x264 - rate_estimate_qscale

// 依据到目前为止编码bit数估算一帧的qscale
// update qscale for 1 frame based on actual bits used so far
static float rate_estimate_qscale( x264_t *h )
{
    float q;
    x264_ratecontrol_t *rcc = h->rc;
    ratecontrol_entry_t UNINIT(rce);
    int pict_type = h->sh.i_type;
    // 获取统计的总bits数
    int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
                          + h->stat.i_frame_size[SLICE_TYPE_P]
                          + h->stat.i_frame_size[SLICE_TYPE_B])
                       - rcc->filler_bits_sum;

    if( rcc->b_2pass )
    {
        // 获取当前帧的rce
        rce = *rcc->rce;
        if( pict_type != rce.pict_type )
        {
            x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
                      slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
        }
    }

    if( pict_type == SLICE_TYPE_B )
    {
        // 当前帧为B帧
        /* B-frames don't have independent ratecontrol, but rather get the
         * average QP of the two adjacent P-frames + an offset */
        // 距当前帧最近的参考帧列表0， 1里的帧的类型是否是I帧
        int i0 = IS_X264_TYPE_I(h->fref_nearest[0]->i_type);
        int i1 = IS_X264_TYPE_I(h->fref_nearest[1]->i_type);
        // 获取当前待编码帧与两最近参考帧的POC之距离
        int dt0 = abs(h->fenc->i_poc - h->fref_nearest[0]->i_poc);
        int dt1 = abs(h->fenc->i_poc - h->fref_nearest[1]->i_poc);
        // 获取两最近参考帧的平均qp
        float q0 = h->fref_nearest[0]->f_qp_avg_rc;
        float q1 = h->fref_nearest[1]->f_qp_avg_rc;
       
        // 如果两参考帧是B帧， 并用作参考帧，将获取的平均qp
        // 各减去pb_offset的一半
        if( h->fref_nearest[0]->i_type == X264_TYPE_BREF )
            q0 -= rcc->pb_offset/2;
        if( h->fref_nearest[1]->i_type == X264_TYPE_BREF )
            q1 -= rcc->pb_offset/2;

        // 如果两最近参考帧都是I帧，则取q0, q1的平均，再加上ip_offset
        // 否则如果list0最近参考帧是I帧,则q = q0
        // 或list1最近参考帧是I帧,则q = q1
        // 如果两帧都不是I帧，则进行插值计算q
        if( i0 && i1 )
            q = (q0 + q1) / 2 + rcc->ip_offset;
        else if( i0 )
            q = q1;
        else if( i1 )
            q = q0;
        else
            q = (q0*dt1 + q1*dt0) / (dt0 + dt1);

        // 依据当前待编码帧是否将作为参考帧，来对
        // 当前帧qp作相应调整
        if( h->fenc->b_kept_as_ref )
            q += rcc->pb_offset/2;
        else
            q += rcc->pb_offset;

        // 对于两阶段, 并且b_vbv生效
        // 根据rce的tex_bits, mv_bits和misc_bits
        // 以及当前帧的qscale，来估算当前帧的bits
        // 见下面 qscale2bits
        if( rcc->b_2pass && rcc->b_vbv )
            rcc->frame_size_planned = qscale2bits( &rce, qp2qscale( q ) );
        else   // 见下面 predict_size
            rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, qp2qscale( q ), h->fref[1][h->i_ref[1]-1]->i_satd );
        /* Limit planned size by MinCR */
        if( rcc->b_vbv )
            rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
        h->rc->frame_size_estimated = rcc->frame_size_planned;

        /* For row SATDs */
        if( rcc->b_vbv )  // 见下面 x264_rc_analyse_slice
            rcc->last_satd = x264_rc_analyse_slice( h );

        rcc->qp_novbv = q;
        return qp2qscale( q );
    }
    else
    {
        // 非B帧情况
        // 自适应缓冲区
        double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
       
        // 两阶段
        if( rcc->b_2pass )
        {  
            double lmin = rcc->lmin[pict_type];
            double lmax = rcc->lmax[pict_type];
            int64_t diff;
            int64_t predicted_bits = total_bits;
           
