macroblock.c

H.264 source codes

                    break;
                case D_L0_8x4:
                    x264_mb_mc_0xywh( h, x, y+0, 2, 1 );
                    x264_mb_mc_0xywh( h, x, y+1, 2, 1 );
                    break;
                case D_L0_4x8:
                    x264_mb_mc_0xywh( h, x+0, y, 1, 2 );
                    x264_mb_mc_0xywh( h, x+1, y, 1, 2 );
                    break;
                case D_L0_4x4:
                    x264_mb_mc_0xywh( h, x+0, y+0, 1, 1 );
                    x264_mb_mc_0xywh( h, x+1, y+0, 1, 1 );
                    x264_mb_mc_0xywh( h, x+0, y+1, 1, 1 );
                    x264_mb_mc_0xywh( h, x+1, y+1, 1, 1 );
                    break;
                case D_L1_8x8:
                    x264_mb_mc_1xywh( h, x, y, 2, 2 );
                    break;
                case D_L1_8x4:
                    x264_mb_mc_1xywh( h, x, y+0, 2, 1 );
                    x264_mb_mc_1xywh( h, x, y+1, 2, 1 );
                    break;
                case D_L1_4x8:
                    x264_mb_mc_1xywh( h, x+0, y, 1, 2 );
                    x264_mb_mc_1xywh( h, x+1, y, 1, 2 );
                    break;
                case D_L1_4x4:
                    x264_mb_mc_1xywh( h, x+0, y+0, 1, 1 );
                    x264_mb_mc_1xywh( h, x+1, y+0, 1, 1 );
                    x264_mb_mc_1xywh( h, x+0, y+1, 1, 1 );
                    x264_mb_mc_1xywh( h, x+1, y+1, 1, 1 );
                    break;
                case D_BI_8x8:
                    x264_mb_mc_01xywh( h, x, y, 2, 2 );
                    break;
                case D_BI_8x4:
                    x264_mb_mc_01xywh( h, x, y+0, 2, 1 );
                    x264_mb_mc_01xywh( h, x, y+1, 2, 1 );
                    break;
                case D_BI_4x8:
                    x264_mb_mc_01xywh( h, x+0, y, 1, 2 );
                    x264_mb_mc_01xywh( h, x+1, y, 1, 2 );
                    break;
                case D_BI_4x4:
                    x264_mb_mc_01xywh( h, x+0, y+0, 1, 1 );
                    x264_mb_mc_01xywh( h, x+1, y+0, 1, 1 );
                    x264_mb_mc_01xywh( h, x+0, y+1, 1, 1 );
                    x264_mb_mc_01xywh( h, x+1, y+1, 1, 1 );
                    break;
                case D_DIRECT_8x8:
                    x264_mb_mc_direct8x8( h, x, y );
                    break;
            }
        }
    }
    else if( h->mb.i_type == B_SKIP || h->mb.i_type == B_DIRECT )
    {
        x264_mb_mc_direct8x8( h, 0, 0 );
        x264_mb_mc_direct8x8( h, 2, 0 );
        x264_mb_mc_direct8x8( h, 0, 2 );
        x264_mb_mc_direct8x8( h, 2, 2 );
    }
    else    /* B_*x* */
    {
        int b_list0[2];
        int b_list1[2];

        int i;

        /* init ref list utilisations */
        for( i = 0; i < 2; i++ )
        {
            b_list0[i] = x264_mb_type_list0_table[h->mb.i_type][i];
            b_list1[i] = x264_mb_type_list1_table[h->mb.i_type][i];
        }
        if( h->mb.i_partition == D_16x16 )
        {
            if( b_list0[0] && b_list1[0] ) x264_mb_mc_01xywh( h, 0, 0, 4, 4 );
            else if( b_list0[0] )          x264_mb_mc_0xywh ( h, 0, 0, 4, 4 );
            else if( b_list1[0] )          x264_mb_mc_1xywh ( h, 0, 0, 4, 4 );
        }
        else if( h->mb.i_partition == D_16x8 )
        {
            if( b_list0[0] && b_list1[0] ) x264_mb_mc_01xywh( h, 0, 0, 4, 2 );
            else if( b_list0[0] )          x264_mb_mc_0xywh ( h, 0, 0, 4, 2 );
            else if( b_list1[0] )          x264_mb_mc_1xywh ( h, 0, 0, 4, 2 );

