/*
* Rate control for video encoders
*
- * Copyright (c) 2002-2003 Michael Niedermayer <michaelni@gmx.at>
+ * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
*
* This library is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
/**
* @file ratecontrol.c
* Rate control for video encoders.
- */
+ */
#include "avcodec.h"
#include "dsputil.h"
static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num);
void ff_write_pass1_stats(MpegEncContext *s){
- sprintf(s->avctx->stats_out, "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d mc-var:%d var:%d icount:%d;\n",
- s->picture_number, s->input_picture_number - s->max_b_frames, s->pict_type,
- s->current_picture.quality, s->i_tex_bits, s->p_tex_bits, s->mv_bits, s->misc_bits,
+ snprintf(s->avctx->stats_out, 256, "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d mc-var:%d var:%d icount:%d;\n",
+ s->current_picture_ptr->display_picture_number, s->current_picture_ptr->coded_picture_number, s->pict_type,
+ s->current_picture.quality, s->i_tex_bits, s->p_tex_bits, s->mv_bits, s->misc_bits,
s->f_code, s->b_code, s->current_picture.mc_mb_var_sum, s->current_picture.mb_var_sum, s->i_count);
}
for(i=0; i<5; i++){
rcc->pred[i].coeff= FF_QP2LAMBDA * 7.0;
rcc->pred[i].count= 1.0;
-
+
rcc->pred[i].decay= 0.4;
rcc->i_cplx_sum [i]=
rcc->p_cplx_sum [i]=
p= strchr(p+1, ';');
}
i+= s->max_b_frames;
+ if(i<=0 || i>=INT_MAX / sizeof(RateControlEntry))
+ return -1;
rcc->entry = (RateControlEntry*)av_mallocz(i*sizeof(RateControlEntry));
rcc->num_entries= i;
-
- /* init all to skiped p frames (with b frames we might have a not encoded frame at the end FIXME) */
+
+ /* init all to skipped p frames (with b frames we might have a not encoded frame at the end FIXME) */
for(i=0; i<rcc->num_entries; i++){
RateControlEntry *rce= &rcc->entry[i];
rce->pict_type= rce->new_pict_type=P_TYPE;
rce->qscale= rce->new_qscale=FF_QP2LAMBDA * 2;
rce->misc_bits= s->mb_num + 10;
rce->mb_var_sum= s->mb_num*100;
- }
-
+ }
+
/* read stats */
p= s->avctx->stats_in;
for(i=0; i<rcc->num_entries - s->max_b_frames; i++){
rce= &rcc->entry[picture_number];
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",
- &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits, &rce->mv_bits, &rce->misc_bits,
+ &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits, &rce->mv_bits, &rce->misc_bits,
&rce->f_code, &rce->b_code, &rce->mc_mb_var_sum, &rce->mb_var_sum, &rce->i_count);
if(e!=12){
av_log(s->avctx, AV_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
}
p= next;
}
-
+
if(init_pass2(s) < 0) return -1;
}
-
+
if(!(s->flags&CODEC_FLAG_PASS2)){
rcc->short_term_qsum=0.001;
rcc->short_term_qcount=0.001;
-
+
rcc->pass1_rc_eq_output_sum= 0.001;
rcc->pass1_wanted_bits=0.001;
-
+
/* init stuff with the user specified complexity */
if(s->avctx->rc_initial_cplx){
for(i=0; i<60*30; i++){
double bits= s->avctx->rc_initial_cplx * (i/10000.0 + 1.0)*s->mb_num;
RateControlEntry rce;
double q;
-
+
if (i%((s->gop_size+3)/4)==0) rce.pict_type= I_TYPE;
else if(i%(s->max_b_frames+1)) rce.pict_type= B_TYPE;
else rce.pict_type= P_TYPE;
bits= rce.i_tex_bits + rce.p_tex_bits;
q= get_qscale(s, &rce, rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum, i);
- rcc->pass1_wanted_bits+= s->bit_rate/(s->avctx->frame_rate / (double)s->avctx->frame_rate_base);
+ rcc->pass1_wanted_bits+= s->bit_rate/(1/av_q2d(s->avctx->time_base)); //FIXME missbehaves a little for variable fps
}
}
}
-
+
return 0;
}
}
return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ bits;
}
-
+
int ff_vbv_update(MpegEncContext *s, int frame_size){
RateControlContext *rcc= &s->rc_context;
- const double fps= (double)s->avctx->frame_rate / (double)s->avctx->frame_rate_base;
+ const double fps= 1/av_q2d(s->avctx->time_base);
const int buffer_size= s->avctx->rc_buffer_size;
const double min_rate= s->avctx->rc_min_rate/fps;
const double max_rate= s->avctx->rc_max_rate/fps;
-
+
//printf("%d %f %d %f %f\n", buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
if(buffer_size){
int left;
if(rcc->buffer_index > buffer_size){
int stuffing= ceil((rcc->buffer_index - buffer_size)/8);
-
+
if(stuffing < 4 && s->codec_id == CODEC_ID_MPEG4)
stuffing=4;
rcc->buffer_index -= 8*stuffing;
-
+
if(s->avctx->debug & FF_DEBUG_RC)
av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
AVCodecContext *a= s->avctx;
double q, bits;
const int pict_type= rce->new_pict_type;
- const double mb_num= s->mb_num;
+ const double mb_num= s->mb_num;
int i;
double const_values[]={
};
bits= ff_eval(s->avctx->rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce);
-
+
rcc->pass1_rc_eq_output_sum+= bits;
bits*=rate_factor;
if(bits<0.0) bits=0.0;
bits+= 1.0; //avoid 1/0 issues
-
+
/* user override */
for(i=0; i<s->avctx->rc_override_count; i++){
RcOverride *rco= s->avctx->rc_override;
if(rco[i].start_frame > frame_num) continue;
if(rco[i].end_frame < frame_num) continue;
-
- if(rco[i].qscale)
+
+ if(rco[i].qscale)
bits= qp2bits(rce, rco[i].qscale); //FIXME move at end to really force it?
