repositories / deb_ffmpeg.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Imported Debian version 2.4.3~trusty1
[deb_ffmpeg.git] / ffmpeg / libavcodec / ratecontrol.c
CommitLineData
2ba45a60
DM		 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 FFmpeg.
7 *
8 * FFmpeg 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 * FFmpeg 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 FFmpeg; 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 "libavutil/attributes.h"
29#include "avcodec.h"
30#include "ratecontrol.h"
31#include "mpegutils.h"
32#include "mpegvideo.h"
33#include "libavutil/eval.h"
34
35#undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
36#include <assert.h>
37
38#ifndef M_E
39#define M_E 2.718281828
40#endif
41
42static int init_pass2(MpegEncContext *s);
43static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
44 double rate_factor, int frame_num);
45
46void ff_write_pass1_stats(MpegEncContext *s)
47{
48 snprintf(s->avctx->stats_out, 256,
49 "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d "
50 "fcode:%d bcode:%d mc-var:%"PRId64" var:%"PRId64" icount:%d skipcount:%d hbits:%d;\n",
51 s->current_picture_ptr->f->display_picture_number,
52 s->current_picture_ptr->f->coded_picture_number,
53 s->pict_type,
54 s->current_picture.f->quality,
55 s->i_tex_bits,
56 s->p_tex_bits,
57 s->mv_bits,
58 s->misc_bits,
59 s->f_code,
60 s->b_code,
61 s->current_picture.mc_mb_var_sum,
62 s->current_picture.mb_var_sum,
63 s->i_count, s->skip_count,
64 s->header_bits);
65}
66
67static double get_fps(AVCodecContext *avctx)
68{
69 return 1.0 / av_q2d(avctx->time_base) / FFMAX(avctx->ticks_per_frame, 1);
70}
71
72static inline double qp2bits(RateControlEntry *rce, double qp)
73{
74 if (qp <= 0.0) {
75 av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
76 }
77 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
78}
79
80static inline double bits2qp(RateControlEntry *rce, double bits)
81{
82 if (bits < 0.9) {
83 av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
84 }
85 return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
86}
87
88av_cold int ff_rate_control_init(MpegEncContext *s)
89{
90 RateControlContext *rcc = &s->rc_context;
91 int i, res;
92 static const char * const const_names[] = {
93 "PI",
94 "E",
95 "iTex",
96 "pTex",
97 "tex",
98 "mv",
99 "fCode",
100 "iCount",
101 "mcVar",
102 "var",
103 "isI",
104 "isP",
105 "isB",
106 "avgQP",
107 "qComp",
108#if 0
109 "lastIQP",
110 "lastPQP",
111 "lastBQP",
112 "nextNonBQP",
113#endif
114 "avgIITex",
115 "avgPITex",
116 "avgPPTex",
117 "avgBPTex",
118 "avgTex",
119 NULL
120 };
121 static double (* const func1[])(void *, double) = {
122 (void *)bits2qp,
123 (void *)qp2bits,
124 NULL
125 };
126 static const char * const func1_names[] = {
127 "bits2qp",
128 "qp2bits",
129 NULL
130 };
131 emms_c();
132
133 if (!s->avctx->rc_max_available_vbv_use && s->avctx->rc_buffer_size) {
134 if (s->avctx->rc_max_rate) {
135 s->avctx->rc_max_available_vbv_use = av_clipf(s->avctx->rc_max_rate/(s->avctx->rc_buffer_size*get_fps(s->avctx)), 1.0/3, 1.0);
136 } else
137 s->avctx->rc_max_available_vbv_use = 1.0;
138 }
139
140 res = av_expr_parse(&rcc->rc_eq_eval,
141 s->avctx->rc_eq ? s->avctx->rc_eq : "tex^qComp",
142 const_names, func1_names, func1,
143 NULL, NULL, 0, s->avctx);
144 if (res < 0) {
145 av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->avctx->rc_eq);
146 return res;
147 }
148
149 for (i = 0; i < 5; i++) {
150 rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
151 rcc->pred[i].count = 1.0;
152 rcc->pred[i].decay = 0.4;
153
154 rcc->i_cplx_sum [i] =
155 rcc->p_cplx_sum [i] =
156 rcc->mv_bits_sum[i] =
157 rcc->qscale_sum [i] =
158 rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
159
160 rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
161 }
162 rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy;
163 if (!rcc->buffer_index)
164 rcc->buffer_index = s->avctx->rc_buffer_size * 3 / 4;
165
166 if (s->flags & CODEC_FLAG_PASS2) {
167 int i;
168 char *p;
169
170 /* find number of pics */
171 p = s->avctx->stats_in;
172 for (i = -1; p; i++)
173 p = strchr(p + 1, ';');
174 i += s->max_b_frames;
175 if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
176 return -1;
177 rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
178 rcc->num_entries = i;
179
180 /* init all to skipped p frames
