x264 - x264_macroblock_encode

最新推荐文章于 2022-09-16 17:26:03 发布

fanbird2008

最新推荐文章于 2022-09-16 17:26:03 发布

阅读量1k

点赞数

分类专栏： Stream Media/H264

本文链接：https://blog.csdn.net/fanbird2008/article/details/28605651

版权

Stream Media/H264 专栏收录该内容

126 篇文章 3 订阅

订阅专栏

void x264_macroblock_encode( x264_t *h )
{
    if( CHROMA444 )
        x264_macroblock_encode_internal( h, 3, 0 );
    else
        x264_macroblock_encode_internal( h, 1, 1 );
}

/*****************************************************************************
* x264_macroblock_encode:
*****************************************************************************/
static ALWAYS_INLINE void x264_macroblock_encode_internal( x264_t *h, int plane_count, int chroma )
{
    // generally, plane_count = 1, chroma = 1
    int i_qp = h->mb.i_qp;
    int b_decimate = h->mb.b_dct_decimate;
    int b_force_no_skip = 0;
    int nz;
    h->mb.i_cbp_luma = 0;
    for( int p = 0; p < plane_count; p++ ) // clear zero dc parts of h->mb.cache.non_zero_count
        h->mb.cache.non_zero_count[x264_scan8[LUMA_DC+p]] = 0;

    if( h->mb.i_type == I_PCM )
    {   // if I_PCM, then do copy directly
        /* if PCM is chosen, we need to store reconstructed frame data */
        for( int p = 0; p < plane_count; p++ ) // copy Y of YUV
            h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[p], FDEC_STRIDE, h->mb.pic.p_fenc[p], FENC_STRIDE, 16 );
        if( chroma )
        {   // copy UV of YUV
            int height = 16 >> CHROMA_V_SHIFT;
            h->mc.copy[PIXEL_8x8] ( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1], FENC_STRIDE, height );
            h->mc.copy[PIXEL_8x8] ( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2], FENC_STRIDE, height );
        }
        return;
    }

    if( !h->mb.b_allow_skip )
    {
        b_force_no_skip = 1;
        if( IS_SKIP(h->mb.i_type) )
        {
            if( h->mb.i_type == P_SKIP )
                h->mb.i_type = P_L0;
            else if( h->mb.i_type == B_SKIP )
                h->mb.i_type = B_DIRECT;
        }
    }

    if( h->mb.i_type == P_SKIP )
    {
        /* don't do pskip motion compensation if it was already done in macroblock_analyse */
        if( !h->mb.b_skip_mc )
        {
            int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
                                  h->mb.mv_min[0], h->mb.mv_max[0] );
            int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
                                  h->mb.mv_min[1], h->mb.mv_max[1] );

            for( int p = 0; p < plane_count; p++ )
                h->mc.mc_luma( h->mb.pic.p_fdec[p], FDEC_STRIDE,
                               &h->mb.pic.p_fref[0][0][p*4], h->mb.pic.i_stride[p],
                               mvx, mvy, 16, 16, &h->sh.weight[0][p] );

            if( chroma )
            {
                int v_shift = CHROMA_V_SHIFT;
                int height = 16 >> v_shift;

                /* Special case for mv0, which is (of course) very common in P-skip mode. */
                if( mvx | mvy )
                    h->mc.mc_chroma( h->mb.pic.p_fdec[1], h->mb.pic.p_fdec[2], FDEC_STRIDE,
                                     h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
                                     mvx, 2*mvy>>v_shift, 8, height );
                else
                    h->mc.load_deinterleave_chroma_fdec( h->mb.pic.p_fdec[1], h->mb.pic.p_fref[0][0][4],
                                                         h->mb.pic.i_stride[1], height );

