openHevc学习笔记：解码器中CTU的TU与PU解码

HEVC解码器的CTU解码（CTU Decoder）部分在整个HEVC解码器中的位置如下图红框所示，在hls_coding_unit()之中。CTU解码（CTU Decoder）部分的函数调用关系如下图右边方框所示。(右键新窗口打开查看大图)

hls_decode_entry()

hls_decode_entry()是FFmpeg HEVC解码器中Slice解码的入口函数。该函数的定义如下所示。

//解码入口函数
static int hls_decode_entry(AVCodecContext *avctxt, void *isFilterThread)
{
    HEVCContext *s  = avctxt->priv_data;
    //CTB尺寸
    int ctb_size    = 1 << s->sps->log2_ctb_size;
    int more_data   = 1;
    int x_ctb       = 0;
    int y_ctb       = 0;
    int ctb_addr_ts = s->pps->ctb_addr_rs_to_ts[s->sh.slice_ctb_addr_rs];

    if (!ctb_addr_ts && s->sh.dependent_slice_segment_flag) {
        av_log(s->avctx, AV_LOG_ERROR, "Impossible initial tile.\n");
        return AVERROR_INVALIDDATA;
    }

    if (s->sh.dependent_slice_segment_flag) {
        int prev_rs = s->pps->ctb_addr_ts_to_rs[ctb_addr_ts - 1];
        if (s->tab_slice_address[prev_rs] != s->sh.slice_addr) {
            av_log(s->avctx, AV_LOG_ERROR, "Previous slice segment missing\n");
            return AVERROR_INVALIDDATA;
        }
    }

    while (more_data && ctb_addr_ts < s->sps->ctb_size) {
        int ctb_addr_rs = s->pps->ctb_addr_ts_to_rs[ctb_addr_ts];
        //CTB的位置x和y
        x_ctb = (ctb_addr_rs % ((s->sps->width + ctb_size - 1) >> s->sps->log2_ctb_size)) << s->sps->log2_ctb_size;
        y_ctb = (ctb_addr_rs / ((s->sps->width + ctb_size - 1) >> s->sps->log2_ctb_size)) << s->sps->log2_ctb_size;
        //初始化周围的参数
        hls_decode_neighbour(s, x_ctb, y_ctb, ctb_addr_ts);
        //初始化CABAC
        ff_hevc_cabac_init(s, ctb_addr_ts);
        //样点自适应补偿参数
        hls_sao_param(s, x_ctb >> s->sps->log2_ctb_size, y_ctb >> s->sps->log2_ctb_size);

        s->deblock[ctb_addr_rs].beta_offset = s->sh.beta_offset;
        s->deblock[ctb_addr_rs].tc_offset   = s->sh.tc_offset;
        s->filter_slice_edges[ctb_addr_rs]  = s->sh.slice_loop_filter_across_slices_enabled_flag;
        
        //解析四叉树结构，并且解码
        more_data = hls_coding_quadtree(s, x_ctb, y_ctb, s->sps->log2_ctb_size, 0);
        if (more_data < 0) {
            s->tab_slice_address[ctb_addr_rs] = -1;
            return more_data;
        }

        ctb_addr_ts++;
        //保存解码信息以供下次使用
        ff_hevc_save_states(s, ctb_addr_ts);
        //去块效应滤波
        ff_hevc_hls_filters(s, x_ctb, y_ctb, ctb_size);
    }

    if (x_ctb + ctb_size >= s->sps->width &&
        y_ctb + ctb_size >= s->sps->height)
        ff_hevc_hls_filter(s, x_ctb, y_ctb, ctb_size);

    return ctb_addr_ts;
}

从源代码可以看出，hls_decode_entry()调用了5个函数进行解码工作：

（1）调用hls_decode_neighbour初始化CTU周围的参数信息。
（2）调用ff_hevc_cabac_init()进行CABAC初始化。

（3）调用hls_sao_param初始化样点自适应补偿参数。

（4）调用hls_coding_quadtree()解码CTU。其中包含了PU和TU的解码。
（5）调用ff_hevc_hls_filters()进行滤波。其中包含了去块效应滤波和SAO滤波。

