* License along with this library; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*/
-#include <math.h>
-#include "common.h"
+
+/**
+ * @file ratecontrol.c
+ * Rate control for video encoders.
+ */
+
#include "avcodec.h"
#include "dsputil.h"
#include "mpegvideo.h"
#undef NDEBUG // allways check asserts, the speed effect is far too small to disable them
#include <assert.h>
+#ifndef M_E
+#define M_E 2.718281828
+#endif
+
static int init_pass2(MpegEncContext *s);
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",
+ sprintf(s->avctx->stats_out, "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;\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);
+ s->frame_qscale, 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);
}
int ff_rate_control_init(MpegEncContext *s)
}
rcc->buffer_index= s->avctx->rc_buffer_size/2;
- rcc->next_non_b_qscale=10;
- rcc->next_p_qscale=10;
-
if(s->flags&CODEC_FLAG_PASS2){
int i;
char *p;
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",
+ 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->f_code, &rce->b_code, &rce->mc_mb_var_sum, &rce->mb_var_sum, &rce->i_count);
if(e!=12){
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 */
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/(s->avctx->frame_rate / (double)s->avctx->frame_rate_base);
}
}
static void update_rc_buffer(MpegEncContext *s, int frame_size){
RateControlContext *rcc= &s->rc_context;
- const double fps= (double)s->frame_rate / FRAME_RATE_BASE;
+ const double fps= (double)s->avctx->frame_rate / (double)s->avctx->frame_rate_base;
const double 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;
const int pict_type= rce->new_pict_type;
const double mb_num= s->mb_num;
int i;
- const double last_q= rcc->last_qscale_for[pict_type];
double const_values[]={
M_PI,
(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",
"avgTex",
NULL
};
- double (*func1[])(void *, double)={
- bits2qp,
- qp2bits,
+ static double (*func1[])(void *, double)={
+ (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;
+ rcc->pass1_rc_eq_output_sum+= bits;
bits*=rate_factor;
if(bits<0.0) bits=0.0;
bits+= 1.0; //avoid 1/0 issues
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;
+}
+
+static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q){
+ RateControlContext *rcc= &s->rc_context;
+ AVCodecContext *a= s->avctx;
+ 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)
+ q= last_non_b_q* a->b_quant_factor + a->b_quant_offset;
+
/* last qscale / qdiff stuff */
- if (q > last_q + s->max_qdiff) q= last_q + s->max_qdiff;
- else if(q < last_q - s->max_qdiff) q= last_q - s->max_qdiff;
+ 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;
+ }
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;
+
return q;
}
* 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->qmin;
+ int qmax= s->avctx->qmax;
if(pict_type==B_TYPE){
qmin= (int)(qmin*ABS(s->avctx->b_quant_factor)+s->avctx->b_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
+ if(qmin==1 && s->avctx->qmin>1) qmin=2; //avoid qmin=1 unless the user wants qmin=1
if(qmin<3 && s->max_qcoeff<=128 && pict_type==I_TYPE) qmin=3; //reduce cliping problems
q*= s->avctx->rc_qmod_amp;
bits= qp2bits(rce, q);
-
+//printf("q:%f\n", q);
/* buffer overflow/underflow protection */
if(buffer_size){
- double expected_size= rcc->buffer_index - bits;
+ double expected_size= rcc->buffer_index;
if(min_rate){
- double d= 2*(buffer_size - (expected_size + min_rate))/buffer_size;
+ double d= 2*(buffer_size - expected_size)/buffer_size;
if(d>1.0) d=1.0;
- q/= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
+ else if(d<0.0001) d=0.0001;
+ q*= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
+
+ q= FFMIN(q, bits2qp(rce, FFMAX((min_rate - buffer_size + rcc->buffer_index)*2, 1)));
}
if(max_rate){
double d= 2*expected_size/buffer_size;
if(d>1.0) d=1.0;
- q*= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
+ else if(d<0.0001) d=0.0001;
+ q/= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
+
+ q= FFMAX(q, bits2qp(rce, FFMAX(rcc->buffer_index/2, 1)));
}
}
-
+//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);
if(s->avctx->rc_qsquish==0.0 || qmin==qmax){
if (q<qmin) q=qmin;
else if(q>qmax) q=qmax;
q= exp(q);
}
-
+
return 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;
+ 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_qmin;
+ const int qmax= s->avctx->mb_qmax;
+ Picture * const pic= &s->current_picture;
+ int last_qscale=0;
+
+ 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]);
+ float spat_cplx= sqrt(pic->mb_var[mb_xy]);
+ const int lumi= pic->mb_mean[mb_xy];
+ float bits, cplx, factor;
+
+ if(spat_cplx < q/3) spat_cplx= q/3; //FIXME finetune
+ if(temp_cplx < q/3) temp_cplx= q/3; //FIXME finetune
+
