_time;
75 };
76
77 #define PPZ { 2000, 2000, 0, 2000 }
78 #define PPR { 2000, 2000, 0, 2000 }
79 static const struct frame_stats ppzs = {PPZ,PPZ,PPZ,PPZ,PPZ,PPZ,PPZ,0,0,9999};
80 static const struct frame_stats pprs = {PPR,PPR,PPR,PPR,PPR,PPR,PPR,0,0,9999};
81
82 extern int opt_screen_size_x;
83 extern int opt_screen_size_y;
84
85 #ifndef MIN
86 #define MIN(a,b) (((a)<(b))?(a):(b))
87 #endif
88 #ifndef MAX
89 #define MAX(a,b) (((a)>(b))?(a):(b))
90 #endif
91
92 #define PDIFFUB(X,Y,T) "movq " #X "," #T "\n\t" \
93 "psubusb " #Y "," #T "\n\t" \
94 "psubusb " #X "," #Y "\n\t" \
95 "paddusb " #Y "," #T "\n\t"
96
97 #define PDIFFUBT(X,Y,T) "movq " #X "," #T "\n\t" \
98 "psubusb " #Y "," #T "\n\t" \
99 "psubusb " #X "," #Y "\n\t" \
100 "paddusb " #T "," #Y "\n\t"
101
102 #define PSUMBW(X,T,Z) "movq " #X "," #T "\n\t" \
103 "punpcklbw " #Z "," #X "\n\t" \
104 "punpckhbw " #Z "," #T "\n\t" \
105 "paddw " #T "," #X "\n\t" \
106 "movq " #X "," #T "\n\t" \
107 "psllq $32, " #T "\n\t" \
108 "paddw " #T "," #X "\n\t" \
109 "movq " #X "," #T "\n\t" \
110 "psllq $16, " #T "\n\t" \
111 "paddw " #T "," #X "\n\t" \
112 "psrlq $48, " #X "\n\t"
113
114 #define PSADBW(X,Y,T,Z) PDIFFUBT(X,Y,T) PSUMBW(Y,T,Z)
115
116 #define PMAXUB(X,Y) "psubusb " #X "," #Y "\n\tpaddusb " #X "," #Y "\n\t"
117 #define PMAXUW(X,Y) "psubusw " #X "," #Y "\n\tpaddusw " #X "," #Y "\n\t"
118 #define PMINUBT(X,Y,T) "movq " #Y "," #T "\n\t" \
119 "psubusb " #X "," #T "\n\t" \
120 "psubusb " #T "," #Y "\n\t"
121 #define PAVGB(X,Y) "pavgusb " #X "," #Y "\n\t"
122
123 static inline void
124 get_metrics_c(unsigned char *a, unsigned char *b, int as, int bs, int lines,
125 struct metrics *m)
126 {
127 a -= as;
128 b -= bs;
129 do {
130 cmmx_t old_po = *(cmmx_t*)(a );
131 cmmx_t po = *(cmmx_t*)(b );
132 cmmx_t e = *(cmmx_t*)(b + bs);
133 cmmx_t old_o = *(cmmx_t*)(a + 2*as);
134 cmmx_t o = *(cmmx_t*)(b + 2*bs);
135 cmmx_t ne = *(cmmx_t*)(b + 3*bs);
136 cmmx_t old_no = *(cmmx_t*)(a + 4*as);
137 cmmx_t no = *(cmmx_t*)(b + 4*bs);
138
139 cmmx_t qup_old_odd = p31avgb(old_o, old_po);
140 cmmx_t qup_odd = p31avgb( o, po);
141 cmmx_t qdown_old_odd = p31avgb(old_o, old_no);
142 cmmx_t qdown_odd = p31avgb( o, no);
143
144 cmmx_t qup_even = p31avgb(ne, e);
145 cmmx_t qdown_even = p31avgb(e, ne);
146
147 cmmx_t temp_up_diff = pdiffub(qdown_even, qup_old_odd);
148 cmmx_t noise_up_diff = pdiffub(qdown_even, qup_odd);
149 cmmx_t temp_down_diff = pdiffub(qup_even, qdown_old_odd);
150 cmmx_t noise_down_diff = pdiffub(qup_even, qdown_odd);
151
152 cmmx_t odd_diff = pdiffub(o, old_o);
153 m->odd += psumbw(odd_diff);
154 m->even += psadbw(e, *(cmmx_t*)(a+as));
155
156 temp_up_diff = pminub(temp_up_diff, temp_down_diff);
157 temp_up_diff = pminub(temp_up_diff, odd_diff);
158 m->temp += psumbw(temp_up_diff);
159 noise_up_diff = pminub(noise_up_diff, odd_diff);
160 noise_up_diff = pminub(noise_up_diff, noise_down_diff);
161
162 m->noise += psumbw(noise_up_diff);
163 a += 2*as;
164 b += 2*bs;
165 } while (--lines);
166 }
167
168 static inline void
169 get_metrics_fast_c(unsigned char *a, unsigned char *b, int as, int bs,
170 int lines, struct metrics *m)
171 {
172 a -= as;
173 b -= bs;
174 do {
175 cmmx_t old_po = (*(cmmx_t*)(a ) >> 1) & ~SIGN_BITS;
176 cmmx_t po = (*(cmmx_t*)(b ) >> 1) & ~SIGN_BITS;
177 cmmx_t old_e = (*(cmmx_t*)(a + as) >> 1) & ~SIGN_BITS;
178 cmmx_t e = (*(cmmx_t*)(b + bs) >> 1) & ~SIGN_BITS;
179 cmmx_t old_o = (*(cmmx_t*)(a + 2*as) >> 1) & ~SIGN_BITS;
180 cmmx_t o = (*(cmmx_t*)(b + 2*bs) >> 1) & ~SIGN_BITS;
181 cmmx_t ne = (*(cmmx_t*)(b + 3*bs) >> 1) & ~SIGN_BITS;
182 cmmx_t old_no = (*(cmmx_t*)(a + 4*as) >> 1) & ~SIGN_BITS;
183 cmmx_t no = (*(cmmx_t*)(b + 4*bs) >> 1) & ~SIGN_BITS;
184
185 cmmx_t qup_old_odd = p31avgb_s(old_o, old_po);
186 cmmx_t qup_odd = p31avgb_s( o, po);
187 cmmx_t qdown_old_odd = p31avgb_s(old_o, old_no);
188 cmmx_t qdown_odd = p31avgb_s( o, no);
189
190 cmmx_t qup_even = p31avgb_s(ne, e);
191 cmmx_t qdown_even = p31avgb_s(e, ne);
192
193 cmmx_t temp_up_diff = pdiffub_s(qdown_even, qup_old_odd);
194 cmmx_t noise_up_diff = pdiffub_s(qdown_even, qup_odd);
195 cmmx_t temp_down_diff = pdiffub_s(qup_even, qdown_old_odd);
196 cmmx_t noise_down_diff = pdiffub_s(qup_even, qdown_odd);
197
198 cmmx_t odd_diff = pdiffub_s(o, old_o);
199 m->odd += psumbw_s(odd_diff) << 1;
200 m->even += psadbw_s(e, old_e) << 1;
201
202 temp_up_diff = pminub_s(temp_up_diff, temp_down_diff);
203 temp_up_diff = pminub_s(temp_up_diff, odd_diff);
204 m->temp += psumbw_s(temp_up_diff) << 1;
205 noise_up_diff = pminub_s(noise_up_diff, odd_diff);
206 noise_up_diff = pminub_s(noise_up_diff, noise_down_diff);
207
208 m->noise += psumbw_s(noise_up_diff) << 1;
209 a += 2*as;
210 b += 2*bs;
211 } while (--lines);
212 }
213
214 static inline void
215 get_metrics_faster_c(unsigned char *a, unsigned char *b, int as, int bs,
216 int lines, struct metrics *m)
217 {
218 a -= as;
219 b -= bs;
220 do {
221 cmmx_t old_po = (*(cmmx_t*)(a )>>1) & ~SIGN_BITS;
222 cmmx_t po = (*(cmmx_t*)(b )>>1) & ~SIGN_BITS;
223 cmmx_t old_e = (*(cmmx_t*)(a + as)>>1) & ~SIGN_BITS;
224 cmmx_t e = (*(cmmx_t*)(b + bs)>>1) & ~SIGN_BITS;
225 cmmx_t old_o = (*(cmmx_t*)(a + 2*as)>>1) & ~SIGN_BITS;