本文分析第四步的PU和TU解码过程

hls_coding_quadtree()

hls_coding_quadtree()用于解析CTU的四叉树句法结构。该函数的定义如下所示。

/*
 * 解析四叉树结构，并且解码
 * 注意该函数是递归调用
 *
 * s：HEVCContext上下文结构体
 * x_ctb：CB位置的x坐标
 * y_ctb：CB位置的y坐标
 * log2_cb_size：CB大小取log2之后的值
 * cb_depth：深度
 *
 */
static int hls_coding_quadtree(HEVCContext *s, int x0, int y0,
                               int log2_cb_size, int cb_depth)
{
    HEVCLocalContext *lc = s->HEVClc;
    //CB的大小,split flag=0
    //log2_cb_size为CB大小取log之后的结果
    const int cb_size    = 1 << log2_cb_size;
    int ret;
    int qp_block_mask = (1<<(s->sps->log2_ctb_size - s->pps->diff_cu_qp_delta_depth)) - 1;
    int split_cu;
    //确定CU是否还会划分
    lc->ct_depth = cb_depth;
    if (x0 + cb_size <= s->sps->width  &&
        y0 + cb_size <= s->sps->height &&
        log2_cb_size > s->sps->log2_min_cb_size) {
        split_cu = ff_hevc_split_coding_unit_flag_decode(s, cb_depth, x0, y0);
    } else {
        split_cu = (log2_cb_size > s->sps->log2_min_cb_size);
    }
    if (s->pps->cu_qp_delta_enabled_flag &&
        log2_cb_size >= s->sps->log2_ctb_size - s->pps->diff_cu_qp_delta_depth) {
        lc->tu.is_cu_qp_delta_coded = 0;
        lc->tu.cu_qp_delta          = 0;
    }

    if (s->sh.cu_chroma_qp_offset_enabled_flag &&
        log2_cb_size >= s->sps->log2_ctb_size - s->pps->diff_cu_chroma_qp_offset_depth) {
        lc->tu.is_cu_chroma_qp_offset_coded = 0;
    }

    if (split_cu) {
    	//如果CU还可以继续划分，则继续解析划分后的CU
    	//注意：这里是递归调用
    	//CB的大小,split flag=1
        const int cb_size_split = cb_size >> 1;
        const int x1 = x0 + cb_size_split;
        const int y1 = y0 + cb_size_split;

        int more_data = 0;

        //注意：
        //CU大小减半，log2_cb_size-1
        //深度d加1，cb_depth+1
        more_data = hls_coding_quadtree(s, x0, y0, log2_cb_size - 1, cb_depth + 1);
        if (more_data < 0)
            return more_data;

        if (more_data && x1 < s->sps->width) {
            more_data = hls_coding_quadtree(s, x1, y0, log2_cb_size - 1, cb_depth + 1);
            if (more_data < 0)
                return more_data;
        }
        if (more_data && y1 < s->sps->height) {
            more_data = hls_coding_quadtree(s, x0, y1, log2_cb_size - 1, cb_depth + 1);
            if (more_data < 0)
                return more_data;
        }
        if (more_data && x1 < s->sps->width &&
            y1 < s->sps->height) {
            more_data = hls_coding_quadtree(s, x1, y1, log2_cb_size - 1, cb_depth + 1);
            if (more_data < 0)
                return more_data;
        }
        if(((x0 + (1<<log2_cb_size)) & qp_block_mask) == 0 &&
            ((y0 + (1<<log2_cb_size)) & qp_block_mask) == 0)
            lc->qPy_pred = lc->qp_y;

        if (more_data)
            return ((x1 + cb_size_split) < s->sps->width ||
                    (y1 + cb_size_split) < s->sps->height);
        else
            return 0;
    } else {
    	//注意处理的是不可划分的CU单元
    	//处理CU单元-真正的解码
        ret = hls_coding_unit(s, x0, y0, log2_cb_size);
        if (ret < 0)
            return ret;
        if ((!((x0 + cb_size) %
               (1 << (s->sps->log2_ctb_size))) ||
             (x0 + cb_size >= s->sps->width)) &&
            (!((y0 + cb_size) %
               (1 << (s->sps->log2_ctb_size))) ||
             (y0 + cb_size >= s->sps->height))) {
            int end_of_slice_flag = ff_hevc_end_of_slice_flag_decode(s);
            return !end_of_slice_flag;
        } else {
            return 1;
        }
    }
    return 0;
}