            // 统计预测的bits
            if( rcc->b_vbv )
            {
                if( h->i_thread_frames > 1 )
                {
                    int j = h->rc - h->thread[0]->rc;
                    for( int i = 1; i < h->i_thread_frames; i++ )
                    {
                        x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
                        double bits = t->rc->frame_size_planned;
                        if( !t->b_thread_active )
                            continue;
                        bits = X264_MAX(bits, t->rc->frame_size_estimated);
                        predicted_bits += (int64_t)bits;
                    }
                }
            }
            else
            {
                if( h->i_frame < h->i_thread_frames )
                    predicted_bits += (int64_t)h->i_frame * rcc->bitrate / rcc->fps;
                else
                    predicted_bits += (int64_t)(h->i_thread_frames - 1) * rcc->bitrate / rcc->fps;
            }

            /* Adjust ABR buffer based on distance to the end of the video. */
            if( rcc->num_entries > h->i_frame )
            {
                double final_bits = rcc->entry[rcc->num_entries-1].expected_bits;
                double video_pos = rce.expected_bits / final_bits;
                double scale_factor = sqrt( (1 - video_pos) * rcc->num_entries );
                abr_buffer *= 0.5 * X264_MAX( scale_factor, 0.5 );
            }

            diff = predicted_bits - (int64_t)rce.expected_bits;
            q = rce.new_qscale;
            q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
            if( ((h->i_frame + 1 - h->i_thread_frames) >= rcc->fps) &&
                (rcc->expected_bits_sum > 0))
            {
                /* Adjust quant based on the difference between
                 * achieved and expected bitrate so far */
                double cur_time = (double)h->i_frame / rcc->num_entries;
                double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
                q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
            }
            rcc->qp_novbv = qscale2qp( q );
            if( rcc->b_vbv )
            {
                /* Do not overflow vbv */
                double expected_size = qscale2bits( &rce, q );
                double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
                double expected_fullness = rce.expected_vbv / rcc->buffer_size;
                double qmax = q*(2 - expected_fullness);
                double size_constraint = 1 + expected_fullness;
                qmax = X264_MAX( qmax, rce.new_qscale );
                if( expected_fullness < .05 )
                    qmax = lmax;
                qmax = X264_MIN(qmax, lmax);
                while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
                        ((expected_vbv < 0) && (q < lmax)))
                {
                    q *= 1.05;
                    expected_size = qscale2bits(&rce, q);
                    expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
                }
                rcc->last_satd = x264_rc_analyse_slice( h );
            }
            q = x264_clip3f( q, lmin, lmax );
        }
        else /* 1pass ABR */
        {
            /* Calculate the quantizer which would have produced the desired
             * average bitrate if it had been applied to all frames so far.
             * Then modulate that quant based on the current frame's complexity
             * relative to the average complexity so far (using the 2pass RCEQ).
             * Then bias the quant up or down if total size so far was far from
             * the target.
             * Result: Depending on the value of rate_tolerance, there is a
             * tradeoff between quality and bitrate precision. But at large
             * tolerances, the bit distribution approaches that of 2pass. */

            double wanted_bits, overflow = 1;

            rcc->last_satd = x264_rc_analyse_slice( h );
            rcc->short_term_cplxsum *= 0.5;
            rcc->short_term_cplxcount *= 0.5;
            rcc->short_term_cplxsum += rcc->last_satd / (CLIP_DURATION(h->fenc->f_duration) / BASE_FRAME_DURATION);
            rcc->short_term_cplxcount ++;

            rce.tex_bits = rcc->last_satd;
            rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
            rce.mv_bits = 0;
            rce.p_count = rcc->nmb;
            rce.i_count = 0;
            rce.s_count = 0;
            rce.qscale = 1;
            rce.pict_type = pict_type;
            rce.i_duration = h->fenc->i_duration;

            if( h->param.rc.i_rc_method == X264_RC_CRF )
            {
                q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
            }
            else
            {
                q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );

                /* ABR code can potentially be counterproductive in CBR, so just don't bother.
                 * Don't run it if the frame complexity is zero either. */
                if( !rcc->b_vbv_min_rate && rcc->last_satd )
                {
                    // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
                    int i_frame_done = h->i_frame + 1 - h->i_thread_frames;
                    double time_done = i_frame_done / rcc->fps;
                    if( h->param.b_vfr_input && i_frame_done > 0 )
                        time_done = ((double)(h->fenc->i_reordered_pts - h->i_reordered_pts_delay)) * h->param.i_timebase_num / h->param.i_timebase_den;
                    wanted_bits = time_done * rcc->bitrate;
                    if( wanted_bits > 0 )
                    {
                        abr_buffer *= X264_MAX( 1, sqrt( time_done ) );
                        overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
                        q *= overflow;
                    }
                }
            }