                if( h->sh.weight[0][1].weightfn )
                    h->sh.weight[0][1].weightfn[8>>2]( h->mb.pic.p_fdec[1], FDEC_STRIDE,
                                                       h->mb.pic.p_fdec[1], FDEC_STRIDE,
                                                       &h->sh.weight[0][1], height );
                if( h->sh.weight[0][2].weightfn )
                    h->sh.weight[0][2].weightfn[8>>2]( h->mb.pic.p_fdec[2], FDEC_STRIDE,
                                                       h->mb.pic.p_fdec[2], FDEC_STRIDE,
                                                       &h->sh.weight[0][2], height );
            }
        }
        // encode mb skipped
        x264_macroblock_encode_skip( h );
        return;
    }
    if( h->mb.i_type == B_SKIP )
    {
        /* don't do bskip motion compensation if it was already done in macroblock_analyse */
        if( !h->mb.b_skip_mc )
            x264_mb_mc( h );
        x264_macroblock_encode_skip( h );
        return;
    }

    if( h->mb.i_type == I_16x16 )
    {
        h->mb.b_transform_8x8 = 0;

        // encode 16x16 block
        for( int p = 0; p < plane_count; p++, i_qp = h->mb.i_chroma_qp )
            x264_mb_encode_i16x16( h, p, i_qp );
    }
    else if( h->mb.i_type == I_8x8 )
    {
        h->mb.b_transform_8x8 = 1;
        /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
        if( h->mb.i_skip_intra )
        {
            h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
            M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i8x8_nnz_buf[0];
            M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i8x8_nnz_buf[1];
            M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i8x8_nnz_buf[2];
            M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i8x8_nnz_buf[3];
            h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
            /* In RD mode, restore the now-overwritten DCT data. */
            if( h->mb.i_skip_intra == 2 )
                h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
        }
        for( int p = 0; p < plane_count; p++, i_qp = h->mb.i_chroma_qp )
        {
            for( int i = (p == 0 && h->mb.i_skip_intra) ? 3 : 0 ; i < 4; i++ )
            {
                int i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
                // encode 4 8x8 blocks
                x264_mb_encode_i8x8( h, p, i, i_qp, i_mode, NULL, 1 );
            }
        }
    }
    else if( h->mb.i_type == I_4x4 )
    {
        h->mb.b_transform_8x8 = 0;
        /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
        if( h->mb.i_skip_intra )
        {
            h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
            M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i4x4_nnz_buf[0];
            M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i4x4_nnz_buf[1];
            M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i4x4_nnz_buf[2];
            M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i4x4_nnz_buf[3];
            h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
            /* In RD mode, restore the now-overwritten DCT data. */
            if( h->mb.i_skip_intra == 2 )
                h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
        }
        for( int p = 0; p < plane_count; p++, i_qp = h->mb.i_chroma_qp )
        {
            for( int i = (p == 0 && h->mb.i_skip_intra) ? 15 : 0 ; i < 16; i++ )
            {
                pixel *p_dst = &h->mb.pic.p_fdec[p][block_idx_xy_fdec[i]];
                int i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];

                if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
                    /* emulate missing topright samples */
                    MPIXEL_X4( &p_dst[4-FDEC_STRIDE] ) = PIXEL_SPLAT_X4( p_dst[3-FDEC_STRIDE] );

                // encode 16 4x4 blocks
                x264_mb_encode_i4x4( h, p, i, i_qp, i_mode, 1 );
            }
        }
    }
    else    /* Inter MB */ /inter encode/
    {
        int i_decimate_mb = 0;

        /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
        if( !h->mb.b_skip_mc )
            x264_mb_mc( h );

        if( h->mb.b_lossless )
        {
            if( h->mb.b_transform_8x8 )
                for( int p = 0; p < plane_count; p++ )
                    for( int i8x8 = 0; i8x8 < 4; i8x8++ )
                    {
                        int x = i8x8&1;
                        int y = i8x8>>1;
                        nz = h->zigzagf.sub_8x8( h->dct.luma8x8[p*4+i8x8], h->mb.pic.p_fenc[p] + 8*x + 8*y*FENC_STRIDE,
                                                                           h->mb.pic.p_fdec[p] + 8*x + 8*y*FDEC_STRIDE );
                        STORE_8x8_NNZ( p, i8x8, nz );
                        h->mb.i_cbp_luma |= nz << i8x8;
                    }
            else
                for( int p = 0; p < plane_count; p++ )
                    for( int i4x4 = 0; i4x4 < 16; i4x4++ )
                    {
                        nz = h->zigzagf.sub_4x4( h->dct.luma4x4[p*16+i4x4],
                                                 h->mb.pic.p_fenc[p]+block_idx_xy_fenc[i4x4],
                                                 h->mb.pic.p_fdec[p]+block_idx_xy_fdec[i4x4] );
                        h->mb.cache.non_zero_count[x264_scan8[p*16+i4x4]] = nz;
                        h->mb.i_cbp_luma |= nz << (i4x4>>2);
                    }
        }
        else if( h->mb.b_transform_8x8 )
        {
            ALIGNED_ARRAY_N( dctcoef, dct8x8,[4],[64] );
            b_decimate &= !h->mb.b_trellis || !h->param.b_cabac; // 8x8 trellis is inherently optimal decimation for CABAC