226 cmmx_t o = (*(cmmx_t*)(b + 2*bs)>>1) & ~SIGN_BITS;
227 cmmx_t ne = (*(cmmx_t*)(b + 3*bs)>>1) & ~SIGN_BITS;
228
229 cmmx_t down_even = p31avgb_s(e, ne);
230 cmmx_t up_odd = p31avgb_s(o, po);
231 cmmx_t up_old_odd = p31avgb_s(old_o, old_po);
232
233 cmmx_t odd_diff = pdiffub_s(o, old_o);
234 cmmx_t temp_diff = pdiffub_s(down_even, up_old_odd);
235 cmmx_t noise_diff = pdiffub_s(down_even, up_odd);
236
237 m->even += psadbw_s(e, old_e) << 1;
238 m->odd += psumbw_s(odd_diff) << 1;
239
240 temp_diff = pminub_s(temp_diff, odd_diff);
241 noise_diff = pminub_s(noise_diff, odd_diff);
242
243 m->noise += psumbw_s(noise_diff) << 1;
244 m->temp += psumbw_s(temp_diff) << 1;
245 a += 2*as;
246 b += 2*bs;
247 } while (--lines);
248
249 }
250
251 static inline void
252 get_block_stats(struct metrics *m, struct vf_priv_s *p, struct frame_stats *s)
253 {
254 unsigned two_e = m->even + MAX(m->even , p->thres.even );
255 unsigned two_o = m->odd + MAX(m->odd , p->thres.odd );
256 unsigned two_n = m->noise + MAX(m->noise, p->thres.noise);
257 unsigned two_t = m->temp + MAX(m->temp , p->thres.temp );
258
259 unsigned e_big = m->even >= (m->odd + two_o + 1)/2;
260 unsigned o_big = m->odd >= (m->even + two_e + 1)/2;
261 unsigned n_big = m->noise >= (m->temp + two_t + 1)/2;
262 unsigned t_big = m->temp >= (m->noise + two_n + 1)/2;
263
264 unsigned e2x = m->even >= two_o;
265 unsigned o2x = m->odd >= two_e;
266 unsigned n2x = m->noise >= two_t;
267 unsigned t2x = m->temp >= two_n;
268
269 unsigned ntiny_e = m->even > p->thres.even ;
270 unsigned ntiny_o = m->odd > p->thres.odd ;
271 unsigned ntiny_n = m->noise > p->thres.noise;
272 unsigned ntiny_t = m->temp > p->thres.temp ;
273
274 unsigned nlow_e = m->even > 2*p->thres.even ;
275 unsigned nlow_o = m->odd > 2*p->thres.odd ;
276 unsigned nlow_n = m->noise > 2*p->thres.noise;
277 unsigned nlow_t = m->temp > 2*p->thres.temp ;
278
279 unsigned high_e = m->even > 4*p->thres.even ;
280 unsigned high_o = m->odd > 4*p->thres.odd ;
281 unsigned high_n = m->noise > 4*p->thres.noise;
282 unsigned high_t = m->temp > 4*p->thres.temp ;
283
284 unsigned low_il = !n_big && !t_big && ntiny_n && ntiny_t;
285 unsigned high_il = !n_big && !t_big && nlow_n && nlow_t;
286
287 if (low_il | high_il) {
288 s->interlaced_low += low_il;
289 s->interlaced_high += high_il;
290 } else {
291 s->tiny.even += ntiny_e;
292 s->tiny.odd += ntiny_o;
293 s->tiny.noise += ntiny_n;
294 s->tiny.temp += ntiny_t;
295
296 s->low .even += nlow_e ;
297 s->low .odd += nlow_o ;
298 s->low .noise += nlow_n ;
299 s->low .temp += nlow_t ;
300
301 s->high.even += high_e ;
302 s->high.odd += high_o ;
303 s->high.noise += high_n ;
304 s->high.temp += high_t ;
305
306 if (m->even >= p->sad_thres) s->sad.even += m->even ;
307 if (m->odd >= p->sad_thres) s->sad.odd += m->odd ;
308 if (m->noise >= p->sad_thres) s->sad.noise += m->noise;
309 if (m->temp >= p->sad_thres) s->sad.temp += m->temp ;
310 }
311 s->num_blocks++;
312 s->max.even = MAX(s->max.even , m->even );
313 s->max.odd = MAX(s->max.odd , m->odd );
314 s->max.noise = MAX(s->max.noise, m->noise);
315 s->max.temp = MAX(s->max.temp , m->temp );
316
317 s->bigger.even += e_big ;
318 s->bigger.odd += o_big ;
319 s->bigger.noise += n_big ;
320 s->bigger.temp += t_big ;
321
322 s->twox.even += e2x ;
323 s->twox.odd += o2x ;
324 s->twox.noise += n2x ;
325 s->twox.temp += t2x ;
326
327 }
328
329 static inline struct metrics
330 block_metrics_c(unsigned char *a, unsigned char *b, int as, int bs,
331 int lines, struct vf_priv_s *p, struct frame_stats *s)
332 {
333 struct metrics tm;
334 tm.even = tm.odd = tm.noise = tm.temp = 0;
335 get_metrics_c(a, b, as, bs, lines, &tm);
336 if (sizeof(cmmx_t) < 8)
337 get_metrics_c(a+4, b+4, as, bs, lines, &tm);
338 get_block_stats(&tm, p, s);
339 return tm;
340 }
341
342 static inline struct metrics
343 block_metrics_fast_c(unsigned char *a, unsigned char *b, int as, int bs,
344 int lines, struct vf_priv_s *p, struct frame_stats *s)
345 {
346 struct metrics tm;
347 tm.even = tm.odd = tm.noise = tm.temp = 0;
348 get_metrics_fast_c(a, b, as, bs, lines, &tm);
349 if (sizeof(cmmx_t) < 8)
350 get_metrics_fast_c(a+4, b+4, as, bs, lines, &tm);
351 get_block_stats(&tm, p, s);
352 return tm;
353 }
354
355 static inline struct metrics
356 block_metrics_faster_c(unsigned char *a, unsigned char *b, int as, int bs,
357 int lines, struct vf_priv_s *p, struct frame_stats *s)
358 {
359 struct metrics tm;
360 tm.even = tm.odd = tm.noise = tm.temp = 0;
361 get_metrics_faster_c(a, b, as, bs, lines, &tm);
362 if (sizeof(cmmx_t) < 8)
363 get_metrics_faster_c(a+4, b+4, as, bs, lines, &tm);
364 get_block_stats(&tm, p, s);
365 return tm;
366 }
367
368 #define MEQ(X,Y) ((X).even == (Y).even && (X).odd == (Y).odd && (X).temp == (Y).temp && (X).noise == (Y).noise)
369
370 #define BLOCK_METRICS_TEMPLATE() \
371 __asm__ volatile("pxor %mm7, %mm7\n\t" /* The result is colleted in mm7 */ \
372 "pxor %mm6, %mm6\n\t" /* Temp to stay at 0 */ \
373 ); \
374 a -= as; \
375 b -= bs; \
376 do { \
377 __asm__ volatile( \
378 "movq (%0,%2), %%mm0\n\t" \
379 "movq (%1,%3), %%mm1\n\t" /* mm1 = even */ \
380 PSADBW(%%mm1, %%mm0, %%mm4, %%mm6) \
381 "paddusw %%mm0, %%mm7\n\t" /* even diff */ \