            if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
                /* should test _next_ pict type, but that isn't decided yet */
                && rcc->last_non_b_pict_type != SLICE_TYPE_I )
            {
                q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
                q /= fabs( h->param.rc.f_ip_factor );
            }
            else if( h->i_frame > 0 )
            {
                if( h->param.rc.i_rc_method != X264_RC_CRF )
                {
                    /* Asymmetric clipping, because symmetric would prevent
                     * overflow control in areas of rapidly oscillating complexity */
                    double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
                    double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
                    if( overflow > 1.1 && h->i_frame > 3 )
                        lmax *= rcc->lstep;
                    else if( overflow < 0.9 )
                        lmin /= rcc->lstep;

                    q = x264_clip3f(q, lmin, lmax);
                }
            }
            else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
            {
                q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
            }
            rcc->qp_novbv = qscale2qp( q );

            //FIXME use get_diff_limited_q() ?
            q = clip_qscale( h, pict_type, q );
        }

        rcc->last_qscale_for[pict_type] =
        rcc->last_qscale = q;

        if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
            rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );

        if( rcc->b_2pass && rcc->b_vbv )
            rcc->frame_size_planned = qscale2bits(&rce, q);
        else
            rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );

        /* Always use up the whole VBV in this case. */
        if( rcc->single_frame_vbv )
            rcc->frame_size_planned = rcc->buffer_rate;
        /* Limit planned size by MinCR */
        if( rcc->b_vbv )
            rcc->frame_size_planned = X264_MIN( rcc->frame_size_planned, rcc->frame_size_maximum );
        h->rc->frame_size_estimated = rcc->frame_size_planned;
        return q;
    }
}

// 根据rce的tex_bits, mv_bits和misc_bits
// 以及当前帧的qscale，来估算当前帧的bits
/* Texture bitrate is not quite inversely proportional to qscale,
 * probably due the the changing number of SKIP blocks.
 * MV bits level off at about qp<=12, because the lambda used
 * for motion estimation is constant there. */
static inline double qscale2bits( ratecontrol_entry_t *rce, double qscale )
{
    if( qscale<0.1 )
        qscale = 0.1;
    return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
           + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
           + rce->misc_bits;
}

// p - 预测器， q - qscale， var - satd
static float predict_size( predictor_t *p, float q, float var )
{
    return (p->coeff*var + p->offset) / (q*p->count);
}

int x264_rc_analyse_slice( x264_t *h )
{
    int p0 = 0, p1, b;
    int cost;
    x264_emms();

    // 如果待编码帧是I帧，那么p1, b = 0
    // 如果是P帧，则p1=b=h->fenc->i_bframes + 1 ???
    // 否则p1取距待编码帧最近的两参考帧(list0, list1中个一个)
    // POC之差的平均， b 取 待编码帧与list0最近的参考帧之距的平均
    if( IS_X264_TYPE_I(h->fenc->i_type) )
        p1 = b = 0;
    else if( h->fenc->i_type == X264_TYPE_P )
        p1 = b = h->fenc->i_bframes + 1;
    else //B帧
    {
        p1 = (h->fref_nearest[1]->i_poc - h->fref_nearest[0]->i_poc)/2;
        b  = (h->fenc->i_poc - h->fref_nearest[0]->i_poc)/2;
    }
   
    // 注意这段代码有点诡异。
    // 下面frames指向h中的某块内存
    // 再下面frames[b] 又指回来了。
    // 即&frames[b] 就是 &h->fenc
    // frames[b] 就是 h->fenc
    /* We don't need to assign p0/p1 since we are not performing any real analysis here. */
    x264_frame_t **frames = &h->fenc - b;

    /* cost should have been already calculated by x264_slicetype_decide */
    //
    cost = frames[b]->i_cost_est[b-p0][p1-b];
    assert( cost >= 0 );

    // 使能mbtree, 但是没有使能stat_read
    if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
    {
        // 统计行积分，计算帧编码代价
        cost = x264_slicetype_frame_cost_recalculate( h, frames, p0, p1, b );
        // ???
        if( b && h->param.rc.i_vbv_buffer_size )
            x264_slicetype_frame_cost_recalculate( h, frames, b, b, b );
    }
    /* In AQ, use the weighted score instead. */
    else if( h->param.rc.i_aq_mode )
        cost = frames[b]->i_cost_est_aq[b-p0][p1-b];