从源代码可以看出，hls_coding_quadtree()首先调用ff_hevc_split_coding_unit_flag_decode()判断当前CU是否还需要划分。如果需要划分的话，就会递归调用4次hls_coding_quadtree()分别对4个子块继续进行四叉树解析；如果不需要划分，就会调用hls_coding_unit()对CU进行解码。

hls_coding_unit()

hls_coding_unit()用于解码一个CU。该函数的定义如下所示

//处理CU单元-真正的解码
static int hls_coding_unit(HEVCContext *s, int x0, int y0, int log2_cb_size)
{
	//CB大小
    int cb_size          = 1 << log2_cb_size;
    HEVCLocalContext *lc = s->HEVClc;
    int log2_min_cb_size = s->sps->log2_min_cb_size;
    int length           = cb_size >> log2_min_cb_size;
    int min_cb_width     = s->sps->min_cb_width;
    //以最小的CB为单位（例如4x4）的时候，当前CB的位置——x坐标和y坐标
    int x_cb             = x0 >> log2_min_cb_size;
    int y_cb             = y0 >> log2_min_cb_size;
    int idx              = log2_cb_size - 2;
    int qp_block_mask    = (1<<(s->sps->log2_ctb_size - s->pps->diff_cu_qp_delta_depth)) - 1;
    int x, y, ret;

    //设置CU的属性值
    lc->cu.x                = x0;
    lc->cu.y                = y0;
    lc->cu.pred_mode        = MODE_INTRA;
    lc->cu.part_mode        = PART_2Nx2N;
    lc->cu.intra_split_flag = 0;

    SAMPLE_CTB(s->skip_flag, x_cb, y_cb) = 0;

    for (x = 0; x < 4; x++)
        lc->pu.intra_pred_mode[x] = 1;
    if (s->pps->transquant_bypass_enable_flag) {
        lc->cu.cu_transquant_bypass_flag = ff_hevc_cu_transquant_bypass_flag_decode(s);
        if (lc->cu.cu_transquant_bypass_flag)
            set_deblocking_bypass(s, x0, y0, log2_cb_size);
    } else
        lc->cu.cu_transquant_bypass_flag = 0;

    if (s->sh.slice_type != I_SLICE) {
    	//Skip类型
        uint8_t skip_flag = ff_hevc_skip_flag_decode(s, x0, y0, x_cb, y_cb);
        //设置到skip_flag缓存中
        x = y_cb * min_cb_width + x_cb;
        for (y = 0; y < length; y++) {
            memset(&s->skip_flag[x], skip_flag, length);
            x += min_cb_width;
        }
        lc->cu.pred_mode = skip_flag ? MODE_SKIP : MODE_INTER;
    } else {
        x = y_cb * min_cb_width + x_cb;
        for (y = 0; y < length; y++) {
            memset(&s->skip_flag[x], 0, length);
            x += min_cb_width;
        }
    }

    if (SAMPLE_CTB(s->skip_flag, x_cb, y_cb)) {
        hls_prediction_unit(s, x0, y0, cb_size, cb_size, log2_cb_size, 0, idx);
        intra_prediction_unit_default_value(s, x0, y0, log2_cb_size);

        if (!s->sh.disable_deblocking_filter_flag)
            ff_hevc_deblocking_boundary_strengths(s, x0, y0, log2_cb_size);
    } else {
        int pcm_flag = 0;

        //读取预测模式（非 I Slice）
        if (s->sh.slice_type != I_SLICE)
            lc->cu.pred_mode = ff_hevc_pred_mode_decode(s);