            for( int p = 0; p < plane_count; p++, i_qp = h->mb.i_chroma_qp )
            {
                CLEAR_16x16_NNZ( p );
                h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[p], h->mb.pic.p_fdec[p] );
                h->nr_count[1+!!p*2] += h->mb.b_noise_reduction * 4;

                int plane_cbp = 0;
                for( int idx = 0; idx < 4; idx++ )
                {
                    nz = x264_quant_8x8( h, dct8x8[idx], i_qp, ctx_cat_plane[DCT_LUMA_8x8][p], 0, p, idx );

                    if( nz )
                    {
                        h->zigzagf.scan_8x8( h->dct.luma8x8[p*4+idx], dct8x8[idx] );
                        if( b_decimate )
                        {
                            int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[p*4+idx] );
                            i_decimate_mb += i_decimate_8x8;
                            if( i_decimate_8x8 >= 4 )
                                plane_cbp |= 1<<idx;
                        }
                        else
                            plane_cbp |= 1<<idx;
                    }
                }

                if( i_decimate_mb >= 6 || !b_decimate )
                {
                    h->mb.i_cbp_luma |= plane_cbp;
                    FOREACH_BIT( idx, 0, plane_cbp )
                    {
                        h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[p?CQM_8PC:CQM_8PY], i_qp );
                        h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[p][8*(idx&1) + 8*(idx>>1)*FDEC_STRIDE], dct8x8[idx] );
                        STORE_8x8_NNZ( p, idx, 1 );
                    }
                }
            }
        }
        else
        {
            ALIGNED_ARRAY_N( dctcoef, dct4x4,[16],[16] );
            for( int p = 0; p < plane_count; p++, i_qp = h->mb.i_chroma_qp )
            {
                CLEAR_16x16_NNZ( p );
                h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[p], h->mb.pic.p_fdec[p] );

                if( h->mb.b_noise_reduction )
                {
                    h->nr_count[0+!!p*2] += 16;
                    for( int idx = 0; idx < 16; idx++ )
                        h->quantf.denoise_dct( dct4x4[idx], h->nr_residual_sum[0+!!p*2], h->nr_offset[0+!!p*2], 16 );
                }

                int plane_cbp = 0;
                for( int i8x8 = 0; i8x8 < 4; i8x8++ )
                {
                    int i_decimate_8x8 = b_decimate ? 0 : 6;
                    int nnz8x8 = 0;
                    if( h->mb.b_trellis )
                    {
                        for( int i4x4 = 0; i4x4 < 4; i4x4++ )
                        {
                            int idx = i8x8*4+i4x4;
                            if( x264_quant_4x4_trellis( h, dct4x4[idx], CQM_4PY, i_qp, ctx_cat_plane[DCT_LUMA_4x4][p], 0, !!p, p*16+idx ) )
                            {
                                h->zigzagf.scan_4x4( h->dct.luma4x4[p*16+idx], dct4x4[idx] );
                                h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[p?CQM_4PC:CQM_4PY], i_qp );
                                if( i_decimate_8x8 < 6 )
                                    i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[p*16+idx] );
                                h->mb.cache.non_zero_count[x264_scan8[p*16+idx]] = 1;
                                nnz8x8 = 1;
                            }
                        }
                    }
                    else
                    {
                        nnz8x8 = nz = h->quantf.quant_4x4x4( &dct4x4[i8x8*4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
                        if( nz )
                        {
                            FOREACH_BIT( idx, i8x8*4, nz )
                            {
                                h->zigzagf.scan_4x4( h->dct.luma4x4[p*16+idx], dct4x4[idx] );
                                h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[p?CQM_4PC:CQM_4PY], i_qp );
                                if( i_decimate_8x8 < 6 )
                                    i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[p*16+idx] );
                                h->mb.cache.non_zero_count[x264_scan8[p*16+idx]] = 1;
                            }
                        }
                    }
                    if( nnz8x8 )
                    {
                        i_decimate_mb += i_decimate_8x8;
                        if( i_decimate_8x8 < 4 )
                            STORE_8x8_NNZ( p, i8x8, 0 );
                        else
                            plane_cbp |= 1<<i8x8;
                    }
                }