    // 获取该编码帧和重构帧的行satd
    h->fenc->i_row_satd = h->fenc->i_row_satds[b-p0][p1-b];
    h->fdec->i_row_satd = h->fdec->i_row_satds[b-p0][p1-b];
    // 将计算的代价赋值给重构帧的i_satd
    h->fdec->i_satd = cost;
    拷贝编码帧的行satd到重构帧的行satd
    memcpy( h->fdec->i_row_satd, h->fenc->i_row_satd, h->mb.i_mb_height * sizeof(int) );
    // 拷贝I帧的行satd???
    if( !IS_X264_TYPE_I(h->fenc->i_type) )
        memcpy( h->fdec->i_row_satds[0][0], h->fenc->i_row_satds[0][0], h->mb.i_mb_height * sizeof(int) );

    // 帧内刷新？？？
    if( h->param.b_intra_refresh && h->param.rc.i_vbv_buffer_size && h->fenc->i_type == X264_TYPE_P )
    {
        int ip_factor = 256 * h->param.rc.f_ip_factor; /* fix8 */
        for( int y = 0; y < h->mb.i_mb_height; y++ )
        {
            int mb_xy = y * h->mb.i_mb_stride + h->fdec->i_pir_start_col;
            for( int x = h->fdec->i_pir_start_col; x <= h->fdec->i_pir_end_col; x++, mb_xy++ )
            {
                int intra_cost = (h->fenc->i_intra_cost[mb_xy] * ip_factor + 128) >> 8;
                int inter_cost = h->fenc->lowres_costs[b-p0][p1-b][mb_xy] & LOWRES_COST_MASK;
                int diff = intra_cost - inter_cost;
                if( h->param.rc.i_aq_mode )
                    h->fdec->i_row_satd[y] += (diff * frames[b]->i_inv_qscale_factor[mb_xy] + 128) >> 8;
                else
                    h->fdec->i_row_satd[y] += diff;
                cost += diff;
            }
        }
    }

    if( BIT_DEPTH > 8 )
        for( int y = 0; y < h->mb.i_mb_height; y++ )
            h->fdec->i_row_satd[y] >>= (BIT_DEPTH - 8);

    return cost >> (BIT_DEPTH - 8);
}

// 如果mbtree 改变了量化器，我们需要在不重新lookahead
// 情况下计算帧编码代价
/* If MB-tree changes the quantizers, we need to recalculate the frame cost without
 * re-running lookahead. */
static int x264_slicetype_frame_cost_recalculate( x264_t *h, x264_frame_t **frames, int p0, int p1, int b )
{
    int i_score = 0;
    // 获得该帧的行satd
    int *row_satd = frames[b]->i_row_satds[b-p0][p1-b];
    // 得到qp offset
    float *qp_offset = IS_X264_TYPE_B(frames[b]->i_type) ? frames[b]->f_qp_offset_aq : frames[b]->f_qp_offset;
    x264_emms();
    // 统计行satd，并且计算总的score
    for( h->mb.i_mb_y = h->mb.i_mb_height - 1; h->mb.i_mb_y >= 0; h->mb.i_mb_y-- )
    {
        row_satd[ h->mb.i_mb_y ] = 0;
        for( h->mb.i_mb_x = h->mb.i_mb_width - 1; h->mb.i_mb_x >= 0; h->mb.i_mb_x-- )
        {
            int i_mb_xy = h->mb.i_mb_x + h->mb.i_mb_y*h->mb.i_mb_stride;
            int i_mb_cost = frames[b]->lowres_costs[b-p0][p1-b][i_mb_xy] & LOWRES_COST_MASK;
            float qp_adj = qp_offset[i_mb_xy];
            i_mb_cost = (i_mb_cost * x264_exp2fix8(qp_adj) + 128) >> 8;
            row_satd[ h->mb.i_mb_y ] += i_mb_cost;
            if( (h->mb.i_mb_y > 0 && h->mb.i_mb_y < h->mb.i_mb_height - 1 &&
                 h->mb.i_mb_x > 0 && h->mb.i_mb_x < h->mb.i_mb_width - 1) ||
                 h->mb.i_mb_width <= 2 || h->mb.i_mb_height <= 2 )
            {
                i_score += i_mb_cost;
            }
        }
    }
    return i_score;
}