/*
* Rate control for video encoders
*
- * Copyright (c) 2002 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
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
- * version 2 of the License, or (at your option) any later version.
+ * version 2.1 of the License, or (at your option) any later version.
*
- * This library is distributed in the hope that it will be useful,
+ * FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
- * License along with this library; if not, write to the Free Software
- * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file ratecontrol.c
+ * Rate control for video encoders.
*/
-#include <math.h>
-#include "common.h"
+
#include "avcodec.h"
#include "dsputil.h"
+#include "ratecontrol.h"
#include "mpegvideo.h"
+#include "eval.h"
#undef NDEBUG // allways check asserts, the speed effect is far too small to disable them
#include <assert.h>
-#ifndef M_PI
-#define M_PI 3.14159265358979323846
-#endif
-
#ifndef M_E
#define M_E 2.718281828
#endif
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->qscale, s->i_tex_bits, s->p_tex_bits, s->mv_bits, s->misc_bits,
- s->f_code, s->b_code, s->mc_mb_var_sum, s->mb_var_sum, s->i_count);
+ 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 skipcount:%d hbits:%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, s->skip_count, s->header_bits);
}
int ff_rate_control_init(MpegEncContext *s)
emms_c();
for(i=0; i<5; i++){
- rcc->pred[i].coeff= 7.0;
+ 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]=
rcc->mv_bits_sum[i]=
rcc->qscale_sum [i]=
rcc->frame_count[i]= 1; // 1 is better cuz of 1/0 and such
- rcc->last_qscale_for[i]=5;
+ rcc->last_qscale_for[i]=FF_QP2LAMBDA * 5;
}
- rcc->buffer_index= s->avctx->rc_buffer_size/2;
+ rcc->buffer_index= s->avctx->rc_initial_buffer_occupancy;
if(s->flags&CODEC_FLAG_PASS2){
int 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=2;
+ 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++){
assert(picture_number < rcc->num_entries);
rce= &rcc->entry[picture_number];
- e+=sscanf(p, " 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",
- &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){
- fprintf(stderr, "statistics are damaged at line %d, parser out=%d\n", i, e);
+ 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",
+ &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, &rce->skip_count, &rce->header_bits);
+ if(e!=14){
+ av_log(s->avctx, AV_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
return -1;
}
+
p= next;
}
-
+
if(init_pass2(s) < 0) return -1;
+
+ //FIXME maybe move to end
+ if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) {
+#ifdef CONFIG_XVID
+ return ff_xvid_rate_control_init(s);
+#else
+ av_log(s->avctx, AV_LOG_ERROR, "XviD ratecontrol requires libavcodec compiled with XviD support\n");
+ return -1;
+#endif
+ }
}
-
+
if(!(s->flags&CODEC_FLAG_PASS2)){
rcc->short_term_qsum=0.001;
rcc->short_term_qcount=0.001;
-
- rcc->pass1_bits =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;
rce.new_pict_type= rce.pict_type;
rce.mc_mb_var_sum= bits*s->mb_num/100000;
rce.mb_var_sum = s->mb_num;
- rce.qscale = 2;
+ rce.qscale = FF_QP2LAMBDA * 2;
rce.f_code = 2;
rce.b_code = 1;
rce.misc_bits= 1;
bits= rce.i_tex_bits + rce.p_tex_bits;
- q= get_qscale(s, &rce, rcc->pass1_wanted_bits/rcc->pass1_bits, i);
- rcc->pass1_wanted_bits+= s->bit_rate/(s->frame_rate / (double)FRAME_RATE_BASE);
+ q= get_qscale(s, &rce, rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum, i);
+ rcc->pass1_wanted_bits+= s->bit_rate/(1/av_q2d(s->avctx->time_base)); //FIXME missbehaves a little for variable fps
}
}
}
-
+
return 0;
}
emms_c();
av_freep(&rcc->entry);
+
+#ifdef CONFIG_XVID
+ if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
+ ff_xvid_rate_control_uninit(s);
+#endif
}
static inline double qp2bits(RateControlEntry *rce, double qp){
if(qp<=0.0){
- fprintf(stderr, "qp<=0.0\n");
+ av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
}