181 * (with b frames we might have a not encoded frame at the end FIXME) */
182 for (i = 0; i < rcc->num_entries; i++) {
183 RateControlEntry *rce = &rcc->entry[i];
184
185 rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P;
186 rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
187 rce->misc_bits = s->mb_num + 10;
188 rce->mb_var_sum = s->mb_num * 100;
189 }
190
191 /* read stats */
192 p = s->avctx->stats_in;
193 for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
194 RateControlEntry *rce;
195 int picture_number;
196 int e;
197 char *next;
198
199 next = strchr(p, ';');
200 if (next) {
201 (*next) = 0; // sscanf in unbelievably slow on looong strings // FIXME copy / do not write
202 next++;
203 }
204 e = sscanf(p, " in:%d ", &picture_number);
205
206 assert(picture_number >= 0);
207 assert(picture_number < rcc->num_entries);
208 rce = &rcc->entry[picture_number];
209
210 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:%"SCNd64" var:%"SCNd64" icount:%d skipcount:%d hbits:%d",
211 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
212 &rce->mv_bits, &rce->misc_bits,
213 &rce->f_code, &rce->b_code,
214 &rce->mc_mb_var_sum, &rce->mb_var_sum,
215 &rce->i_count, &rce->skip_count, &rce->header_bits);
216 if (e != 14) {
217 av_log(s->avctx, AV_LOG_ERROR,
218 "statistics are damaged at line %d, parser out=%d\n",
219 i, e);
220 return -1;
221 }
222
223 p = next;
224 }
225
226 if (init_pass2(s) < 0)
227 return -1;
228
229 // FIXME maybe move to end
230 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) {
231#if CONFIG_LIBXVID
232 return ff_xvid_rate_control_init(s);
233#else
234 av_log(s->avctx, AV_LOG_ERROR,
235 "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
236 return -1;
237#endif
238 }
239 }
240
241 if (!(s->flags & CODEC_FLAG_PASS2)) {
242 rcc->short_term_qsum = 0.001;
243 rcc->short_term_qcount = 0.001;
244
245 rcc->pass1_rc_eq_output_sum = 0.001;
246 rcc->pass1_wanted_bits = 0.001;
247
248 if (s->avctx->qblur > 1.0) {
249 av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
250 return -1;
251 }
252 /* init stuff with the user specified complexity */
253 if (s->avctx->rc_initial_cplx) {
254 for (i = 0; i < 60 * 30; i++) {
255 double bits = s->avctx->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
256 RateControlEntry rce;
257
258 if (i % ((s->gop_size + 3) / 4) == 0)
259 rce.pict_type = AV_PICTURE_TYPE_I;
260 else if (i % (s->max_b_frames + 1))
261 rce.pict_type = AV_PICTURE_TYPE_B;
262 else
263 rce.pict_type = AV_PICTURE_TYPE_P;
264
265 rce.new_pict_type = rce.pict_type;
266 rce.mc_mb_var_sum = bits * s->mb_num / 100000;
267 rce.mb_var_sum = s->mb_num;
268
269 rce.qscale = FF_QP2LAMBDA * 2;
270 rce.f_code = 2;
271 rce.b_code = 1;
272 rce.misc_bits = 1;
273
274 if (s->pict_type == AV_PICTURE_TYPE_I) {
275 rce.i_count = s->mb_num;
276 rce.i_tex_bits = bits;
277 rce.p_tex_bits = 0;
278 rce.mv_bits = 0;
279 } else {
280 rce.i_count = 0; // FIXME we do know this approx
281 rce.i_tex_bits = 0;
282 rce.p_tex_bits = bits * 0.9;
283 rce.mv_bits = bits * 0.1;
284 }
285 rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
286 rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
287 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
288 rcc->frame_count[rce.pict_type]++;
289
290 get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
291
292 // FIXME misbehaves a little for variable fps
293 rcc->pass1_wanted_bits += s->bit_rate / get_fps(s->avctx);
294 }
295 }
296 }
297
298 return 0;
299}
300
301av_cold void ff_rate_control_uninit(MpegEncContext *s)
302{
303 RateControlContext *rcc = &s->rc_context;
304 emms_c();
305
306 av_expr_free(rcc->rc_eq_eval);
307 av_freep(&rcc->entry);
308
309#if CONFIG_LIBXVID
310 if ((s->flags & CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
311 ff_xvid_rate_control_uninit(s);
312#endif
313}
314
315int ff_vbv_update(MpegEncContext *s, int frame_size)
316{
317 RateControlContext *rcc = &s->rc_context;
318 const double fps = get_fps(s->avctx);
319 const int buffer_size = s->avctx->rc_buffer_size;
320 const double min_rate = s->avctx->rc_min_rate / fps;
321 const double max_rate = s->avctx->rc_max_rate / fps;
322
323 av_dlog(s, "%d %f %d %f %f\n",
324 buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
325
326 if (buffer_size) {
327 int left;
328
329 rcc->buffer_index -= frame_size;
330 if (rcc->buffer_index < 0) {
331 av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
