?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
lavfi: switch to AVFrame.
[FFMpeg-mirror/mplayer-patches.git] / libavcodec / ratecontrol.c
blob21fc8c5b23faa673b8d89050d510ba3ae79544b7
1 /*
2 * Rate control for video encoders
3 *
4 * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
5 *
6 * This file is part of Libav.
7 *
8 * Libav is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * Libav is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with Libav; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23 /**
24 * @file
25 * Rate control for video encoders.
26 */
27
28 #include "avcodec.h"
29 #include "ratecontrol.h"
30 #include "mpegvideo.h"
31 #include "libavutil/eval.h"
32
33 #undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
34 #include <assert.h>
35
36 #ifndef M_E
37 #define M_E 2.718281828
38 #endif
39
40 static int init_pass2(MpegEncContext *s);
41 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
42 double rate_factor, int frame_num);
43
44 void ff_write_pass1_stats(MpegEncContext *s)
45 {
46 snprintf(s->avctx->stats_out, 256,
47 "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d "
48 "fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d;\n",
49 s->current_picture_ptr->f.display_picture_number,
50 s->current_picture_ptr->f.coded_picture_number,
51 s->pict_type,
52 s->current_picture.f.quality,
53 s->i_tex_bits,
54 s->p_tex_bits,
55 s->mv_bits,
56 s->misc_bits,
57 s->f_code,
58 s->b_code,
59 s->current_picture.mc_mb_var_sum,
60 s->current_picture.mb_var_sum,
61 s->i_count, s->skip_count,
62 s->header_bits);
63 }
64
65 static inline double qp2bits(RateControlEntry *rce, double qp)
66 {
67 if (qp <= 0.0) {
68 av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
69 }
70 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
71 }
72
73 static inline double bits2qp(RateControlEntry *rce, double bits)
74 {
75 if (bits < 0.9) {
76 av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
77 }
78 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
79 }
80
81 int ff_rate_control_init(MpegEncContext *s)
82 {
83 RateControlContext *rcc = &s->rc_context;
84 int i, res;
85 static const char * const const_names[] = {
86 "PI",
87 "E",
88 "iTex",
89 "pTex",
90 "tex",
91 "mv",
92 "fCode",
93 "iCount",
94 "mcVar",
95 "var",
96 "isI",
97 "isP",
98 "isB",
99 "avgQP",
100 "qComp",
101 #if 0
102 "lastIQP",
103 "lastPQP",
104 "lastBQP",
105 "nextNonBQP",
106 #endif
107 "avgIITex",
108 "avgPITex",
109 "avgPPTex",
110 "avgBPTex",
111 "avgTex",
112 NULL
113 };
114 static double (* const func1[])(void *, double) = {
115 (void *)bits2qp,
116 (void *)qp2bits,
117 NULL
118 };
119 static const char * const func1_names[] = {
120 "bits2qp",
121 "qp2bits",
122 NULL
123 };
124 emms_c();
125
126 res = av_expr_parse(&rcc->rc_eq_eval,
127 s->avctx->rc_eq ? s->avctx->rc_eq : "tex^qComp",
128 const_names, func1_names, func1,
129 NULL, NULL, 0, s->avctx);
130 if (res < 0) {
131 av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->avctx->rc_eq);
132 return res;
133 }
134
135 for (i = 0; i < 5; i++) {
136 rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
137 rcc->pred[i].count = 1.0;
138 rcc->pred[i].decay = 0.4;
139
140 rcc->i_cplx_sum [i] =
141 rcc->p_cplx_sum [i] =
142 rcc->mv_bits_sum[i] =
143 rcc->qscale_sum [i] =
144 rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
145
146 rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
147 }
148 rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
149
150 if (s->flags & CODEC_FLAG_PASS2) {
151 int i;
152 char *p;
153
154 /* find number of pics */
155 p = s->avctx->stats_in;
156 for (i = -1; p; i++)
157 p = strchr(p + 1, ';');
158 i += s->max_b_frames;
159 if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
160 return -1;
161 rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
162 rcc->num_entries = i;
163
164 /* init all to skipped p frames
165 * (with b frames we might have a not encoded frame at the end FIXME) */
166 for (i = 0; i < rcc->num_entries; i++) {
167 RateControlEntry *rce = &rcc->entry[i];
168
169 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
170 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