return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ qp;
}
static inline double bits2qp(RateControlEntry *rce, double bits){
if(bits<0.9){
- fprintf(stderr, "bits<0.9\n");
+ av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
}
return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ bits;
}
-
-static void update_rc_buffer(MpegEncContext *s, int frame_size){
+
+int ff_vbv_update(MpegEncContext *s, int frame_size){
RateControlContext *rcc= &s->rc_context;
- const double fps= (double)s->frame_rate / FRAME_RATE_BASE;
- const double buffer_size= s->avctx->rc_buffer_size;
+ 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;
+
rcc->buffer_index-= frame_size;
- if(rcc->buffer_index < buffer_size/2 /*FIXME /2 */ || min_rate==0){
- rcc->buffer_index+= max_rate;
- if(rcc->buffer_index >= buffer_size)
- rcc->buffer_index= buffer_size-1;
- }else{
- rcc->buffer_index+= min_rate;
+ if(rcc->buffer_index < 0){
+ av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
+ rcc->buffer_index= 0;
+ }
+
+ left= buffer_size - rcc->buffer_index - 1;
+ rcc->buffer_index += clip(left, min_rate, max_rate);
+
+ 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);
+
+ return stuffing;
}
-
- if(rcc->buffer_index < 0)
- fprintf(stderr, "rc buffer underflow\n");
- if(rcc->buffer_index >= s->avctx->rc_buffer_size)
- fprintf(stderr, "rc buffer overflow\n");
}
+ return 0;
}
/**
*/
static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num){
RateControlContext *rcc= &s->rc_context;
+ 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;
+ char *error = NULL;
double const_values[]={
M_PI,
rce->pict_type == P_TYPE,
rce->pict_type == B_TYPE,
rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
- s->qcompress,
+ a->qcompress,
/* rcc->last_qscale_for[I_TYPE],
rcc->last_qscale_for[P_TYPE],
rcc->last_qscale_for[B_TYPE],
(rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
0
};
- char *const_names[]={
+ static const char *const_names[]={
"PI",
"E",
"iTex",
NULL
};
static double (*func1[])(void *, double)={
- bits2qp,
- qp2bits,
+ (void *)bits2qp,
+ (void *)qp2bits,
NULL
};
- char *func1_names[]={
+ static const char *func1_names[]={
"bits2qp",
"qp2bits",
NULL
};
- bits= ff_eval(s->avctx->rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce);
-
- rcc->pass1_bits+= bits;
+ bits= ff_eval2(s->avctx->rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce, &error);
+ if (isnan(bits)) {
+ av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\": %s\n", s->avctx->rc_eq, error? error : "");
+ return -1;
+ }
+
+ 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 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;
+ q= last_p_q *FFABS(a->i_quant_factor) + a->i_quant_offset;
else if(pict_type==B_TYPE && a->b_quant_factor>0.0)
q= last_non_b_q* a->b_quant_factor + a->b_quant_offset;
/* last qscale / qdiff stuff */
if(rcc->last_non_b_pict_type==pict_type || pict_type!=I_TYPE){
double last_q= rcc->last_qscale_for[pict_type];
- if (q > last_q + a->max_qdiff) q= last_q + a->max_qdiff;
- else if(q < last_q - a->max_qdiff) q= last_q - a->max_qdiff;
+ const int maxdiff= FF_QP2LAMBDA * a->max_qdiff;
+
+ if (q > last_q + maxdiff) q= last_q + maxdiff;
+ else if(q < last_q - maxdiff) q= last_q - maxdiff;
}
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->qmin;
- int qmax= s->qmax;
+ int qmin= s->avctx->lmin;
+ int qmax= s->avctx->lmax;
+
+ assert(qmin <= qmax);
if(pict_type==B_TYPE){
- qmin= (int)(qmin*ABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
- qmax= (int)(qmax*ABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
+ qmin= (int)(qmin*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
+ qmax= (int)(qmax*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
}else if(pict_type==I_TYPE){
- qmin= (int)(qmin*ABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
- qmax= (int)(qmax*ABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
+ qmin= (int)(qmin*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
