X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fratecontrol.c;h=d3bee6a457934e26ff96dde6d82ea4eb7acab264;hb=817015e4e2e8d9efbec446033341c9117f889bb8;hp=94fca1ea47048f48b8ed99d49e8f58d53c4db9f9;hpb=160d679c075e9f575f96ea2c2c000545a49916a0;p=ffmpeg diff --git a/libavcodec/ratecontrol.c b/libavcodec/ratecontrol.c index 94fca1ea470..d3bee6a4579 100644 --- a/libavcodec/ratecontrol.c +++ b/libavcodec/ratecontrol.c @@ -3,31 +3,35 @@ * * Copyright (c) 2002-2004 Michael Niedermayer * - * 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 "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 +#undef NDEBUG // Always check asserts, the speed effect is far too small to disable them. #include #ifndef M_E @@ -38,28 +42,86 @@ 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){ - 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); + 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); +} + +static inline double qp2bits(RateControlEntry *rce, double qp){ + if(qp<=0.0){ + 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){ + av_log(NULL, AV_LOG_ERROR, "bits<0.9\n"); + } + return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ bits; } int ff_rate_control_init(MpegEncContext *s) { RateControlContext *rcc= &s->rc_context; int i; + const char *error = NULL; + static const char * const const_names[]={ + "PI", + "E", + "iTex", + "pTex", + "tex", + "mv", + "fCode", + "iCount", + "mcVar", + "var", + "isI", + "isP", + "isB", + "avgQP", + "qComp", +/* "lastIQP", + "lastPQP", + "lastBQP", + "nextNonBQP",*/ + "avgIITex", + "avgPITex", + "avgPPTex", + "avgBPTex", + "avgTex", + NULL + }; + static double (*func1[])(void *, double)={ + (void *)bits2qp, + (void *)qp2bits, + NULL + }; + static const char * const func1_names[]={ + "bits2qp", + "qp2bits", + NULL + }; emms_c(); + rcc->rc_eq_eval = ff_parse(s->avctx->rc_eq, const_names, func1, func1_names, NULL, NULL, &error); + if (!rcc->rc_eq_eval) { + av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\": %s\n", s->avctx->rc_eq, error? error : ""); + return -1; + } + 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]= rcc->mv_bits_sum[i]= rcc->qscale_sum [i]= - rcc->frame_count[i]= 1; // 1 is better cuz of 1/0 and such + rcc->frame_count[i]= 1; // 1 is better because of 1/0 and such rcc->last_qscale_for[i]=FF_QP2LAMBDA * 5; } rcc->buffer_index= s->avctx->rc_initial_buffer_occupancy; @@ -76,18 +138,18 @@ int ff_rate_control_init(MpegEncContext *s) i+= s->max_b_frames; if(i<=0 || i>=INT_MAX / sizeof(RateControlEntry)) return -1; - rcc->entry = (RateControlEntry*)av_mallocz(i*sizeof(RateControlEntry)); + rcc->entry = av_mallocz(i*sizeof(RateControlEntry)); rcc->num_entries= i; - + /* init all to skipped p frames (with b frames we might have a not encoded frame at the end FIXME) */ for(i=0; inum_entries; i++){ RateControlEntry *rce= &rcc->entry[i]; - rce->pict_type= rce->new_pict_type=P_TYPE; + rce->pict_type= rce->new_pict_type=FF_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; inum_entries - s->max_b_frames; i++){ @@ -98,7 +160,7 @@ int ff_rate_control_init(MpegEncContext *s) next= strchr(p, ';'); if(next){ - (*next)=0; //sscanf in unbelieavle slow on looong strings //FIXME copy / dont write + (*next)=0; //sscanf in unbelievably slow on looong strings //FIXME copy / do not write next++; } e= sscanf(p, " in:%d ", &picture_number); @@ -107,37 +169,52 @@ int ff_rate_control_init(MpegEncContext *s) assert(picture_number < rcc->num_entries); 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->f_code, &rce->b_code, &rce->mc_mb_var_sum, &rce->mb_var_sum, &rce->i_count); - if(e!=12){ + 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_LIBXVID + 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_rc_eq_output_sum= 0.001; rcc->pass1_wanted_bits=0.001; - + + if(s->avctx->qblur > 1.0){ + av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n"); + return -1; + } /* 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; + + if (i%((s->gop_size+3)/4)==0) rce.pict_type= FF_I_TYPE; + else if(i%(s->max_b_frames+1)) rce.pict_type= FF_B_TYPE; + else rce.pict_type= FF_P_TYPE; rce.new_pict_type= rce.pict_type; rce.mc_mb_var_sum= bits*s->mb_num/100000; @@ -147,7 +224,7 @@ int ff_rate_control_init(MpegEncContext *s) rce.b_code = 1; rce.misc_bits= 1; - if(s->pict_type== I_TYPE){ + if(s->pict_type== FF_I_TYPE){ rce.i_count = s->mb_num; rce.i_tex_bits= bits; rce.p_tex_bits= 0; @@ -166,12 +243,12 @@ int ff_rate_control_init(MpegEncContext *s) 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 misbehaves a little for variable fps } } } - + return 0; } @@ -180,30 +257,22 @@ void ff_rate_control_uninit(MpegEncContext *s) RateControlContext *rcc= &s->rc_context; emms_c(); + ff_eval_free(rcc->rc_eq_eval); av_freep(&rcc->entry); -} -static inline double qp2bits(RateControlEntry *rce, double qp){ - if(qp<=0.0){ - av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n"); - } - return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ qp; +#ifdef CONFIG_LIBXVID + if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) + ff_xvid_rate_control_uninit(s); +#endif } -static inline double bits2qp(RateControlEntry *rce, double bits){ - if(bits<0.9){ - av_log(NULL, AV_LOG_ERROR, "bits<0.9\n"); - } - 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; @@ -215,15 +284,15 @@ int ff_vbv_update(MpegEncContext *s, int frame_size){ } left= buffer_size - rcc->buffer_index - 1; - rcc->buffer_index += clip(left, min_rate, max_rate); + rcc->buffer_index += av_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); @@ -241,7 +310,7 @@ static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_f 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[]={ @@ -251,91 +320,59 @@ static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_f rce->p_tex_bits*rce->qscale, (rce->i_tex_bits + rce->p_tex_bits)*(double)rce->qscale, rce->mv_bits/mb_num, - rce->pict_type == B_TYPE ? (rce->f_code + rce->b_code)*0.5 : rce->f_code, + rce->pict_type == FF_B_TYPE ? (rce->f_code + rce->b_code)*0.5 : rce->f_code, rce->i_count/mb_num, rce->mc_mb_var_sum/mb_num, rce->mb_var_sum/mb_num, - rce->pict_type == I_TYPE, - rce->pict_type == P_TYPE, - rce->pict_type == B_TYPE, + rce->pict_type == FF_I_TYPE, + rce->pict_type == FF_P_TYPE, + rce->pict_type == FF_B_TYPE, rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type], a->qcompress, -/* rcc->last_qscale_for[I_TYPE], - rcc->last_qscale_for[P_TYPE], - rcc->last_qscale_for[B_TYPE], +/* rcc->last_qscale_for[FF_I_TYPE], + rcc->last_qscale_for[FF_P_TYPE], + rcc->last_qscale_for[FF_B_TYPE], rcc->next_non_b_qscale,*/ - rcc->i_cplx_sum[I_TYPE] / (double)rcc->frame_count[I_TYPE], - rcc->i_cplx_sum[P_TYPE] / (double)rcc->frame_count[P_TYPE], - rcc->p_cplx_sum[P_TYPE] / (double)rcc->frame_count[P_TYPE], - rcc->p_cplx_sum[B_TYPE] / (double)rcc->frame_count[B_TYPE], + rcc->i_cplx_sum[FF_I_TYPE] / (double)rcc->frame_count[FF_I_TYPE], + rcc->i_cplx_sum[FF_P_TYPE] / (double)rcc->frame_count[FF_P_TYPE], + rcc->p_cplx_sum[FF_P_TYPE] / (double)rcc->frame_count[FF_P_TYPE], + rcc->p_cplx_sum[FF_B_TYPE] / (double)rcc->frame_count[FF_B_TYPE], (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type], 0 }; - static const char *const_names[]={ - "PI", - "E", - "iTex", - "pTex", - "tex", - "mv", - "fCode", - "iCount", - "mcVar", - "var", - "isI", - "isP", - "isB", - "avgQP", - "qComp", -/* "lastIQP", - "lastPQP", - "lastBQP", - "nextNonBQP",*/ - "avgIITex", - "avgPITex", - "avgPPTex", - "avgBPTex", - "avgTex", - NULL - }; - static double (*func1[])(void *, double)={ - (void *)bits2qp, - (void *)qp2bits, - NULL - }; - 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); - + bits= ff_parse_eval(rcc->rc_eq_eval, const_values, rce); + if (isnan(bits)) { + av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq); + 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; iavctx->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) + if (pict_type==FF_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) + else if(pict_type==FF_B_TYPE && s->avctx->b_quant_factor<0.0) q= -q*s->avctx->b_quant_factor + s->avctx->b_quant_offset; - + if(q<1) q=1; + return q; } @@ -343,16 +380,17 @@ static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, doubl 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_p_q = rcc->last_qscale_for[FF_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) + + if (pict_type==FF_I_TYPE && (a->i_quant_factor>0.0 || rcc->last_non_b_pict_type==FF_P_TYPE)) + q= last_p_q *FFABS(a->i_quant_factor) + a->i_quant_offset; + else if(pict_type==FF_B_TYPE && a->b_quant_factor>0.0) q= last_non_b_q* a->b_quant_factor + a->b_quant_offset; + if(q<1) q=1; /* last qscale / qdiff stuff */ - if(rcc->last_non_b_pict_type==pict_type || pict_type!