171 rce->misc_bits = s->mb_num + 10;
172 rce->mb_var_sum = s->mb_num * 100;
173 }
174
175 /* read stats */
176 p = s->avctx->stats_in;
177 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
178 RateControlEntry *rce;
179 int picture_number;
180 int e;
181 char *next;
182
183 next = strchr(p, ';');
184 if (next) {
185 (*next) = 0; // sscanf in unbelievably slow on looong strings // FIXME copy / do not write
186 next++;
187 }
188 e = sscanf(p, " in:%d ", &picture_number);
189
190 assert(picture_number >= 0);
191 assert(picture_number < rcc->num_entries);
192 rce = &rcc->entry[picture_number];
193
194 e += sscanf(p, " in:%*d out:%*d type:%d q:%f itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d",
195 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
196 &rce->mv_bits, &rce->misc_bits,
197 &rce->f_code, &rce->b_code,
198 &rce->mc_mb_var_sum, &rce->mb_var_sum,
199 &rce->i_count, &rce->skip_count, &rce->header_bits);
200 if (e != 14) {
201 av_log(s->avctx, AV_LOG_ERROR,
202 "statistics are damaged at line %d, parser out=%d\n",
203 i, e);
204 return -1;
205 }
206
207 p = next;
208 }
209
210 if (init_pass2(s) < 0)
211 return -1;
212
213 // FIXME maybe move to end
214 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) {
215 #if CONFIG_LIBXVID
216 return ff_xvid_rate_control_init(s);
217 #else
218 av_log(s->avctx, AV_LOG_ERROR,
219 "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
220 return -1;
221 #endif
222 }
223 }
224
225 if (!(s->flags & CODEC_FLAG_PASS2)) {
226 rcc->short_term_qsum = 0.001;
227 rcc->short_term_qcount = 0.001;
228
229 rcc->pass1_rc_eq_output_sum = 0.001;
230 rcc->pass1_wanted_bits = 0.001;
231
232 if (s->avctx->qblur > 1.0) {
233 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
234 return -1;
235 }
236 /* init stuff with the user specified complexity */
237 if (s->avctx->rc_initial_cplx) {
238 for (i = 0; i < 60 * 30; i++) {
239 double bits = s->avctx->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
240 RateControlEntry rce;
241
242 if (i % ((s->gop_size + 3) / 4) == 0)
243 rce.pict_type = AV_PICTURE_TYPE_I;
244 else if (i % (s->max_b_frames + 1))
245 rce.pict_type = AV_PICTURE_TYPE_B;
246 else
247 rce.pict_type = AV_PICTURE_TYPE_P;
248
249 rce.new_pict_type = rce.pict_type;
250 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
251 rce.mb_var_sum = s->mb_num;
252
253 rce.qscale = FF_QP2LAMBDA * 2;
254 rce.f_code = 2;
255 rce.b_code = 1;
256 rce.misc_bits = 1;
257
258 if (s->pict_type == AV_PICTURE_TYPE_I) {
259 rce.i_count = s->mb_num;
260 rce.i_tex_bits = bits;
261 rce.p_tex_bits = 0;
262 rce.mv_bits = 0;
263 } else {
264 rce.i_count = 0; // FIXME we do know this approx
265 rce.i_tex_bits = 0;
266 rce.p_tex_bits = bits * 0.9;
267 rce.mv_bits = bits * 0.1;
268 }
269 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
270 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
271 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
272 rcc->frame_count[rce.pict_type]++;
273
274 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
275
276 // FIXME misbehaves a little for variable fps
277 rcc->pass1_wanted_bits += s->bit_rate / (1 / av_q2d(s->avctx->time_base));
278 }
279 }
280 }
281
282 return 0;
283 }
284
285 void ff_rate_control_uninit(MpegEncContext *s)
286 {
287 RateControlContext *rcc = &s->rc_context;
288 emms_c();
289
290 av_expr_free(rcc->rc_eq_eval);
291 av_freep(&rcc->entry);
292
293 #if CONFIG_LIBXVID
294 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
295 ff_xvid_rate_control_uninit(s);
296 #endif
297 }
298
299 int ff_vbv_update(MpegEncContext *s, int frame_size)
300 {
301 RateControlContext *rcc = &s->rc_context;
302 const double fps = 1 / av_q2d(s->avctx->time_base);
303 const int buffer_size = s->avctx->rc_buffer_size;
304 const double min_rate = s->avctx->rc_min_rate / fps;
305 const double max_rate = s->avctx->rc_max_rate / fps;
306
307 av_dlog(s, "%d %f %d %f %f\n",
308 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
309
310 if (buffer_size) {
311 int left;
312
313 rcc->buffer_index -= frame_size;
314 if (rcc->buffer_index < 0) {
315 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
