comparison h264.c @ 5226:65bffcc5571a libavcodec

Precompute a chroma_qp table with index offset for each pps, this saves speed for the upcoming secondqp fix. Patch by Andreas �ªÓman % andreas A olebyn P nu % Original thread: Date: Jun 26, 2007 8:48 PM subject: [FFmpeg-devel] Color corruption and seeking errors with H264 disc sources

5225:099c495618c1

#endif

/**
 * gets the chroma qp.
 */
static inline int get_chroma_qp(int chroma_qp_index_offset, int qscale){

    return chroma_qp[av_clip(qscale + chroma_qp_index_offset, 0, 51)];
}

//FIXME need to check that this does not overflow signed 32 bit for low qp, i am not sure, it's very close
//FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){

            // deblock a pair
            // top
            s->mb_y--;
            tprintf(h->s.avctx, "call mbaff filter_mb mb_x:%d mb_y:%d pair_dest_y = %p, dest_y = %p\n", mb_x, mb_y, pair_dest_y, dest_y);
            fill_caches(h, mb_type_top, 1); //FIXME don't fill stuff which isn't used by filter_mb
            h->chroma_qp = get_chroma_qp(h->pps.chroma_qp_index_offset, s->current_picture.qscale_table[mb_xy]);
            filter_mb(h, mb_x, mb_y, pair_dest_y, pair_dest_cb, pair_dest_cr, linesize, uvlinesize);
            // bottom
            s->mb_y++;
            tprintf(h->s.avctx, "call mbaff filter_mb\n");
            fill_caches(h, mb_type_bottom, 1); //FIXME don't fill stuff which isn't used by filter_mb
            h->chroma_qp = get_chroma_qp(h->pps.chroma_qp_index_offset, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
            filter_mb(h, mb_x, mb_y+1, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
        } else {
            tprintf(h->s.avctx, "call filter_mb\n");
            backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
            fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb

    if(tmp>51){
        av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
        return -1;
    }
    s->qscale= tmp;
    h->chroma_qp = get_chroma_qp(h->pps.chroma_qp_index_offset, s->qscale);
    //FIXME qscale / qp ... stuff
    if(h->slice_type == SP_TYPE){
        get_bits1(&s->gb); /* sp_for_switch_flag */
    }
    if(h->slice_type==SP_TYPE || h->slice_type == SI_TYPE){

            }
        }

        // In deblocking, the quantizer is 0
        s->current_picture.qscale_table[mb_xy]= 0;
        h->chroma_qp = get_chroma_qp(h->pps.chroma_qp_index_offset, 0);
        // All coeffs are present
        memset(h->non_zero_count[mb_xy], 16, 16);

        s->current_picture.mb_type[mb_xy]= mb_type;
        return 0;

        if(((unsigned)s->qscale) > 51){
            if(s->qscale<0) s->qscale+= 52;
            else            s->qscale-= 52;
        }

        h->chroma_qp= chroma_qp= get_chroma_qp(h->pps.chroma_qp_index_offset, s->qscale);
        if(IS_INTRA16x16(mb_type)){
            if( decode_residual(h, h->intra_gb_ptr, h->mb, LUMA_DC_BLOCK_INDEX, dc_scan, h->dequant4_coeff[0][s->qscale], 16) < 0){
                return -1; //FIXME continue if partitioned and other return -1 too
            }

        // All blocks are present
        h->cbp_table[mb_xy] = 0x1ef;
        h->chroma_pred_mode_table[mb_xy] = 0;
        // In deblocking, the quantizer is 0
        s->current_picture.qscale_table[mb_xy]= 0;
        h->chroma_qp = get_chroma_qp(h->pps.chroma_qp_index_offset, 0);
        // All coeffs are present
        memset(h->non_zero_count[mb_xy], 16, 16);
        s->current_picture.mb_type[mb_xy]= mb_type;
        return 0;
    }