                if( i_decimate_mb < 6 )
                {
                    plane_cbp = 0;
                    CLEAR_16x16_NNZ( p );
                }
                else
                {
                    h->mb.i_cbp_luma |= plane_cbp;
                    FOREACH_BIT( i8x8, 0, plane_cbp )
                    {
                        h->dctf.add8x8_idct( &h->mb.pic.p_fdec[p][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
                    }
                }
            }
        }
    }

    /* encode chroma */
    if( chroma )
    {
        if( IS_INTRA( h->mb.i_type ) )
        {   // in current testing, i_mode = 6
            int i_mode = h->mb.i_chroma_pred_mode;
            if( h->mb.b_lossless )
                x264_predict_lossless_chroma( h, i_mode );
            else
            {   // chroma U predict, in current test,
                // h->predict_chroma[6] = x264_predict_8x8c_dc_128_c
                h->predict_chroma[i_mode]( h->mb.pic.p_fdec[1] );
                // chroma V predict
                h->predict_chroma[i_mode]( h->mb.pic.p_fdec[2] );
            }
        }

        /* encode the 8x8 blocks */
        // encode 8x8 chroma block
        x264_mb_encode_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
    }
    else
        h->mb.i_cbp_chroma = 0;

    /* store cbp */
    int cbp = h->mb.i_cbp_chroma << 4 | h->mb.i_cbp_luma;
    if( h->param.b_cabac )
        cbp |= h->mb.cache.non_zero_count[x264_scan8[LUMA_DC    ]] << 8
            | h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+0]] << 9
            | h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+1]] << 10;
    h->mb.cbp[h->mb.i_mb_xy] = cbp;

    /* Check for P_SKIP
     * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
     *      (if multiple mv give same result)*/
    if( !b_force_no_skip )
    {
        if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
            !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
            M32( h->mb.cache.mv[0][x264_scan8[0]] ) == M32( h->mb.cache.pskip_mv )
            && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
        {
            h->mb.i_type = P_SKIP;
        }

        /* Check for B_SKIP */
        if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
        {
            h->mb.i_type = B_SKIP;
        }
    }
}

void x264_mb_encode_chroma( x264_t *h, int b_inter, int i_qp )
{
    // in current testing, b_inter = 0, i_qp = 25

    if( CHROMA_FORMAT == CHROMA_420 )
        x264_mb_encode_chroma_internal( h, b_inter, i_qp, 0 );
    else
        x264_mb_encode_chroma_internal( h, b_inter, i_qp, 1 );
}

static ALWAYS_INLINE void x264_mb_encode_chroma_internal( x264_t *h, int b_inter, int i_qp, int chroma422 )
{
    // in current testing, b_inter = 0, i_qp = 25, chroma422 = 0

    int nz, nz_dc;
    int b_decimate = b_inter && h->mb.b_dct_decimate;
    int (*dequant_mf)[16] = h->dequant4_mf[CQM_4IC + b_inter];
    ALIGNED_ARRAY_16( dctcoef, dct_dc,[8] );
    h->mb.i_cbp_chroma = 0;
    h->nr_count[2] += h->mb.b_noise_reduction * 4;