316 rcc->buffer_index = 0;
317 }
318
319 left = buffer_size - rcc->buffer_index - 1;
320 rcc->buffer_index += av_clip(left, min_rate, max_rate);
321
322 if (rcc->buffer_index > buffer_size) {
323 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
324
325 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
326 stuffing = 4;
327 rcc->buffer_index -= 8 * stuffing;
328
329 if (s->avctx->debug & FF_DEBUG_RC)
330 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
331
332 return stuffing;
333 }
334 }
335 return 0;
336 }
337
338 /**
339 * Modify the bitrate curve from pass1 for one frame.
340 */
341 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
342 double rate_factor, int frame_num)
343 {
344 RateControlContext *rcc = &s->rc_context;
345 AVCodecContext *a = s->avctx;
346 const int pict_type = rce->new_pict_type;
347 const double mb_num = s->mb_num;
348 double q, bits;
349 int i;
350
351 double const_values[] = {
352 M_PI,
353 M_E,
354 rce->i_tex_bits * rce->qscale,
355 rce->p_tex_bits * rce->qscale,
356 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
357 rce->mv_bits / mb_num,
358 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
359 rce->i_count / mb_num,
360 rce->mc_mb_var_sum / mb_num,
361 rce->mb_var_sum / mb_num,
362 rce->pict_type == AV_PICTURE_TYPE_I,
363 rce->pict_type == AV_PICTURE_TYPE_P,
364 rce->pict_type == AV_PICTURE_TYPE_B,
365 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
366 a->qcompress,
367 #if 0
368 rcc->last_qscale_for[AV_PICTURE_TYPE_I],
369 rcc->last_qscale_for[AV_PICTURE_TYPE_P],
370 rcc->last_qscale_for[AV_PICTURE_TYPE_B],
371 rcc->next_non_b_qscale,
372 #endif
373 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
374 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
375 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
376 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
377 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
378 0
379 };
380
381 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
382 if (isnan(bits)) {
383 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
384 return -1;
385 }
386
387 rcc->pass1_rc_eq_output_sum += bits;
388 bits *= rate_factor;
389 if (bits < 0.0)
390 bits = 0.0;
391 bits += 1.0; // avoid 1/0 issues
392
393 /* user override */
394 for (i = 0; i < s->avctx->rc_override_count; i++) {
395 RcOverride *rco = s->avctx->rc_override;
396 if (rco[i].start_frame > frame_num)
397 continue;
398 if (rco[i].end_frame < frame_num)
399 continue;
400
401 if (rco[i].qscale)
402 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
403 else
404 bits *= rco[i].quality_factor;
405 }
406
407 q = bits2qp(rce, bits);
408
409 /* I/B difference */
410 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
411 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
412 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
413 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
414 if (q < 1)
415 q = 1;
416
417 return q;
418 }
419
420 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
421 {
422 RateControlContext *rcc = &s->rc_context;
423 AVCodecContext *a = s->avctx;
424 const int pict_type = rce->new_pict_type;
425 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
426 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
427
428 if (pict_type == AV_PICTURE_TYPE_I &&
429 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
430 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
431 else if (pict_type == AV_PICTURE_TYPE_B &&
432 a->b_quant_factor > 0.0)
433 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
434 if (q < 1)
435 q = 1;
436
437 /* last qscale / qdiff stuff */
438 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
439 double last_q = rcc->last_qscale_for[pict_type];
440 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
441
442 if (q > last_q + maxdiff)
443 q = last_q + maxdiff;
444 else if (q < last_q - maxdiff)
445 q = last_q - maxdiff;
446 }
447
448 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
449
450 if (pict_type != AV_PICTURE_TYPE_B)
451 rcc->last_non_b_pict_type = pict_type;
452
453 return q;
454 }
455
456 /**
457 * Get the qmin & qmax for pict_type.