382 "movq (%0,%2,2), %%mm0\n\t" /* mm0 = old odd */ \
383 "movq (%1,%3,2), %%mm2\n\t" /* mm2 = odd */ \
384 "movq (%0), %%mm3\n\t" \
385 "psubusb %4, %%mm3\n\t" \
386 PAVGB(%%mm0, %%mm3) \
387 PAVGB(%%mm0, %%mm3) /* mm3 = qup old odd */ \
388 "movq %%mm0, %%mm5\n\t" \
389 PSADBW(%%mm2, %%mm0, %%mm4, %%mm6) \
390 "psllq $16, %%mm0\n\t" \
391 "paddusw %%mm0, %%mm7\n\t" \
392 "movq (%1), %%mm4\n\t" \
393 "lea (%0,%2,2), %0\n\t" \
394 "lea (%1,%3,2), %1\n\t" \
395 "psubusb %4, %%mm4\n\t" \
396 PAVGB(%%mm2, %%mm4) \
397 PAVGB(%%mm2, %%mm4) /* mm4 = qup odd */ \
398 PDIFFUBT(%%mm5, %%mm2, %%mm0) /* mm2 =abs(oldodd-odd) */ \
399 "movq (%1,%3), %%mm5\n\t" \
400 "psubusb %4, %%mm5\n\t" \
401 PAVGB(%%mm1, %%mm5) \
402 PAVGB(%%mm5, %%mm1) /* mm1 = qdown even */ \
403 PAVGB((%1,%3), %%mm5) /* mm5 = qup next even */ \
404 PDIFFUBT(%%mm1, %%mm3, %%mm0) /* mm3 = abs(qupoldo-qde) */ \
405 PDIFFUBT(%%mm1, %%mm4, %%mm0) /* mm4 = abs(qupodd-qde) */ \
406 PMINUBT(%%mm2, %%mm3, %%mm0) /* limit temp to odd diff */ \
407 PMINUBT(%%mm2, %%mm4, %%mm0) /* limit noise to odd diff */ \
408 "movq (%1,%3,2), %%mm2\n\t" \
409 "psubusb %4, %%mm2\n\t" \
410 PAVGB((%1), %%mm2) \
411 PAVGB((%1), %%mm2) /* mm2 = qdown odd */ \
412 "movq (%0,%2,2), %%mm1\n\t" \
413 "psubusb %4, %%mm1\n\t" \
414 PAVGB((%0), %%mm1) \
415 PAVGB((%0), %%mm1) /* mm1 = qdown old odd */ \
416 PDIFFUBT(%%mm5, %%mm2, %%mm0) /* mm2 = abs(qdo-qune) */ \
417 PDIFFUBT(%%mm5, %%mm1, %%mm0) /* mm1 = abs(qdoo-qune) */ \
418 PMINUBT(%%mm4, %%mm2, %%mm0) /* current */ \
419 PMINUBT(%%mm3, %%mm1, %%mm0) /* old */ \
420 PSUMBW(%%mm2, %%mm0, %%mm6) \
421 PSUMBW(%%mm1, %%mm0, %%mm6) \
422 "psllq $32, %%mm2\n\t" \
423 "psllq $48, %%mm1\n\t" \
424 "paddusw %%mm2, %%mm7\n\t" \
425 "paddusw %%mm1, %%mm7\n\t" \
426 : "=r" (a), "=r" (b) \
427 : "r"((long)as), "r"((long)bs), "m" (ones), "0"(a), "1"(b), "X"(*a), "X"(*b) \
428 ); \
429 } while (--lines);
430
431 static inline struct metrics
432 block_metrics_3dnow(unsigned char *a, unsigned char *b, int as, int bs,
433 int lines, struct vf_priv_s *p, struct frame_stats *s)
434 {
435 struct metrics tm;
436 #if !HAVE_AMD3DNOW
437 mp_msg(MSGT_VFILTER, MSGL_FATAL, "block_metrics_3dnow: internal error\n");
438 #else
439 static const unsigned long long ones = 0x0101010101010101ull;
440
441 BLOCK_METRICS_TEMPLATE();
442 __asm__ volatile("movq %%mm7, %0\n\temms" : "=m" (tm));
443 get_block_stats(&tm, p, s);
444 #endif
445 return tm;
446 }
447
448 #undef PSUMBW
449 #undef PSADBW
450 #undef PMAXUB
451 #undef PMINUBT
452 #undef PAVGB
453
454 #define PSUMBW(X,T,Z) "psadbw " #Z "," #X "\n\t"
455 #define PSADBW(X,Y,T,Z) "psadbw " #X "," #Y "\n\t"
456 #define PMAXUB(X,Y) "pmaxub " #X "," #Y "\n\t"
457 #define PMINUBT(X,Y,T) "pminub " #X "," #Y "\n\t"
458 #define PAVGB(X,Y) "pavgb " #X "," #Y "\n\t"
459
460 static inline struct metrics
461 block_metrics_mmx2(unsigned char *a, unsigned char *b, int as, int bs,
462 int lines, struct vf_priv_s *p, struct frame_stats *s)
463 {
464 struct metrics tm;
465 #if !HAVE_MMX
466 mp_msg(MSGT_VFILTER, MSGL_FATAL, "block_metrics_mmx2: internal error\n");
467 #else
468 static const unsigned long long ones = 0x0101010101010101ull;
469 unsigned long interlaced;
470 unsigned long prefetch_line = (((long)a>>3) & 7) + 10;
471 #ifdef DEBUG
472 struct frame_stats ts = *s;
473 #endif
474 __asm__ volatile("prefetcht0 (%0,%2)\n\t"
475 "prefetcht0 (%1,%3)\n\t" :
476 : "r" (a), "r" (b),
477 "r" (prefetch_line * as), "r" (prefetch_line * bs));
478
479 BLOCK_METRICS_TEMPLATE();
480
481 s->num_blocks++;
482 __asm__ volatile(
483 "movq %3, %%mm0\n\t"
484 "movq %%mm7, %%mm1\n\t"
485 "psubusw %%mm0, %%mm1\n\t"
486 "movq %%mm1, %%mm2\n\t"
487 "paddusw %%mm0, %%mm2\n\t"
488 "paddusw %%mm7, %%mm2\n\t"
489 "pshufw $0xb1, %%mm2, %%mm3\n\t"
490 "pavgw %%mm7, %%mm2\n\t"
491 "pshufw $0xb1, %%mm2, %%mm2\n\t"
492 "psubusw %%mm7, %%mm2\n\t"
493 "pcmpeqw %%mm6, %%mm2\n\t" /* 1 if >= 1.5x */
494 "psubusw %%mm7, %%mm3\n\t"
495 "pcmpeqw %%mm6, %%mm3\n\t" /* 1 if >= 2x */
496 "movq %1, %%mm4\n\t"
497 "movq %2, %%mm5\n\t"
498 "psubw %%mm2, %%mm4\n\t"
499 "psubw %%mm3, %%mm5\n\t"
500 "movq %%mm4, %1\n\t"
501 "movq %%mm5, %2\n\t"
502 "pxor %%mm4, %%mm4\n\t"
503 "pcmpeqw %%mm1, %%mm4\n\t" /* 1 if <= t */
504 "psubusw %%mm0, %%mm1\n\t"
505 "pxor %%mm5, %%mm5\n\t"
506 "pcmpeqw %%mm1, %%mm5\n\t" /* 1 if <= 2t */
507 "psubusw %%mm0, %%mm1\n\t"
508 "psubusw %%mm0, %%mm1\n\t"
509 "pcmpeqw %%mm6, %%mm1\n\t" /* 1 if <= 4t */
510 "pshufw $0xb1, %%mm2, %%mm0\n\t"
511 "por %%mm2, %%mm0\n\t" /* 1 if not close */
512 "punpckhdq %%mm0, %%mm0\n\t"
513 "movq %%mm4, %%mm2\n\t" /* tttt */
514 "punpckhdq %%mm5, %%mm2\n\t" /* ttll */
515 "por %%mm2, %%mm0\n\t"
516 "pcmpeqd %%mm6, %%mm0\n\t" /* close && big */
517 "psrlq $16, %%mm0\n\t"
518 "psrlw $15, %%mm0\n\t"
519 "movd %%mm0, %0\n\t"
520 : "=r" (interlaced), "=m" (s->bigger), "=m" (s->twox)
521 : "m" (p->thres)
522 );
523
524 if (interlaced) {
525 s->interlaced_high += interlaced >> 16;
526 s->interlaced_low += interlaced;
527 } else {
528 __asm__ volatile(
529 "pcmpeqw %%mm0, %%mm0\n\t" /* -1 */
530 "psubw %%mm0, %%mm4\n\t"
531 "psubw %%mm0, %%mm5\n\t"
532 "psubw %%mm0, %%mm1\n\t"