332 if (frame_size > max_rate && s->qscale == s->avctx->qmax) {
333 av_log(s->avctx, AV_LOG_ERROR, "max bitrate possibly too small or try trellis with large lmax or increase qmax\n");
334 }
335 rcc->buffer_index = 0;
336 }
337
338 left = buffer_size - rcc->buffer_index - 1;
339 rcc->buffer_index += av_clip(left, min_rate, max_rate);
340
341 if (rcc->buffer_index > buffer_size) {
342 int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
343
344 if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
345 stuffing = 4;
346 rcc->buffer_index -= 8 * stuffing;
347
348 if (s->avctx->debug & FF_DEBUG_RC)
349 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
350
351 return stuffing;
352 }
353 }
354 return 0;
355}
356
357/**
358 * Modify the bitrate curve from pass1 for one frame.
359 */
360static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
361 double rate_factor, int frame_num)
362{
363 RateControlContext *rcc = &s->rc_context;
364 AVCodecContext *a = s->avctx;
365 const int pict_type = rce->new_pict_type;
366 const double mb_num = s->mb_num;
367 double q, bits;
368 int i;
369
370 double const_values[] = {
371 M_PI,
372 M_E,
373 rce->i_tex_bits * rce->qscale,
374 rce->p_tex_bits * rce->qscale,
375 (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
376 rce->mv_bits / mb_num,
377 rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
378 rce->i_count / mb_num,
379 rce->mc_mb_var_sum / mb_num,
380 rce->mb_var_sum / mb_num,
381 rce->pict_type == AV_PICTURE_TYPE_I,
382 rce->pict_type == AV_PICTURE_TYPE_P,
383 rce->pict_type == AV_PICTURE_TYPE_B,
384 rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
385 a->qcompress,
386#if 0
387 rcc->last_qscale_for[AV_PICTURE_TYPE_I],
388 rcc->last_qscale_for[AV_PICTURE_TYPE_P],
389 rcc->last_qscale_for[AV_PICTURE_TYPE_B],
390 rcc->next_non_b_qscale,
391#endif
392 rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
393 rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
394 rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
395 rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
396 (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
397 0
398 };
399
400 bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
401 if (isnan(bits)) {
402 av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
403 return -1;
404 }
405
406 rcc->pass1_rc_eq_output_sum += bits;
407 bits *= rate_factor;
408 if (bits < 0.0)
409 bits = 0.0;
410 bits += 1.0; // avoid 1/0 issues
411
412 /* user override */
413 for (i = 0; i < s->avctx->rc_override_count; i++) {
414 RcOverride *rco = s->avctx->rc_override;
415 if (rco[i].start_frame > frame_num)
416 continue;
417 if (rco[i].end_frame < frame_num)
418 continue;
419
420 if (rco[i].qscale)
421 bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
422 else
423 bits *= rco[i].quality_factor;
424 }
425
426 q = bits2qp(rce, bits);
427
428 /* I/B difference */
429 if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
430 q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
431 else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
432 q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
433 if (q < 1)
434 q = 1;
435
436 return q;
437}
438
439static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
440{
441 RateControlContext *rcc = &s->rc_context;
442 AVCodecContext *a = s->avctx;
443 const int pict_type = rce->new_pict_type;
444 const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
445 const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
446
447 if (pict_type == AV_PICTURE_TYPE_I &&
448 (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P))
449 q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
450 else if (pict_type == AV_PICTURE_TYPE_B &&
451 a->b_quant_factor > 0.0)
452 q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
453 if (q < 1)
454 q = 1;
455
456 /* last qscale / qdiff stuff */
457 if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
458 double last_q = rcc->last_qscale_for[pict_type];
459 const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
460
461 if (q > last_q + maxdiff)
462 q = last_q + maxdiff;
463 else if (q < last_q - maxdiff)
464 q = last_q - maxdiff;
465 }
466
467 rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
468
469 if (pict_type != AV_PICTURE_TYPE_B)
470 rcc->last_non_b_pict_type = pict_type;
471
472 return q;
473}
474
475/**
476 * Get the qmin & qmax for pict_type.