458 */
459 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
460 {
461 int qmin = s->avctx->lmin;
462 int qmax = s->avctx->lmax;
463
464 assert(qmin <= qmax);
465
466 switch (pict_type) {
467 case AV_PICTURE_TYPE_B:
468 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
469 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
470 break;
471 case AV_PICTURE_TYPE_I:
472 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
473 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
474 break;
475 }
476
477 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
478 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
479
480 if (qmax < qmin)
481 qmax = qmin;
482
483 *qmin_ret = qmin;
484 *qmax_ret = qmax;
485 }
486
487 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
488 double q, int frame_num)
489 {
490 RateControlContext *rcc = &s->rc_context;
491 const double buffer_size = s->avctx->rc_buffer_size;
492 const double fps = 1 / av_q2d(s->avctx->time_base);
493 const double min_rate = s->avctx->rc_min_rate / fps;
494 const double max_rate = s->avctx->rc_max_rate / fps;
495 const int pict_type = rce->new_pict_type;
496 int qmin, qmax;
497
498 get_qminmax(&qmin, &qmax, s, pict_type);
499
500 /* modulation */
501 if (s->avctx->rc_qmod_freq &&
502 frame_num % s->avctx->rc_qmod_freq == 0 &&
503 pict_type == AV_PICTURE_TYPE_P)
504 q *= s->avctx->rc_qmod_amp;
505
506 /* buffer overflow/underflow protection */
507 if (buffer_size) {
508 double expected_size = rcc->buffer_index;
509 double q_limit;
510
511 if (min_rate) {
512 double d = 2 * (buffer_size - expected_size) / buffer_size;
513 if (d > 1.0)
514 d = 1.0;
515 else if (d < 0.0001)
516 d = 0.0001;
517 q *= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
518
519 q_limit = bits2qp(rce,
520 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
521 s->avctx->rc_min_vbv_overflow_use, 1));
522
523 if (q > q_limit) {
524 if (s->avctx->debug & FF_DEBUG_RC)
525 av_log(s->avctx, AV_LOG_DEBUG,
526 "limiting QP %f -> %f\n", q, q_limit);
527 q = q_limit;
528 }
529 }
530
531 if (max_rate) {
532 double d = 2 * expected_size / buffer_size;
533 if (d > 1.0)
534 d = 1.0;
535 else if (d < 0.0001)
536 d = 0.0001;
537 q /= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
538
539 q_limit = bits2qp(rce,
540 FFMAX(rcc->buffer_index *
541 s->avctx->rc_max_available_vbv_use,
542 1));
543 if (q < q_limit) {
544 if (s->avctx->debug & FF_DEBUG_RC)
545 av_log(s->avctx, AV_LOG_DEBUG,
546 "limiting QP %f -> %f\n", q, q_limit);
547 q = q_limit;
548 }
549 }
550 }
551 av_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
552 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
553 s->avctx->rc_buffer_aggressivity);
554 if (s->avctx->rc_qsquish == 0.0 || qmin == qmax) {
555 if (q < qmin)
556 q = qmin;
557 else if (q > qmax)
558 q = qmax;
559 } else {
560 double min2 = log(qmin);
561 double max2 = log(qmax);
562
563 q = log(q);
564 q = (q - min2) / (max2 - min2) - 0.5;
565 q *= -4.0;
566 q = 1.0 / (1.0 + exp(q));
567 q = q * (max2 - min2) + min2;
568
569 q = exp(q);
570 }
571
572 return q;
573 }
574
575 // ----------------------------------
576 // 1 Pass Code
577
578 static double predict_size(Predictor *p, double q, double var)
579 {
580 return p->coeff * var / (q * p->count);
581 }
582
583 static void update_predictor(Predictor *p, double q, double var, double size)
584 {
585 double new_coeff = size * q / (var + 1);
586 if (var < 10)
