fixed incorrect buffer handling
[libav.git] / libavcodec / ratecontrol.c
CommitLineData
8b4c7dbc
MN
1/*
2 Copyright (C) 2002 Michael Niedermayer <michaelni@gmx.at>
3
4 This program is free software; you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation; either version 2 of the License, or
7 (at your option) any later version.
8
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
13
14 You should have received a copy of the GNU General Public License
15 along with this program; if not, write to the Free Software
16 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
17*/
18
19#include <inttypes.h>
20#include <stdio.h>
21
22#include "common.h" //needed for mpegvideo.h to compile
23#include "dsputil.h" //needed for mpegvideo.h to compile
24#include "avcodec.h"
25#include "mpegvideo.h"
26
27#define STATS_FILE "lavc_stats.txt"
28
29static int init_pass2(MpegEncContext *s);
30
31void ff_write_pass1_stats(MpegEncContext *s){
32 RateControlContext *rcc= &s->rc_context;
33// fprintf(c->stats_file, "type:%d q:%d icount:%d pcount:%d scount:%d itex:%d ptex%d mv:%d misc:%d fcode:%d bcode:%d\")
34 fprintf(rcc->stats_file, "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d\n",
35 s->picture_number, s->input_picture_number - s->max_b_frames, s->pict_type,
36 s->qscale, s->i_tex_bits, s->p_tex_bits, s->mv_bits, s->misc_bits, s->f_code, s->b_code);
37}
38
39int ff_rate_control_init(MpegEncContext *s)
40{
41 RateControlContext *rcc= &s->rc_context;
42 emms_c();
43
44 if(s->flags&CODEC_FLAG_PASS1){
45 rcc->stats_file= fopen(STATS_FILE, "w");
46 if(!rcc->stats_file){
47 fprintf(stderr, "failed to open " STATS_FILE "\n");
48 return -1;
49 }
50 } else if(s->flags&CODEC_FLAG_PASS2){
51 int size;
52 int i;
53
54 rcc->stats_file= fopen(STATS_FILE, "r");
55 if(!rcc->stats_file){
56 fprintf(stderr, "failed to open " STATS_FILE "\n");
57 return -1;
58 }
59
60 /* find number of pics without reading the file twice :) */
61 fseek(rcc->stats_file, 0, SEEK_END);
62 size= ftell(rcc->stats_file);
63 fseek(rcc->stats_file, 0, SEEK_SET);
64
65 size/= 64; // we need at least 64 byte to store a line ...
66 rcc->entry = (RateControlEntry*)av_mallocz(size*sizeof(RateControlEntry));
67
68 for(i=0; !feof(rcc->stats_file); i++){
69 RateControlEntry *rce;
70 int picture_number;
71 int e;
72
73 e= fscanf(rcc->stats_file, "in:%d ", &picture_number);
74 rce= &rcc->entry[picture_number];
75 e+=fscanf(rcc->stats_file, "out:%*d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d fcode:%*d bcode:%*d\n",
76 &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits, &rce->mv_bits, &rce->misc_bits);
77 if(e!=7){
78 fprintf(stderr, STATS_FILE " is damaged\n");
79 return -1;
80 }
81 }
82 rcc->num_entries= i;
83
84 if(init_pass2(s) < 0) return -1;
85 }
86
87 /* no 2pass stuff, just normal 1-pass */
88 //initial values, they dont really matter as they will be totally different within a few frames
89 s->i_pred.coeff= s->p_pred.coeff= 7.0;
90 s->i_pred.count= s->p_pred.count= 1.0;
91
92 s->i_pred.decay= s->p_pred.decay= 0.4;
93
94 // use more bits at the beginning, otherwise high motion at the begin will look like shit
95 s->qsum=100 * s->qmin;
96 s->qcount=100;
97
98 s->short_term_qsum=0.001;
99 s->short_term_qcount=0.001;
100
101 return 0;
102}
103