    // clear zero against 4 4x4 U blocks, 4x4 U block 0, 1
    M16( &h->mb.cache.non_zero_count[x264_scan8[16]] ) = 0;
    // 4x4 U block 2 ,3
    M16( &h->mb.cache.non_zero_count[x264_scan8[18]] ) = 0;
    // clear zero against 4 4x4 V blocks, 4x4 V block 0, 1
    M16( &h->mb.cache.non_zero_count[x264_scan8[32]] ) = 0;
    // 4x4 V block 2, 3
    M16( &h->mb.cache.non_zero_count[x264_scan8[34]] ) = 0;
    if( chroma422 )
    {
        // for chroma422 = true, one 8x16 block will be splited 8 4x4 block
        // its layout as the followings
        //   0   1
        //   2   3
        //   4   5
        //   6   7

        // if chroma422, height will be invariant.
        // 4x4 U block 4, 5
        M16( &h->mb.cache.non_zero_count[x264_scan8[24]] ) = 0;
        // 4x4 U block 6, 7
        M16( &h->mb.cache.non_zero_count[x264_scan8[26]] ) = 0;
        // 4x4 V block 4, 5
        M16( &h->mb.cache.non_zero_count[x264_scan8[40]] ) = 0;
        // 4x4 V block 6, 7
        M16( &h->mb.cache.non_zero_count[x264_scan8[42]] ) = 0;
    }

    /* Early termination: check variance of chroma residual before encoding.
     * Don't bother trying early termination at low QPs.
     * Values are experimentally derived. */
    if( b_decimate && i_qp >= (h->mb.b_trellis ? 12 : 18) && !h->mb.b_noise_reduction )
    {
        int thresh = chroma422 ? (x264_lambda2_tab[i_qp] + 16) >> 5 : (x264_lambda2_tab[i_qp] + 32) >> 6;
        int ssd[2];
        int chromapix = chroma422 ? PIXEL_8x16 : PIXEL_8x8;

        int score = h->pixf.var2[chromapix]( h->mb.pic.p_fenc[1], FENC_STRIDE, h->mb.pic.p_fdec[1], FDEC_STRIDE, &ssd[0] );
        if( score < thresh*4 )
            score += h->pixf.var2[chromapix]( h->mb.pic.p_fenc[2], FENC_STRIDE, h->mb.pic.p_fdec[2], FDEC_STRIDE, &ssd[1] );
        if( score < thresh*4 )
        {
            h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+0]] = 0;
            h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+1]] = 0;

            for( int ch = 0; ch < 2; ch++ )
            {
                if( ssd[ch] > thresh )
                {
                    pixel *p_src = h->mb.pic.p_fenc[1+ch];
                    pixel *p_dst = h->mb.pic.p_fdec[1+ch];

                    if( chroma422 )
                        /* Cannot be replaced by two calls to sub8x8_dct_dc since the hadamard transform is different */
                        h->dctf.sub8x16_dct_dc( dct_dc, p_src, p_dst );
                    else
                        h->dctf.sub8x8_dct_dc( dct_dc, p_src, p_dst );

                    if( h->mb.b_trellis )
                        nz_dc = x264_quant_chroma_dc_trellis( h, dct_dc, i_qp+3*chroma422, !b_inter, CHROMA_DC+ch );
                    else
                    {
                        nz_dc = 0;
                        for( int i = 0; i <= chroma422; i++ )
                            nz_dc |= h->quantf.quant_2x2_dc( &dct_dc[4*i], h->quant4_mf[CQM_4IC+b_inter][i_qp+3*chroma422][0] >> 1,
                                                             h->quant4_bias[CQM_4IC+b_inter][i_qp+3*chroma422][0] << 1 );
                    }

                    if( nz_dc )
                    {
                        if( !x264_mb_optimize_chroma_dc( h, dct_dc, dequant_mf, i_qp+3*chroma422, chroma422 ) )
                            continue;
                        h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+ch]] = 1;
                        if( chroma422 )
                        {
                            zigzag_scan_2x4_dc( h->dct.chroma_dc[ch], dct_dc );
                            h->quantf.idct_dequant_2x4_dconly( dct_dc, dequant_mf, i_qp+3 );
                        }
                        else
                        {
                            zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct_dc );
                            idct_dequant_2x2_dconly( dct_dc, dequant_mf, i_qp );
                        }

                        for( int i = 0; i <= chroma422; i++ )
                            h->dctf.add8x8_idct_dc( p_dst + 8*i*FDEC_STRIDE, &dct_dc[4*i] );
                        h->mb.i_cbp_chroma = 1;
                    }
                }
            }
            return;
        }
    }

