hvirtual/mpeg2enc/ratectl.c

Go to the documentation of this file.

/* ratectl.c, bitrate control routines (linear quantization only currently) */

/* Copyright (C) 1996, MPEG Software Simulation Group. All Rights Reserved. */

/*
 * Disclaimer of Warranty
 *
 * These software programs are available to the user without any license fee or
 * royalty on an "as is" basis.  The MPEG Software Simulation Group disclaims
 * any and all warranties, whether express, implied, or statuary, including any
 * implied warranties or merchantability or of fitness for a particular
 * purpose.  In no event shall the copyright-holder be liable for any
 * incidental, punitive, or consequential damages of any kind whatsoever
 * arising from the use of these programs.
 *
 * This disclaimer of warranty extends to the user of these programs and user's
 * customers, employees, agents, transferees, successors, and assigns.
 *
 * The MPEG Software Simulation Group does not represent or warrant that the
 * programs furnished hereunder are free of infringement of any third-party
 * patents.
 *
 * Commercial implementations of MPEG-1 and MPEG-2 video, including shareware,
 * are subject to royalty fees to patent holders.  Many of these patents are
 * general enough such that they are unavoidable regardless of implementation
 * design.
 *
 */

#include <stdio.h>
#include <limits.h>
#include <math.h>
#include <pthread.h>

#include "config.h"
#include "global.h"
#include "fastintfns.h"

/* rate control variables */
/*
 * static double R, T, d;
 * static double actsum;
 * static int Np, Nb;
 * static double S, Q;
 * static int prev_mquant;
 * static double bitcnt_EOP;
 * static double next_ip_delay; // due to frame reordering delay
 * static double decoding_time;
 * static int Xi, Xp, Xb, r, d0i, d0p, d0b;
 * static double avg_act;
 */

void ratectl_init_seq(ratectl_t *ratectl)
{
        pthread_mutexattr_t mutex_attr;
        pthread_mutexattr_init(&mutex_attr);
        pthread_mutex_init(&(ratectl->ratectl_lock), &mutex_attr);

        ratectl->avg_KI = 2.5;   /* TODO: These values empirically determined            */
        ratectl->avg_KB = 10.0;   /* for MPEG-1, may need tuning for MPEG-2   */
        ratectl->avg_KP = 10.0;

        ratectl->bits_per_mb = (double)bit_rate / (mb_per_pict);
        /* reaction parameter (constant) decreased to increase response
           rate as encoder is currently tending to under/over-shoot... in
           rate TODO: Reaction parameter is *same* for every frame type
           despite different weightings...  */

        if (ratectl->r == 0)  
                ratectl->r = (int)floor(2.0 * bit_rate / frame_rate + 0.5);

        ratectl->Ki = 1.2;  /* EXPERIMENT: ADJUST activities for I MB's */
        ratectl->Kb = 1.4;
        ratectl->Kp = 1.1;

/* average activity */
        if (ratectl->avg_act == 0.0)  ratectl->avg_act = 400.0;

/* remaining # of bits in GOP */
        ratectl->R = 0;
        ratectl->IR = 0;

        /* Heuristic: In constant bit-rate streams we assume buffering
           will allow us to pre-load some (probably small) fraction of the
           buffers size worth of following data if earlier data was
           undershot its bit-rate allocation
         
        */

        ratectl->CarryR = 0;
        ratectl->CarryRLim = video_buffer_size / 3;
        /* global complexity (Chi! not X!) measure of different frame types */
        /* These are just some sort-of sensible initial values for start-up */

        ratectl->Xi = 1500*mb_per_pict;   /* Empirically derived values */
        ratectl->Xp = 550*mb_per_pict;
        ratectl->Xb = 170*mb_per_pict;
        ratectl->d0i = -1;                              /* Force initial Quant prediction */
        ratectl->d0pb = -1;

        ratectl->current_quant = 1;
}

void ratectl_init_GOP(ratectl_t *ratectl, int np, int nb)
{
        double per_gop_bits = 
                (double)(1 + np + nb) * (double)bit_rate / frame_rate;

        /* A.Stevens Aug 2000: at last I've found the wretched
           rate-control overshoot bug...  Simply "topping up" R here means
           that we can accumulate an indefinately large pool of bits
           "saved" from previous low-activity frames.  This is of
           course nonsense.

           In CBR we can only accumulate as much as our buffer allows, after that
           the eventual system stream will have to be padded.  The automatic padding
           will make this calculation fairly reasonable but since that's based on 
           estimates we again impose our rough and ready heuristic that we can't
           accumulate more than approximately half  a video buffer full.

