x264 - x264_ratecontrol_new

int x264_ratecontrol_new( x264_t *h )
{
    x264_ratecontrol_t *rc;

    x264_emms();

    // 依据线程数， 为h->rc分配内存空间
    CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
    rc = h->rc;

    // 确定是否为自适应码率控制模式
    rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
    rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;

    /* FIXME: use integers */
    // 计算帧率
    if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
        rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
    else
        rc->fps = 25.0;

    // 如果 mbtree 码率控制， 则pb_factor = 1, 即B帧等同P帧的qp
    // 同时， 量化器曲线压缩因子等于 1
    if( h->param.rc.b_mb_tree )
    {
        h->param.rc.f_pb_factor = 1;
        rc->qcompress = 1;
    }
    else
        rc->qcompress = h->param.rc.f_qcompress;

    rc->bitrate = h->param.rc.i_bitrate * (h->param.i_avcintra_class ? 1024. : 1000.);
    // bitrate 容许误差
    rc->rate_tolerance = h->param.rc.f_rate_tolerance;
    rc->nmb = h->mb.i_mb_count;
    rc->last_non_b_pict_type = -1;
    // cbr 衰减系数
    rc->cbr_decay = 1.0;

    if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
    {
        x264_log( h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n" );
        return -1;
    }
    // 见 x264_ratecontrol_init_reconfigurable
    x264_ratecontrol_init_reconfigurable( h, 1 );

    if( h->param.i_nal_hrd )
    {
        uint64_t denom = (uint64_t)h->sps->vui.hrd.i_bit_rate_unscaled * h->sps->vui.i_time_scale;
        uint64_t num = 180000;
        x264_reduce_fraction64( &num, &denom );
        rc->hrd_multiply_denom = 180000 / num;

        double bits_required = log2( 180000 / rc->hrd_multiply_denom )
                             + log2( h->sps->vui.i_time_scale )
                             + log2( h->sps->vui.hrd.i_cpb_size_unscaled );
        if( bits_required >= 63 )
        {
            x264_log( h, X264_LOG_ERROR, "HRD with very large timescale and bufsize not supported\n" );
            return -1;
        }
    }

    // 容许误差太小
    if( rc->rate_tolerance < 0.01 )
    {
        x264_log( h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n" );
        rc->rate_tolerance = 0.01;
    }

    // 可变qp标志取决于两个变量的或， b_vbv 和rc.i_aq_mode
    h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;

    // 如果自适应码率控制
    if( rc->b_abr )
    {
        /* FIXME ABR_INIT_QP is actually used only in CRF */
#define ABR_INIT_QP (( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 ) + QP_BD_OFFSET)
        rc->accum_p_norm = .01;
        rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
        /* estimated ratio that produces a reasonable QP for the first I-frame */
        // 预计复杂度之和
        rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
        // 分摊到每帧bit数窗口
        rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
        // 最后非B帧图片类型为I
        rc->last_non_b_pict_type = SLICE_TYPE_I;
    }

    // 缺省 param.rc.f_ip_factor = 1.4
    // h->param.rc.f_ip_factor = param.rc.f_ip_factor
    // I帧qp偏移 = 6.0 * log2f( h->param.rc.f_ip_factor )
    // 缺省 param.rc.f_pb_factor = 1.3
    // h->param.rc.f_pb_factor = param.rc.f_pb_factor
    // B帧qp偏移 = 6.0 * log2f( h->param.rc.f_pb_factor );
   
    rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
    rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
    rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
    rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, QP_MAX );
    rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, QP_MAX );
    h->mb.ip_offset = rc->ip_offset + 0.5;

    // lstep=>每帧qscale的最大变化系数
    // 初始化last_qscale
    rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
    rc->last_qscale = qp2qscale( 26 );
   
    // 初始化pred, row_preds, lmin, lmax, last_qscale_for
    // lmin, lmax 表示与帧类型相关的最小，最大qscale
    // last_qscale_for 表示与帧类型相关的上一次的qscale
    // 帧（图像类型）
    // enum slice_type_e
    // {
    //    SLICE_TYPE_P  = 0,
    //    SLICE_TYPE_B  = 1,
    //    SLICE_TYPE_I  = 2,
    // };
   
    int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
    CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
    CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
    for( int i = 0; i < 3; i++ )
    {
        rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
        rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
        rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
        for( int j = 0; j < num_preds; j++ )
        {
            rc->pred[i+j*5].coeff_min = 2.0 / 4;
            rc->pred[i+j*5].coeff = 2.0;
            rc->pred[i+j*5].count = 1.0;
            rc->pred[i+j*5].decay = 0.5;
            rc->pred[i+j*5].offset = 0.0;
        }
        for( int j = 0; j < 2; j++ )
        {
            rc->row_preds[i][j].coeff_min = .25 / 4;
            rc->row_preds[i][j].coeff = .25;
            rc->row_preds[i][j].count = 1.0;
            rc->row_preds[i][j].decay = 0.5;
            rc->row_preds[i][j].offset = 0.0;
        }
    }
    *rc->pred_b_from_p = rc->pred[0];