+ if((s->mb_type[mb_xy]&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(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){
+ for(i=0; i<s->mb_num; i++){
+ float newq= q*cplx_tab[i]/bits_tab[i];
+ newq*= bits_sum/cplx_sum;
+
+ 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;
+ }
+ }
+ }
+
+ 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;
+ }
+
+ if(i && ABS(last_qscale - newq)<0.75)
+ intq= last_qscale;
+ else
+ 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]));
+ last_qscale=
+ pic->qscale_table[mb_xy]= intq;
+ }
+}
+
+float ff_rate_estimate_qscale(MpegEncContext *s)
{
float q;
- int qscale, qmin, qmax;
+ int qmin, qmax;
float br_compensation;
double diff;
double short_term_q;
double rate_factor;
int var;
const int pict_type= s->pict_type;
+ Picture * const pic= &s->current_picture;
emms_c();
get_qminmax(&qmin, &qmax, s, pict_type);
- fps= (double)s->frame_rate / FRAME_RATE_BASE;
-//printf("input_picture_number:%d picture_number:%d\n", s->input_picture_number, s->picture_number);
+ fps= (double)s->avctx->frame_rate / (double)s->avctx->frame_rate_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){
const int last_var= s->last_pict_type == I_TYPE ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
br_compensation= (s->bit_rate_tolerance - diff)/s->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;
if(s->flags&CODEC_FLAG_PASS2){
if(pict_type!=I_TYPE)
}else{
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->mc_mb_var_sum= pic->mc_mb_var_sum;
+ rce->mb_var_sum = pic-> mb_var_sum;
rce->qscale = 2;
rce->f_code = s->f_code;
rce->b_code = s->b_code;
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);
assert(q>0.0);
//printf("%f ", q);
- if (pict_type==I_TYPE && s->avctx->i_quant_factor>0.0)
- q= rcc->next_p_qscale*s->avctx->i_quant_factor + s->avctx->i_quant_offset;
- else if(pict_type==B_TYPE && s->avctx->b_quant_factor>0.0)
- q= rcc->next_non_b_qscale*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
+ q= get_diff_limited_q(s, rce, q);
//printf("%f ", q);
assert(q>0.0);
q= short_term_q= rcc->short_term_qsum/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);
+ }
- if(pict_type != B_TYPE) rcc->next_non_b_qscale= q;
- if(pict_type == P_TYPE) rcc->next_p_qscale= q;
+ if(s->avctx->debug&FF_DEBUG_RC){
+ printf("%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
+ );
}
-//printf("qmin:%d, qmax:%d, q:%f\n", qmin, qmax, q);
-
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("%d ", qscale);
+ if(s->adaptive_quant)
+ adaptive_quantization(s, q);
+ else
+ q= (int)(q + 0.5);
-//printf("q:%d diff:%d comp:%f rate_q:%d st_q:%f fvar:%d last_size:%d\n", qscale, (int)diff, br_compensation,
-// rate_q, short_term_q, s->mc_mb_var, s->frame_bits);
-//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;
+ 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;
}
//----------------------------------------------
{
RateControlContext *rcc= &s->rc_context;
int i;
- double fps= (double)s->frame_rate / FRAME_RATE_BASE;
+ double fps= (double)s->avctx->frame_rate / (double)s->avctx->frame_rate_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 all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
double rate_factor=0;
double step;
- int last_i_frame=-10000000;
+ //int last_i_frame=-10000000;
const int filter_size= (int)(s->qblur*4) | 1;
double expected_bits;
double *qscale, *blured_qscale;
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;
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]);
+//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);
for(step=256*256; step>0.0000001; step*=0.5){
expected_bits=0;
assert(filter_size%2==1);
/* fixed I/B QP relative to P mode */
- rcc->next_non_b_qscale= 10;
- rcc->next_p_qscale= 10;
for(i=rcc->num_entries-1; i>=0; i--){
RateControlEntry *rce= &rcc->entry[i];
- const int pict_type= rce->new_pict_type;
-
- if (pict_type==I_TYPE && s->avctx->i_quant_factor>0.0)
- qscale[i]= rcc->next_p_qscale*s->avctx->i_quant_factor + s->avctx->i_quant_offset;
- else if(pict_type==B_TYPE && s->avctx->b_quant_factor>0.0)
- qscale[i]= rcc->next_non_b_qscale*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
-
- if(pict_type!=B_TYPE)
- rcc->next_non_b_qscale= qscale[i];
- if(pict_type==P_TYPE)
- rcc->next_p_qscale= qscale[i];
+
+ qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
}
/* smooth curve */
RateControlEntry *rce= &rcc->entry[i];
double bits;
rce->new_qscale= modify_qscale(s, rce, blured_qscale[i], i);
- bits= qp2bits(rce, rce->new_qscale);
+ 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);
// printf("%f %d %f\n", expected_bits, (int)all_available_bits, rate_factor);
if(expected_bits > all_available_bits) rate_factor-= step;
}
- 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");
projects / ffmpeg / blobdiff