477 */
478static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
479{
480 int qmin = s->avctx->lmin;
481 int qmax = s->avctx->lmax;
482
483 assert(qmin <= qmax);
484
485 switch (pict_type) {
486 case AV_PICTURE_TYPE_B:
487 qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
488 qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
489 break;
490 case AV_PICTURE_TYPE_I:
491 qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
492 qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
493 break;
494 }
495
496 qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
497 qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
498
499 if (qmax < qmin)
500 qmax = qmin;
501
502 *qmin_ret = qmin;
503 *qmax_ret = qmax;
504}
505
506static double modify_qscale(MpegEncContext *s, RateControlEntry *rce,
507 double q, int frame_num)
508{
509 RateControlContext *rcc = &s->rc_context;
510 const double buffer_size = s->avctx->rc_buffer_size;
511 const double fps = get_fps(s->avctx);
512 const double min_rate = s->avctx->rc_min_rate / fps;
513 const double max_rate = s->avctx->rc_max_rate / fps;
514 const int pict_type = rce->new_pict_type;
515 int qmin, qmax;
516
517 get_qminmax(&qmin, &qmax, s, pict_type);
518
519 /* modulation */
520 if (s->avctx->rc_qmod_freq &&
521 frame_num % s->avctx->rc_qmod_freq == 0 &&
522 pict_type == AV_PICTURE_TYPE_P)
523 q *= s->avctx->rc_qmod_amp;
524
525 /* buffer overflow/underflow protection */
526 if (buffer_size) {
527 double expected_size = rcc->buffer_index;
528 double q_limit;
529
530 if (min_rate) {
531 double d = 2 * (buffer_size - expected_size) / buffer_size;
532 if (d > 1.0)
533 d = 1.0;
534 else if (d < 0.0001)
535 d = 0.0001;
536 q *= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
537
538 q_limit = bits2qp(rce,
539 FFMAX((min_rate - buffer_size + rcc->buffer_index) *
540 s->avctx->rc_min_vbv_overflow_use, 1));
541
542 if (q > q_limit) {
543 if (s->avctx->debug & FF_DEBUG_RC)
544 av_log(s->avctx, AV_LOG_DEBUG,
545 "limiting QP %f -> %f\n", q, q_limit);
546 q = q_limit;
547 }
548 }
549
550 if (max_rate) {
551 double d = 2 * expected_size / buffer_size;
552 if (d > 1.0)
553 d = 1.0;
554 else if (d < 0.0001)
555 d = 0.0001;
556 q /= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
557
558 q_limit = bits2qp(rce,
559 FFMAX(rcc->buffer_index *
560 s->avctx->rc_max_available_vbv_use,
561 1));
562 if (q < q_limit) {
563 if (s->avctx->debug & FF_DEBUG_RC)
564 av_log(s->avctx, AV_LOG_DEBUG,
565 "limiting QP %f -> %f\n", q, q_limit);
566 q = q_limit;
567 }
568 }
569 }
570 av_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
571 q, max_rate, min_rate, buffer_size, rcc->buffer_index,
572 s->avctx->rc_buffer_aggressivity);
573 if (s->avctx->rc_qsquish == 0.0 || qmin == qmax) {
574 if (q < qmin)
575 q = qmin;
576 else if (q > qmax)
577 q = qmax;
578 } else {
579 double min2 = log(qmin);
580 double max2 = log(qmax);
581
582 q = log(q);
583 q = (q - min2) / (max2 - min2) - 0.5;
584 q *= -4.0;
585 q = 1.0 / (1.0 + exp(q));
586 q = q * (max2 - min2) + min2;
587
588 q = exp(q);
589 }
590
591 return q;
592}
593
594// ----------------------------------
595// 1 Pass Code
596
597static double predict_size(Predictor *p, double q, double var)
598{
599 return p->coeff * var / (q * p->count);
600}
601
602static void update_predictor(Predictor *p, double q, double var, double size)
603{
604 double new_coeff = size * q / (var + 1);
605 if (var < 10)
606 return;
607
608 p->count *= p->decay;
609 p->coeff *= p->decay;
610 p->count++;