+ qmax= (int)(qmax*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
}
- if(qmin<1) qmin=1;
- if(qmin==1 && s->qmin>1) qmin=2; //avoid qmin=1 unless the user wants qmin=1
+ qmin= clip(qmin, 1, FF_LAMBDA_MAX);
+ qmax= clip(qmax, 1, FF_LAMBDA_MAX);
- if(qmin<3 && s->max_qcoeff<=128 && pict_type==I_TYPE) qmin=3; //reduce cliping problems
+ if(qmax<qmin) qmax= qmin;
- if(qmax>31) qmax=31;
- if(qmax<=qmin) qmax= qmin= (qmax+qmin+1)>>1;
-
*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 min_rate= s->avctx->rc_min_rate;
- const double max_rate= s->avctx->rc_max_rate;
-
+ 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 */
/* buffer overflow/underflow protection */
if(buffer_size){
double expected_size= rcc->buffer_index;
+ double q_limit;
if(min_rate){
double d= 2*(buffer_size - expected_size)/buffer_size;
else if(d<0.0001) d=0.0001;
q*= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
- q= MIN(q, bits2qp(rce, MAX((min_rate - buffer_size + rcc->buffer_index)*2, 1)));
+ q_limit= bits2qp(rce, FFMAX((min_rate - buffer_size + rcc->buffer_index)*3, 1));
+ if(q > q_limit){
+ if(s->avctx->debug&FF_DEBUG_RC){
+ av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
+ }
+ q= q_limit;
+ }
}
if(max_rate){
else if(d<0.0001) d=0.0001;
q/= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
- q= MAX(q, bits2qp(rce, MAX(rcc->buffer_index/2, 1)));
+ q_limit= bits2qp(rce, FFMAX(rcc->buffer_index/3, 1));
+ if(q < q_limit){
+ if(s->avctx->debug&FF_DEBUG_RC){
+ av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
+ }
+ q= q_limit;
+ }
}
}
//printf("q:%f max:%f min:%f size:%f index:%d bits:%f agr:%f\n", q,max_rate, min_rate, buffer_size, rcc->buffer_index, bits, s->avctx->rc_buffer_aggressivity);
}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 p->coeff*var / (q*p->count);
}
+/*
static double predict_qp(Predictor *p, double size, double var)
{
//printf("coeff:%f, count:%f, var:%f, size:%f//\n", p->coeff, p->count, var, size);
return p->coeff*var / (size*p->count);
}
+*/
static void update_predictor(Predictor *p, double q, double var, double size)
{
p->coeff+= new_coeff;
}
-int ff_rate_estimate_qscale(MpegEncContext *s)
+static void adaptive_quantization(MpegEncContext *s, double q){
+ int i;
+ const float lumi_masking= s->avctx->lumi_masking / (128.0*128.0);
+ const float dark_masking= s->avctx->dark_masking / (128.0*128.0);
+ 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->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;
+ 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
+ if(spat_cplx < 4) spat_cplx= 4; //FIXME finetune
+ if(temp_cplx < 4) temp_cplx= 4; //FIXME finetune
+
+ if((s->mb_type[mb_xy]&CANDIDATE_MB_TYPE_INTRA)){//FIXME hq mode
+ cplx= spat_cplx;
+ factor= 1.0 + p_masking;
+ }else{
+ cplx= temp_cplx;
+ factor= pow(temp_cplx, - temp_cplx_masking);
+ }
+ factor*=pow(spat_cplx, - spatial_cplx_masking);
+
+ if(lumi>127)
+ 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;
+ cplx_tab[i]= cplx;
+ bits_tab[i]= 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*= factor;
+
+ if (newq > qmax){
+ bits_sum -= bits_tab[i];
+ cplx_sum -= cplx_tab[i]*q/qmax;
+ }
+ else if(newq < qmin){
+ 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];
+ int intq;
+
+ if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
+ newq*= bits_sum/cplx_sum;
+ }
+
+ intq= (int)(newq + 0.5);
+
+ if (intq > qmax) intq= qmax;
+ else if(intq < qmin) intq= qmin;
+//if(i%s->mb_width==0) printf("\n");
+//printf("%2d%3d ", intq, ff_sqrt(s->mc_mb_var[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, int dry_run)
{
float q;
- int qscale, qmin, qmax;
+ int qmin, qmax;
float br_compensation;
double diff;
double short_term_q;
int picture_number= s->picture_number;
int64_t wanted_bits;
RateControlContext *rcc= &s->rc_context;
+ AVCodecContext *a= s->avctx;
RateControlEntry local_rce, *rce;
double bits;
double rate_factor;
int var;
const int pict_type= s->pict_type;
+ Picture * const pic= &s->current_picture;