104void ff_rate_control_uninit(MpegEncContext *s)
105{
106 RateControlContext *rcc= &s->rc_context;
107 emms_c();
108
6000abfa
FB
109 if(rcc->stats_file)
110 fclose(rcc->stats_file);
111 rcc->stats_file = NULL;
112 av_freep(&rcc->entry);
8b4c7dbc
MN
113}
114
115//----------------------------------
116// 1 Pass Code
117
118static double predict(Predictor *p, double q, double var)
119{
120 return p->coeff*var / (q*p->count);
121}
122
123static void update_predictor(Predictor *p, double q, double var, double size)
124{
125 double new_coeff= size*q / (var + 1);
126 if(var<1000) return;
127
128 p->count*= p->decay;
129 p->coeff*= p->decay;
130 p->count++;
131 p->coeff+= new_coeff;
132}
133
134int ff_rate_estimate_qscale(MpegEncContext *s)
135{
136 int qmin= s->qmin;
137 int qmax= s->qmax;
138 int rate_q=5;
139 float q;
140 int qscale;
141 float br_compensation;
142 double diff;
143 double short_term_q;
144 double long_term_q;
145 double fps;
146 int picture_number= s->input_picture_number - s->max_b_frames;
147 int64_t wanted_bits;
148 emms_c();
149
150 fps= (double)s->frame_rate / FRAME_RATE_BASE;
151 wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
152// printf("%d %d %d\n", picture_number, (int)wanted_bits, (int)s->total_bits);
153
154 if(s->pict_type==B_TYPE){
155 qmin= (int)(qmin*s->b_quant_factor+0.5);
156 qmax= (int)(qmax*s->b_quant_factor+0.5);
157 }
4b7b1960 158 if(qmin<1) qmin=1;
8b4c7dbc
MN
159 if(qmax>31) qmax=31;
160 if(qmax<=qmin) qmax= qmin;
161
162 /* update predictors */
163 if(picture_number>2){
164 if(s->pict_type!=B_TYPE && s->last_non_b_pict_type == P_TYPE){
165//printf("%d %d %d %f\n", s->qscale, s->last_mc_mb_var, s->frame_bits, s->p_pred.coeff);
166 update_predictor(&s->p_pred, s->last_non_b_qscale, s->last_non_b_mc_mb_var, s->pb_frame_bits);
167 }
168 }
169
170 if(s->pict_type == I_TYPE){
171 short_term_q= s->short_term_qsum/s->short_term_qcount;
172
173 long_term_q= s->qsum/s->qcount*(s->total_bits+1)/(wanted_bits+1); //+1 to avoid nan & 0
174
175 q= 1/((1/long_term_q - 1/short_term_q)*s->qcompress + 1/short_term_q);
176 }else if(s->pict_type==B_TYPE){
177 q= (int)(s->last_non_b_qscale*s->b_quant_factor+0.5);
178 }else{ //P Frame
179 int i;
180 int diff, best_diff=1000000000;
181 for(i=1; i<=31; i++){
182 diff= predict(&s->p_pred, i, s->mc_mb_var) - (double)s->bit_rate/fps;
183 if(diff<0) diff= -diff;
184 if(diff<best_diff){
185 best_diff= diff;
186 rate_q= i;
187 }
188 }
189 s->short_term_qsum*=s->qblur;
190 s->short_term_qcount*=s->qblur;
191
192 s->short_term_qsum+= rate_q;
193 s->short_term_qcount++;
194 short_term_q= s->short_term_qsum/s->short_term_qcount;
195
196 long_term_q= s->qsum/s->qcount*(s->total_bits+1)/(wanted_bits+1); //+1 to avoid nan & 0
197
198// q= (long_term_q - short_term_q)*s->qcompress + short_term_q;
199 q= 1/((1/long_term_q - 1/short_term_q)*s->qcompress + 1/short_term_q);
200 }
201
202 diff= s->total_bits - wanted_bits;
203 br_compensation= (s->bit_rate_tolerance - diff)/s->bit_rate_tolerance;
204 if(br_compensation<=0.0) br_compensation=0.001;
205 q/=br_compensation;
206//printf("%f %f %f\n", q, br_compensation, short_term_q);
207 qscale= (int)(q + 0.5);
208 if (qscale<qmin) qscale=qmin;
209 else if(qscale>qmax) qscale=qmax;
210
211 if(s->pict_type!=B_TYPE){
