comparison mpegvideo.c @ 1784:65f7bd09f37b libavcodec

quantizer noise shaping optimization

1783:66ae3c109d90

    }

    return last_non_zero;
}

#define BASIS_SHIFT 16
#define RECON_SHIFT 6
//#define REFINE_STATS 1
static int16_t basis[64][64];

static void build_basis(uint8_t *perm){
    int i, j, x, y;

                    basis[perm_index][8*x + y]= lrintf(s * cos((M_PI/8.0)*i*(x+0.5)) * cos((M_PI/8.0)*j*(y+0.5)));
                }
            }
        }
    }
}

static int try_basis(int16_t rem[64], int16_t weight[64], int index, int scale){
    int i;
    unsigned int sum=0;

    for(i=0; i<8*8; i++){
        int b= rem[i] - ((basis[index][i]*scale + (1<<(BASIS_SHIFT - RECON_SHIFT-1)))>>(BASIS_SHIFT - RECON_SHIFT));
        int w= weight[i];
        b= (b + (1<<(RECON_SHIFT-1))) >> RECON_SHIFT;
        assert(-512<b && b<512);

        sum += (w*b)*(w*b)>>4;
    }
    return sum>>2;
}

static void add_basis(int16_t rem[64], int index, int scale){
    int i;

    for(i=0; i<8*8; i++){
        rem[i] -= (basis[index][i]*scale + (1<<(BASIS_SHIFT - RECON_SHIFT-1)))>>(BASIS_SHIFT - RECON_SHIFT);
    }    
}

static int dct_quantize_refine(MpegEncContext *s, //FIXME breaks denoise?
                        DCTELEM *block, int16_t *weight, DCTELEM *orig,
                        int n, int qscale){

    last_non_zero = s->block_last_index[n];

#ifdef REFINE_STATS
{START_TIMER
#endif
    for(i=0; i<64; i++){ //FIXME memsetw or similar
        rem[i]= (orig[i]<<RECON_SHIFT) - dc; //FIXME  use orig dirrectly insteadof copying to rem[]
    }
#ifdef REFINE_STATS
STOP_TIMER("memset rem[]")}
#endif
    sum=0;

            if(level<0) coeff= qmul*level - qadd;
            else        coeff= qmul*level + qadd;
            run_tab[rle_index++]=run;
            run=0;

            add_basis(rem, j, coeff);
        }else{
            run++;
        }
    }
#ifdef REFINE_STATS

}

{START_TIMER
#endif
    for(;;){
        int best_score=try_basis(rem, weight, 0, 0);
        int nochange_score= best_score;
        int best_coeff=0;
        int best_change=0;
        int run2, best_unquant_change;
#ifdef REFINE_STATS

                
                new_coeff= q*new_level;
                if(new_coeff >= 2048 || new_coeff < 0)
                    continue;

                score= try_basis(rem, weight, 0, new_coeff - old_coeff);
                if(score<best_score){
                    best_score= score;
                    best_coeff= 0;
                    best_change= change;
                    best_unquant_change= new_coeff - old_coeff;

                score *= lambda;

                unquant_change= new_coeff - old_coeff;
                assert((score < 100*lambda && score > -100*lambda) || lambda==0);
                
                score+= try_basis(rem, weight, j, unquant_change);
                if(score<best_score){
                    best_score= score;
                    best_coeff= i;
                    best_change= change;
                    best_unquant_change= unquant_change;

                 }else{
                     run++;
                 }
            }
            
            add_basis(rem, j, best_unquant_change);
        }else{
            break;
        }
    }
#ifdef REFINE_STATS

1784:65f7bd09f37b

    }

    return last_non_zero;
}

//#define REFINE_STATS 1
static int16_t basis[64][64];

static void build_basis(uint8_t *perm){
    int i, j, x, y;

                    basis[perm_index][8*x + y]= lrintf(s * cos((M_PI/8.0)*i*(x+0.5)) * cos((M_PI/8.0)*j*(y+0.5)));
                }
            }
        }
    }
}

static int dct_quantize_refine(MpegEncContext *s, //FIXME breaks denoise?
                        DCTELEM *block, int16_t *weight, DCTELEM *orig,
                        int n, int qscale){

    last_non_zero = s->block_last_index[n];

#ifdef REFINE_STATS
{START_TIMER
#endif
    dc += (1<<(RECON_SHIFT-1));
    for(i=0; i<64; i++){
        rem[i]= dc - (orig[i]<<RECON_SHIFT); //FIXME  use orig dirrectly insteadof copying to rem[]
    }
#ifdef REFINE_STATS
STOP_TIMER("memset rem[]")}
#endif
    sum=0;

            if(level<0) coeff= qmul*level - qadd;
            else        coeff= qmul*level + qadd;
            run_tab[rle_index++]=run;
            run=0;

            s->dsp.add_8x8basis(rem, basis[j], coeff);
        }else{
            run++;
        }
    }
#ifdef REFINE_STATS

}

{START_TIMER
#endif
    for(;;){
        int best_score=s->dsp.try_8x8basis(rem, weight, basis[0], 0);
        int nochange_score= best_score;
        int best_coeff=0;
        int best_change=0;
        int run2, best_unquant_change;
#ifdef REFINE_STATS

                
                new_coeff= q*new_level;
                if(new_coeff >= 2048 || new_coeff < 0)
                    continue;

                score= s->dsp.try_8x8basis(rem, weight, basis[0], new_coeff - old_coeff);
                if(score<best_score){
                    best_score= score;
                    best_coeff= 0;
                    best_change= change;
                    best_unquant_change= new_coeff - old_coeff;

                score *= lambda;

                unquant_change= new_coeff - old_coeff;
                assert((score < 100*lambda && score > -100*lambda) || lambda==0);
                
                score+= s->dsp.try_8x8basis(rem, weight, basis[j], unquant_change);
                if(score<best_score){
                    best_score= score;
                    best_coeff= i;
                    best_change= change;
                    best_unquant_change= unquant_change;

                 }else{
                     run++;
                 }
            }
            
            s->dsp.add_8x8basis(rem, basis[j], best_unquant_change);
        }else{
            break;
        }
    }
#ifdef REFINE_STATS