611 p->coeff += new_coeff;
612}
613
614static void adaptive_quantization(MpegEncContext *s, double q)
615{
616 int i;
617 const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
618 const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
619 const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
620 const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
621 const float p_masking = s->avctx->p_masking;
622 const float border_masking = s->avctx->border_masking;
623 float bits_sum = 0.0;
624 float cplx_sum = 0.0;
625 float *cplx_tab = s->cplx_tab;
626 float *bits_tab = s->bits_tab;
627 const int qmin = s->avctx->mb_lmin;
628 const int qmax = s->avctx->mb_lmax;
629 Picture *const pic = &s->current_picture;
630 const int mb_width = s->mb_width;
631 const int mb_height = s->mb_height;
632
633 for (i = 0; i < s->mb_num; i++) {
634 const int mb_xy = s->mb_index2xy[i];
635 float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
636 float spat_cplx = sqrt(pic->mb_var[mb_xy]);
637 const int lumi = pic->mb_mean[mb_xy];
638 float bits, cplx, factor;
639 int mb_x = mb_xy % s->mb_stride;
640 int mb_y = mb_xy / s->mb_stride;
641 int mb_distance;
642 float mb_factor = 0.0;
643 if (spat_cplx < 4)
644 spat_cplx = 4; // FIXME finetune
645 if (temp_cplx < 4)
646 temp_cplx = 4; // FIXME finetune
647
648 if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
649 cplx = spat_cplx;
650 factor = 1.0 + p_masking;
651 } else {
652 cplx = temp_cplx;
653 factor = pow(temp_cplx, -temp_cplx_masking);
654 }
655 factor *= pow(spat_cplx, -spatial_cplx_masking);
656
657 if (lumi > 127)
658 factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
659 else
660 factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
661
662 if (mb_x < mb_width / 5) {
663 mb_distance = mb_width / 5 - mb_x;
664 mb_factor = (float)mb_distance / (float)(mb_width / 5);
665 } else if (mb_x > 4 * mb_width / 5) {
666 mb_distance = mb_x - 4 * mb_width / 5;
667 mb_factor = (float)mb_distance / (float)(mb_width / 5);
668 }
669 if (mb_y < mb_height / 5) {
670 mb_distance = mb_height / 5 - mb_y;
671 mb_factor = FFMAX(mb_factor,
672 (float)mb_distance / (float)(mb_height / 5));
673 } else if (mb_y > 4 * mb_height / 5) {
674 mb_distance = mb_y - 4 * mb_height / 5;
675 mb_factor = FFMAX(mb_factor,
676 (float)mb_distance / (float)(mb_height / 5));
677 }
678
679 factor *= 1.0 - border_masking * mb_factor;
680
681 if (factor < 0.00001)
682 factor = 0.00001;
683
684 bits = cplx * factor;
685 cplx_sum += cplx;
686 bits_sum += bits;
687 cplx_tab[i] = cplx;
688 bits_tab[i] = bits;
689 }
690
691 /* handle qmin/qmax clipping */
692 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
693 float factor = bits_sum / cplx_sum;
694 for (i = 0; i < s->mb_num; i++) {
695 float newq = q * cplx_tab[i] / bits_tab[i];
696 newq *= factor;
697
698 if (newq > qmax) {
699 bits_sum -= bits_tab[i];
700 cplx_sum -= cplx_tab[i] * q / qmax;
701 } else if (newq < qmin) {
702 bits_sum -= bits_tab[i];
703 cplx_sum -= cplx_tab[i] * q / qmin;
704 }
705 }
706 if (bits_sum < 0.001)
707 bits_sum = 0.001;
708 if (cplx_sum < 0.001)
709 cplx_sum = 0.001;
710 }
711
712 for (i = 0; i < s->mb_num; i++) {
713 const int mb_xy = s->mb_index2xy[i];
714 float newq = q * cplx_tab[i] / bits_tab[i];
715 int intq;
716
717 if (s->mpv_flags & FF_MPV_FLAG_NAQ) {
718 newq *= bits_sum / cplx_sum;
719 }
720
721 intq = (int)(newq + 0.5);
722
723 if (intq > qmax)
724 intq = qmax;
725 else if (intq < qmin)
726 intq = qmin;
727 s->lambda_table[mb_xy] = intq;
728 }