587 return;
588
589 p->count *= p->decay;
590 p->coeff *= p->decay;
591 p->count++;
592 p->coeff += new_coeff;
593 }
594
595 static void adaptive_quantization(MpegEncContext *s, double q)
596 {
597 int i;
598 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
599 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
600 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
601 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
602 const float p_masking = s->avctx->p_masking;
603 const float border_masking = s->avctx->border_masking;
604 float bits_sum = 0.0;
605 float cplx_sum = 0.0;
606 float *cplx_tab = s->cplx_tab;
607 float *bits_tab = s->bits_tab;
608 const int qmin = s->avctx->mb_lmin;
609 const int qmax = s->avctx->mb_lmax;
610 Picture *const pic = &s->current_picture;
611 const int mb_width = s->mb_width;
612 const int mb_height = s->mb_height;
613
614 for (i = 0; i < s->mb_num; i++) {
615 const int mb_xy = s->mb_index2xy[i];
616 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
617 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
618 const int lumi = pic->mb_mean[mb_xy];
619 float bits, cplx, factor;
620 int mb_x = mb_xy % s->mb_stride;
621 int mb_y = mb_xy / s->mb_stride;
622 int mb_distance;
623 float mb_factor = 0.0;
624 if (spat_cplx < 4)
625 spat_cplx = 4; // FIXME finetune
626 if (temp_cplx < 4)
627 temp_cplx = 4; // FIXME finetune
628
629 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
630 cplx = spat_cplx;
631 factor = 1.0 + p_masking;
632 } else {
633 cplx = temp_cplx;
634 factor = pow(temp_cplx, -temp_cplx_masking);
635 }
636 factor *= pow(spat_cplx, -spatial_cplx_masking);
637
638 if (lumi > 127)
639 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
640 else
641 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
642
643 if (mb_x < mb_width / 5) {
644 mb_distance = mb_width / 5 - mb_x;
645 mb_factor = (float)mb_distance / (float)(mb_width / 5);
646 } else if (mb_x > 4 * mb_width / 5) {
647 mb_distance = mb_x - 4 * mb_width / 5;
648 mb_factor = (float)mb_distance / (float)(mb_width / 5);
649 }
650 if (mb_y < mb_height / 5) {
651 mb_distance = mb_height / 5 - mb_y;
652 mb_factor = FFMAX(mb_factor,
653 (float)mb_distance / (float)(mb_height / 5));
654 } else if (mb_y > 4 * mb_height / 5) {
655 mb_distance = mb_y - 4 * mb_height / 5;
656 mb_factor = FFMAX(mb_factor,
657 (float)mb_distance / (float)(mb_height / 5));
658 }
659
660 factor *= 1.0 - border_masking * mb_factor;
661
662 if (factor < 0.00001)
663 factor = 0.00001;
664
665 bits = cplx * factor;
666 cplx_sum += cplx;
667 bits_sum += bits;
668 cplx_tab[i] = cplx;
669 bits_tab[i] = bits;
670 }
671
672 /* handle qmin/qmax clipping */
673 if (s->flags & CODEC_FLAG_NORMALIZE_AQP) {
674 float factor = bits_sum / cplx_sum;
675 for (i = 0; i < s->mb_num; i++) {
676 float newq = q * cplx_tab[i] / bits_tab[i];
677 newq *= factor;
678
679 if (newq > qmax) {
680 bits_sum -= bits_tab[i];
681 cplx_sum -= cplx_tab[i] * q / qmax;
682 } else if (newq < qmin) {
683 bits_sum -= bits_tab[i];
684 cplx_sum -= cplx_tab[i] * q / qmin;
685 }
686 }
687 if (bits_sum < 0.001)
688 bits_sum = 0.001;
689 if (cplx_sum < 0.001)
690 cplx_sum = 0.001;
691 }
692
693 for (i = 0; i < s->mb_num; i++) {
694 const int mb_xy = s->mb_index2xy[i];
695 float newq = q * cplx_tab[i] / bits_tab[i];
696 int intq;
697
698 if (s->flags & CODEC_FLAG_NORMALIZE_AQP) {
699 newq *= bits_sum / cplx_sum;
700 }
701