           In VBR we actually do nothing different.  Here the bitrate is
           simply a ceiling rate which we expect to undershoot a lot as
           our quantisation floor cuts in. We specify a great big buffer
           and simply don't pad when we undershoot.

           However, we don't want to carry over absurd undershoots as when it
           does get hectic we'll breach our maximum.
                
           TODO: For CBR we should do a proper video buffer model and use
           it to make bit allocation decisions.

        */

        if( ratectl->R > 0  )
        {
                /* We replacing running estimate of undershoot with
                   *exact* value and use that for calculating how much we
                   may "carry over"
                */
                ratectl->gop_undershoot = intmin( video_buffer_size/2, (int)ratectl->R );

                ratectl->R = ratectl->gop_undershoot + per_gop_bits;            
        }
        else 
        {
                /* Overshoots are easy - we have to make up the bits */
                ratectl->R +=  per_gop_bits;
                ratectl->gop_undershoot = 0;
        }
        ratectl->IR = ratectl->R;
        ratectl->Np = fieldpic ? 2 * np + 1 : np;
        ratectl->Nb = fieldpic ? 2 * nb : nb;
}

static int scale_quant(pict_data_s *picture, double quant )
{
        int iquant;
        
        if (picture->q_scale_type  )
        {
                iquant = (int) floor(quant+0.5);

                /* clip mquant to legal (linear) range */
                if (iquant<1)
                        iquant = 1;
                if (iquant>112)
                        iquant = 112;

                iquant =
                        non_linear_mquant_table_hv[map_non_linear_mquant_hv[iquant]];
        }
        else
        {
                /* clip mquant to legal (linear) range */
                iquant = (int)floor(quant+0.5);
                if (iquant<2)
                        iquant = 2;
                if (iquant>62)
                        iquant = 62;
                iquant = (iquant/2)*2; // Must be *even*
        }
        return iquant;
}




/* compute variance of 8x8 block */
static double var_sblk(p, lx)
unsigned char *p;
int lx;
{
        int j;
        register unsigned int v, s, s2;

        s = s2 = 0;

        for (j=0; j<8; j++)
        {
                v = p[0];   s += v;    s2 += v * v;
                v = p[1];   s += v;    s2 += v * v;
                v = p[2];   s += v;    s2 += v * v;
                v = p[3];   s += v;    s2 += v * v;
                v = p[4];   s += v;    s2 += v * v;
                v = p[5];   s += v;    s2 += v * v;
                v = p[6];   s += v;    s2 += v * v;
                v = p[7];   s += v;    s2 += v * v;
                p += lx;
        }

        return (double)s2 / 64.0 - ((double)s / 64.0) * ((double)s / 64.0);
}


static double calc_actj(pict_data_s *picture)
{
        int i,j,k,l;
        double actj,sum;
        uint16_t *i_q_mat;
        int actsum;
        sum = 0.0;
        k = 0;

        for (j=0; j<height2; j+=16)
                for (i=0; i<width; i+=16)
                {
                        /* A.Stevens Jul 2000 Luminance variance *has* to be a rotten measure
                           of how active a block in terms of bits needed to code a lossless DCT.
                           E.g. a half-white half-black block has a maximal variance but 
                           pretty small DCT coefficients.

                           So.... we use the absolute sum of DCT coefficients as our
                           variance measure.  
                        */
                        if( picture->mbinfo[k].mb_type  & MB_INTRA )
                        {
                                i_q_mat = i_intra_q;
                                /* EXPERIMENT: See what happens if we compensate for
                                 the wholly disproprotionate weight of the DC
                                 coefficients.  Shold produce more sensible results...  */
                                actsum =  -80*COEFFSUM_SCALE;
                        }
                        else
                        {
                                i_q_mat = i_inter_q;
                                actsum = 0;
                        }

                        /* It takes some bits to code even an entirely zero block...
                           It also makes a lot of calculations a lot better conditioned
                           if it can be guaranteed that activity is always distinctly
                           non-zero.
                         */


                        for( l = 0; l < 6; ++l )
                                actsum += 
                                        (*pquant_weight_coeff_sum)
                                            ( cur_picture.mbinfo[k].dctblocks[l], i_q_mat ) ;
                        actj = (double)actsum / (double)COEFFSUM_SCALE;
                        if( actj < 12.0 )
                                actj = 12.0;

                        picture->mbinfo[k].act = (double)actj;
                        sum += (double)actj;
                        ++k;
                }
        return sum;
}

/* Note: we need to substitute K for the 1.4 and 1.0 constants -- this can
   be modified to fit image content */

/* Step 1: compute target bits for current picture being coded */
void ratectl_init_pict(ratectl_t *ratectl, pict_data_s *picture)
{
        double avg_K;
        double target_Q;
        double current_Q;
        double Si, Sp, Sb;
        /* TODO: A.Stevens  Nov 2000 - This modification needs testing visually.