212 s->qsum+= qscale;
213 s->qcount++;
214 if (qscale<s->last_non_b_qscale-s->max_qdiff) qscale=s->last_non_b_qscale-s->max_qdiff;
215 else if(qscale>s->last_non_b_qscale+s->max_qdiff) qscale=s->last_non_b_qscale+s->max_qdiff;
216 }
217//printf("q:%d diff:%d comp:%f rate_q:%d st_q:%f fvar:%d last_size:%d\n", qscale, (int)diff, br_compensation,
218// rate_q, short_term_q, s->mc_mb_var, s->frame_bits);
219//printf("%d %d\n", s->bit_rate, (int)fps);
220 return qscale;
221}
222
223//----------------------------------------------
224// 2-Pass code
225
226static int init_pass2(MpegEncContext *s)
227{
228 RateControlContext *rcc= &s->rc_context;
229 int i;
230 double fps= (double)s->frame_rate / FRAME_RATE_BASE;
231 double complexity[5]={0,0,0,0,0}; // aproximate bits at quant=1
232 double avg_quantizer[5];
233 uint64_t const_bits[5]={0,0,0,0,0}; // quantizer idependant bits
234 uint64_t available_bits[5];
235 uint64_t all_const_bits;
236 uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
237 int num_frames[5]={0,0,0,0,0};
238 double rate_factor=0;
239 double step;
240 int last_i_frame=-10000000;
241
242 /* find complexity & const_bits & decide the pict_types */
243 for(i=0; i<rcc->num_entries; i++){
244 RateControlEntry *rce= &rcc->entry[i];
245
246 if(s->b_frame_strategy==0 || s->max_b_frames==0){
247 rce->new_pict_type= rce->pict_type;
248 }else{
249 int j;
250 int next_non_b_type=P_TYPE;
251
252 switch(rce->pict_type){
253 case I_TYPE:
254 if(i-last_i_frame>s->gop_size/2){ //FIXME this is not optimal
255 rce->new_pict_type= I_TYPE;
256 last_i_frame= i;
257 }else{
258 rce->new_pict_type= P_TYPE; // will be caught by the scene detection anyway
259 }
260 break;
261 case P_TYPE:
262 rce->new_pict_type= P_TYPE;
263 break;
264 case B_TYPE:
265 for(j=i+1; j<i+s->max_b_frames+2 && j<rcc->num_entries; j++){
266 if(rcc->entry[j].pict_type != B_TYPE){
267 next_non_b_type= rcc->entry[j].pict_type;
268 break;
269 }
270 }
271 if(next_non_b_type==I_TYPE)
272 rce->new_pict_type= P_TYPE;
273 else
274 rce->new_pict_type= B_TYPE;
275 break;
276 }
277 }
278
279 complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)rce->qscale;
280 const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
281 num_frames[rce->new_pict_type]++;
282 }
283 all_const_bits= const_bits[I_TYPE] + const_bits[P_TYPE] + const_bits[B_TYPE];
284
285 if(all_available_bits < all_const_bits){
286 fprintf(stderr, "requested bitrate is to low\n");
287 return -1;
288 }
289
290// avg_complexity= complexity/rcc->num_entries;
291 avg_quantizer[P_TYPE]=
292 avg_quantizer[I_TYPE]= (complexity[I_TYPE]+complexity[P_TYPE] + complexity[B_TYPE]/s->b_quant_factor)
293 / (all_available_bits - all_const_bits);
294 avg_quantizer[B_TYPE]= avg_quantizer[P_TYPE]*s->b_quant_factor;
295//printf("avg quantizer: %f %f\n", avg_quantizer[P_TYPE], avg_quantizer[B_TYPE]);
296
297 for(i=0; i<5; i++){
298 available_bits[i]= const_bits[i] + complexity[i]/avg_quantizer[i];
299 }
300//printf("%lld %lld %lld %lld\n", available_bits[I_TYPE], available_bits[P_TYPE], available_bits[B_TYPE], all_available_bits);
301
302 for(step=256*256; step>0.0000001; step*=0.5){
303 uint64_t expected_bits=0;
304 rate_factor+= step;
305 /* find qscale */
306 for(i=0; i<rcc->num_entries; i++){