702 intq = (int)(newq + 0.5);
703
704 if (intq > qmax)
705 intq = qmax;
706 else if (intq < qmin)
707 intq = qmin;
708 s->lambda_table[mb_xy] = intq;
709 }
710 }
711
712 void ff_get_2pass_fcode(MpegEncContext *s)
713 {
714 RateControlContext *rcc = &s->rc_context;
715 RateControlEntry *rce = &rcc->entry[s->picture_number];
716
717 s->f_code = rce->f_code;
718 s->b_code = rce->b_code;
719 }
720
721 // FIXME rd or at least approx for dquant
722
723 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
724 {
725 float q;
726 int qmin, qmax;
727 float br_compensation;
728 double diff;
729 double short_term_q;
730 double fps;
731 int picture_number = s->picture_number;
732 int64_t wanted_bits;
733 RateControlContext *rcc = &s->rc_context;
734 AVCodecContext *a = s->avctx;
735 RateControlEntry local_rce, *rce;
736 double bits;
737 double rate_factor;
738 int var;
739 const int pict_type = s->pict_type;
740 Picture * const pic = &s->current_picture;
741 emms_c();
742
743 #if CONFIG_LIBXVID
744 if ((s->flags & CODEC_FLAG_PASS2) &&
745 s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
746 return ff_xvid_rate_estimate_qscale(s, dry_run);
747 #endif
748
749 get_qminmax(&qmin, &qmax, s, pict_type);
750
751 fps = 1 / av_q2d(s->avctx->time_base);
752 /* update predictors */
753 if (picture_number > 2 && !dry_run) {
754 const int last_var = s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
755 : rcc->last_mc_mb_var_sum;
756 update_predictor(&rcc->pred[s->last_pict_type],
757 rcc->last_qscale,
758 sqrt(last_var), s->frame_bits);
759 }
760
761 if (s->flags & CODEC_FLAG_PASS2) {
762 assert(picture_number >= 0);
763 assert(picture_number < rcc->num_entries);
764 rce = &rcc->entry[picture_number];
765 wanted_bits = rce->expected_bits;
766 } else {
767 Picture *dts_pic;
768 rce = &local_rce;
769
770 /* FIXME add a dts field to AVFrame and ensure it is set and use it
771 * here instead of reordering but the reordering is simpler for now
772 * until H.264 B-pyramid must be handled. */
773 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
774 dts_pic = s->current_picture_ptr;
775 else
776 dts_pic = s->last_picture_ptr;
777
778 if (!dts_pic || dts_pic->f.pts == AV_NOPTS_VALUE)
779 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
780 else
781 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f.pts / fps);
782 }
783
784 diff = s->total_bits - wanted_bits;
785 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
786 if (br_compensation <= 0.0)
787 br_compensation = 0.001;
788
789 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
790
791 short_term_q = 0; /* avoid warning */
792 if (s->flags & CODEC_FLAG_PASS2) {
793 if (pict_type != AV_PICTURE_TYPE_I)
794 assert(pict_type == rce->new_pict_type);
795
796 q = rce->new_qscale / br_compensation;
797 av_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale,
798 br_compensation, s->frame_bits, var, pict_type);
799 } else {
800 rce->pict_type =
801 rce->new_pict_type = pict_type;
802 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
803 rce->mb_var_sum = pic->mb_var_sum;
804 rce->qscale = FF_QP2LAMBDA * 2;
805 rce->f_code = s->f_code;
806 rce->b_code = s->b_code;
807 rce->misc_bits = 1;
808
809 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
810 if (pict_type == AV_PICTURE_TYPE_I) {
811 rce->i_count = s->mb_num;
812 rce->i_tex_bits = bits;
813 rce->p_tex_bits = 0;
814 rce->mv_bits = 0;
815 } else {