           Weird.  The original code used the average activity of the
           *previous* frame as the basis for quantisation calculations for
           rather than the activity in the *current* frame.  That *has* to
           be a bad idea..., surely, here we try to be smarter by using the
           current values and keeping track of how much of the frames
           activitity has been covered as we go along.

           We also guesstimate the relationship between  (sum
           of DCT coefficients) and actual quantisation weighted activty.
           We use this to try to predict the activity of each frame.
        */

        ratectl->actsum =  calc_actj(picture );
        ratectl->avg_act = (double)ratectl->actsum/(double)(mb_per_pict);
        ratectl->sum_avg_act += ratectl->avg_act;
        ratectl->actcovered = 0.0;

        /* Allocate target bits for frame based on frames numbers in GOP
           weighted by global complexity estimates and B-frame scale factor
           T = (Nx * Xx/Kx) / Sigma_j (Nj * Xj / Kj)
        */
        ratectl->min_q = ratectl->min_d = INT_MAX;
        ratectl->max_q = ratectl->max_d = INT_MIN;
        switch (picture->pict_type)
        {
        case I_TYPE:
                
                /* There is little reason to rely on the *last* I-frame
                   as they're not closely related.  The slow correction of
                   K should be enough to fine-tune...
                */

                ratectl->d = ratectl->d0i;
                avg_K = ratectl->avg_KI;
                Si = (ratectl->Xi + 3.0*avg_K*ratectl->actsum)/4.0;
                ratectl->T = ratectl->R/(1.0+ratectl->Np*ratectl->Xp*ratectl->Ki/(Si*ratectl->Kp)+ratectl->Nb*ratectl->Xb*ratectl->Ki/(Si*ratectl->Kb));

                break;
        case P_TYPE:
                ratectl->d = ratectl->d0pb;
                avg_K = ratectl->avg_KP;
                Sp = (ratectl->Xp + avg_K*ratectl->actsum) / 2.0;
                ratectl->T =  ratectl->R/(ratectl->Np+ratectl->Nb*ratectl->Kp*ratectl->Xb/(ratectl->Kb*Sp)) + 0.5;
                break;
        case B_TYPE:
                ratectl->d = ratectl->d0pb;            // I and P frame share ratectl virtual buffer
                avg_K = ratectl->avg_KB;
                Sb = ratectl->Xb /* + avg_K * ratectl->actsum) / 2.0 */;
                ratectl->T =  ratectl->R/(ratectl->Nb+ratectl->Np*ratectl->Kb*ratectl->Xp/(ratectl->Kp*Sb));
                break;
        }

        /* Undershot bits have been "returned" via R */
        if( ratectl->d < 0 )
                ratectl->d = 0;

        /* We don't let the target volume get absurdly low as it makes some
           of the prediction maths ill-condtioned.  At these levels quantisation
           is always minimum anyway
        */
        if( ratectl->T < 4000.0 )
        {
                ratectl->T = 4000.0;
        }
        target_Q = scale_quant(picture, 
                                                   avg_K * ratectl->avg_act *(mb_per_pict) / ratectl->T);
        current_Q = scale_quant(picture,62.0*ratectl->d / ratectl->r);
#ifdef DEBUG
        if( !quiet )
        {
                /* printf( "AA=%3.4f T=%6.0f K=%.1f ",avg_act, (double)T, avg_K  ); */
                printf( "AA=%3.4f SA==%3.4f ",avg_act, sum_avg_act  ); 
        }
#endif  
        
        if ( current_Q < 3 && target_Q > 12 )
        {
                /* We're undershooting and a serious surge in the data_flow
                   due to lagging adjustment is possible...
                */
                ratectl->d = (int) (target_Q * ratectl->r / 62.0);
        }

        ratectl->S = bitcount();
        ratectl->frame_start = bitcount();
//      ratectl->current_quant = ratectl->d * 62.0 / ratectl->r;
        if(ratectl->current_quant < 1) ratectl->current_quant = 1;
        if(ratectl->current_quant > 100) ratectl->current_quant = 100;
}