    // for each chroma, encode it repectively.
    for( int ch = 0; ch < 2; ch++ )
    {
        pixel *p_src = h->mb.pic.p_fenc[1+ch];
        pixel *p_dst = h->mb.pic.p_fdec[1+ch];
        int i_decimate_score = b_decimate ? 0 : 7;
        int nz_ac = 0;

ALIGNED_ARRAY_N( dctcoef, dct4x4,[8],[16] );

        if( h->mb.b_lossless )
        {
            static const uint8_t chroma422_scan[8] = { 0, 2, 1, 5, 3, 6, 4, 7 };

            for( int i = 0; i < (chroma422?8:4); i++ )
            {
                int oe = 4*(i&1) + 4*(i>>1)*FENC_STRIDE;
                int od = 4*(i&1) + 4*(i>>1)*FDEC_STRIDE;
                nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i+(chroma422?i&4:0)+ch*16], p_src+oe, p_dst+od,
                                           &h->dct.chroma_dc[ch][chroma422?chroma422_scan[i]:i] );
                h->mb.cache.non_zero_count[x264_scan8[16+i+(chroma422?i&4:0)+ch*16]] = nz;
                h->mb.i_cbp_chroma |= nz;
            }
            h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+ch]] = array_non_zero( h->dct.chroma_dc[ch], chroma422?8:4 );
            continue;
        }

        // if chroma422 true, then encode 2 8x8 blocks for each chroma
        // else encode one 8x8 block for each chroma
        // h->dctf.sub8x8_dct = sub8x8_dct, do 8x8 dct transform against diff of src pixels and predict pixels
        for( int i = 0; i <= chroma422; i++ )
            h->dctf.sub8x8_dct( &dct4x4[4*i], p_src + 8*i*FENC_STRIDE, p_dst + 8*i*FDEC_STRIDE );

        // if denoise enabled, do denoise against dct coef
        // if chroma422, 8 4x4 block will be denoised,
        // else          4 4x4 block will be denoised.
        if( h->mb.b_noise_reduction )
            for( int i = 0; i < (chroma422?8:4); i++ )
                h->quantf.denoise_dct( dct4x4[i], h->nr_residual_sum[2], h->nr_offset[2], 16 );

        // if chroma422, do H2x4 against dc of dct4x4
        // else          do H2x2 against dc of dct4x4
        //
        if( chroma422 )
            h->dctf.dct2x4dc( dct_dc, dct4x4 );
        else
            dct2x2dc( dct_dc, dct4x4 );

        /* calculate dct coeffs */
        for( int i8x8 = 0; i8x8 < (chroma422?2:1); i8x8++ )
        {
            if( h->mb.b_trellis )
            {
                for( int i4x4 = 0; i4x4 < 4; i4x4++ )
                {
                    if( x264_quant_4x4_trellis( h, dct4x4[i8x8*4+i4x4], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 1, 0 ) )
                    {
                        int idx = 16+ch*16+i8x8*8+i4x4;
                        h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[i8x8*4+i4x4] );
                        h->quantf.dequant_4x4( dct4x4[i8x8*4+i4x4], dequant_mf, i_qp );
                        if( i_decimate_score < 7 )
                            i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[idx] );
                        h->mb.cache.non_zero_count[x264_scan8[idx]] = 1;
                        nz_ac = 1;
                    }
                }
            }
            else
            {
                // for one 8x8 block, quant ac coef of 4 4x4 block
                // nz is or-shifted
                // h->quantf.quant_4x4x4 = quant_4x4x4
                nz = h->quantf.quant_4x4x4( &dct4x4[i8x8*4], h->quant4_mf[CQM_4IC+b_inter][i_qp],
                                            h->quant4_bias[CQM_4IC+b_inter][i_qp] );
                nz_ac |= nz;