816 rce->i_count = 0; // FIXME we do know this approx
817 rce->i_tex_bits = 0;
818 rce->p_tex_bits = bits * 0.9;
819 rce->mv_bits = bits * 0.1;
820 }
821 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
822 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
823 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
824 rcc->frame_count[pict_type]++;
825
826 bits = rce->i_tex_bits + rce->p_tex_bits;
827 rate_factor = rcc->pass1_wanted_bits /
828 rcc->pass1_rc_eq_output_sum * br_compensation;
829
830 q = get_qscale(s, rce, rate_factor, picture_number);
831 if (q < 0)
832 return -1;
833
834 assert(q > 0.0);
835 q = get_diff_limited_q(s, rce, q);
836 assert(q > 0.0);
837
838 // FIXME type dependent blur like in 2-pass
839 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
840 rcc->short_term_qsum *= a->qblur;
841 rcc->short_term_qcount *= a->qblur;
842
843 rcc->short_term_qsum += q;
844 rcc->short_term_qcount++;
845 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
846 }
847 assert(q > 0.0);
848
849 q = modify_qscale(s, rce, q, picture_number);
850
851 rcc->pass1_wanted_bits += s->bit_rate / fps;
852
853 assert(q > 0.0);
854 }
855
856 if (s->avctx->debug & FF_DEBUG_RC) {
857 av_log(s->avctx, AV_LOG_DEBUG,
858 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
859 "size:%d var:%d/%d br:%d fps:%d\n",
860 av_get_picture_type_char(pict_type),
861 qmin, q, qmax, picture_number,
862 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
863 br_compensation, short_term_q, s->frame_bits,
864 pic->mb_var_sum, pic->mc_mb_var_sum,
865 s->bit_rate / 1000, (int)fps);
866 }
867
868 if (q < qmin)
869 q = qmin;
870 else if (q > qmax)
871 q = qmax;
872
873 if (s->adaptive_quant)
874 adaptive_quantization(s, q);
875 else
876 q = (int)(q + 0.5);
877
878 if (!dry_run) {
879 rcc->last_qscale = q;
880 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
881 rcc->last_mb_var_sum = pic->mb_var_sum;
882 }
883 return q;
884 }
885
886 // ----------------------------------------------
887 // 2-Pass code
888
889 static int init_pass2(MpegEncContext *s)
890 {
891 RateControlContext *rcc = &s->rc_context;
892 AVCodecContext *a = s->avctx;
893 int i, toobig;
894 double fps = 1 / av_q2d(s->avctx->time_base);
895 double complexity[5] = { 0 }; // approximate bits at quant=1
896 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
897 uint64_t all_const_bits;
898 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
899 (double)rcc->num_entries / fps);
900 double rate_factor = 0;
901 double step;
902 const int filter_size = (int)(a->qblur * 4) | 1;
903 double expected_bits;
904 double *qscale, *blurred_qscale, qscale_sum;
905
906 /* find complexity & const_bits & decide the pict_types */
907 for (i = 0; i < rcc->num_entries; i++) {
908 RateControlEntry *rce = &rcc->entry[i];
909
910 rce->new_pict_type = rce->pict_type;
911 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
912 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
913 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
914 rcc->frame_count[rce->pict_type]++;
915
916 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
917 (double)rce->qscale;
918 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
919 }
920
921 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
922 const_bits[AV_PICTURE_TYPE_P] +
923 const_bits[AV_PICTURE_TYPE_B];
924
925 if (all_available_bits < all_const_bits) {
926 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