            if( b_list0[1] && b_list1[1] ) x264_mb_mc_01xywh( h, 0, 2, 4, 2 );
            else if( b_list0[1] )          x264_mb_mc_0xywh ( h, 0, 2, 4, 2 );
            else if( b_list1[1] )          x264_mb_mc_1xywh ( h, 0, 2, 4, 2 );
        }
        else if( h->mb.i_partition == D_8x16 )
        {
            if( b_list0[0] && b_list1[0] ) x264_mb_mc_01xywh( h, 0, 0, 2, 4 );
            else if( b_list0[0] )          x264_mb_mc_0xywh ( h, 0, 0, 2, 4 );
            else if( b_list1[0] )          x264_mb_mc_1xywh ( h, 0, 0, 2, 4 );

            if( b_list0[1] && b_list1[1] ) x264_mb_mc_01xywh( h, 2, 0, 2, 4 );
            else if( b_list0[1] )          x264_mb_mc_0xywh ( h, 2, 0, 2, 4 );
            else if( b_list1[1] )          x264_mb_mc_1xywh ( h, 2, 0, 2, 4 );
        }
    }
}

void x264_macroblock_cache_init( x264_t *h )
{
    int i, j;
    int i_mb_count = h->mb.i_mb_count;

    h->mb.i_mb_stride = h->sps->i_mb_width;
    h->mb.i_b8_stride = h->sps->i_mb_width * 2;
    h->mb.i_b4_stride = h->sps->i_mb_width * 4;

    h->mb.qp  = x264_malloc( i_mb_count * sizeof( int8_t) );
    h->mb.cbp = x264_malloc( i_mb_count * sizeof( int16_t) );
    h->mb.skipbp = x264_malloc( i_mb_count * sizeof( int8_t) );

    /* 0 -> 3 top(4), 4 -> 6 : left(3) */
    h->mb.intra4x4_pred_mode = x264_malloc( i_mb_count * 7 * sizeof( int8_t ) );

    /* all coeffs */
    h->mb.non_zero_count = x264_malloc( i_mb_count * 24 * sizeof( uint8_t ) );

    if( h->param.b_cabac )
    {
        h->mb.chroma_pred_mode = x264_malloc( i_mb_count * sizeof( int8_t) );
        h->mb.mvd[0] = x264_malloc( 2*16 * i_mb_count * sizeof( int16_t ) );
        h->mb.mvd[1] = x264_malloc( 2*16 * i_mb_count * sizeof( int16_t ) );
    }

    for( i=0; i<2; i++ )
    {
        int i_refs = i ? 1 + h->param.b_bframe_pyramid : h->param.i_frame_reference;
        for( j=0; j < i_refs; j++ )
            h->mb.mvr[i][j] = x264_malloc( 2 * i_mb_count * sizeof( int16_t ) );
    }

    /* init with not avaiable (for top right idx=7,15) */
    memset( h->mb.cache.ref[0], -2, X264_SCAN8_SIZE * sizeof( int8_t ) );
    memset( h->mb.cache.ref[1], -2, X264_SCAN8_SIZE * sizeof( int8_t ) );
}
void x264_macroblock_cache_end( x264_t *h )
{
    int i, j;
    for( i=0; i<2; i++ )
    {
        int i_refs = i ? 1 + h->param.b_bframe_pyramid : h->param.i_frame_reference;
        for( j=0; j < i_refs; j++ )
            x264_free( h->mb.mvr[i][j] );
    }
    if( h->param.b_cabac )
    {
        x264_free( h->mb.chroma_pred_mode );
        x264_free( h->mb.mvd[0] );
        x264_free( h->mb.mvd[1] );
    }
    x264_free( h->mb.intra4x4_pred_mode );
    x264_free( h->mb.non_zero_count );
    x264_free( h->mb.skipbp );
    x264_free( h->mb.cbp );
    x264_free( h->mb.qp );
}
void x264_macroblock_slice_init( x264_t *h )
{
    int i, j;

    h->mb.mv[0] = h->fdec->mv[0];
    h->mb.mv[1] = h->fdec->mv[1];
    h->mb.ref[0] = h->fdec->ref[0];
    h->mb.ref[1] = h->fdec->ref[1];
    h->mb.type = h->fdec->mb_type;

    h->fdec->i_ref[0] = h->i_ref0;
    h->fdec->i_ref[1] = h->i_ref1;
    for( i = 0; i < h->i_ref0; i++ )
        h->fdec->ref_poc[0][i] = h->fref0[i]->i_poc;
    if( h->sh.i_type == SLICE_TYPE_B )
    {
        for( i = 0; i < h->i_ref1; i++ )
            h->fdec->ref_poc[1][i] = h->fref1[i]->i_poc;