/* compute initial quantization stepsize (at the beginning of picture) */
int ratectl_start_mb(ratectl_t *ratectl, pict_data_s *picture)
{
        double Qj;
        int mquant;
        
        if(fixed_mquant) 
                Qj = fixed_mquant;
        else
                Qj = ratectl->current_quant;
//              Qj = ratectl->d * 62.0 / ratectl->r;

        mquant = scale_quant( picture, Qj);
        mquant = intmax(mquant, quant_floor);

        return mquant;
}

void ratectl_update_pict(ratectl_t *ratectl, pict_data_s *picture)
{
        double X;
        double K;
        int64_t AP,PP;          /* Actual and padded picture bit counts */
        int    i;
        int    Qsum;
        int frame_overshoot;
        double avg_bitrate;
        int last_size;
        double new_weight;
        double old_weight;
        
        if(fixed_mquant) return;
        
        AP = bitcount() - ratectl->S;
        frame_overshoot = (int)AP-(int)ratectl->T;

        /* For the virtual buffers for quantisation feedback it is the
           actual under/overshoot that counts, not what's left after padding
        */
        ratectl->d += frame_overshoot;
        
        /* If the cummulative undershoot is getting too large (as
           a rough and ready heuristic we use 1/2 video buffer size)
           we start padding the stream.  Or, in the case of VBR,
           we pretend we're padding but don't actually write anything!

         */

        if( ratectl->gop_undershoot-frame_overshoot > video_buffer_size/2 )
        {
                int padding_bytes = 
                        ((ratectl->gop_undershoot - frame_overshoot) - video_buffer_size/2)/8;
                if( quant_floor != 0 )  /* VBR case pretend to pad */
                {
                        PP = AP + padding_bytes;
                }
                else
                {
//                      printf( "PAD" );
//                      alignbits();
                        for( i = 0; i < padding_bytes/2; ++i )
                        {
//                              putbits(0, 16);
                        }
                        PP = bitcount() - ratectl->S;                   /* total # of bits in picture */
                }
                frame_overshoot = (int)PP - (int)ratectl->T;
        }
        else
                PP = AP;
        
        /* Estimate cummulative undershoot within this gop. 
           This is only an estimate because T includes an allocation
           from earlier undershoots causing multiple counting.
           Padding and an exact calculation each gop prevent the error
           in the estimate growing too excessive...
           */
        ratectl->gop_undershoot -= frame_overshoot;
        ratectl->gop_undershoot = ratectl->gop_undershoot > 0 ? ratectl->gop_undershoot : 0;
        ratectl->R -= PP;                                               /* remaining # of bits in GOP */

        Qsum = 0;
        for( i = 0; i < mb_per_pict; ++i )
        {
                Qsum += picture->mbinfo[i].mquant;
        }


        ratectl->AQ = (double)Qsum/(double)mb_per_pict;
        /* TODO: The X are used as relative activity measures... so why
           bother dividing by 2?  
           Note we have to be careful to measure the actual data not the
           padding too!
        */
        ratectl->SQ += ratectl->AQ;
        X = (double)AP*(ratectl->AQ/2.0);
        
        K = X / ratectl->actsum;
#ifdef DEBUG
        if( !quiet )
        {
                printf( "AQ=%.1f SQ=%.2f",  AQ,SQ);
        }
#endif
        /* Bits that never got used in the past can't be resurrected
           now...  We use an average of past (positive) virtual buffer
           fullness in the event of an under-shoot as otherwise we will
           tend to over-shoot heavily when activity picks up.
         
           TODO: We should really use our estimate K[IBP] of
           bit_usage*activity / quantisation ratio to set a "sensible"
           initial d to achieve a reasonable initial quantisation. Rather
           than have to cut in a huge (lagging correction).

           Alternatively, simply requantising with mean buffer if there is
           a big buffer swing would work nicely...