else
bits*= rco[i].quality_factor;
}
q= bits2qp(rce, bits);
-
+
/* I/B difference */
if (pict_type==I_TYPE && s->avctx->i_quant_factor<0.0)
q= -q*s->avctx->i_quant_factor + s->avctx->i_quant_offset;
else if(pict_type==B_TYPE && s->avctx->b_quant_factor<0.0)
q= -q*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
-
+
return q;
}
const int pict_type= rce->new_pict_type;
const double last_p_q = rcc->last_qscale_for[P_TYPE];
const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
-
+
if (pict_type==I_TYPE && (a->i_quant_factor>0.0 || rcc->last_non_b_pict_type==P_TYPE))
q= last_p_q *ABS(a->i_quant_factor) + a->i_quant_offset;
else if(pict_type==B_TYPE && a->b_quant_factor>0.0)
}
rcc->last_qscale_for[pict_type]= q; //Note we cant do that after blurring
-
+
if(pict_type!=B_TYPE)
rcc->last_non_b_pict_type= pict_type;
* gets the qmin & qmax for pict_type
*/
static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type){
- int qmin= s->avctx->lmin;
+ int qmin= s->avctx->lmin;
int qmax= s->avctx->lmax;
-
+
assert(qmin <= qmax);
if(pict_type==B_TYPE){
qmax= clip(qmax, 1, FF_LAMBDA_MAX);
if(qmax<qmin) qmax= qmin;
-
+
*qmin_ret= qmin;
*qmax_ret= qmax;
}
double bits;
const int pict_type= rce->new_pict_type;
const double buffer_size= s->avctx->rc_buffer_size;
- const double fps= (double)s->avctx->frame_rate / (double)s->avctx->frame_rate_base;
+ const double fps= 1/av_q2d(s->avctx->time_base);
const double min_rate= s->avctx->rc_min_rate / fps;
const double max_rate= s->avctx->rc_max_rate / fps;
-
+
get_qminmax(&qmin, &qmax, s, pict_type);
/* modulation */
}else{
double min2= log(qmin);
double max2= log(qmax);
-
+
q= log(q);
q= (q - min2)/(max2-min2) - 0.5;
q*= -4.0;
q= 1.0/(1.0 + exp(q));
q= q*(max2-min2) + min2;
-
+
q= exp(q);
}
-
+
return q;
}
const float temp_cplx_masking= s->avctx->temporal_cplx_masking;
const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
const float p_masking = s->avctx->p_masking;
+ const float border_masking = s->avctx->border_masking;
float bits_sum= 0.0;
float cplx_sum= 0.0;
float cplx_tab[s->mb_num];
float bits_tab[s->mb_num];
- const int qmin= s->avctx->lmin;
- const int qmax= s->avctx->lmax;
+ const int qmin= s->avctx->mb_lmin;
+ const int qmax= s->avctx->mb_lmax;
Picture * const pic= &s->current_picture;
-
+ const int mb_width = s->mb_width;
+ const int mb_height = s->mb_height;
+
for(i=0; i<s->mb_num; i++){
const int mb_xy= s->mb_index2xy[i];
float temp_cplx= sqrt(pic->mc_mb_var[mb_xy]); //FIXME merge in pow()
float spat_cplx= sqrt(pic->mb_var[mb_xy]);
const int lumi= pic->mb_mean[mb_xy];
float bits, cplx, factor;
-#if 0
+ int mb_x = mb_xy % s->mb_stride;
+ int mb_y = mb_xy / s->mb_stride;
+ int mb_distance;
+ float mb_factor = 0.0;
+#if 0
if(spat_cplx < q/3) spat_cplx= q/3; //FIXME finetune
if(temp_cplx < q/3) temp_cplx= q/3; //FIXME finetune
-#endif
+#endif
if(spat_cplx < 4) spat_cplx= 4; //FIXME finetune
if(temp_cplx < 4) temp_cplx= 4; //FIXME finetune
- if((s->mb_type[mb_xy]&MB_TYPE_INTRA)){//FIXME hq mode
+ if((s->mb_type[mb_xy]&CANDIDATE_MB_TYPE_INTRA)){//FIXME hq mode
cplx= spat_cplx;
factor= 1.0 + p_masking;
}else{
factor*= (1.0 - (lumi-128)*(lumi-128)*lumi_masking);
else
factor*= (1.0 - (lumi-128)*(lumi-128)*dark_masking);
-
+
+ if(mb_x < mb_width/5){