    // 见 parse_zones
    if( parse_zones( h ) < 0 )
    {
        x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
        return -1;
    }

    /* Load stat file and init 2pass algo */
    if( h->param.rc.b_stat_read )
    {
        char *p, *stats_in, *stats_buf;

        /* read 1st pass stats */
        assert( h->param.rc.psz_stat_in );
        stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
        if( !stats_buf )
        {
            x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
            return -1;
        }
        if( h->param.rc.b_mb_tree )
        {
            char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
            if( !mbtree_stats_in )
                return -1;
            rc->p_mbtree_stat_file_in = x264_fopen( mbtree_stats_in, "rb" );
            x264_free( mbtree_stats_in );
            if( !rc->p_mbtree_stat_file_in )
            {
                x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
                return -1;
            }
        }

        /* check whether 1st pass options were compatible with current options */
        if( strncmp( stats_buf, "#options:", 9 ) )
        {
            x264_log( h, X264_LOG_ERROR, "options list in stats file not valid\n" );
            return -1;
        }

        float res_factor, res_factor_bits;
        {
            int i, j;
            uint32_t k, l;
            char *opts = stats_buf;
            stats_in = strchr( stats_buf, '\n' );
            if( !stats_in )
                return -1;
            *stats_in = '\0';
            stats_in++;
            if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
            {
                x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
                return -1;
            }
            else if( h->param.rc.b_mb_tree )
            {
                rc->mbtree.srcdim[0] = i;
                rc->mbtree.srcdim[1] = j;
            }
            res_factor = (float)h->param.i_width * h->param.i_height / (i*j);
            /* Change in bits relative to resolution isn't quite linear on typical sources,
             * so we'll at least try to roughly approximate this effect. */
            res_factor_bits = powf( res_factor, 0.7 );

            if( !( p = strstr( opts, "timebase=" ) ) || sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
            {
                x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
                return -1;
            }
            if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
            {
                x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
                          h->param.i_timebase_num, h->param.i_timebase_den, k, l );
                return -1;
            }

            CMP_OPT_FIRST_PASS( "bitdepth", BIT_DEPTH );
            CMP_OPT_FIRST_PASS( "weightp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
            CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
            CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
            CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
            CMP_OPT_FIRST_PASS( "open_gop", h->param.b_open_gop );
            CMP_OPT_FIRST_PASS( "bluray_compat", h->param.b_bluray_compat );

            if( (p = strstr( opts, "interlaced=" )) )
            {
                char *current = h->param.b_interlaced ? h->param.b_tff ? "tff" : "bff" : h->param.b_fake_interlaced ? "fake" : "0";
                char buf[5];
                sscanf( p, "interlaced=%4s", buf );
                if( strcmp( current, buf ) )
                {
                    x264_log( h, X264_LOG_ERROR, "different interlaced setting than first pass (%s vs %s)\n", current, buf );
                    return -1;
                }
            }

            if( (p = strstr( opts, "keyint=" )) )
            {
                p += 7;
                char buf[13] = "infinite ";
                if( h->param.i_keyint_max != X264_KEYINT_MAX_INFINITE )
                    sprintf( buf, "%d ", h->param.i_keyint_max );
                if( strncmp( p, buf, strlen(buf) ) )
                {
                    x264_log( h, X264_LOG_ERROR, "different keyint setting than first pass (%.*s vs %.*s)\n",
                              strlen(buf)-1, buf, strcspn(p, " "), p );
                    return -1;
                }
            }

            if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
                x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );

            if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
            {
                x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
                h->mb.b_direct_auto_write = 1;
            }

            if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
                h->param.i_bframe_adaptive = i;
            else if( h->param.i_bframe )
            {
                x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
                return -1;
            }

            if( (h->param.rc.b_mb_tree || h->param.rc.i_vbv_buffer_size) && ( p = strstr( opts, "rc_lookahead=" ) ) && sscanf( p, "rc_lookahead=%d", &i ) )
                h->param.rc.i_lookahead = i;
        }

        /* find number of pics */
        p = stats_in;
        int num_entries;
        for( num_entries = -1; p; num_entries++ )
            p = strchr( p + 1, ';' );
        if( !num_entries )
        {
            x264_log( h, X264_LOG_ERROR, "empty stats file\n" );
            return -1;
        }
        rc->num_entries = num_entries;

        if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
        {
            x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
                      h->param.i_frame_total, rc->num_entries );
        }
        if( h->param.i_frame_total > rc->num_entries )
        {
            x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
                      h->param.i_frame_total, rc->num_entries );
            return -1;
        }

        CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );

        /* init all to skipped p frames */
        for( int i = 0; i < rc->num_entries; i++ )
        {
            ratecontrol_entry_t *rce = &rc->entry[i];
            rce->pict_type = SLICE_TYPE_P;
            rce->qscale = rce->new_qscale = qp2qscale( 20 );
            rce->misc_bits = rc->nmb + 10;
            rce->new_qp = 0;
        }