533 "paddw %0, %%mm4\n\t"
534 "paddw %1, %%mm5\n\t"
535 "paddw %2, %%mm1\n\t"
536 "movq %%mm4, %0\n\t"
537 "movq %%mm5, %1\n\t"
538 "movq %%mm1, %2\n\t"
539 : "=m" (s->tiny), "=m" (s->low), "=m" (s->high)
540 );
541
542 __asm__ volatile(
543 "pshufw $0, %2, %%mm0\n\t"
544 "psubusw %%mm7, %%mm0\n\t"
545 "pcmpeqw %%mm6, %%mm0\n\t" /* 0 if below sad_thres */
546 "pand %%mm7, %%mm0\n\t"
547 "movq %%mm0, %%mm1\n\t"
548 "punpcklwd %%mm6, %%mm0\n\t" /* sad even, odd */
549 "punpckhwd %%mm6, %%mm1\n\t" /* sad noise, temp */
550 "paddd %0, %%mm0\n\t"
551 "paddd %1, %%mm1\n\t"
552 "movq %%mm0, %0\n\t"
553 "movq %%mm1, %1\n\t"
554 : "=m" (s->sad.even), "=m" (s->sad.noise)
555 : "m" (p->sad_thres)
556 );
557 }
558
559 __asm__ volatile(
560 "movq %%mm7, (%1)\n\t"
561 PMAXUW((%0), %%mm7)
562 "movq %%mm7, (%0)\n\t"
563 "emms"
564 : : "r" (&s->max), "r" (&tm), "X" (s->max)
565 : "memory"
566 );
567 #ifdef DEBUG
568 if (1) {
569 struct metrics cm;
570 a -= 7*as;
571 b -= 7*bs;
572 cm = block_metrics_c(a, b, as, bs, 4, p, &ts);
573 if (!MEQ(tm, cm))
574 mp_msg(MSGT_VFILTER, MSGL_WARN, "Bad metrics\n");
575 if (s) {
576 # define CHECK(X) if (!MEQ(s->X, ts.X)) \
577 mp_msg(MSGT_VFILTER, MSGL_WARN, "Bad " #X "\n");
578 CHECK(tiny);
579 CHECK(low);
580 CHECK(high);
581 CHECK(sad);
582 CHECK(max);
583 }
584 }
585 #endif
586 #endif
587 return tm;
588 }
589
590 static inline int
591 dint_copy_line_mmx2(unsigned char *dst, unsigned char *a, long bos,
592 long cos, int ds, int ss, int w, int t)
593 {
594 #if !HAVE_MMX
595 mp_msg(MSGT_VFILTER, MSGL_FATAL, "dint_copy_line_mmx2: internal error\n");
596 return 0;
597 #else
598 unsigned long len = (w+7) >> 3;
599 int ret;
600 __asm__ volatile (
601 "pxor %%mm6, %%mm6 \n\t" /* deinterlaced pixel counter */
602 "movd %0, %%mm7 \n\t"
603 "punpcklbw %%mm7, %%mm7 \n\t"
604 "punpcklwd %%mm7, %%mm7 \n\t"
605 "punpckldq %%mm7, %%mm7 \n\t" /* mm7 = threshold */
606 : /* no output */
607 : "rm" (t)
608 );
609 do {
610 __asm__ volatile (
611 "movq (%0), %%mm0\n\t"
612 "movq (%0,%3,2), %%mm1\n\t"
613 "movq %%mm0, (%2)\n\t"
614 "pmaxub %%mm1, %%mm0\n\t"
615 "pavgb (%0), %%mm1\n\t"
616 "psubusb %%mm1, %%mm0\n\t"
617 "paddusb %%mm7, %%mm0\n\t" /* mm0 = max-avg+thr */
618 "movq (%0,%1), %%mm2\n\t"
619 "movq (%0,%5), %%mm3\n\t"
620 "movq %%mm2, %%mm4\n\t"
621 PDIFFUBT(%%mm1, %%mm2, %%mm5)
622 PDIFFUBT(%%mm1, %%mm3, %%mm5)
623 "pminub %%mm2, %%mm3\n\t"
624 "pcmpeqb %%mm3, %%mm2\n\t" /* b = min */
625 "pand %%mm2, %%mm4\n\t"
626 "pandn (%0,%5), %%mm2\n\t"
627 "por %%mm4, %%mm2\n\t"
628 "pminub %%mm0, %%mm3\n\t"
629 "pcmpeqb %%mm0, %%mm3\n\t" /* set to 1s if >= threshold */
630 "psubb %%mm3, %%mm6\n\t" /* count pixels above thr. */
631 "pand %%mm3, %%mm1 \n\t"
632 "pandn %%mm2, %%mm3 \n\t"
633 "por %%mm3, %%mm1 \n\t" /* avg if >= threshold */
634 "movq %%mm1, (%2,%4) \n\t"
635 : /* no output */
636 : "r" (a), "r" (bos), "r" (dst), "r" ((long)ss), "r" ((long)ds), "r" (cos)
637 );
638 a += 8;
639 dst += 8;
640 } while (--len);
641
642 __asm__ volatile ("pxor %%mm7, %%mm7 \n\t"
643 "psadbw %%mm6, %%mm7 \n\t"
644 "movd %%mm7, %0 \n\t"
645 "emms \n\t"
646 : "=r" (ret)
647 );
648 return ret;
649 #endif
650 }
651
652 static inline int
653 dint_copy_line(unsigned char *dst, unsigned char *a, long bos,
654 long cos, int ds, int ss, int w, int t)
655 {
656 unsigned long len = ((unsigned long)w+sizeof(cmmx_t)-1) / sizeof(cmmx_t);
657 cmmx_t dint_count = 0;
658 cmmx_t thr;
659 t |= t << 8;
660 thr = t | (t << 16);
661 if (sizeof(cmmx_t) > 4)
662 thr |= thr << (sizeof(cmmx_t)*4);
663 do {
664 cmmx_t e = *(cmmx_t*)a;
665 cmmx_t ne = *(cmmx_t*)(a+2*ss);
666 cmmx_t o = *(cmmx_t*)(a+bos);
667 cmmx_t oo = *(cmmx_t*)(a+cos);
668 cmmx_t maxe = pmaxub(e, ne);
669 cmmx_t avge = pavgb(e, ne);
670 cmmx_t max_diff = maxe - avge + thr; /* 0<=max-avg<128, thr<128 */
671 cmmx_t diffo = pdiffub(avge, o);
672 cmmx_t diffoo = pdiffub(avge, oo);
673 cmmx_t diffcmp = pcmpgtub(diffo, diffoo);
674 cmmx_t bo = ((oo ^ o) & diffcmp) ^ o;
675 cmmx_t diffbo = ((diffoo ^ diffo) & diffcmp) ^ diffo;
676 cmmx_t above_thr = ~pcmpgtub(max_diff, diffbo);
677 cmmx_t bo_or_avg = ((avge ^ bo) & above_thr) ^ bo;
678 dint_count += above_thr & ONE_BYTES;
679 *(cmmx_t*)(dst) = e;
680 *(cmmx_t*)(dst+ds) = bo_or_avg;
681 a += sizeof(cmmx_t);
682 dst += sizeof(cmmx_t);
683 } while (--len);
684 return psumbw(dint_count);
685 }
686
687 static int
688 dint_copy_plane(unsigned char *d, unsigned char *a, unsigned char *b,
689 unsigned char *c, unsigned long w, unsigned long h,
690 unsigned long ds, unsigned long ss, unsigned long threshold,
691 long field, long mmx2)
692 {
693 unsigned long ret = 0;
694 long bos = b - a;
695 long cos = c - a;
696 if (field) {
697 fast_memcpy(d, b, w);
698 h--;
699 d += ds;
700 a += ss;
701 }
702 bos += ss;
703 cos += ss;
704 while (h > 2) {
705 if (threshold >= 128) {
706 fast_memcpy(d, a, w);
707 fast_memcpy(d+ds, a+bos, w);
708 } else if (mmx2 == 1) {
709 ret += dint_copy_line_mmx2(d, a, bos, cos, ds, ss, w, threshold);
710 } else
711 ret += dint_copy_line(d, a, bos, cos, ds, ss, w, threshold);
712 h -= 2;
713 d += 2*ds;
714 a += 2*ss;
715 }
716 fast_memcpy(d, a, w);
717 if (h == 2)
718 fast_memcpy(d+ds, a+bos, w);
