ratecontrol.c

来自「从FFMPEG转换而来的H264解码程序,VC下编译..」· C语言 代码 · 共 987 行 · 第 1/3 页

            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;
        avg+= intq;
    }
    return avg/s->mb_num;
}

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 qmin, qmax;
    float br_compensation;
    double diff;
    double short_term_q;
    double fps;
    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, picture_number);

    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 && !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(s->flags&CODEC_FLAG_PASS2){
        assert(picture_number>=0);
        assert(picture_number<rcc->num_entries);
        rce= &rcc->entry[picture_number];
        wanted_bits= rce->expected_bits;
    }else{
        rce= &local_rce;
        wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/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;

    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->new_pict_type= pict_type;
        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;

        bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
        if(pict_type== I_TYPE){
            rce->i_count   = s->mb_num;
            rce->i_tex_bits= bits;
            rce->p_tex_bits= 0;
            rce->mv_bits= 0;
        }else{
            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->p_cplx_sum [pict_type] += rce->p_tex_bits*rce->qscale;
        rcc->mv_bits_sum[pict_type] += rce->mv_bits;
        rcc->frame_count[pict_type] ++;

        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);
        q= get_diff_limited_q(s, rce, q);
//printf("%f ", q);
        //assert(q>0.0);

        if(pict_type==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;

            rcc->short_term_qsum+= q;
            rcc->short_term_qcount++;
//printf("%f ", q);
            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(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;

    if(s->adaptive_quant)
        ff_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;
}

//----------------------------------------------
// 2-Pass code

static int init_pass2(MpegEncContext *s)
{
    RateControlContext *rcc= &s->rc_context;
    AVCodecContext *a= s->avctx;
    int i;
    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)(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;
        rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
        rcc->frame_count[rce->pict_type] ++;

        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];

    if(all_available_bits < all_const_bits){
        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]*FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset;
        avg_quantizer[B_TYPE]= avg_quantizer[P_TYPE]*FFABS(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);

    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 */
        for(i=0; i<rcc->num_entries; i++){
            qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
        }
        assert(filter_size%2==1);

        /* 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]);
        }

        /* smooth curve */
        for(i=0; i<rcc->num_entries; i++){
            RateControlEntry *rce= &rcc->entry[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;
                sum+= coeff;
            }
            blured_qscale[i]= q/sum;
        }

        /* find expected bits */
        for(i=0; i<rcc->num_entries; i++){
            RateControlEntry *rce= &rcc->entry[i];
            double bits;
            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]);
            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_free(qscale);
    av_free(blured_qscale);

    if(fabs(expected_bits/all_available_bits - 1.0) > 0.01 ){
        av_log(s->avctx, AV_LOG_ERROR, "Error: 2pass curve failed to converge\n");
        return -1;
    }

    return 0;
}