+ mb_distance = mb_width/5 - mb_x;
+ mb_factor = (float)mb_distance / (float)(mb_width/5);
+ }else if(mb_x > 4*mb_width/5){
+ mb_distance = mb_x - 4*mb_width/5;
+ mb_factor = (float)mb_distance / (float)(mb_width/5);
+ }
+ if(mb_y < mb_height/5){
+ mb_distance = mb_height/5 - mb_y;
+ mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
+ }else if(mb_y > 4*mb_height/5){
+ mb_distance = mb_y - 4*mb_height/5;
+ mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
+ }
+
+ factor*= 1.0 - border_masking*mb_factor;
+
if(factor<0.00001) factor= 0.00001;
-
+
bits= cplx*factor;
cplx_sum+= cplx;
bits_sum+= bits;
/* handle qmin/qmax cliping */
if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
+ float factor= bits_sum/cplx_sum;
for(i=0; i<s->mb_num; i++){
float newq= q*cplx_tab[i]/bits_tab[i];
- newq*= bits_sum/cplx_sum;
+ newq*= factor;
if (newq > qmax){
bits_sum -= bits_tab[i];
cplx_sum -= cplx_tab[i]*q/qmin;
}
}
+ if(bits_sum < 0.001) bits_sum= 0.001;
+ if(cplx_sum < 0.001) cplx_sum= 0.001;
}
-
+
for(i=0; i<s->mb_num; i++){
const int mb_xy= s->mb_index2xy[i];
float newq= q*cplx_tab[i]/bits_tab[i];
s->lambda_table[mb_xy]= intq;
}
}
+
+void ff_get_2pass_fcode(MpegEncContext *s){
+ RateControlContext *rcc= &s->rc_context;
+ int picture_number= s->picture_number;
+ RateControlEntry *rce;
+
+ rce= &rcc->entry[picture_number];
+ s->f_code= rce->f_code;
+ s->b_code= rce->b_code;
+}
+
//FIXME rd or at least approx for dquant
-float ff_rate_estimate_qscale(MpegEncContext *s)
+float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
{
float q;
int qmin, qmax;
get_qminmax(&qmin, &qmax, s, pict_type);
- fps= (double)s->avctx->frame_rate / (double)s->avctx->frame_rate_base;
+ fps= 1/av_q2d(s->avctx->time_base);
//printf("input_pic_num:%d pic_num:%d frame_rate:%d\n", s->input_picture_number, s->picture_number, s->frame_rate);
/* update predictors */
- if(picture_number>2){
+ if(picture_number>2 && !dry_run){
const int last_var= s->last_pict_type == I_TYPE ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits);
}
if(br_compensation<=0.0) br_compensation=0.001;
var= pict_type == I_TYPE ? pic->mb_var_sum : pic->mc_mb_var_sum;
-
+
short_term_q = 0; /* avoid warning */
if(s->flags&CODEC_FLAG_PASS2){
if(pict_type!=I_TYPE)
q= rce->new_qscale / br_compensation;
//printf("%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale, br_compensation, s->frame_bits, var, pict_type);
}else{
- rce->pict_type=
+ rce->pict_type=
rce->new_pict_type= pict_type;
rce->mc_mb_var_sum= pic->mc_mb_var_sum;
rce->mb_var_sum = pic-> mb_var_sum;
rce->i_count = 0; //FIXME we do know this approx
rce->i_tex_bits= 0;
rce->p_tex_bits= bits*0.9;
-
+
rce->mv_bits= bits*0.1;
}
rcc->i_cplx_sum [pict_type] += rce->i_tex_bits*rce->qscale;
bits= rce->i_tex_bits + rce->p_tex_bits;
rate_factor= rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum * br_compensation;
-
+
q= get_qscale(s, rce, rate_factor, picture_number);
assert(q>0.0);
//printf("%f ", q);
}
assert(q>0.0);
-
+
q= modify_qscale(s, rce, q, picture_number);
rcc->pass1_wanted_bits+= s->bit_rate/fps;
);
}
- if (q<qmin) q=qmin;
+ if (q<qmin) q=qmin;
else if(q>qmax) q=qmax;
if(s->adaptive_quant)
adaptive_quantization(s, q);
else
q= (int)(q + 0.5);
-
- rcc->last_qscale= q;