        */
        
        /* EXPERIMENT: Xi are used as a guesstimate of likely *future* frame
           activities based on the past.  Thus we don't want anomalous outliers
           due to scene changes swinging things too much.  Introduce moving averages
           for the Xi...
           TODO: The averaging constants should be adjust to suit relative frame
           frequencies...
        */
        switch (picture->pict_type)
        {
        case I_TYPE:
                ratectl->avg_KI = (K + ratectl->avg_KI * K_AVG_WINDOW_I) / (K_AVG_WINDOW_I+1.0) ;
                ratectl->d0i = ratectl->d;
                ratectl->Xi = (X + 3.0 * ratectl->Xi) / 4.0;
                break;
        case P_TYPE:
                ratectl->avg_KP = (K + ratectl->avg_KP * K_AVG_WINDOW_P) / (K_AVG_WINDOW_P+1.0) ;
                ratectl->d0pb = ratectl->d;
                ratectl->Xp = (X + ratectl->Xp * 12.0) / 13.0;
                ratectl->Np--;
                break;
        case B_TYPE:
                ratectl->avg_KB = (K + ratectl->avg_KB * K_AVG_WINDOW_B) / (K_AVG_WINDOW_B + 1.0) ;
                ratectl->d0pb = ratectl->d;
                ratectl->Xb = (X + ratectl->Xb * 24.0) / 25.0;
                ratectl->Nb--;
                break;
        }

        ratectl->frame_end = bitcount();

        last_size = ratectl->frame_end - ratectl->frame_start;
        avg_bitrate = (double)last_size * frame_rate;
        switch(picture->pict_type)
        {
                case I_TYPE:
                        new_weight = avg_bitrate / bit_rate * 1 / N;
                        old_weight = (double)(N - 1) / N;
                        break;

                default:
                case P_TYPE:
                        new_weight = avg_bitrate / bit_rate * (N - 1) / N;
                        old_weight = (double)1 / N;
                        break;
        }
        ratectl->current_quant *= (old_weight + new_weight);

/*
 * printf("ratectl_update_pict %f %f\n", 
 *      avg_bitrate,
 *      ratectl->current_quant);
 */

}

/* Step 2: measure virtual buffer - estimated buffer discrepancy */
int ratectl_calc_mquant(ratectl_t *ratectl, pict_data_s *picture, int j)
{
        int mquant;
        double dj, Qj, actj, N_actj; 

//      pthread_mutex_lock(&(ratectl->ratectl_lock));
        /* A.Stevens 2000 : we measure how much *information* (total activity)
           has been covered and aim to release bits in proportion.  Indeed,
           complex blocks get an disproprortionate boost of allocated bits.
           To avoid visible "ringing" effects...

        */

        actj = picture->mbinfo[j].act;
        /* Guesstimate a virtual buffer fullness based on
           bits used vs. bits in proportion to activity encoded
        */


        dj = ((double)ratectl->d) + 
                ((double)(bitcount() - ratectl->S) - ratectl->actcovered * ((double)ratectl->T) / ratectl->actsum);



        /* scale against dynamic range of mquant and the bits/picture count.
           quant_floor != 0.0 is the VBR case where we set a bitrate as a (high)
           maximum and then put a floor on quantisation to achieve a reasonable
           overall size.
           Not that this *is* baseline quantisation.  Not adjust for local activity.
           Otherwise we end up blurring active macroblocks. Silly in a VBR context.
        */

        Qj = dj * 62.0 / ratectl->r;

//printf("ratectl_calc_mquant %f\n", Qj);
        if(fixed_mquant)
                Qj = fixed_mquant;
        else
                Qj = ratectl->current_quant;

        Qj = (Qj > quant_floor) ? Qj : quant_floor;
        /*  Heuristic: Decrease quantisation for blocks with lots of
                sizeable coefficients.  We assume we just get a mess if
                a complex texture's coefficients get chopped...
        */
                
        N_actj =  actj < ratectl->avg_act ? 
                1.0 : 
                (actj + act_boost * ratectl->avg_act) /
                (act_boost * actj + ratectl->avg_act);
   
        mquant = scale_quant(picture, Qj * N_actj);


        /* Update activity covered */

        ratectl->actcovered += actj;
//      pthread_mutex_unlock(&(ratectl->ratectl_lock));

        return mquant;
}

/* VBV calculations
 *
 * generates warnings if underflow or overflow occurs
 */

/* vbv_end_of_picture
 *
 * - has to be called directly after writing picture_data()
 * - needed for accurate VBV buffer overflow calculation
 * - assumes there is no byte stuffing prior to the next start code
 */

void vbv_end_of_picture()
{
}

/* calc_vbv_delay
 *
 * has to be called directly after writing the picture start code, the
 * reference point for vbv_delay
 */

void calc_vbv_delay()
{
}

void stop_ratectl(ratectl_t *ratectl)
{
        pthread_mutex_destroy(&(ratectl->ratectl_lock));
}

---