729}
730
731void ff_get_2pass_fcode(MpegEncContext *s)
732{
733 RateControlContext *rcc = &s->rc_context;
734 RateControlEntry *rce = &rcc->entry[s->picture_number];
735
736 s->f_code = rce->f_code;
737 s->b_code = rce->b_code;
738}
739
740// FIXME rd or at least approx for dquant
741
742float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
743{
744 float q;
745 int qmin, qmax;
746 float br_compensation;
747 double diff;
748 double short_term_q;
749 double fps;
750 int picture_number = s->picture_number;
751 int64_t wanted_bits;
752 RateControlContext *rcc = &s->rc_context;
753 AVCodecContext *a = s->avctx;
754 RateControlEntry local_rce, *rce;
755 double bits;
756 double rate_factor;
757 int64_t var;
758 const int pict_type = s->pict_type;
759 Picture * const pic = &s->current_picture;
760 emms_c();
761
762#if CONFIG_LIBXVID
763 if ((s->flags & CODEC_FLAG_PASS2) &&
764 s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
765 return ff_xvid_rate_estimate_qscale(s, dry_run);
766#endif
767
768 get_qminmax(&qmin, &qmax, s, pict_type);
769
770 fps = get_fps(s->avctx);
771 /* update predictors */
772 if (picture_number > 2 && !dry_run) {
773 const int64_t last_var =
774 s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
775 : rcc->last_mc_mb_var_sum;
776 av_assert1(s->frame_bits >= s->stuffing_bits);
777 update_predictor(&rcc->pred[s->last_pict_type],
778 rcc->last_qscale,
779 sqrt(last_var),
780 s->frame_bits - s->stuffing_bits);
781 }
782
783 if (s->flags & CODEC_FLAG_PASS2) {
784 assert(picture_number >= 0);
785 if (picture_number >= rcc->num_entries) {
786 av_log(s, AV_LOG_ERROR, "Input is longer than 2-pass log file\n");
787 return -1;
788 }
789 rce = &rcc->entry[picture_number];
790 wanted_bits = rce->expected_bits;
791 } else {
792 Picture *dts_pic;
793 rce = &local_rce;
794
795 /* FIXME add a dts field to AVFrame and ensure it is set and use it
796 * here instead of reordering but the reordering is simpler for now
797 * until H.264 B-pyramid must be handled. */
798 if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
799 dts_pic = s->current_picture_ptr;
800 else
801 dts_pic = s->last_picture_ptr;
802
803 if (!dts_pic || dts_pic->f->pts == AV_NOPTS_VALUE)
804 wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
805 else
806 wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f->pts / fps);
807 }
808
809 diff = s->total_bits - wanted_bits;
810 br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
811 if (br_compensation <= 0.0)
812 br_compensation = 0.001;
813
814 var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
815
816 short_term_q = 0; /* avoid warning */
817 if (s->flags & CODEC_FLAG_PASS2) {
818 if (pict_type != AV_PICTURE_TYPE_I)
819 assert(pict_type == rce->new_pict_type);
820
821 q = rce->new_qscale / br_compensation;
822 av_dlog(s, "%f %f %f last:%d var:%"PRId64" type:%d//\n", q, rce->new_qscale,
823 br_compensation, s->frame_bits, var, pict_type);
824 } else {
825 rce->pict_type =
826 rce->new_pict_type = pict_type;
827 rce->mc_mb_var_sum = pic->mc_mb_var_sum;
828 rce->mb_var_sum = pic->mb_var_sum;
829 rce->qscale = FF_QP2LAMBDA * 2;
830 rce->f_code = s->f_code;
831 rce->b_code = s->b_code;
832 rce->misc_bits = 1;
833
834 bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
835 if (pict_type == AV_PICTURE_TYPE_I) {
836 rce->i_count = s->mb_num;
837 rce->i_tex_bits = bits;
838 rce->p_tex_bits = 0;
839 rce->mv_bits = 0;
840 } else {
841 rce->i_count = 0; // FIXME we do know this approx
842 rce->i_tex_bits = 0;