- rcc->last_mc_mb_var_sum= pic->mc_mb_var_sum;
- rcc->last_mb_var_sum= pic->mb_var_sum;
+
+ if(!dry_run){
+ rcc->last_qscale= q;
+ rcc->last_mc_mb_var_sum= pic->mc_mb_var_sum;
+ rcc->last_mb_var_sum= pic->mb_var_sum;
+ }
#if 0
{
static int mvsum=0, texsum=0;
RateControlContext *rcc= &s->rc_context;
AVCodecContext *a= s->avctx;
int i;
- double fps= (double)s->avctx->frame_rate / (double)s->avctx->frame_rate_base;
+ double fps= 1/av_q2d(s->avctx->time_base);
double complexity[5]={0,0,0,0,0}; // aproximate bits at quant=1
double avg_quantizer[5];
uint64_t const_bits[5]={0,0,0,0,0}; // quantizer idependant bits
double rate_factor=0;
double step;
//int last_i_frame=-10000000;
- const int filter_size= (int)(a->qblur*4) | 1;
+ const int filter_size= (int)(a->qblur*4) | 1;
double expected_bits;
double *qscale, *blured_qscale;
/* find complexity & const_bits & decide the pict_types */
for(i=0; i<rcc->num_entries; i++){
RateControlEntry *rce= &rcc->entry[i];
-
+
rce->new_pict_type= rce->pict_type;
rcc->i_cplx_sum [rce->pict_type] += rce->i_tex_bits*rce->qscale;
rcc->p_cplx_sum [rce->pict_type] += rce->p_tex_bits*rce->qscale;
const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
}
all_const_bits= const_bits[I_TYPE] + const_bits[P_TYPE] + const_bits[B_TYPE];
-
+
if(all_available_bits < all_const_bits){
av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is to low\n");
return -1;
}
-
+
/* find average quantizers */
avg_quantizer[P_TYPE]=0;
for(step=256*256; step>0.0000001; step*=0.5){
double expected_bits=0;
avg_quantizer[P_TYPE]+= step;
-
+
avg_quantizer[I_TYPE]= avg_quantizer[P_TYPE]*ABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset;
avg_quantizer[B_TYPE]= avg_quantizer[P_TYPE]*ABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset;
-
- expected_bits=
- + all_const_bits
+
+ expected_bits=
+ + all_const_bits
+ complexity[I_TYPE]/avg_quantizer[I_TYPE]
+ complexity[P_TYPE]/avg_quantizer[P_TYPE]
+ complexity[B_TYPE]/avg_quantizer[B_TYPE];
-
+
if(expected_bits < all_available_bits) avg_quantizer[P_TYPE]-= step;
//printf("%f %lld %f\n", expected_bits, all_available_bits, avg_quantizer[P_TYPE]);
}
available_bits[i]= const_bits[i] + complexity[i]/avg_quantizer[i];
}
//printf("%lld %lld %lld %lld\n", available_bits[I_TYPE], available_bits[P_TYPE], available_bits[B_TYPE], all_available_bits);
-
+
qscale= av_malloc(sizeof(double)*rcc->num_entries);
blured_qscale= av_malloc(sizeof(double)*rcc->num_entries);
for(step=256*256; step>0.0000001; step*=0.5){
expected_bits=0;
rate_factor+= step;
-
+
rcc->buffer_index= s->avctx->rc_buffer_size/2;
/* find qscale */
/* fixed I/B QP relative to P mode */
for(i=rcc->num_entries-1; i>=0; i--){
RateControlEntry *rce= &rcc->entry[i];
-
+
qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
}
const int pict_type= rce->new_pict_type;
int j;
double q=0.0, sum=0.0;
-
+
for(j=0; j<filter_size; j++){
int index= i+j-filter_size/2;
double d= index-i;
double coeff= a->qblur==0 ? 1.0 : exp(-d*d/(a->qblur * a->qblur));
-
+
if(index < 0 || index >= rcc->num_entries) continue;
if(pict_type != rcc->entry[index].new_pict_type) continue;
q+= qscale[index] * coeff;
}
blured_qscale[i]= q/sum;
}
-
+
/* find expected bits */
for(i=0; i<rcc->num_entries; i++){
RateControlEntry *rce= &rcc->entry[i];