ratecontrol.c

H.264 source codes

// apply VBV constraints and clip qscale to between lmin and lmax
static double clip_qscale( x264_t *h, int pict_type, double q )
{
    x264_ratecontrol_t *rcc = h->rc;
    double lmin = rcc->lmin[pict_type];
    double lmax = rcc->lmax[pict_type];
    double q0 = q;

    /* B-frames are not directly subject to VBV,
     * since they are controlled by the P-frames' QPs.
     * FIXME: in 2pass we could modify previous frames' QP too,
     *        instead of waiting for the buffer to fill */
    if( rcc->buffer_size &&
        ( pict_type == SLICE_TYPE_P ||
          ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) )
    {
        if( rcc->buffer_fill/rcc->buffer_size < 0.5 )
            q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
    }
    /* Now a hard threshold to make sure the frame fits in VBV.
     * This one is mostly for I-frames. */
    if( rcc->buffer_size && rcc->last_satd > 0 )
    {
        double bits = predict_size( &rcc->pred[rcc->slice_type], q, rcc->last_satd );
        double qf = 1.0;
        if( bits > rcc->buffer_fill/2 )
            qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 );
        q /= qf;
        bits *= qf;
        if( bits < rcc->buffer_rate/2 )
            q *= bits*2/rcc->buffer_rate;
        q = X264_MAX( q0, q );
    }

    if(lmin==lmax)
        return lmin;
    else if(rcc->b_2pass)
    {
        double min2 = log(lmin);
        double max2 = log(lmax);
        q = (log(q) - min2)/(max2-min2) - 0.5;
        q = 1.0/(1.0 + exp(-4*q));
        q = q*(max2-min2) + min2;
        return exp(q);
    }
    else
        return x264_clip3f(q, lmin, lmax);
}

// update qscale for 1 frame based on actual bits used so far
static float rate_estimate_qscale(x264_t *h, int pict_type)
{
    float q;
    x264_ratecontrol_t *rcc = h->rc;
    ratecontrol_entry_t rce;
    double lmin = rcc->lmin[pict_type];
    double lmax = rcc->lmax[pict_type];
    int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I]
                          + h->stat.i_slice_size[SLICE_TYPE_P]
                          + h->stat.i_slice_size[SLICE_TYPE_B]);

    if( rcc->b_2pass )
    {
        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 )
    {
        rcc->last_satd = 0;
        if(h->fenc->b_kept_as_ref)
            q = rcc->last_qscale * sqrtf(h->param.rc.f_pb_factor);
        else
            q = rcc->last_qscale * h->param.rc.f_pb_factor;
        return x264_clip3f(q, lmin, lmax);
    }
    else
    {
        double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
        if( rcc->b_2pass )
        {
            //FIXME adjust abr_buffer based on distance to the end of the video
            int64_t diff = total_bits - (int64_t)rce.expected_bits;
            q = rce.new_qscale;
            q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
            if( h->fenc->i_frame > 30 )
            {
                /* Adjust quant based on the difference between
                 * achieved and expected bitrate so far */
                double time = (double)h->fenc->i_frame / rcc->num_entries;
                double w = x264_clip3f( time*100, 0.0, 1.0 );
                q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
            }
            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, lmin, lmax;

            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;
            rcc->short_term_cplxcount ++;

            rce.p_tex_bits = rcc->last_satd;
            rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
            rce.i_tex_bits = 0;
            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;
            rcc->last_rceq = get_qscale(h, &rce, 1);

            wanted_bits = h->fenc->i_frame * rcc->bitrate / rcc->fps;
            abr_buffer *= X264_MAX( 1, sqrt(h->fenc->i_frame/25) );
            overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );

            q = rcc->last_rceq * overflow * rcc->cplxr_sum / rcc->wanted_bits_window;

            if( pict_type == SLICE_TYPE_I
                /* 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 );
                q = clip_qscale( h, pict_type, q );
            }
            else
            {
                if( h->stat.i_slice_count[SLICE_TYPE_P] < 5 )
                {
                    float w = h->stat.i_slice_count[SLICE_TYPE_P] / 5.;
                    float q2 = qp2qscale(ABR_INIT_QP);
                    q = q*w + q2*(1-w);
                }

                /* Asymmetric clipping, because symmetric would prevent
                 * overflow control in areas of rapidly oscillating complexity */
                lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
                lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
                if( overflow > 1.1 )
                    lmax *= rcc->lstep;
                else if( overflow < 0.9 )
                    lmin /= rcc->lstep;

                q = x264_clip3f(q, lmin, lmax);
                q = clip_qscale(h, pict_type, q);
                //FIXME use get_diff_limited_q() ?
            }
        }

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

        return q;
    }
}

static int init_pass2( x264_t *h )
{
    x264_ratecontrol_t *rcc = h->rc;
    uint64_t all_const_bits = 0;
    uint64_t all_available_bits = (uint64_t)(h->param.rc.i_bitrate * 1000 * (double)rcc->num_entries / rcc->fps);
    double rate_factor, step, step_mult;
    double qblur = h->param.rc.f_qblur;
    double cplxblur = h->param.rc.f_complexity_blur;
    const int filter_size = (int)(qblur*4) | 1;
    double expected_bits;
    double *qscale, *blurred_qscale;
    int i;