719 return ret;
720 }
721
722 static void
723 copy_merge_fields(struct vf_priv_s *p, mp_image_t *dmpi,
724 unsigned char **old, unsigned char **new, unsigned long show)
725 {
726 unsigned long threshold = 256;
727 unsigned long field = p->swapped;
728 unsigned long dint_pixels = 0;
729 unsigned char **other = old;
730 if (show >= 12 || !(show & 3))
731 show >>= 2, other = new, new = old;
732 if (show <= 2) { /* Single field: de-interlace */
733 threshold = p->dint_thres;
734 field ^= show & 1;
735 old = new;
736 } else if (show == 3)
737 old = new;
738 else
739 field ^= 1;
740 dint_pixels +=dint_copy_plane(dmpi->planes[0], old[0], new[0],
741 other[0], p->w, p->h, dmpi->stride[0],
742 p->stride, threshold, field, p->mmx2);
743 if (dmpi->flags & MP_IMGFLAG_PLANAR) {
744 if (p->luma_only)
745 old = new, other = new;
746 else
747 threshold = threshold/2 + 1;
748 field ^= p->chroma_swapped;
749 dint_copy_plane(dmpi->planes[1], old[1], new[1],
750 other[1], p->cw, p->ch, dmpi->stride[1],
751 p->chroma_stride, threshold, field, p->mmx2);
752 dint_copy_plane(dmpi->planes[2], old[2], new[2],
753 other[2], p->cw, p->ch, dmpi->stride[2],
754 p->chroma_stride, threshold, field, p->mmx2);
755 }
756 if (dint_pixels > 0 && p->verbose)
757 mp_msg(MSGT_VFILTER,MSGL_INFO,"Deinterlaced %lu pixels\n",dint_pixels);
758 }
759
760 static void diff_planes(struct vf_priv_s *p, struct frame_stats *s,
761 unsigned char *of, unsigned char *nf,
762 int w, int h, int os, int ns, int swapped)
763 {
764 int i, y;
765 int align = -(long)nf & 7;
766 of += align;
767 nf += align;
768 w -= align;
769 if (swapped)
770 of -= os, nf -= ns;
771 i = (h*3 >> 7) & ~1;
772 of += i*os + 8;
773 nf += i*ns + 8;
774 h -= i;
775 w -= 16;
776
777 memset(s, 0, sizeof(*s));
778
779 for (y = (h-8) >> 3; y; y--) {
780 if (p->mmx2 == 1) {
781 for (i = 0; i < w; i += 8)
782 block_metrics_mmx2(of+i, nf+i, os, ns, 4, p, s);
783 } else if (p->mmx2 == 2) {
784 for (i = 0; i < w; i += 8)
785 block_metrics_3dnow(of+i, nf+i, os, ns, 4, p, s);
786 } else if (p->fast > 3) {
787 for (i = 0; i < w; i += 8)
788 block_metrics_faster_c(of+i, nf+i, os, ns, 4, p, s);
789 } else if (p->fast > 1) {
790 for (i = 0; i < w; i += 8)
791 block_metrics_fast_c(of+i, nf+i, os, ns, 4, p, s);
792 } else {
793 for (i = 0; i < w; i += 8)
794 block_metrics_c(of+i, nf+i, os, ns, 4, p, s);
795 }
796 of += 8*os;
797 nf += 8*ns;
798 }
799 }
800
801 #define METRICS(X) (X).even, (X).odd, (X).noise, (X).temp
802
803 static void diff_fields(struct vf_priv_s *p, struct frame_stats *s,
804 unsigned char **old, unsigned char **new)
805 {
806 diff_planes(p, s, old[0], new[0], p->w, p->h,
807 p->stride, p->stride, p->swapped);
808 s->sad.even = (s->sad.even * 16ul) / s->num_blocks;
809 s->sad.odd = (s->sad.odd * 16ul) / s->num_blocks;
810 s->sad.noise = (s->sad.noise * 16ul) / s->num_blocks;
811 s->sad.temp = (s->sad.temp * 16ul) / s->num_blocks;
812 if (p->verbose)
813 mp_msg(MSGT_VFILTER, MSGL_INFO, "%lu%c M:%d/%d/%d/%d - %d, "
814 "t:%d/%d/%d/%d, l:%d/%d/%d/%d, h:%d/%d/%d/%d, bg:%d/%d/%d/%d, "
815 "2x:%d/%d/%d/%d, sad:%d/%d/%d/%d, lil:%d, hil:%d, ios:%.1f\n",
816 p->inframes, p->chflag, METRICS(s->max), s->num_blocks,
817 METRICS(s->tiny), METRICS(s->low), METRICS(s->high),
818 METRICS(s->bigger), METRICS(s->twox), METRICS(s->sad),
819 s->interlaced_low, s->interlaced_high,
820 p->iosync / (double) p->in_inc);
821 }
822
823 static const char *parse_args(struct vf_priv_s *p, const char *args)
824 {
825 args--;
826 while (args && *++args &&
827 (sscanf(args, "io=%lu:%lu", &p->out_dec, &p->in_inc) == 2 ||
828 sscanf(args, "diff_thres=%hu", &p->thres.even ) == 1 ||
829 sscanf(args, "comb_thres=%hu", &p->thres.noise) == 1 ||
830 sscanf(args, "sad_thres=%lu", &p->sad_thres ) == 1 ||
831 sscanf(args, "dint_thres=%lu", &p->dint_thres ) == 1 ||
832 sscanf(args, "fast=%u", &p->fast ) == 1 ||
833 sscanf(args, "mmx2=%lu", &p->mmx2 ) == 1 ||
834 sscanf(args, "luma_only=%u", &p->luma_only ) == 1 ||
835 sscanf(args, "verbose=%u", &p->verbose ) == 1 ||
836 sscanf(args, "crop=%lu:%lu:%lu:%lu", &p->w,
837 &p->h, &p->crop_x, &p->crop_y) == 4))
838 args = strchr(args, '/');
839 return args;
840 }
841
842 static unsigned long gcd(unsigned long x, unsigned long y)
843 {
844 unsigned long t;
845 if (x > y)
846 t = x, x = y, y = t;
847
848 while (x) {
849 t = y % x;
850 y = x;
851 x = t;
852 }
853 return y;
854 }
855
856 static void init(struct vf_priv_s *p, mp_image_t *mpi)
857 {
858 unsigned long i;
859 unsigned long plane_size, chroma_plane_size;
860 unsigned char *plane;
861 unsigned long cos, los;
862 p->crop_cx = p->crop_x >> mpi->chroma_x_shift;
863 p->crop_cy = p->crop_y >> mpi->chroma_y_shift;
864 if (mpi->flags & MP_IMGFLAG_ACCEPT_STRIDE) {
865 p->stride = (mpi->w + 15) & ~15;
866 p->chroma_stride = p->stride >> mpi->chroma_x_shift;
867 } else {
868 p->stride = mpi->width;
869 p->chroma_stride = mpi->chroma_width;
870 }
871 p->cw = p->w >> mpi->chroma_x_shift;
872 p->ch = p->h >> mpi->chroma_y_shift;
873 p->nplanes = 1;
874 p->static_idx = 0;
875 p->temp_idx = 0;
876 p->old_planes = p->planes[0];
877 plane_size = mpi->h * p->stride;
878 chroma_plane_size = mpi->flags & MP_IMGFLAG_PLANAR ?