=I_TYPE){ + if(rcc->last_non_b_pict_type==pict_type || pict_type!=FF_I_TYPE){ double last_q= rcc->last_qscale_for[pict_type]; const int maxdiff= FF_QP2LAMBDA * a->max_qdiff; @@ -360,9 +398,9 @@ static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, doubl 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_qscale_for[pict_type]= q; //Note we cannot do that after blurring + + if(pict_type!=FF_B_TYPE) rcc->last_non_b_pict_type= pict_type; return q; @@ -372,24 +410,24 @@ static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, doubl * 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){ - 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); - }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); + if(pict_type==FF_B_TYPE){ + 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==FF_I_TYPE){ + 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); } - qmin= clip(qmin, 1, FF_LAMBDA_MAX); - qmax= clip(qmax, 1, FF_LAMBDA_MAX); + qmin= av_clip(qmin, 1, FF_LAMBDA_MAX); + qmax= av_clip(qmax, 1, FF_LAMBDA_MAX); if(qmaxnew_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 */ - if(s->avctx->rc_qmod_freq && frame_num%s->avctx->rc_qmod_freq==0 && pict_type==P_TYPE) + if(s->avctx->rc_qmod_freq && frame_num%s->avctx->rc_qmod_freq==0 && pict_type==FF_P_TYPE) q*= s->avctx->rc_qmod_amp; bits= qp2bits(rce, q); @@ -454,16 +492,16 @@ static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, double q, }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; } @@ -511,7 +549,7 @@ static void adaptive_quantization(MpegEncContext *s, double q){ Picture * const pic= &s->current_picture; const int mb_width = s->mb_width; const int mb_height = s->mb_height; - + for(i=0; imb_num; i++){ const int mb_xy= s->mb_index2xy[i]; float temp_cplx= sqrt(pic->mc_mb_var[mb_xy]); //FIXME merge in pow() @@ -522,14 +560,14 @@ static void adaptive_quantization(MpegEncContext *s, double q){ int mb_y = mb_xy / s->mb_stride; int mb_distance; float mb_factor = 0.0; -#if 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]&CANDIDATE_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{ @@ -559,9 +597,9 @@ static void adaptive_quantization(MpegEncContext *s, double q){ } factor*= 1.0 - border_masking*mb_factor; - + if(factor<0.00001) factor= 0.00001; - + bits= cplx*factor; cplx_sum+= cplx; bits_sum+= bits; @@ -569,7 +607,7 @@ static void adaptive_quantization(MpegEncContext *s, double q){ bits_tab[i]= bits; } - /* handle qmin/qmax cliping */ + /* handle qmin/qmax clipping */ if(s->flags&CODEC_FLAG_NORMALIZE_AQP){ float factor= bits_sum/cplx_sum; for(i=0; imb_num; i++){ @@ -588,7 +626,7 @@ static void adaptive_quantization(MpegEncContext *s, double q){ if(bits_sum < 0.001) bits_sum= 0.001; if(cplx_sum < 0.001) cplx_sum= 0.001; } - + for(i=0; imb_num; i++){ const int mb_xy= s->mb_index2xy[i]; float newq= q*cplx_tab[i]/bits_tab[i]; @@ -607,9 +645,20 @@ static void adaptive_quantization(MpegEncContext *s, double q){ 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; @@ -629,13 +678,18 @@ float ff_rate_estimate_qscale(MpegEncContext *s) Picture * const pic= &s->current_picture; emms_c(); +#ifdef CONFIG_LIBXVID + 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->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){ - const int last_var= s->last_pict_type == I_TYPE ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum; + if(picture_number>2 && !dry_run){ + const int last_var= s->last_pict_type == FF_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); } @@ -645,25 +699,40 @@ float ff_rate_estimate_qscale(MpegEncContext *s) rce= &rcc->entry[picture_number]; wanted_bits= rce->expected_bits; }else{ + Picture *dts_pic; rce= &local_rce; - wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps); + + //FIXME add a dts field to AVFrame and ensure its set and use it here instead of reordering + //but the reordering is simpler for now until h.264 b pyramid must be handeld + if(s->pict_type == FF_B_TYPE || s->low_delay) + dts_pic= s->current_picture_ptr; + else + dts_pic= s->last_picture_ptr; + +//if(dts_pic) +// av_log(NULL, AV_LOG_ERROR, "%Ld %Ld %Ld %d\n", s->current_picture_ptr->pts, s->user_specified_pts, dts_pic->pts, picture_number); + + if(!dts_pic || dts_pic->pts == AV_NOPTS_VALUE) + wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps); + else + wanted_bits= (uint64_t)(s->bit_rate*(double)dts_pic->pts/fps); } diff= s->total_bits - wanted_bits; 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 ? pic->mb_var_sum : pic->mc_mb_var_sum; - + var= pict_type == FF_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) + if(pict_type!