emms_c();
+#ifdef CONFIG_XVID
+ if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
+ return ff_xvid_rate_estimate_qscale(s, dry_run);
+#endif
+
get_qminmax(&qmin, &qmax, s, pict_type);
- fps= (double)s->frame_rate / 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);
}
}
diff= s->total_bits - wanted_bits;
- br_compensation= (s->bit_rate_tolerance - diff)/s->bit_rate_tolerance;
+ br_compensation= (a->bit_rate_tolerance - diff)/a->bit_rate_tolerance;
if(br_compensation<=0.0) br_compensation=0.001;
- var= pict_type == I_TYPE ? s->mb_var_sum : s->mc_mb_var_sum;
-
+ 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)
assert(pict_type == rce->new_pict_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= s->mc_mb_var_sum;
- rce->mb_var_sum = s-> mb_var_sum;
- rce->qscale = 2;
+ rce->mc_mb_var_sum= pic->mc_mb_var_sum;
+ rce->mb_var_sum = pic-> mb_var_sum;
+ rce->qscale = FF_QP2LAMBDA * 2;
rce->f_code = s->f_code;
rce->b_code = s->b_code;
rce->misc_bits= 1;
- if(picture_number>0)
- update_rc_buffer(s, s->frame_bits);
-
bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
if(pict_type== I_TYPE){
rce->i_count = s->mb_num;
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;
rcc->frame_count[pict_type] ++;
bits= rce->i_tex_bits + rce->p_tex_bits;
- rate_factor= rcc->pass1_wanted_bits/rcc->pass1_bits * br_compensation;
-
+ rate_factor= rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum * br_compensation;
+
q= get_qscale(s, rce, rate_factor, picture_number);
+ if (q < 0)
+ return -1;
assert(q>0.0);
//printf("%f ", q);
assert(q>0.0);
if(pict_type==P_TYPE || s->intra_only){ //FIXME type dependant blur like in 2-pass
- rcc->short_term_qsum*=s->qblur;
- rcc->short_term_qcount*=s->qblur;
+ rcc->short_term_qsum*=a->qblur;
+ rcc->short_term_qcount*=a->qblur;
rcc->short_term_qsum+= q;
rcc->short_term_qcount++;
//printf("%f ", q);
}
assert(q>0.0);
-
+
q= modify_qscale(s, rce, q, picture_number);
rcc->pass1_wanted_bits+= s->bit_rate/fps;
assert(q>0.0);
}
-//printf("qmin:%d, qmax:%d, q:%f\n", qmin, qmax, q);
-
- if (q<qmin) q=qmin;
+ if(s->avctx->debug&FF_DEBUG_RC){
+ av_log(s->avctx, AV_LOG_DEBUG, "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f size:%d var:%d/%d br:%d fps:%d\n",
+ av_get_pict_type_char(pict_type), qmin, q, qmax, picture_number, (int)wanted_bits/1000, (int)s->total_bits/1000,
+ br_compensation, short_term_q, s->frame_bits, pic->mb_var_sum, pic->mc_mb_var_sum, s->bit_rate/1000, (int)fps
+ );
+ }
+
+ if (q<qmin) q=qmin;
else if(q>qmax) q=qmax;
-
-// printf("%f %d %d %d\n", q, picture_number, (int)wanted_bits, (int)s->total_bits);
-
-//printf("%f %f %f\n", q, br_compensation, short_term_q);
- qscale= (int)(q + 0.5);
-
-//printf("q:%d diff:%d comp:%f st_q:%f last_size:%d type:%d\n", qscale, (int)diff, br_compensation,
-// short_term_q, s->frame_bits, pict_type);
-//printf("%d %d\n", s->bit_rate, (int)fps);
-
- rcc->last_qscale= qscale;
- rcc->last_mc_mb_var_sum= s->mc_mb_var_sum;
- rcc->last_mb_var_sum= s->mb_var_sum;
- return qscale;
+
+ if(s->adaptive_quant)
+ adaptive_quantization(s, q);
+ else
+ q= (int)(q + 0.5);
+
+ 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;
+ mvsum += s->mv_bits;
+ texsum += s->i_tex_bits + s->p_tex_bits;
+ printf("%d %d//\n\n", mvsum, texsum);
+}
+#endif
+ return q;
}
//----------------------------------------------
static int init_pass2(MpegEncContext *s)
{
RateControlContext *rcc= &s->rc_context;
- int i;
- double fps= (double)s->frame_rate / FRAME_RATE_BASE;
+ AVCodecContext *a= s->avctx;
+ int i, toobig;
+ 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
- uint64_t available_bits[5];
uint64_t all_const_bits;
uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
double rate_factor=0;
double step;
- int last_i_frame=-10000000;
- const int filter_size= (int)(s->qblur*4) | 1;
+ //int last_i_frame=-10000000;
+ const int filter_size= (int)(a->qblur*4) | 1;
double expected_bits;
- double *qscale, *blured_qscale;