927 return -1;
928 }
929
930 qscale = av_malloc(sizeof(double) * rcc->num_entries);
931 blurred_qscale = av_malloc(sizeof(double) * rcc->num_entries);
932 toobig = 0;
933
934 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
935 expected_bits = 0;
936 rate_factor += step;
937
938 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
939
940 /* find qscale */
941 for (i = 0; i < rcc->num_entries; i++) {
942 RateControlEntry *rce = &rcc->entry[i];
943
944 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
945 rcc->last_qscale_for[rce->pict_type] = qscale[i];
946 }
947 assert(filter_size % 2 == 1);
948
949 /* fixed I/B QP relative to P mode */
950 for (i = rcc->num_entries - 1; i >= 0; i--) {
951 RateControlEntry *rce = &rcc->entry[i];
952
953 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
954 }
955
956 /* smooth curve */
957 for (i = 0; i < rcc->num_entries; i++) {
958 RateControlEntry *rce = &rcc->entry[i];
959 const int pict_type = rce->new_pict_type;
960 int j;
961 double q = 0.0, sum = 0.0;
962
963 for (j = 0; j < filter_size; j++) {
964 int index = i + j - filter_size / 2;
965 double d = index - i;
966 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
967
968 if (index < 0 || index >= rcc->num_entries)
969 continue;
970 if (pict_type != rcc->entry[index].new_pict_type)
971 continue;
972 q += qscale[index] * coeff;
973 sum += coeff;
974 }
975 blurred_qscale[i] = q / sum;
976 }
977
978 /* find expected bits */
979 for (i = 0; i < rcc->num_entries; i++) {
980 RateControlEntry *rce = &rcc->entry[i];
981 double bits;
982
983 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
984
985 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
986 bits += 8 * ff_vbv_update(s, bits);
987
988 rce->expected_bits = expected_bits;
989 expected_bits += bits;
990 }
991
992 av_dlog(s->avctx,
993 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
994 expected_bits, (int)all_available_bits, rate_factor);
995 if (expected_bits > all_available_bits) {
996 rate_factor -= step;
997 ++toobig;
998 }
999 }
1000 av_free(qscale);
1001 av_free(blurred_qscale);
1002
1003 /* check bitrate calculations and print info */
1004 qscale_sum = 0.0;
1005 for (i = 0; i < rcc->num_entries; i++) {
1006 av_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
1007 i,
1008 rcc->entry[i].new_qscale,
1009 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
1010 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
1011 s->avctx->qmin, s->avctx->qmax);
1012 }
1013 assert(toobig <= 40);
1014 av_log(s->avctx, AV_LOG_DEBUG,
1015 "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
1016 s->bit_rate,
1017 (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
1018 av_log(s->avctx, AV_LOG_DEBUG,
1019 "[lavc rc] estimated target average qp: %.3f\n",
1020 (float)qscale_sum / rcc->num_entries);
1021 if (toobig == 0) {
1022 av_log(s->avctx, AV_LOG_INFO,
1023 "[lavc rc] Using all of requested bitrate is not "
1024 "necessary for this video with these parameters.\n");
1025 } else if (toobig == 40) {
1026 av_log(s->avctx, AV_LOG_ERROR,
1027 "[lavc rc] Error: bitrate too low for this video "
1028 "with these parameters.\n");
1029 return -1;
1030 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
1031 av_log(s->avctx, AV_LOG_ERROR,
1032 "[lavc rc] Error: 2pass curve failed to converge\n");
1033 return -1;
1034 }
1035
1036 return 0;
1037 }
A couple of patches to make building with MPlayer easier