879 mpi->chroma_height * p->chroma_stride : 0;
880 p->memory_allocated =
881 malloc(NUM_STORED * (plane_size+2*chroma_plane_size) +
882 8*p->chroma_stride + 4096);
883 /* align to page boundary */
884 plane = p->memory_allocated + (-(long)p->memory_allocated & 4095);
885 memset(plane, 0, NUM_STORED * plane_size);
886 los = p->crop_x + p->crop_y * p->stride;
887 cos = p->crop_cx + p->crop_cy * p->chroma_stride;
888 for (i = 0; i != NUM_STORED; i++, plane += plane_size) {
889 p->planes[i][0] = plane;
890 p->planes[NUM_STORED + i][0] = plane + los;
891 }
892 if (mpi->flags & MP_IMGFLAG_PLANAR) {
893 p->nplanes = 3;
894 memset(plane, 0x80, NUM_STORED * 2 * chroma_plane_size);
895 for (i = 0; i != NUM_STORED; i++) {
896 p->planes[i][1] = plane;
897 p->planes[NUM_STORED + i][1] = plane + cos;
898 plane += chroma_plane_size;
899 p->planes[i][2] = plane;
900 p->planes[NUM_STORED + i][2] = plane + cos;
901 plane += chroma_plane_size;
902 }
903 }
904 p->out_dec <<= 2;
905 i = gcd(p->in_inc, p->out_dec);
906 p->in_inc /= i;
907 p->out_dec /= i;
908 p->iosync = 0;
909 p->num_fields = 3;
910 }
911
912 static inline double get_time(void)
913 {
914 struct timeval tv;
915 gettimeofday(&tv, 0);
916 return tv.tv_sec + tv.tv_usec * 1e-6;
917 }
918
919 static void get_image(struct vf_instance_s* vf, mp_image_t *mpi)
920 {
921 struct vf_priv_s *p = vf->priv;
922 static unsigned char **planes, planes_idx;
923
924 if (mpi->type == MP_IMGTYPE_STATIC) return;
925
926 if (!p->planes[0][0]) init(p, mpi);
927
928 if (mpi->type == MP_IMGTYPE_TEMP ||
929 (mpi->type == MP_IMGTYPE_IPB && !(mpi->flags & MP_IMGFLAG_READABLE)))
930 planes_idx = NUM_STORED/2 + (++p->temp_idx % (NUM_STORED/2));
931 else
932 planes_idx = ++p->static_idx % (NUM_STORED/2);
933 planes = p->planes[planes_idx];
934 mpi->priv = p->planes[NUM_STORED + planes_idx];
935 if (mpi->priv == p->old_planes) {
936 unsigned char **old_planes =
937 p->planes[NUM_STORED + 2 + (++p->temp_idx & 1)];
938 my_memcpy_pic(old_planes[0], p->old_planes[0],
939 p->w, p->h, p->stride, p->stride);
940 if (mpi->flags & MP_IMGFLAG_PLANAR) {
941 my_memcpy_pic(old_planes[1], p->old_planes[1],
942 p->cw, p->ch, p->chroma_stride, p->chroma_stride);
943 my_memcpy_pic(old_planes[2], p->old_planes[2],
944 p->cw, p->ch, p->chroma_stride, p->chroma_stride);
945 }
946 p->old_planes = old_planes;
947 p->num_copies++;
948 }
949 mpi->planes[0] = planes[0];
950 mpi->stride[0] = p->stride;
951 if (mpi->flags & MP_IMGFLAG_PLANAR) {
952 mpi->planes[1] = planes[1];
953 mpi->planes[2] = planes[2];
954 mpi->stride[1] = mpi->stride[2] = p->chroma_stride;
955 }
956 mpi->width = p->stride;
957
958 mpi->flags |= MP_IMGFLAG_DIRECT;
959 mpi->flags &= ~MP_IMGFLAG_DRAW_CALLBACK;
960 }
961
962 static inline long
963 cmpe(unsigned long x, unsigned long y, unsigned long err, unsigned long e)
964 {
965 long diff = x-y;
966 long unit = ((x+y+err) >> e);
967 long ret = (diff > unit) - (diff < -unit);
968 unit >>= 1;
969 return ret + (diff > unit) - (diff < -unit);
970 }
971
972 static unsigned long
973 find_breaks(struct vf_priv_s *p, struct frame_stats *s)
974 {
975 struct frame_stats *ps = &p->stats[(p->inframes-1) & 1];
976 long notfilm = 5*p->in_inc - p->out_dec;
977 unsigned long n = s->num_blocks >> 8;
978 unsigned long sad_comb_cmp = cmpe(s->sad.temp, s->sad.noise, 512, 1);
979 unsigned long ret = 8;
980
981 if (cmpe(s->sad.temp, s->sad.even, 512, 1) > 0)
982 mp_msg(MSGT_VFILTER, MSGL_WARN,
983 "@@@@@@@@ Bottom-first field??? @@@@@@@@\n");
984 if (s->sad.temp > 1000 && s->sad.noise > 1000)
985 return 3;
986 if (s->interlaced_high >= 2*n && s->sad.temp > 256 && s->sad.noise > 256)
987 return 3;
988 if (s->high.noise > s->num_blocks/4 && s->sad.noise > 10000 &&
989 s->sad.noise > 2*s->sad.even && s->sad.noise > 2*ps->sad.odd) {
990 // Mid-frame scene change
991 if (s->tiny.temp + s->interlaced_low < n ||
992 s->low.temp + s->interlaced_high < n/4 ||
993 s->high.temp + s->interlaced_high < n/8 ||
994 s->sad.temp < 160)
995 return 1;
996 return 3;
997 }
998 if (s->high.temp > s->num_blocks/4 && s->sad.temp > 10000 &&
999 s->sad.temp > 2*ps->sad.odd && s->sad.temp > 2*ps->sad.even) {
1000 // Start frame scene change
1001 if (s->tiny.noise + s->interlaced_low < n ||
1002 s->low.noise + s->interlaced_high < n/4 ||
1003 s->high.noise + s->interlaced_high < n/8 ||
1004 s->sad.noise < 160)
1005 return 2;
1006 return 3;
1007 }
1008 if (sad_comb_cmp == 2)
1009 return 2;
1010 if (sad_comb_cmp == -2)
1011 return 1;
1012
1013 if (s->tiny.odd > 3*MAX(n,s->tiny.even) + s->interlaced_low)
1014 return 1;
1015 if (s->tiny.even > 3*MAX(n,s->tiny.odd)+s->interlaced_low &&
1016 (!sad_comb_cmp || (s->low.noise <= n/4 && s->low.temp <= n/4)))
1017 return 4;
1018
1019 if (s->sad.noise < 64 && s->sad.temp < 64 &&
1020 s->low.noise <= n/2 && s->high.noise <= n/4 &&
1021 s->low.temp <= n/2 && s->high.temp <= n/4)
1022 goto still;
1023
1024 if (s->tiny.temp > 3*MAX(n,s->tiny.noise) + s->interlaced_low)
1025 return 2;
1026 if (s->tiny.noise > 3*MAX(n,s->tiny.temp) + s->interlaced_low)
1027 return 1;
1028
1029 if (s->low.odd > 3*MAX(n/4,s->low.even) + s->interlaced_high)
1030 return 1;
1031 if (s->low.even > 3*MAX(n/4,s->low.odd)+s->interlaced_high &&
1032 s->sad.even > 2*s->sad.odd &&
1033 (!sad_comb_cmp || (s->low.noise <= n/4 && s->low.temp <= n/4)))
1034 return 4;
1035
1036 if (s->low.temp > 3*MAX(n/4,s->low.noise) + s->interlaced_high)
1037 return 2;
1038 if (s->low.noise > 3*MAX(n/4,s->low.temp) + s->interlaced_high)
1039 return 1;
1040
1041 if (sad_comb_cmp == 1 && s->sad.noise < 64)
1042 return 2;
1043 if (sad_comb_cmp == -1 && s->sad.temp < 64)
1044 return 1;
1045
1046 if (s->tiny.odd <= n || (s->tiny.noise <= n/2 && s->tiny.temp <= n/2)) {
1047 if (s->interlaced_low <= n) {
1048 if (p->num_fields == 1)
1049 goto still;
1050 if (s->tiny.even <= n || ps->tiny.noise <= n/2)
1051 /* Still frame */
1052 goto still;
1053 if (s->bigger.even >= 2*MAX(n,s->bigger.odd) + s->interlaced_low)