        s->qscale += dqp;
        if(((unsigned)s->qscale) > 51){
            if(s->qscale<0) s->qscale+= 52;
            else            s->qscale-= 52;
        }
        h->chroma_qp = get_chroma_qp(h->pps.chroma_qp_index_offset, s->qscale);

        if( IS_INTRA16x16( mb_type ) ) {
            int i;
            //av_log( s->avctx, AV_LOG_ERROR, "INTRA16x16 DC\n" );
            if( decode_cabac_residual( h, h->mb, 0, 0, dc_scan, NULL, 16) < 0)

    mb_type = s->current_picture.mb_type[mb_xy];
    qp = s->current_picture.qscale_table[mb_xy];
    qp0 = s->current_picture.qscale_table[mb_xy-1];
    qp1 = s->current_picture.qscale_table[h->top_mb_xy];
    qpc = get_chroma_qp( h->pps.chroma_qp_index_offset, qp );
    qpc0 = get_chroma_qp( h->pps.chroma_qp_index_offset, qp0 );
    qpc1 = get_chroma_qp( h->pps.chroma_qp_index_offset, qp1 );
    qp0 = (qp + qp0 + 1) >> 1;
    qp1 = (qp + qp1 + 1) >> 1;
    qpc0 = (qpc + qpc0 + 1) >> 1;
    qpc1 = (qpc + qpc1 + 1) >> 1;
    qp_thresh = 15 - h->slice_alpha_c0_offset;

        mb_qp = s->current_picture.qscale_table[mb_xy];
        mbn0_qp = s->current_picture.qscale_table[left_mb_xy[0]];
        mbn1_qp = s->current_picture.qscale_table[left_mb_xy[1]];
        qp[0] = ( mb_qp + mbn0_qp + 1 ) >> 1;
        chroma_qp[0] = ( get_chroma_qp( h->pps.chroma_qp_index_offset, mb_qp ) +
                         get_chroma_qp( h->pps.chroma_qp_index_offset, mbn0_qp ) + 1 ) >> 1;
        qp[1] = ( mb_qp + mbn1_qp + 1 ) >> 1;
        chroma_qp[1] = ( get_chroma_qp( h->pps.chroma_qp_index_offset, mb_qp ) +
                         get_chroma_qp( h->pps.chroma_qp_index_offset, mbn1_qp ) + 1 ) >> 1;

        /* Filter edge */
        tprintf(s->avctx, "filter mb:%d/%d MBAFF, QPy:%d/%d, QPc:%d/%d ls:%d uvls:%d", mb_x, mb_y, qp[0], qp[1], chroma_qp[0], chroma_qp[1], linesize, uvlinesize);
        { int i; for (i = 0; i < 8; i++) tprintf(s->avctx, " bS[%d]:%d", i, bS[i]); tprintf(s->avctx, "\n"); }
        filter_mb_mbaff_edgev ( h, &img_y [0], linesize,   bS, qp );

                qp = ( s->current_picture.qscale_table[mb_xy] + s->current_picture.qscale_table[mbn_xy] + 1 ) >> 1;
                tprintf(s->avctx, "filter mb:%d/%d dir:%d edge:%d, QPy:%d ls:%d uvls:%d", mb_x, mb_y, dir, edge, qp, tmp_linesize, tmp_uvlinesize);
                { int i; for (i = 0; i < 4; i++) tprintf(s->avctx, " bS[%d]:%d", i, bS[i]); tprintf(s->avctx, "\n"); }
                filter_mb_edgeh( h, &img_y[j*linesize], tmp_linesize, bS, qp );
                chroma_qp = ( h->chroma_qp +
                              get_chroma_qp( h->pps.chroma_qp_index_offset, s->current_picture.qscale_table[mbn_xy] ) + 1 ) >> 1;
                filter_mb_edgech( h, &img_cb[j*uvlinesize], tmp_uvlinesize, bS, chroma_qp );
                filter_mb_edgech( h, &img_cr[j*uvlinesize], tmp_uvlinesize, bS, chroma_qp );
            }

            start = 1;