843 rce->p_tex_bits = bits * 0.9;
844 rce->mv_bits = bits * 0.1;
845 }
846 rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
847 rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
848 rcc->mv_bits_sum[pict_type] += rce->mv_bits;
849 rcc->frame_count[pict_type]++;
850
851 rate_factor = rcc->pass1_wanted_bits /
852 rcc->pass1_rc_eq_output_sum * br_compensation;
853
854 q = get_qscale(s, rce, rate_factor, picture_number);
855 if (q < 0)
856 return -1;
857
858 assert(q > 0.0);
859 q = get_diff_limited_q(s, rce, q);
860 assert(q > 0.0);
861
862 // FIXME type dependent blur like in 2-pass
863 if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
864 rcc->short_term_qsum *= a->qblur;
865 rcc->short_term_qcount *= a->qblur;
866
867 rcc->short_term_qsum += q;
868 rcc->short_term_qcount++;
869 q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
870 }
871 assert(q > 0.0);
872
873 q = modify_qscale(s, rce, q, picture_number);
874
875 rcc->pass1_wanted_bits += s->bit_rate / fps;
876
877 assert(q > 0.0);
878 }
879
880 if (s->avctx->debug & FF_DEBUG_RC) {
881 av_log(s->avctx, AV_LOG_DEBUG,
882 "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
883 "size:%d var:%"PRId64"/%"PRId64" br:%d fps:%d\n",
884 av_get_picture_type_char(pict_type),
885 qmin, q, qmax, picture_number,
886 (int)wanted_bits / 1000, (int)s->total_bits / 1000,
887 br_compensation, short_term_q, s->frame_bits,
888 pic->mb_var_sum, pic->mc_mb_var_sum,
889 s->bit_rate / 1000, (int)fps);
890 }
891
892 if (q < qmin)
893 q = qmin;
894 else if (q > qmax)
895 q = qmax;
896
897 if (s->adaptive_quant)
898 adaptive_quantization(s, q);
899 else
900 q = (int)(q + 0.5);
901
902 if (!dry_run) {
903 rcc->last_qscale = q;
904 rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
905 rcc->last_mb_var_sum = pic->mb_var_sum;
906 }
907 return q;
908}
909
910// ----------------------------------------------
911// 2-Pass code
912
913static int init_pass2(MpegEncContext *s)
914{
915 RateControlContext *rcc = &s->rc_context;
916 AVCodecContext *a = s->avctx;
917 int i, toobig;
918 double fps = get_fps(s->avctx);
919 double complexity[5] = { 0 }; // approximate bits at quant=1
920 uint64_t const_bits[5] = { 0 }; // quantizer independent bits
921 uint64_t all_const_bits;
922 uint64_t all_available_bits = (uint64_t)(s->bit_rate *
923 (double)rcc->num_entries / fps);
924 double rate_factor = 0;
925 double step;
926 const int filter_size = (int)(a->qblur * 4) | 1;
927 double expected_bits = 0; // init to silence gcc warning
928 double *qscale, *blurred_qscale, qscale_sum;
929
930 /* find complexity & const_bits & decide the pict_types */
931 for (i = 0; i < rcc->num_entries; i++) {
932 RateControlEntry *rce = &rcc->entry[i];
933
934 rce->new_pict_type = rce->pict_type;
935 rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
936 rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
937 rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
938 rcc->frame_count[rce->pict_type]++;
939
940 complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
941 (double)rce->qscale;
942 const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
943 }
944
945 all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
946 const_bits[AV_PICTURE_TYPE_P] +
947 const_bits[AV_PICTURE_TYPE_B];
948
949 if (all_available_bits < all_const_bits) {
950 av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
951 return -1;
952 }
953
954 qscale = av_malloc_array(rcc->num_entries, sizeof(double));
955 blurred_qscale = av_malloc_array(rcc->num_entries, sizeof(double));