=FF_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= pic->mc_mb_var_sum; rce->mb_var_sum = pic-> mb_var_sum; @@ -673,7 +742,7 @@ float ff_rate_estimate_qscale(MpegEncContext *s) rce->misc_bits= 1; bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var)); - if(pict_type== I_TYPE){ + if(pict_type== FF_I_TYPE){ rce->i_count = s->mb_num; rce->i_tex_bits= bits; rce->p_tex_bits= 0; @@ -682,7 +751,7 @@ float ff_rate_estimate_qscale(MpegEncContext *s) 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; @@ -692,8 +761,10 @@ float ff_rate_estimate_qscale(MpegEncContext *s) 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); + if (q < 0) + return -1; assert(q>0.0); //printf("%f ", q); @@ -701,7 +772,7 @@ float ff_rate_estimate_qscale(MpegEncContext *s) //printf("%f ", q); assert(q>0.0); - if(pict_type==P_TYPE || s->intra_only){ //FIXME type dependant blur like in 2-pass + if(pict_type==FF_P_TYPE || s->intra_only){ //FIXME type dependent blur like in 2-pass rcc->short_term_qsum*=a->qblur; rcc->short_term_qcount*=a->qblur; @@ -712,7 +783,7 @@ float ff_rate_estimate_qscale(MpegEncContext *s) //printf("%f ", q); } assert(q>0.0); - + q= modify_qscale(s, rce, q, picture_number); rcc->pass1_wanted_bits+= s->bit_rate/fps; @@ -727,17 +798,19 @@ float ff_rate_estimate_qscale(MpegEncContext *s) ); } - if (qqmax) 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; @@ -756,25 +829,23 @@ static int init_pass2(MpegEncContext *s) { RateControlContext *rcc= &s->rc_context; AVCodecContext *a= s->avctx; - int i; - double fps= (double)s->avctx->frame_rate / (double)s->avctx->frame_rate_base; + 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 const_bits[5]={0,0,0,0,0}; // quantizer independent bits 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)(a->qblur*4) | 1; + const int filter_size= (int)(a->qblur*4) | 1; double expected_bits; - double *qscale, *blured_qscale; + double *qscale, *blurred_qscale, qscale_sum; /* find complexity & const_bits & decide the pict_types */ for(i=0; inum_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; @@ -784,45 +855,21 @@ static int init_pass2(MpegEncContext *s) complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)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]; - + all_const_bits= const_bits[FF_I_TYPE] + const_bits[FF_P_TYPE] + const_bits[FF_B_TYPE]; + if(all_available_bits < all_const_bits){ - av_log(s->avctx, AV_LOG_ERROR, "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= av_malloc(sizeof(double)*rcc->num_entries); - blured_qscale= av_malloc(sizeof(double)*rcc->num_entries); + blurred_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 */ @@ -834,7 +881,7 @@ static int init_pass2(MpegEncContext *s) /* 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]); } @@ -844,41 +891,73 @@ static int init_pass2(MpegEncContext *s) const int pict_type= rce->new_pict_type; int j; double q=0.0, sum=0.0; - + for(j=0; jqblur==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; sum+= coeff; } - blured_qscale[i]= q/sum; + blurred_qscale[i]= q/sum; } - + /* find expected bits */ for(i=0; inum_entries; i++){ RateControlEntry *rce= &rcc->entry[i]; double bits; - rce->new_qscale= modify_qscale(s, rce, blured_qscale[i], i); + rce->new_qscale= modify_qscale(s, rce, blurred_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]); +//printf("%d %f\n", rce->new_bits, blurred_qscale[i]); 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; + } } av_free(qscale); - av_free(blured_qscale); + av_free(blurred_qscale); - if(abs(expected_bits/all_available_bits - 1.0) > 0.01 ){ - av_log(s->avctx, AV_LOG_ERROR, "Error: 2pass curve failed to converge\n"); + /* check bitrate calculations and print info */ + qscale_sum = 0.0; + for(i=0; inum_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 += av_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; }