        h->mb.map_col_to_list0[-1] = -1;
        h->mb.map_col_to_list0[-2] = -2;
        for( i = 0; i < h->fref1[0]->i_ref[0]; i++ )
        {
            int poc = h->fref1[0]->ref_poc[0][i];
            h->mb.map_col_to_list0[i] = -2;
            for( j = 0; j < h->i_ref0; j++ )
                if( h->fref0[j]->i_poc == poc )
                {
                    h->mb.map_col_to_list0[i] = j;
                    break;
                }
        }
    }
}


void x264_macroblock_cache_load( x264_t *h, int i_mb_x, int i_mb_y )
{
    const int i_mb_4x4 = 4*(i_mb_y * h->mb.i_b4_stride + i_mb_x);
    const int i_mb_8x8 = 2*(i_mb_y * h->mb.i_b8_stride + i_mb_x);

    int i_top_xy = -1;
    int i_left_xy = -1;
    int i_top_type = -1;    /* gcc warn */
    int i_left_type= -1;

    int i;

    /* init index */
    h->mb.i_mb_x = i_mb_x;
    h->mb.i_mb_y = i_mb_y;
    h->mb.i_mb_xy = i_mb_y * h->mb.i_mb_stride + i_mb_x;
    h->mb.i_b8_xy = i_mb_8x8;
    h->mb.i_b4_xy = i_mb_4x4;
    h->mb.i_neighbour = 0;

    /* load picture pointers */
    for( i = 0; i < 3; i++ )
    {
        const int w = (i == 0 ? 16 : 8);
        const int i_stride = h->fdec->i_stride[i];
        int   j;

        h->mb.pic.i_stride[i] = i_stride;

        h->mb.pic.p_fenc[i] = &h->fenc->plane[i][ w * ( i_mb_x + i_mb_y * i_stride )];

        h->mb.pic.p_fdec[i] = &h->fdec->plane[i][ w * ( i_mb_x + i_mb_y * i_stride )];

        for( j = 0; j < h->i_ref0; j++ )
        {
            h->mb.pic.p_fref[0][j][i==0 ? 0:i+3] = &h->fref0[j]->plane[i][ w * ( i_mb_x + i_mb_y * i_stride )];
            h->mb.pic.p_fref[0][j][i+1] = &h->fref0[j]->filtered[i+1][ 16 * ( i_mb_x + i_mb_y * h->fdec->i_stride[0] )];
        }
        for( j = 0; j < h->i_ref1; j++ )
        {
            h->mb.pic.p_fref[1][j][i==0 ? 0:i+3] = &h->fref1[j]->plane[i][ w * ( i_mb_x + i_mb_y * i_stride )];
            h->mb.pic.p_fref[1][j][i+1] = &h->fref1[j]->filtered[i+1][ 16 * ( i_mb_x + i_mb_y * h->fdec->i_stride[0] )];
        }
    }

    /* load cache */
    if( i_mb_y > 0 )
    {
        i_top_xy  = h->mb.i_mb_xy - h->mb.i_mb_stride;
        i_top_type= h->mb.type[i_top_xy];

        h->mb.i_neighbour |= MB_TOP;

        /* load intra4x4 */
        h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8] = h->mb.intra4x4_pred_mode[i_top_xy][0];
        h->mb.cache.intra4x4_pred_mode[x264_scan8[1] - 8] = h->mb.intra4x4_pred_mode[i_top_xy][1];
        h->mb.cache.intra4x4_pred_mode[x264_scan8[4] - 8] = h->mb.intra4x4_pred_mode[i_top_xy][2];
        h->mb.cache.intra4x4_pred_mode[x264_scan8[5] - 8] = h->mb.intra4x4_pred_mode[i_top_xy][3];

        /* load non_zero_count */
        h->mb.cache.non_zero_count[x264_scan8[0] - 8] = h->mb.non_zero_count[i_top_xy][10];
        h->mb.cache.non_zero_count[x264_scan8[1] - 8] = h->mb.non_zero_count[i_top_xy][11];
        h->mb.cache.non_zero_count[x264_scan8[4] - 8] = h->mb.non_zero_count[i_top_xy][14];
        h->mb.cache.non_zero_count[x264_scan8[5] - 8] = h->mb.non_zero_count[i_top_xy][15];

        h->mb.cache.non_zero_count[x264_scan8[16+0] - 8] = h->mb.non_zero_count[i_top_xy][16+2];
        h->mb.cache.non_zero_count[x264_scan8[16+1] - 8] = h->mb.non_zero_count[i_top_xy][16+3];