            { int i; for (i = 0; i < 4; i++) tprintf(s->avctx, " bS[%d]:%d", i, bS[i]); tprintf(s->avctx, "\n"); }
            if( dir == 0 ) {
                filter_mb_edgev( h, &img_y[4*edge], linesize, bS, qp );
                if( (edge&1) == 0 ) {
                    int chroma_qp = ( h->chroma_qp +
                                      get_chroma_qp( h->pps.chroma_qp_index_offset, s->current_picture.qscale_table[mbn_xy] ) + 1 ) >> 1;
                    filter_mb_edgecv( h, &img_cb[2*edge], uvlinesize, bS, chroma_qp );
                    filter_mb_edgecv( h, &img_cr[2*edge], uvlinesize, bS, chroma_qp );
                }
            } else {
                filter_mb_edgeh( h, &img_y[4*edge*linesize], linesize, bS, qp );
                if( (edge&1) == 0 ) {
                    int chroma_qp = ( h->chroma_qp +
                                      get_chroma_qp( h->pps.chroma_qp_index_offset, s->current_picture.qscale_table[mbn_xy] ) + 1 ) >> 1;
                    filter_mb_edgech( h, &img_cb[2*edge*uvlinesize], uvlinesize, bS, chroma_qp );
                    filter_mb_edgech( h, &img_cr[2*edge*uvlinesize], uvlinesize, bS, chroma_qp );
                }
            }
        }

               sps->crop_top, sps->crop_bottom,
               sps->vui_parameters_present_flag ? "VUI" : ""
               );
    }
    return 0;
}

static inline int decode_picture_parameter_set(H264Context *h, int bit_length){
    MpegEncContext * const s = &h->s;
    unsigned int tmp, pps_id= get_ue_golomb(&s->gb);

    pps->weighted_pred= get_bits1(&s->gb);
    pps->weighted_bipred_idc= get_bits(&s->gb, 2);
    pps->init_qp= get_se_golomb(&s->gb) + 26;
    pps->init_qs= get_se_golomb(&s->gb) + 26;
    pps->chroma_qp_index_offset= get_se_golomb(&s->gb);
    pps->deblocking_filter_parameters_present= get_bits1(&s->gb);
    pps->constrained_intra_pred= get_bits1(&s->gb);
    pps->redundant_pic_cnt_present = get_bits1(&s->gb);

    pps->transform_8x8_mode= 0;

5226:65bffcc5571a

#endif

/**
 * gets the chroma qp.
 */
static inline int get_chroma_qp(H264Context *h, int qscale){
    return h->pps.chroma_qp_table[qscale & 0xff];
}

//FIXME need to check that this does not overflow signed 32 bit for low qp, i am not sure, it's very close
//FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){

            // deblock a pair
            // top
            s->mb_y--;
            tprintf(h->s.avctx, "call mbaff filter_mb mb_x:%d mb_y:%d pair_dest_y = %p, dest_y = %p\n", mb_x, mb_y, pair_dest_y, dest_y);
            fill_caches(h, mb_type_top, 1); //FIXME don't fill stuff which isn't used by filter_mb
            h->chroma_qp = get_chroma_qp(h, s->current_picture.qscale_table[mb_xy]);
            filter_mb(h, mb_x, mb_y, pair_dest_y, pair_dest_cb, pair_dest_cr, linesize, uvlinesize);
            // bottom
            s->mb_y++;
            tprintf(h->s.avctx, "call mbaff filter_mb\n");
            fill_caches(h, mb_type_bottom, 1); //FIXME don't fill stuff which isn't used by filter_mb
            h->chroma_qp = get_chroma_qp(h, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
            filter_mb(h, mb_x, mb_y+1, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
        } else {
            tprintf(h->s.avctx, "call filter_mb\n");
            backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
            fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb

    if(tmp>51){
        av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
        return -1;
    }
    s->qscale= tmp;
    h->chroma_qp = get_chroma_qp(h, s->qscale);
    //FIXME qscale / qp ... stuff
    if(h->slice_type == SP_TYPE){
        get_bits1(&s->gb); /* sp_for_switch_flag */
    }
    if(h->slice_type==SP_TYPE || h->slice_type == SI_TYPE){