        //不是帧内预测模式的时候
        //或者已经是最小CB的时候
        if (lc->cu.pred_mode != MODE_INTRA ||
            log2_cb_size == s->sps->log2_min_cb_size) {
        	//读取CU分割模式
            lc->cu.part_mode        = ff_hevc_part_mode_decode(s, log2_cb_size);
            lc->cu.intra_split_flag = lc->cu.part_mode == PART_NxN &&
                                      lc->cu.pred_mode == MODE_INTRA;
        }

        if (lc->cu.pred_mode == MODE_INTRA) {
        	//帧内预测模式

        	//PCM方式编码，不常见
            if (lc->cu.part_mode == PART_2Nx2N && s->sps->pcm_enabled_flag &&
                log2_cb_size >= s->sps->pcm.log2_min_pcm_cb_size &&
                log2_cb_size <= s->sps->pcm.log2_max_pcm_cb_size) {
                pcm_flag = ff_hevc_pcm_flag_decode(s);
            }
            if (pcm_flag) {
                intra_prediction_unit_default_value(s, x0, y0, log2_cb_size);
                ret = hls_pcm_sample(s, x0, y0, log2_cb_size);
                if (s->sps->pcm.loop_filter_disable_flag)
                    set_deblocking_bypass(s, x0, y0, log2_cb_size);

                if (ret < 0)
                    return ret;
            } else {
                //获取帧内预测模式
                intra_prediction_unit(s, x0, y0, log2_cb_size);
            }
        } else {
        	//帧间预测模式
            intra_prediction_unit_default_value(s, x0, y0, log2_cb_size);

            //帧间模式一共有8种划分模式

            switch (lc->cu.part_mode) {
            case PART_2Nx2N:
            	//处理PU单元-运动补偿
				 /*
            	 * hls_prediction_unit()参数：
            	 * x0 : PU左上角x坐标
            	 * y0 : PU左上角y坐标
            	 * nPbW : PU宽度
            	 * nPbH : PU高度
            	 * log2_cb_size : CB大小取log2()的值
            	 * partIdx : PU的索引号-分成4个块的时候取0-3，分成两个块的时候取0和1
            	 */
                hls_prediction_unit(s, x0, y0, cb_size, cb_size, log2_cb_size, 0, idx);
                break;
            case PART_2NxN:
                hls_prediction_unit(s, x0, y0,               cb_size, cb_size / 2, log2_cb_size, 0, idx);
                hls_prediction_unit(s, x0, y0 + cb_size / 2, cb_size, cb_size / 2, log2_cb_size, 1, idx);
                break;
            case PART_Nx2N:
                hls_prediction_unit(s, x0,               y0, cb_size / 2, cb_size, log2_cb_size, 0, idx - 1);
                hls_prediction_unit(s, x0 + cb_size / 2, y0, cb_size / 2, cb_size, log2_cb_size, 1, idx - 1);
                break;
            case PART_2NxnU:
                hls_prediction_unit(s, x0, y0,               cb_size, cb_size     / 4, log2_cb_size, 0, idx);
                hls_prediction_unit(s, x0, y0 + cb_size / 4, cb_size, cb_size * 3 / 4, log2_cb_size, 1, idx);
                break;
            case PART_2NxnD:
                hls_prediction_unit(s, x0, y0,                   cb_size, cb_size * 3 / 4, log2_cb_size, 0, idx);
                hls_prediction_unit(s, x0, y0 + cb_size * 3 / 4, cb_size, cb_size     / 4, log2_cb_size, 1, idx);
                break;
            case PART_nLx2N:
                hls_prediction_unit(s, x0,               y0, cb_size     / 4, cb_size, log2_cb_size, 0, idx - 2);
                hls_prediction_unit(s, x0 + cb_size / 4, y0, cb_size * 3 / 4, cb_size, log2_cb_size, 1, idx - 2);
                break;
            case PART_nRx2N:
                hls_prediction_unit(s, x0,                   y0, cb_size * 3 / 4, cb_size, log2_cb_size, 0, idx - 2);
                hls_prediction_unit(s, x0 + cb_size * 3 / 4, y0, cb_size     / 4, cb_size, log2_cb_size, 1, idx - 2);
                break;
            case PART_NxN:
                hls_prediction_unit(s, x0,               y0,               cb_size / 2, cb_size / 2, log2_cb_size, 0, idx - 1);
                hls_prediction_unit(s, x0 + cb_size / 2, y0,               cb_size / 2, cb_size / 2, log2_cb_size, 1, idx - 1);
                hls_prediction_unit(s, x0,               y0 + cb_size / 2, cb_size / 2, cb_size / 2, log2_cb_size, 2, idx - 1);
                hls_prediction_unit(s, x0 + cb_size / 2, y0 + cb_size / 2, cb_size / 2, cb_size / 2, log2_cb_size, 3, idx - 1);
                break;
            }
        }