    /* find total/average complexity & const_bits */
    for(i=0; i<rcc->num_entries; i++){
        ratecontrol_entry_t *rce = &rcc->entry[i];
        all_const_bits += rce->misc_bits;
        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 * rce->qscale;
        rcc->frame_count[rce->pict_type] ++;
    }

    if( all_available_bits < all_const_bits)
    {
        x264_log(h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
                 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000)));
        return -1;
    }

    /* Blur complexities, to reduce local fluctuation of QP.
     * We don't blur the QPs directly, because then one very simple frame
     * could drag down the QP of a nearby complex frame and give it more
     * bits than intended. */
    for(i=0; i<rcc->num_entries; i++){
        ratecontrol_entry_t *rce = &rcc->entry[i];
        double weight_sum = 0;
        double cplx_sum = 0;
        double weight = 1.0;
        int j;
        /* weighted average of cplx of future frames */
        for(j=1; j<cplxblur*2 && j<rcc->num_entries-i; j++){
            ratecontrol_entry_t *rcj = &rcc->entry[i+j];
            weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
            if(weight < .0001)
                break;
            weight_sum += weight;
            cplx_sum += weight * qscale2bits(rcj, 1);
        }
        /* weighted average of cplx of past frames */
        weight = 1.0;
        for(j=0; j<=cplxblur*2 && j<=i; j++){
            ratecontrol_entry_t *rcj = &rcc->entry[i-j];
            weight_sum += weight;
            cplx_sum += weight * qscale2bits(rcj, 1);
            weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
            if(weight < .0001)
                break;
        }
        rce->blurred_complexity = cplx_sum / weight_sum;
    }

    qscale = x264_malloc(sizeof(double)*rcc->num_entries);
    if(filter_size > 1)
        blurred_qscale = x264_malloc(sizeof(double)*rcc->num_entries);
    else
        blurred_qscale = qscale;

    /* Search for a factor which, when multiplied by the RCEQ values from
     * each frame, adds up to the desired total size.
     * There is no exact closed-form solution because of VBV constraints and
     * because qscale2bits is not invertible, but we can start with the simple
     * approximation of scaling the 1st pass by the ratio of bitrates.
     * The search range is probably overkill, but speed doesn't matter here. */

    expected_bits = 1;
    for(i=0; i<rcc->num_entries; i++)
        expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0));
    step_mult = all_available_bits / expected_bits;

    rate_factor = 0;
    for(step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5){
        expected_bits = 0;
        rate_factor += step;

        rcc->last_non_b_pict_type = -1;
        rcc->last_accum_p_norm = 1;
        rcc->accum_p_norm = 0;
        rcc->buffer_fill = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;

        /* find qscale */
        for(i=0; i<rcc->num_entries; i++){
            qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor);
        }

        /* fixed I/B qscale relative to P */
        for(i=rcc->num_entries-1; i>=0; i--){
            qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]);
            assert(qscale[i] >= 0);
        }

        /* smooth curve */
        if(filter_size > 1){
            assert(filter_size%2==1);
            for(i=0; i<rcc->num_entries; i++){
                ratecontrol_entry_t *rce = &rcc->entry[i];
                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 = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur));
                    if(index < 0 || index >= rcc->num_entries) continue;
                    if(rce->pict_type != rcc->entry[index].pict_type) continue;
                    q += qscale[index] * coeff;
                    sum += coeff;
                }
                blurred_qscale[i] = q/sum;
            }
        }

        /* find expected bits */
        for(i=0; i<rcc->num_entries; i++){
            ratecontrol_entry_t *rce = &rcc->entry[i];
            double bits;
            rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]);
            assert(rce->new_qscale >= 0);
            bits = qscale2bits(rce, rce->new_qscale) + rce->misc_bits;

            rce->expected_bits = expected_bits;
            expected_bits += bits;
            update_vbv(h, bits);
        }

//printf("expected:%llu available:%llu factor:%lf avgQ:%lf\n", (uint64_t)expected_bits, all_available_bits, rate_factor);
        if(expected_bits > all_available_bits) rate_factor -= step;
    }

    x264_free(qscale);
    if(filter_size > 1)
        x264_free(blurred_qscale);

    if(fabs(expected_bits/all_available_bits - 1.0) > 0.01)
    {
        double avgq = 0;
        for(i=0; i<rcc->num_entries; i++)
            avgq += rcc->entry[i].new_qscale;
        avgq = qscale2qp(avgq / rcc->num_entries);

        x264_log(h, X264_LOG_ERROR, "Error: 2pass curve failed to converge\n");
        x264_log(h, X264_LOG_ERROR, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
                 (float)h->param.rc.i_bitrate,
                 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
                 avgq);
        if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2)
        {
            if(h->param.rc.i_qp_min > 0)
                x264_log(h, X264_LOG_ERROR, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min);
            else
                x264_log(h, X264_LOG_ERROR, "try reducing target bitrate\n");
        }
        else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2)
        {
            if(h->param.rc.i_qp_max < 51)
                x264_log(h, X264_LOG_ERROR, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max);
            else
                x264_log(h, X264_LOG_ERROR, "try increasing target bitrate\n");
        }
        else
            x264_log(h, X264_LOG_ERROR, "internal error\n");
    }

    return 0;
}