            }
        }

        // In deblocking, the quantizer is 0
        s->current_picture.qscale_table[mb_xy]= 0;
        h->chroma_qp = get_chroma_qp(h, 0);
        // All coeffs are present
        memset(h->non_zero_count[mb_xy], 16, 16);

        s->current_picture.mb_type[mb_xy]= mb_type;
        return 0;

        if(((unsigned)s->qscale) > 51){
            if(s->qscale<0) s->qscale+= 52;
            else            s->qscale-= 52;
        }

        h->chroma_qp= chroma_qp= get_chroma_qp(h, s->qscale);
        if(IS_INTRA16x16(mb_type)){
            if( decode_residual(h, h->intra_gb_ptr, h->mb, LUMA_DC_BLOCK_INDEX, dc_scan, h->dequant4_coeff[0][s->qscale], 16) < 0){
                return -1; //FIXME continue if partitioned and other return -1 too
            }

        // All blocks are present
        h->cbp_table[mb_xy] = 0x1ef;
        h->chroma_pred_mode_table[mb_xy] = 0;
        // In deblocking, the quantizer is 0
        s->current_picture.qscale_table[mb_xy]= 0;
        h->chroma_qp = get_chroma_qp(h, 0);
        // All coeffs are present
        memset(h->non_zero_count[mb_xy], 16, 16);
        s->current_picture.mb_type[mb_xy]= mb_type;
        return 0;
    }

        s->qscale += dqp;
        if(((unsigned)s->qscale) > 51){
            if(s->qscale<0) s->qscale+= 52;
            else            s->qscale-= 52;
        }
        h->chroma_qp = get_chroma_qp(h, s->qscale);

        if( IS_INTRA16x16( mb_type ) ) {
            int i;
            //av_log( s->avctx, AV_LOG_ERROR, "INTRA16x16 DC\n" );
            if( decode_cabac_residual( h, h->mb, 0, 0, dc_scan, NULL, 16) < 0)

    mb_type = s->current_picture.mb_type[mb_xy];
    qp = s->current_picture.qscale_table[mb_xy];
    qp0 = s->current_picture.qscale_table[mb_xy-1];
    qp1 = s->current_picture.qscale_table[h->top_mb_xy];
    qpc = get_chroma_qp( h, qp );
    qpc0 = get_chroma_qp( h, qp0 );
    qpc1 = get_chroma_qp( h, qp1 );
    qp0 = (qp + qp0 + 1) >> 1;
    qp1 = (qp + qp1 + 1) >> 1;
    qpc0 = (qpc + qpc0 + 1) >> 1;
    qpc1 = (qpc + qpc1 + 1) >> 1;
    qp_thresh = 15 - h->slice_alpha_c0_offset;

        mb_qp = s->current_picture.qscale_table[mb_xy];
        mbn0_qp = s->current_picture.qscale_table[left_mb_xy[0]];
        mbn1_qp = s->current_picture.qscale_table[left_mb_xy[1]];
        qp[0] = ( mb_qp + mbn0_qp + 1 ) >> 1;
        chroma_qp[0] = ( get_chroma_qp( h, mb_qp ) +
                         get_chroma_qp( h, mbn0_qp ) + 1 ) >> 1;
        qp[1] = ( mb_qp + mbn1_qp + 1 ) >> 1;
        chroma_qp[1] = ( get_chroma_qp( h, mb_qp ) +
                         get_chroma_qp( h, mbn1_qp ) + 1 ) >> 1;