1054 return 4;
1055 if (s->low.even >= 2*n + s->interlaced_low)
1056 return 4;
1057 goto still;
1058 }
1059 }
1060 if (s->low.odd <= n/4) {
1061 if (s->interlaced_high <= n/4) {
1062 if (p->num_fields == 1)
1063 goto still;
1064 if (s->low.even <= n/4)
1065 /* Still frame */
1066 goto still;
1067 if (s->bigger.even >= 2*MAX(n/4,s->bigger.odd)+s->interlaced_high)
1068 return 4;
1069 if (s->low.even >= n/2 + s->interlaced_high)
1070 return 4;
1071 goto still;
1072 }
1073 }
1074 if (s->bigger.temp > 2*MAX(n,s->bigger.noise) + s->interlaced_low)
1075 return 2;
1076 if (s->bigger.noise > 2*MAX(n,s->bigger.temp) + s->interlaced_low)
1077 return 1;
1078 if (s->bigger.temp > 2*MAX(n,s->bigger.noise) + s->interlaced_high)
1079 return 2;
1080 if (s->bigger.noise > 2*MAX(n,s->bigger.temp) + s->interlaced_high)
1081 return 1;
1082 if (s->twox.temp > 2*MAX(n,s->twox.noise) + s->interlaced_high)
1083 return 2;
1084 if (s->twox.noise > 2*MAX(n,s->twox.temp) + s->interlaced_high)
1085 return 1;
1086 if (s->bigger.even > 2*MAX(n,s->bigger.odd) + s->interlaced_low &&
1087 s->bigger.temp < n && s->bigger.noise < n)
1088 return 4;
1089 if (s->interlaced_low > MIN(2*n, s->tiny.odd))
1090 return 3;
1091 ret = 8 + (1 << (s->sad.temp > s->sad.noise));
1092 still:
1093 if (p->num_fields == 1 && p->prev_fields == 3 && notfilm >= 0 &&
1094 (s->tiny.temp <= s->tiny.noise || s->sad.temp < s->sad.noise+16))
1095 return 1;
1096 if (p->notout < p->num_fields && p->iosync > 2*p->in_inc && notfilm < 0)
1097 notfilm = 0;
1098 if (p->num_fields < 2 ||
1099 (p->num_fields == 2 && p->prev_fields == 2 && notfilm < 0))
1100 return ret;
1101 if (!notfilm && (p->prev_fields&~1) == 2) {
1102 if (p->prev_fields + p->num_fields == 5) {
1103 if (s->tiny.noise <= s->tiny.temp ||
1104 s->low.noise == 0 || s->low.noise < s->low.temp ||
1105 s->sad.noise < s->sad.temp+16)
1106 return 2;
1107 }
1108 if (p->prev_fields + p->num_fields == 4) {
1109 if (s->tiny.temp <= s->tiny.noise ||
1110 s->low.temp == 0 || s->low.temp < s->low.noise ||
1111 s->sad.temp < s->sad.noise+16)
1112 return 1;
1113 }
1114 }
1115 if (p->num_fields > 2 &&
1116 ps->sad.noise > s->sad.noise && ps->sad.noise > s->sad.temp)
1117 return 4;
1118 return 2 >> (s->sad.noise > s->sad.temp);
1119 }
1120
1121 #define ITOC(X) (!(X) ? ' ' : (X) + ((X)>9 ? 'a'-10 : '0'))
1122
1123 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts)
1124 {
1125 mp_image_t *dmpi;
1126 struct vf_priv_s *p = vf->priv;
1127 unsigned char **planes, **old_planes;
1128 struct frame_stats *s = &p->stats[p->inframes & 1];
1129 struct frame_stats *ps = &p->stats[(p->inframes-1) & 1];
1130 int swapped = 0;
1131 const int flags = mpi->fields;
1132 int breaks, prev;
1133 int show_fields = 0;
1134 int dropped_fields = 0;
1135 double start_time, diff_time;
1136 char prev_chflag = p->chflag;
1137 int keep_rate;
1138
1139 if (!p->planes[0][0]) init(p, mpi);
1140
1141 old_planes = p->old_planes;
1142
1143 if ((mpi->flags & MP_IMGFLAG_DIRECT) && mpi->priv) {
1144 planes = mpi->priv;
1145 mpi->priv = 0;
1146 } else {
1147 planes = p->planes[2 + (++p->temp_idx & 1)];
1148 my_memcpy_pic(planes[0],
1149 mpi->planes[0] + p->crop_x + p->crop_y * mpi->stride[0],
1150 p->w, p->h, p->stride, mpi->stride[0]);
1151 if (mpi->flags & MP_IMGFLAG_PLANAR) {
1152 my_memcpy_pic(planes[1],
1153 mpi->planes[1] + p->crop_cx + p->crop_cy * mpi->stride[1],
1154 p->cw, p->ch, p->chroma_stride, mpi->stride[1]);
1155 my_memcpy_pic(planes[2],
1156 mpi->planes[2] + p->crop_cx + p->crop_cy * mpi->stride[2],
1157 p->cw, p->ch, p->chroma_stride, mpi->stride[2]);
1158 p->num_copies++;
1159 }
1160 }
1161
1162 p->old_planes = planes;
1163 p->chflag = ';';
1164 if (flags & MP_IMGFIELD_ORDERED) {
1165 swapped = !(flags & MP_IMGFIELD_TOP_FIRST);
1166 p->chflag = (flags & MP_IMGFIELD_REPEAT_FIRST ? '|' :
1167 flags & MP_IMGFIELD_TOP_FIRST ? ':' : '.');
1168 }
1169 p->swapped = swapped;
1170
1171 start_time = get_time();
1172 if (p->chflag == '|') {
1173 *s = ppzs;
1174 p->iosync += p->in_inc;
1175 } else if ((p->fast & 1) && prev_chflag == '|')
1176 *s = pprs;
1177 else
1178 diff_fields(p, s, old_planes, planes);
1179 diff_time = get_time();
1180 p->diff_time += diff_time - start_time;
1181 breaks = p->inframes ? find_breaks(p, s) : 2;
1182 p->inframes++;
1183 keep_rate = 4*p->in_inc == p->out_dec;
1184
1185 switch (breaks) {
1186 case 0:
1187 case 8:
1188 case 9:
1189 case 10:
1190 if (!keep_rate && p->notout < p->num_fields && p->iosync < 2*p->in_inc)
1191 break;
1192 if (p->notout < p->num_fields)
1193 dropped_fields = -2;
1194 case 4:
1195 if (keep_rate || p->iosync >= -2*p->in_inc)
1196 show_fields = (4<<p->num_fields)-1;
1197 break;
1198 case 3:
1199 if (keep_rate)
1200 show_fields = 2;
1201 else if (p->iosync > 0) {
1202 if (p->notout >= p->num_fields && p->iosync > 2*p->in_inc) {
1203 show_fields = 4; /* prev odd only */
1204 if (p->num_fields > 1)
1205 show_fields |= 8; /* + prev even */
1206 } else {
1207 show_fields = 2; /* even only */
1208 if (p->notout >= p->num_fields)
1209 dropped_fields += p->num_fields;
1210 }
1211 }
1212 break;
1213 case 2:
1214 if (p->iosync <= -3*p->in_inc) {
1215 if (p->notout >= p->num_fields)
1216 dropped_fields = p->num_fields;
1217 break;
1218 }
1219 if (p->num_fields == 1) {
1220 int prevbreak = ps->sad.noise >= 128;
1221 if (p->iosync < 4*p->in_inc) {
1222 show_fields = 3;
1223 dropped_fields = prevbreak;
1224 } else {
1225 show_fields = 4 | (!prevbreak << 3);
1226 if (p->notout < 1 + p->prev_fields)
1227 dropped_fields = -!prevbreak;
1228 }
1229 break;
1230 }
1231 default:
1232 if (keep_rate)
1233 show_fields = 3 << (breaks & 1);
1234 else if (p->notout >= p->num_fields &&
1235 p->iosync >= (breaks == 1 ? -p->in_inc :
1236 p->in_inc << (p->num_fields == 1))) {
1237 show_fields = (1 << (2 + p->num_fields)) - (1<<breaks);
1238 } else {
1239 if (p->notout >= p->num_fields)
1240 dropped_fields += p->num_fields + 2 - breaks;
1241 if (breaks == 1) {
1242 if (p->iosync >= 4*p->in_inc)