956 toobig = 0;
957
958 for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
959 expected_bits = 0;
960 rate_factor += step;
961
962 rcc->buffer_index = s->avctx->rc_buffer_size / 2;
963
964 /* find qscale */
965 for (i = 0; i < rcc->num_entries; i++) {
966 RateControlEntry *rce = &rcc->entry[i];
967
968 qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
969 rcc->last_qscale_for[rce->pict_type] = qscale[i];
970 }
971 assert(filter_size % 2 == 1);
972
973 /* fixed I/B QP relative to P mode */
974 for (i = FFMAX(0, rcc->num_entries - 300); i < rcc->num_entries; i++) {
975 RateControlEntry *rce = &rcc->entry[i];
976
977 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
978 }
979
980 for (i = rcc->num_entries - 1; i >= 0; i--) {
981 RateControlEntry *rce = &rcc->entry[i];
982
983 qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
984 }
985
986 /* smooth curve */
987 for (i = 0; i < rcc->num_entries; i++) {
988 RateControlEntry *rce = &rcc->entry[i];
989 const int pict_type = rce->new_pict_type;
990 int j;
991 double q = 0.0, sum = 0.0;
992
993 for (j = 0; j < filter_size; j++) {
994 int index = i + j - filter_size / 2;
995 double d = index - i;
996 double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
997
998 if (index < 0 || index >= rcc->num_entries)
999 continue;
1000 if (pict_type != rcc->entry[index].new_pict_type)
1001 continue;
1002 q += qscale[index] * coeff;
1003 sum += coeff;
1004 }
1005 blurred_qscale[i] = q / sum;
1006 }
1007
1008 /* find expected bits */
1009 for (i = 0; i < rcc->num_entries; i++) {
1010 RateControlEntry *rce = &rcc->entry[i];
1011 double bits;
1012
1013 rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
1014
1015 bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
1016 bits += 8 * ff_vbv_update(s, bits);
1017
1018 rce->expected_bits = expected_bits;
1019 expected_bits += bits;
1020 }
1021
1022 av_dlog(s->avctx,
1023 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
1024 expected_bits, (int)all_available_bits, rate_factor);
1025 if (expected_bits > all_available_bits) {
1026 rate_factor -= step;
1027 ++toobig;
1028 }
1029 }
1030 av_free(qscale);
1031 av_free(blurred_qscale);
1032
1033 /* check bitrate calculations and print info */
1034 qscale_sum = 0.0;
1035 for (i = 0; i < rcc->num_entries; i++) {
1036 av_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
1037 i,
1038 rcc->entry[i].new_qscale,
1039 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
1040 qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
1041 s->avctx->qmin, s->avctx->qmax);
1042 }
1043 assert(toobig <= 40);
1044 av_log(s->avctx, AV_LOG_DEBUG,
1045 "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
1046 s->bit_rate,
1047 (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
1048 av_log(s->avctx, AV_LOG_DEBUG,
1049 "[lavc rc] estimated target average qp: %.3f\n",
1050 (float)qscale_sum / rcc->num_entries);
1051 if (toobig == 0) {
1052 av_log(s->avctx, AV_LOG_INFO,
1053 "[lavc rc] Using all of requested bitrate is not "
1054 "necessary for this video with these parameters.\n");
1055 } else if (toobig == 40) {
1056 av_log(s->avctx, AV_LOG_ERROR,
1057 "[lavc rc] Error: bitrate too low for this video "
1058 "with these parameters.\n");
1059 return -1;
1060 } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
1061 av_log(s->avctx, AV_LOG_ERROR,
1062 "[lavc rc] Error: 2pass curve failed to converge\n");
1063 return -1;
1064 }
1065
1066 return 0;
1067}
ffmpeg Piment Noir Debian package repository