307 RateControlEntry *rce= &rcc->entry[i];
308 double short_term_q, q, bits_left;
309 const int pict_type= rce->new_pict_type;
310 int qmin= s->qmin;
311 int qmax= s->qmax;
312
313 if(pict_type==B_TYPE){
314 qmin= (int)(qmin*s->b_quant_factor+0.5);
315 qmax= (int)(qmax*s->b_quant_factor+0.5);
316 }
4b7b1960 317 if(qmin<1) qmin=1;
8b4c7dbc
MN
318 if(qmax>31) qmax=31;
319 if(qmax<=qmin) qmax= qmin;
320
321 switch(s->rc_strategy){
322 case 0:
323 bits_left= available_bits[pict_type]/num_frames[pict_type]*rate_factor - rce->misc_bits - rce->mv_bits;
324 if(bits_left<1.0) bits_left=1.0;
325 short_term_q= rce->qscale*(rce->i_tex_bits + rce->p_tex_bits)/bits_left;
326 break;
327 case 1:
328 bits_left= (available_bits[pict_type] - const_bits[pict_type])/num_frames[pict_type]*rate_factor;
329 if(bits_left<1.0) bits_left=1.0;
330 short_term_q= rce->qscale*(rce->i_tex_bits + rce->p_tex_bits)/bits_left;
331 break;
332 case 2:
333 bits_left= available_bits[pict_type]/num_frames[pict_type]*rate_factor;
334 if(bits_left<1.0) bits_left=1.0;
335 short_term_q= rce->qscale*(rce->i_tex_bits + rce->p_tex_bits + rce->misc_bits + rce->mv_bits)/bits_left;
336 break;
337 default:
338 fprintf(stderr, "unknown strategy\n");
339 short_term_q=3; //gcc warning fix
340 }
341
342 if(short_term_q>31.0) short_term_q=31.0;
343 else if (short_term_q<1.0) short_term_q=1.0;
344
345 q= 1/((1/avg_quantizer[pict_type] - 1/short_term_q)*s->qcompress + 1/short_term_q);
346 if (q<qmin) q=qmin;
347 else if(q>qmax) q=qmax;
348//printf("lq:%f, sq:%f t:%f q:%f\n", avg_quantizer[rce->pict_type], short_term_q, bits_left, q);
349 rce->new_qscale= q;
350 }
351
352 /* smooth curve */
353
354 /* find expected bits */
355 for(i=0; i<rcc->num_entries; i++){
356 RateControlEntry *rce= &rcc->entry[i];
357 double factor= rce->qscale / rce->new_qscale;
358
359 rce->expected_bits= expected_bits;
360 expected_bits += (int)(rce->misc_bits + rce->mv_bits + (rce->i_tex_bits + rce->p_tex_bits)*factor + 0.5);
361 }
362
363// printf("%d %d %f\n", (int)expected_bits, (int)all_available_bits, rate_factor);
364 if(expected_bits > all_available_bits) rate_factor-= step;
365 }
366
367 return 0;
368}
369
370int ff_rate_estimate_qscale_pass2(MpegEncContext *s)
371{
372 int qmin= s->qmin;
373 int qmax= s->qmax;
374 float q;
375 int qscale;
376 float br_compensation;
377 double diff;
378 int picture_number= s->picture_number;
379 RateControlEntry *rce= &s->rc_context.entry[picture_number];
380 int64_t wanted_bits= rce->expected_bits;
381 emms_c();
382
383// printf("%d %d %d\n", picture_number, (int)wanted_bits, (int)s->total_bits);
384
385 if(s->pict_type==B_TYPE){
386 qmin= (int)(qmin*s->b_quant_factor+0.5);
387 qmax= (int)(qmax*s->b_quant_factor+0.5);
388 }
4b7b1960 389 if(qmin<1) qmin=1;
8b4c7dbc
MN
390 if(qmax>31) qmax=31;
391 if(qmax<=qmin) qmax= qmin;
392
393 q= rce->new_qscale;
394
395 diff= s->total_bits - wanted_bits;
396 br_compensation= (s->bit_rate_tolerance - diff)/s->bit_rate_tolerance;
397 if(br_compensation<=0.0) br_compensation=0.001;
398 q/=br_compensation;
399
400 qscale= (int)(q + 0.5);
401 if (qscale<qmin) qscale=qmin;
402 else if(qscale>qmax) qscale=qmax;
403// printf("%d %d %d %d type:%d\n", qmin, qscale, qmax, picture_number, s->pict_type); fflush(stdout);
404 return qscale;
405}