1243 show_fields = 6;
1244 } else if (p->iosync > -3*p->in_inc)
1245 show_fields = 3; /* odd+even */
1246 }
1247 break;
1248 }
1249
1250 show_fields &= 15;
1251 prev = p->prev_fields;
1252 if (breaks < 8) {
1253 if (p->num_fields == 1)
1254 breaks &= ~4;
1255 if (breaks)
1256 p->num_breaks++;
1257 if (breaks == 3)
1258 p->prev_fields = p->num_fields = 1;
1259 else if (breaks) {
1260 p->prev_fields = p->num_fields + (breaks==1) - (breaks==4);
1261 p->num_fields = breaks - (breaks == 4) + (p->chflag == '|');
1262 } else
1263 p->num_fields += 2;
1264 } else
1265 p->num_fields += 2;
1266
1267 p->iosync += 4 * p->in_inc;
1268 if (p->chflag == '|')
1269 p->iosync += p->in_inc;
1270
1271 if (show_fields) {
1272 p->iosync -= p->out_dec;
1273 p->notout = !(show_fields & 1) + !(show_fields & 3);
1274 if (((show_fields & 3) == 3 &&
1275 (s->low.noise + s->interlaced_low < (s->num_blocks>>8) ||
1276 s->sad.noise < 160)) ||
1277 ((show_fields & 12) == 12 &&
1278 (ps->low.noise + ps->interlaced_low < (s->num_blocks>>8) ||
1279 ps->sad.noise < 160))) {
1280 p->export_count++;
1281 dmpi = vf_get_image(vf->next, mpi->imgfmt, MP_IMGTYPE_EXPORT,
1282 MP_IMGFLAG_PRESERVE|MP_IMGFLAG_READABLE,
1283 p->w, p->h);
1284 if ((show_fields & 3) != 3) planes = old_planes;
1285 dmpi->planes[0] = planes[0];
1286 dmpi->stride[0] = p->stride;
1287 dmpi->width = mpi->width;
1288 if (mpi->flags & MP_IMGFLAG_PLANAR) {
1289 dmpi->planes[1] = planes[1];
1290 dmpi->planes[2] = planes[2];
1291 dmpi->stride[1] = p->chroma_stride;
1292 dmpi->stride[2] = p->chroma_stride;
1293 }
1294 } else {
1295 p->merge_count++;
1296 dmpi = vf_get_image(vf->next, mpi->imgfmt,
1297 MP_IMGTYPE_TEMP, MP_IMGFLAG_ACCEPT_STRIDE,
1298 p->w, p->h);
1299 copy_merge_fields(p, dmpi, old_planes, planes, show_fields);
1300 }
1301 p->outframes++;
1302 } else
1303 p->notout += 2;
1304
1305 if (p->verbose)
1306 mp_msg(MSGT_VFILTER, MSGL_INFO, "%lu %lu: %x %c %c %lu%s%s%c%s\n",
1307 p->inframes, p->outframes,
1308 breaks, breaks<8 && breaks>0 ? (int) p->prev_fields+'0' : ' ',
1309 ITOC(show_fields),
1310 p->num_breaks, 5*p->in_inc == p->out_dec && breaks<8 &&
1311 breaks>0 && ((prev&~1)!=2 || prev+p->prev_fields!=5) ?
1312 " ######## bad telecine ########" : "",
1313 dropped_fields ? " ======== dropped ":"", ITOC(dropped_fields),
1314 !show_fields || (show_fields & (show_fields-1)) ?
1315 "" : " @@@@@@@@@@@@@@@@@");
1316
1317 p->merge_time += get_time() - diff_time;
1318 return show_fields ? vf_next_put_image(vf, dmpi, MP_NOPTS_VALUE) : 0;
1319 }
1320
1321 static int query_format(struct vf_instance_s* vf, unsigned int fmt)
1322 {
1323 /* FIXME - support more formats */
1324 switch (fmt) {
1325 case IMGFMT_YV12:
1326 case IMGFMT_IYUV:
1327 case IMGFMT_I420:
1328 case IMGFMT_411P:
1329 case IMGFMT_422P:
1330 case IMGFMT_444P:
1331 return vf_next_query_format(vf, fmt);
1332 }
1333 return 0;
1334 }
1335
1336 static int config(struct vf_instance_s* vf,
1337 int width, int height, int d_width, int d_height,
1338 unsigned int flags, unsigned int outfmt)
1339 {
1340 unsigned long cxm = 0;
1341 unsigned long cym = 0;
1342 struct vf_priv_s *p = vf->priv;
1343 // rounding:
1344 if(!IMGFMT_IS_RGB(outfmt) && !IMGFMT_IS_BGR(outfmt)){
1345 switch(outfmt){
1346 case IMGFMT_444P:
1347 case IMGFMT_Y800:
1348 case IMGFMT_Y8:
1349 break;
1350 case IMGFMT_YVU9:
1351 case IMGFMT_IF09:
1352 cym = 3;
1353 case IMGFMT_411P:
1354 cxm = 3;
1355 break;
1356 case IMGFMT_YV12:
1357 case IMGFMT_I420:
1358 case IMGFMT_IYUV:
1359 cym = 1;
1360 default:
1361 cxm = 1;
1362 }
1363 }
1364 p->chroma_swapped = !!(p->crop_y & (cym+1));
1365 if (p->w) p->w += p->crop_x & cxm;
1366 if (p->h) p->h += p->crop_y & cym;
1367 p->crop_x &= ~cxm;
1368 p->crop_y &= ~cym;
1369 if (!p->w || p->w > width ) p->w = width;
1370 if (!p->h || p->h > height) p->h = height;
1371 if (p->crop_x + p->w > width ) p->crop_x = 0;
1372 if (p->crop_y + p->h > height) p->crop_y = 0;
1373
1374 if(!opt_screen_size_x && !opt_screen_size_y){
1375 d_width = d_width * p->w/width;
1376 d_height = d_height * p->h/height;
1377 }
1378 return vf_next_config(vf, p->w, p->h, d_width, d_height, flags, outfmt);
1379 }
1380
1381 static void uninit(struct vf_instance_s* vf)
1382 {
1383 struct vf_priv_s *p = vf->priv;
1384 mp_msg(MSGT_VFILTER, MSGL_INFO, "diff_time: %.3f, merge_time: %.3f, "
1385 "export: %lu, merge: %lu, copy: %lu\n", p->diff_time, p->merge_time,
1386 p->export_count, p->merge_count, p->num_copies);
1387 free(p->memory_allocated);
1388 free(p);
1389 }
1390
1391 static int open(vf_instance_t *vf, char* args)
1392 {
1393 struct vf_priv_s *p;
1394 vf->get_image = get_image;
1395 vf->put_image = put_image;
1396 vf->config = config;
1397 vf->query_format = query_format;
1398 vf->uninit = uninit;
1399 vf->default_reqs = VFCAP_ACCEPT_STRIDE;
1400 vf->priv = p = calloc(1, sizeof(struct vf_priv_s));
1401 p->out_dec = 5;
1402 p->in_inc = 4;
1403 p->thres.noise = 128;
1404 p->thres.even = 128;
1405 p->sad_thres = 64;
1406 p->dint_thres = 4;
1407 p->luma_only = 0;
1408 p->fast = 3;
1409 p->mmx2 = gCpuCaps.hasMMX2 ? 1 : gCpuCaps.has3DNow ? 2 : 0;
1410 if (args) {
1411 const char *args_remain = parse_args(p, args);
1412 if (args_remain) {
1413 mp_msg(MSGT_VFILTER, MSGL_FATAL,
1414 "filmdint: unknown suboption: %s\n", args_remain);
1415 return 0;
1416 }
1417 if (p->out_dec < p->in_inc) {
1418 mp_msg(MSGT_VFILTER, MSGL_FATAL,
1419 "filmdint: increasing the frame rate is not supported\n");
1420 return 0;
1421 }
1422 }
1423 if (p->mmx2 > 2)
1424 p->mmx2 = 0;
1425 #if !HAVE_MMX
1426 p->mmx2 = 0;
1427 #endif
1428 #if !HAVE_AMD3DNOW
1429 p->mmx2 &= 1;
1430 #endif
1431 p->thres.odd = p->thres.even;
1432 p->thres.temp = p->thres.noise;
1433 p->diff_time = 0;
1434 p->merge_time = 0;
1435 return 1;
1436 }
1437
1438 const vf_info_t vf_info_filmdint = {
1439 "Advanced inverse telecine filer",
1440 "filmdint",
1441 "Zoltan Hidvegi",
1442 "",
1443 open,
1444 NULL
1445 };
mplayer with SMedia Glamo drivers