        /* read stats */
        p = stats_in;
        double total_qp_aq = 0;
        for( int i = 0; i < rc->num_entries; i++ )
        {
            ratecontrol_entry_t *rce;
            int frame_number;
            char pict_type;
            int e;
            char *next;
            float qp_rc, qp_aq;
            int ref;

            next= strchr(p, ';');
            if( next )
                *next++ = 0; //sscanf is unbelievably slow on long strings
            e = sscanf( p, " in:%d ", &frame_number );

            if( frame_number < 0 || frame_number >= rc->num_entries )
            {
                x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
                return -1;
            }
            rce = &rc->entry[frame_number];
            rce->direct_mode = 0;

            e += sscanf( p, " in:%*d out:%*d type:%c dur:%"SCNd64" cpbdur:%"SCNd64" q:%f aq:%f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c",
                   &pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp_rc, &qp_aq, &rce->tex_bits,
                   &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
                   &rce->s_count, &rce->direct_mode );
            rce->tex_bits  *= res_factor_bits;
            rce->mv_bits   *= res_factor_bits;
            rce->misc_bits *= res_factor_bits;
            rce->i_count   *= res_factor;
            rce->p_count   *= res_factor;
            rce->s_count   *= res_factor;

            p = strstr( p, "ref:" );
            if( !p )
                goto parse_error;
            p += 4;
            for( ref = 0; ref < 16; ref++ )
            {
                if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
                    break;
                p = strchr( p+1, ' ' );
                if( !p )
                    goto parse_error;
            }
            rce->refs = ref;

            /* find weights */
            rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
            char *w = strchr( p, 'w' );
            if( w )
            {
                int count = sscanf( w, "w:%hd,%hd,%hd,%hd,%hd,%hd,%hd,%hd",
                                    &rce->i_weight_denom[0], &rce->weight[0][0], &rce->weight[0][1],
                                    &rce->i_weight_denom[1], &rce->weight[1][0], &rce->weight[1][1],
                                    &rce->weight[2][0], &rce->weight[2][1] );
                if( count == 3 )
                    rce->i_weight_denom[1] = -1;
                else if ( count != 8 )
                    rce->i_weight_denom[0] = rce->i_weight_denom[1] = -1;
            }

            if( pict_type != 'b' )
                rce->kept_as_ref = 1;
            switch( pict_type )
            {
                case 'I':
                    rce->frame_type = X264_TYPE_IDR;
                    rce->pict_type  = SLICE_TYPE_I;
                    break;
                case 'i':
                    rce->frame_type = X264_TYPE_I;
                    rce->pict_type  = SLICE_TYPE_I;
                    break;
                case 'P':
                    rce->frame_type = X264_TYPE_P;
                    rce->pict_type  = SLICE_TYPE_P;
                    break;
                case 'B':
                    rce->frame_type = X264_TYPE_BREF;
                    rce->pict_type  = SLICE_TYPE_B;
                    break;
                case 'b':
                    rce->frame_type = X264_TYPE_B;
                    rce->pict_type  = SLICE_TYPE_B;
                    break;
                default:  e = -1; break;
            }
            if( e < 13 )
            {
parse_error:
                x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
                return -1;
            }
            rce->qscale = qp2qscale( qp_rc );
            total_qp_aq += qp_aq;
            p = next;
        }
        if( !h->param.b_stitchable )
            h->pps->i_pic_init_qp = SPEC_QP( (int)(total_qp_aq / rc->num_entries + 0.5) );

        x264_free( stats_buf );

        if( h->param.rc.i_rc_method == X264_RC_ABR )
        {
            if( init_pass2( h ) < 0 )
                return -1;
        } /* else we're using constant quant, so no need to run the bitrate allocation */
    }

    /* Open output file */
    /* If input and output files are the same, output to a temp file
     * and move it to the real name only when it's complete */
    if( h->param.rc.b_stat_write )
    {
        char *p;
        rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
        if( !rc->psz_stat_file_tmpname )
            return -1;

        rc->p_stat_file_out = x264_fopen( rc->psz_stat_file_tmpname, "wb" );
        if( rc->p_stat_file_out == NULL )
        {
            x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n" );
            return -1;
        }

        p = x264_param2string( &h->param, 1 );
        if( p )
            fprintf( rc->p_stat_file_out, "#options: %s\n", p );
        x264_free( p );
        if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
        {
            rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
            rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
            if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
                return -1;

            rc->p_mbtree_stat_file_out = x264_fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
            if( rc->p_mbtree_stat_file_out == NULL )
            {
                x264_log( h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n" );
                return -1;
            }
        }
    }

    if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
    {
        if( !h->param.rc.b_stat_read )
        {
            rc->mbtree.srcdim[0] = h->param.i_width;
            rc->mbtree.srcdim[1] = h->param.i_height;
        }
        if( x264_macroblock_tree_rescale_init( h, rc ) < 0 )
            return -1;
    }

    // 对多线程而言，为每个线程分配一个rc
    for( int i = 0; i<h->param.i_threads; i++ )
    {
        h->thread[i]->rc = rc+i;
        if( i )
        {
            对其他rc, 拷贝第一个rc对象
            rc[i] = rc[0];
            h->thread[i]->param = h->param;
            h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
            h->thread[i]->mb.ip_offset = h->mb.ip_offset;
        }
    }

    return 0;
fail:
    return -1;
}