                // for each nz(nonzero is true), do following actions
                FOREACH_BIT( i4x4, 0, nz )
                {
                    int idx = 16+ch*16+i8x8*8+i4x4;
                    // zigzag scan ac dct coef
                    h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[i8x8*4+i4x4] );
                    // dequant ac dct coef
                    h->quantf.dequant_4x4( dct4x4[i8x8*4+i4x4], dequant_mf, i_qp );
                    if( i_decimate_score < 7 )
                        i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[idx] );
                    // save ac nz mark
                    h->mb.cache.non_zero_count[x264_scan8[idx]] = 1;
                }
            }
        }

        if( h->mb.b_trellis )
            nz_dc = x264_quant_chroma_dc_trellis( h, dct_dc, i_qp+3*chroma422, !b_inter, CHROMA_DC+ch );
        else
        {
            nz_dc = 0;
            // quant dc coef
            // if chroma422, do 2 quant against each 2x2 dc block
            // else          do one quant
            // h->quantf.quant_2x2_dc = quant_2x2_dc
            for( int i = 0; i <= chroma422; i++ )
                nz_dc |= h->quantf.quant_2x2_dc( &dct_dc[4*i], h->quant4_mf[CQM_4IC+b_inter][i_qp+3*chroma422][0] >> 1,
                                                 h->quant4_bias[CQM_4IC+b_inter][i_qp+3*chroma422][0] << 1 );
        }

        // save dc nz mark
        h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+ch]] = nz_dc;

        if( i_decimate_score < 7 || !nz_ac )
        {
            // differences(差异) of src pixel block and predict pixel block is too small.
            /* Decimate the block */
            // clear 0 against 4x4 block 0, 1
            M16( &h->mb.cache.non_zero_count[x264_scan8[16+16*ch]] ) = 0;
            // clear 0 against 4x4 block 2, 3
            M16( &h->mb.cache.non_zero_count[x264_scan8[18+16*ch]] ) = 0;
            if( chroma422 )
            {
                // clear 0 against 4x4 block 4, 5
                M16( &h->mb.cache.non_zero_count[x264_scan8[24+16*ch]] ) = 0;
                // clear 0 against 4x4 block 6, 7
                M16( &h->mb.cache.non_zero_count[x264_scan8[26+16*ch]] ) = 0;
            }

            if( !nz_dc ) /* Whole block is empty */
                continue;
            if( !x264_mb_optimize_chroma_dc( h, dct_dc, dequant_mf, i_qp+3*chroma422, chroma422 ) )
            {
                h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+ch]] = 0;
                continue;
            }
            /* DC-only */
            if( chroma422 )
            {
                zigzag_scan_2x4_dc( h->dct.chroma_dc[ch], dct_dc );
                h->quantf.idct_dequant_2x4_dconly( dct_dc, dequant_mf, i_qp+3 );
            }
            else
            {
                zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct_dc );
                idct_dequant_2x2_dconly( dct_dc, dequant_mf, i_qp );
            }

            for( int i = 0; i <= chroma422; i++ )
                h->dctf.add8x8_idct_dc( p_dst + 8*i*FDEC_STRIDE, &dct_dc[4*i] );
        }
        else
        {
            h->mb.i_cbp_chroma = 1;

            if( nz_dc )
            {
                if( chroma422 )
                {
                    // zigzag scan dc coef of chroma
                    zigzag_scan_2x4_dc( h->dct.chroma_dc[ch], dct_dc );
                    // do idct(2x4), then dequant against dc coef of chroma
                    h->quantf.idct_dequant_2x4_dc( dct_dc, dct4x4, dequant_mf, i_qp+3 );
                }
                else
                {   // zigzag scan dc coef of chroma
                    zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct_dc );
                    // do idct(2x2), then dequant against dc coef of chroma
                    idct_dequant_2x2_dc( dct_dc, dct4x4, dequant_mf, i_qp );
                }
            }

            // reconstruct whole chroma block
            // if chroma422, do twice 8x8 block reconstruction
            // else          do one 8x8 block reconstruction
            for( int i = 0; i <= chroma422; i++ )
                h->dctf.add8x8_idct( p_dst + 8*i*FDEC_STRIDE, &dct4x4[4*i] );
        }
    }

    /* 0 = none, 1 = DC only, 2 = DC+AC */
    h->mb.i_cbp_chroma += (h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+0]] |
                           h->mb.cache.non_zero_count[x264_scan8[CHROMA_DC+1]] | h->mb.i_cbp_chroma);
}