        if (!pcm_flag) {
            int rqt_root_cbf = 1;

            if (lc->cu.pred_mode != MODE_INTRA &&
                !(lc->cu.part_mode == PART_2Nx2N && lc->pu.merge_flag)) {
                rqt_root_cbf = ff_hevc_no_residual_syntax_flag_decode(s);
            }
            if (rqt_root_cbf) {
                const static int cbf[2] = { 0 };
                lc->cu.max_trafo_depth = lc->cu.pred_mode == MODE_INTRA ?
                                         s->sps->max_transform_hierarchy_depth_intra + lc->cu.intra_split_flag :
                                         s->sps->max_transform_hierarchy_depth_inter;
                //处理TU四叉树
                ret = hls_transform_tree(s, x0, y0, x0, y0, x0, y0,
                                         log2_cb_size,
                                         log2_cb_size, 0, 0, cbf, cbf);
                if (ret < 0)
                    return ret;
            } else {
                if (!s->sh.disable_deblocking_filter_flag)
                    ff_hevc_deblocking_boundary_strengths(s, x0, y0, log2_cb_size);
            }
        }
    }

    if (s->pps->cu_qp_delta_enabled_flag && lc->tu.is_cu_qp_delta_coded == 0)
        ff_hevc_set_qPy(s, x0, y0, log2_cb_size);

    x = y_cb * min_cb_width + x_cb;
    for (y = 0; y < length; y++) {
        memset(&s->qp_y_tab[x], lc->qp_y, length);
        x += min_cb_width;
    }

    if(((x0 + (1<<log2_cb_size)) & qp_block_mask) == 0 &&
       ((y0 + (1<<log2_cb_size)) & qp_block_mask) == 0) {
        lc->qPy_pred = lc->qp_y;
    }

    set_ct_depth(s, x0, y0, log2_cb_size, lc->ct_depth);

    return 0;
}

从源代码可以看出，hls_coding_unit()主要进行了两个方面的处理：

（1）调用hls_prediction_unit()处理PU。
（2）调用hls_transform_tree()处理TU树。

其中， hls_prediction_unit()完成了以下两步工作：

（1）解析码流得到运动矢量。HEVC中包含了Merge和AMVP两种运动矢量预测技术。对于使用Merge的码流，调用ff_hevc_luma_mv_merge_mode()；对于使用AMVP的码流，调用hevc_luma_mv_mpv_mode()。
（2）根据运动矢量进行运动补偿。对于单向预测亮度运动补偿，调用luma_mc_uni()，对于单向预测色度运动补偿，调用chroma_mc_uni()；对于双向预测亮度运动补偿，调用luma_mc_bi()，对于双向预测色度运动补偿，调用chroma_mc_bi()。

而 hls_transform_tree()，

首先调用ff_hevc_split_transform_flag_decode()判断当前TU是否还需要划分。

如果需要划分的话，就会递归调用4次hls_transform_tree()分别对4个子块继续进行四叉树解析；如果不需要划分，就会调用hls_transform_unit()对TU进行解码。和前面递归划分CTU至CU的思路是一致的。最终会对每一个TU逐一调用hls_transform_unit()进行解码。

hls_transform_unit()

hls_transform_unit()用于解码一个TU，如果是帧内CU的话，hls_transform_unit()会调用HEVCPredContext的intra_pred[]()汇编函数进行帧内预测；然后不论帧内预测还是帧间CU都会调用ff_hevc_hls_residual_coding()解码残差数据，并叠加在预测数据上。