        h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 8] = h->mb.non_zero_count[i_top_xy][16+4+2];
        h->mb.cache.non_zero_count[x264_scan8[16+4+1] - 8] = h->mb.non_zero_count[i_top_xy][16+4+3];
    }
    else
    {
        /* load intra4x4 */
        h->mb.cache.intra4x4_pred_mode[x264_scan8[0] - 8] =
        h->mb.cache.intra4x4_pred_mode[x264_scan8[1] - 8] =
        h->mb.cache.intra4x4_pred_mode[x264_scan8[4] - 8] =
        h->mb.cache.intra4x4_pred_mode[x264_scan8[5] - 8] = -1;

        /* load non_zero_count */
        h->mb.cache.non_zero_count[x264_scan8[0] - 8] =
        h->mb.cache.non_zero_count[x264_scan8[1] - 8] =
        h->mb.cache.non_zero_count[x264_scan8[4] - 8] =
        h->mb.cache.non_zero_count[x264_scan8[5] - 8] =
        h->mb.cache.non_zero_count[x264_scan8[16+0] - 8] =
        h->mb.cache.non_zero_count[x264_scan8[16+1] - 8] =
        h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 8] =
        h->mb.cache.non_zero_count[x264_scan8[16+4+1] - 8] = 0x80;

    }

    if( i_mb_x > 0 )
    {
        i_left_xy  = h->mb.i_mb_xy - 1;
        i_left_type= h->mb.type[i_left_xy];

        h->mb.i_neighbour |= MB_LEFT;

        /* load intra4x4 */
        h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][4];
        h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][5];
        h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][6];
        h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = h->mb.intra4x4_pred_mode[i_left_xy][3];

        /* load non_zero_count */
        h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] = h->mb.non_zero_count[i_left_xy][5];
        h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] = h->mb.non_zero_count[i_left_xy][7];
        h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] = h->mb.non_zero_count[i_left_xy][13];
        h->mb.cache.non_zero_count[x264_scan8[10] - 1] = h->mb.non_zero_count[i_left_xy][15];

        h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] = h->mb.non_zero_count[i_left_xy][16+1];
        h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] = h->mb.non_zero_count[i_left_xy][16+3];

        h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] = h->mb.non_zero_count[i_left_xy][16+4+1];
        h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = h->mb.non_zero_count[i_left_xy][16+4+3];
    }
    else
    {
        h->mb.cache.intra4x4_pred_mode[x264_scan8[0 ] - 1] =
        h->mb.cache.intra4x4_pred_mode[x264_scan8[2 ] - 1] =
        h->mb.cache.intra4x4_pred_mode[x264_scan8[8 ] - 1] =
        h->mb.cache.intra4x4_pred_mode[x264_scan8[10] - 1] = -1;

        /* load non_zero_count */
        h->mb.cache.non_zero_count[x264_scan8[0 ] - 1] =
        h->mb.cache.non_zero_count[x264_scan8[2 ] - 1] =
        h->mb.cache.non_zero_count[x264_scan8[8 ] - 1] =
        h->mb.cache.non_zero_count[x264_scan8[10] - 1] =
        h->mb.cache.non_zero_count[x264_scan8[16+0] - 1] =
        h->mb.cache.non_zero_count[x264_scan8[16+2] - 1] =
        h->mb.cache.non_zero_count[x264_scan8[16+4+0] - 1] =
        h->mb.cache.non_zero_count[x264_scan8[16+4+2] - 1] = 0x80;
    }

    if( i_mb_y > 0 && i_mb_x < h->sps->i_mb_width - 1 )
    {
        h->mb.i_neighbour |= MB_TOPRIGHT;
    }

    /* load ref/mv/mvd */
    if( h->sh.i_type != SLICE_TYPE_I )
    {
        const int s8x8 = h->mb.i_b8_stride;
        const int s4x4 = h->mb.i_b4_stride;

        int i_top_left_xy   = -1;
        int i_top_right_xy  = -1;

        int i_list;

        if( h->mb.i_mb_y > 0 && h->mb.i_mb_x > 0 )
        {
            i_top_left_xy   = i_top_xy - 1;
        }
        if( h->mb.i_mb_y > 0 && h->mb.i_mb_x < h->sps->i_mb_width - 1 )
        {
            i_top_right_xy = i_top_xy + 1;
        }

        for( i_list = 0; i_list < (h->sh.i_type == SLICE_TYPE_B ? 2  : 1 ); i_list++ )
        {