+ double *qscale, *blured_qscale, qscale_sum;
/* find complexity & const_bits & decide the pict_types */
for(i=0; i<rcc->num_entries; i++){
RateControlEntry *rce= &rcc->entry[i];
-
- if(s->b_frame_strategy==0 || s->max_b_frames==0){
- rce->new_pict_type= rce->pict_type;
- }else{
- int j;
- int next_non_b_type=P_TYPE;
- switch(rce->pict_type){
- case I_TYPE:
- if(i-last_i_frame>s->gop_size/2){ //FIXME this is not optimal
- rce->new_pict_type= I_TYPE;
- last_i_frame= i;
- }else{
- rce->new_pict_type= P_TYPE; // will be caught by the scene detection anyway
- }
- break;
- case P_TYPE:
- rce->new_pict_type= P_TYPE;
- break;
- case B_TYPE:
- for(j=i+1; j<i+s->max_b_frames+2 && j<rcc->num_entries; j++){
- if(rcc->entry[j].pict_type != B_TYPE){
- next_non_b_type= rcc->entry[j].pict_type;
- break;
- }
- }
- if(next_non_b_type==I_TYPE)
- rce->new_pict_type= P_TYPE;
- else
- rce->new_pict_type= B_TYPE;
- break;
- }
- }
+ 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;
rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
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){
- fprintf(stderr, "requested bitrate is to low\n");
+ av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too 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
- + 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]);
- }
-//printf("qp_i:%f, qp_p:%f, qp_b:%f\n", avg_quantizer[I_TYPE],avg_quantizer[P_TYPE],avg_quantizer[B_TYPE]);
- for(i=0; i<5; i++){
- 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= malloc(sizeof(double)*rcc->num_entries);
- blured_qscale= malloc(sizeof(double)*rcc->num_entries);
+ qscale= av_malloc(sizeof(double)*rcc->num_entries);
+ blured_qscale= av_malloc(sizeof(double)*rcc->num_entries);
+ toobig = 0;
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= s->qblur==0 ? 1.0 : exp(-d*d/(s->qblur * s->qblur));
-
+ 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];
rce->new_qscale= modify_qscale(s, rce, blured_qscale[i], i);
bits= qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
//printf("%d %f\n", rce->new_bits, blured_qscale[i]);
- update_rc_buffer(s, bits);
+ bits += 8*ff_vbv_update(s, bits);
rce->expected_bits= expected_bits;
expected_bits += bits;
}
-// printf("%f %d %f\n", expected_bits, (int)all_available_bits, rate_factor);
- if(expected_bits > all_available_bits) rate_factor-= step;
+ /*
+ av_log(s->avctx, AV_LOG_INFO,
+ "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
+ expected_bits, (int)all_available_bits, rate_factor);
+ */
+ if(expected_bits > all_available_bits) {
+ rate_factor-= step;
+ ++toobig;
+ }
}
- free(qscale);
- free(blured_qscale);
+ av_free(qscale);
+ av_free(blured_qscale);
- if(abs(expected_bits/all_available_bits - 1.0) > 0.01 ){
- fprintf(stderr, "Error: 2pass curve failed to converge\n");
+ /* check bitrate calculations and print info */
+ qscale_sum = 0.0;
+ for(i=0; i<rcc->num_entries; i++){
+ /* av_log(s->avctx, AV_LOG_DEBUG, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
+ i, rcc->entry[i].new_qscale, rcc->entry[i].new_qscale / FF_QP2LAMBDA); */
+ qscale_sum += clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA, s->avctx->qmin, s->avctx->qmax);
+ }
+ assert(toobig <= 40);
+ av_log(s->avctx, AV_LOG_DEBUG,
+ "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
+ s->bit_rate,
+ (int)(expected_bits / ((double)all_available_bits/s->bit_rate)));
+ av_log(s->avctx, AV_LOG_DEBUG,
+ "[lavc rc] estimated target average qp: %.3f\n",
+ (float)qscale_sum / rcc->num_entries);
+ if (toobig == 0) {
+ av_log(s->avctx, AV_LOG_INFO,
+ "[lavc rc] Using all of requested bitrate is not "
+ "necessary for this video with these parameters.\n");
+ } else if (toobig == 40) {
+ av_log(s->avctx, AV_LOG_ERROR,
+ "[lavc rc] Error: bitrate too low for this video "
+ "with these parameters.\n");
+ return -1;
+ } else if (fabs(expected_bits/all_available_bits - 1.0) > 0.01) {
+ av_log(s->avctx, AV_LOG_ERROR,
+ "[lavc rc] Error: 2pass curve failed to converge\n");
return -1;
}