        /* Filter edge */
        tprintf(s->avctx, "filter mb:%d/%d MBAFF, QPy:%d/%d, QPc:%d/%d ls:%d uvls:%d", mb_x, mb_y, qp[0], qp[1], chroma_qp[0], chroma_qp[1], linesize, uvlinesize);
        { int i; for (i = 0; i < 8; i++) tprintf(s->avctx, " bS[%d]:%d", i, bS[i]); tprintf(s->avctx, "\n"); }
        filter_mb_mbaff_edgev ( h, &img_y [0], linesize,   bS, qp );

                qp = ( s->current_picture.qscale_table[mb_xy] + s->current_picture.qscale_table[mbn_xy] + 1 ) >> 1;
                tprintf(s->avctx, "filter mb:%d/%d dir:%d edge:%d, QPy:%d ls:%d uvls:%d", mb_x, mb_y, dir, edge, qp, tmp_linesize, tmp_uvlinesize);
                { int i; for (i = 0; i < 4; i++) tprintf(s->avctx, " bS[%d]:%d", i, bS[i]); tprintf(s->avctx, "\n"); }
                filter_mb_edgeh( h, &img_y[j*linesize], tmp_linesize, bS, qp );
                chroma_qp = ( h->chroma_qp +
                              get_chroma_qp( h, s->current_picture.qscale_table[mbn_xy] ) + 1 ) >> 1;
                filter_mb_edgech( h, &img_cb[j*uvlinesize], tmp_uvlinesize, bS, chroma_qp );
                filter_mb_edgech( h, &img_cr[j*uvlinesize], tmp_uvlinesize, bS, chroma_qp );
            }

            start = 1;

            { int i; for (i = 0; i < 4; i++) tprintf(s->avctx, " bS[%d]:%d", i, bS[i]); tprintf(s->avctx, "\n"); }
            if( dir == 0 ) {
                filter_mb_edgev( h, &img_y[4*edge], linesize, bS, qp );
                if( (edge&1) == 0 ) {
                    int chroma_qp = ( h->chroma_qp +
                                      get_chroma_qp( h, s->current_picture.qscale_table[mbn_xy] ) + 1 ) >> 1;
                    filter_mb_edgecv( h, &img_cb[2*edge], uvlinesize, bS, chroma_qp );
                    filter_mb_edgecv( h, &img_cr[2*edge], uvlinesize, bS, chroma_qp );
                }
            } else {
                filter_mb_edgeh( h, &img_y[4*edge*linesize], linesize, bS, qp );
                if( (edge&1) == 0 ) {
                    int chroma_qp = ( h->chroma_qp +
                                      get_chroma_qp( h, s->current_picture.qscale_table[mbn_xy] ) + 1 ) >> 1;
                    filter_mb_edgech( h, &img_cb[2*edge*uvlinesize], uvlinesize, bS, chroma_qp );
                    filter_mb_edgech( h, &img_cr[2*edge*uvlinesize], uvlinesize, bS, chroma_qp );
                }
            }
        }

               sps->crop_top, sps->crop_bottom,
               sps->vui_parameters_present_flag ? "VUI" : ""
               );
    }
    return 0;
}

static void
build_qp_table(PPS *pps, int index)
{
    int i;
    for(i = 0; i < 255; i++)
        pps->chroma_qp_table[i & 0xff] = chroma_qp[av_clip(i + index, 0, 51)];
    pps->chroma_qp_index_offset = index;
}

static inline int decode_picture_parameter_set(H264Context *h, int bit_length){
    MpegEncContext * const s = &h->s;
    unsigned int tmp, pps_id= get_ue_golomb(&s->gb);

    pps->weighted_pred= get_bits1(&s->gb);
    pps->weighted_bipred_idc= get_bits(&s->gb, 2);
    pps->init_qp= get_se_golomb(&s->gb) + 26;
    pps->init_qs= get_se_golomb(&s->gb) + 26;
    build_qp_table(pps, get_se_golomb(&s->gb));
    pps->deblocking_filter_parameters_present= get_bits1(&s->gb);
    pps->constrained_intra_pred= get_bits1(&s->gb);
    pps->redundant_pic_cnt_present = get_bits1(&s->gb);

    pps->transform_8x8_mode= 0;