comparison vp8.c @ 12217:f6b229456bdf libavcodec

Much faster VP8 mv and mode prediction

12216:160eceee6c3d

    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
}

static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
                          VP56mv near[2], VP56mv *best, int cnt[4])
{
    VP8Macroblock *mb_edge[3] = { mb - s->mb_stride     /* top */,
                                  mb - 1                /* left */,
                                  mb - s->mb_stride - 1 /* top-left */ };
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
    VP56mv near_mv[4]  = {{ 0 }};
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
    int idx = CNT_ZERO, n;
    int best_idx = CNT_ZERO;

    /* Process MB on top, left and top-left */
    for (n = 0; n < 3; n++) {
        VP8Macroblock *edge = mb_edge[n];
        if (edge->ref_frame != VP56_FRAME_CURRENT) {
            if (edge->mv.x | edge->mv.y) {
                VP56mv tmp = edge->mv;
                if (s->sign_bias[mb->ref_frame] != s->sign_bias[edge->ref_frame]) {
                    tmp.x *= -1;
                    tmp.y *= -1;
                }
                if ((tmp.x ^ near_mv[idx].x) | (tmp.y ^ near_mv[idx].y))
                    near_mv[++idx] = tmp;
                cnt[idx]       += 1 + (n != 2);
            } else
                cnt[CNT_ZERO] += 1 + (n != 2);
        }
    }

    /* If we have three distinct MV's, merge first and last if they're the same */
    if (cnt[CNT_SPLITMV] &&
        !((near_mv[1+EDGE_TOP].x ^ near_mv[1+EDGE_TOPLEFT].x) |
          (near_mv[1+EDGE_TOP].y ^ near_mv[1+EDGE_TOPLEFT].y)))
        cnt[CNT_NEAREST] += 1;

    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);

    /* Swap near and nearest if necessary */
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
        FFSWAP(int,    cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
        FFSWAP(VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
    }

    /* Choose the best mv out of 0,0 and the nearest mv */
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
        best_idx = CNT_NEAREST;

static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
                           uint8_t *intra4x4)
{
    VP56RangeCoder *c = &s->c;
    int n;

    if (s->segmentation.update_map)
        mb->segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);

    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;

        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
        mb->ref_frame = VP56_FRAME_CURRENT;
    } else if (vp56_rac_get_prob(c, s->prob->intra)) {
        VP56mv near[2], best;
        int cnt[4] = { 0 };
        uint8_t p[4];

        // inter MB, 16.2
        if (vp56_rac_get_prob(c, s->prob->last))
            mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?

        else
            mb->ref_frame = VP56_FRAME_PREVIOUS;

        // motion vectors, 16.3
        find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
        for (n = 0; n < 4; n++)
            p[n] = vp8_mode_contexts[cnt[n]][n];
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
        switch (mb->mode) {
        case VP8_MVMODE_SPLIT:
            mb->mv = mb->bmv[decode_splitmvs(s, c, mb, &best) - 1];
            break;

12217:f6b229456bdf

    dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN),
                     ((s->mb_height - 1 - mb_y) << 6) + MARGIN);
}

static void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
                          VP56mv near[2], VP56mv *best, uint8_t cnt[4])
{
    int mb_stride = s->mb_stride;
    VP8Macroblock *mb_edge[3] = { mb - mb_stride     /* top */,
                                  mb - 1             /* left */,
                                  mb - mb_stride - 1 /* top-left */ };
    enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT };
    VP56mv near_mv[4]  = {{ 0 }};
    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };
    int idx = CNT_ZERO;
    int best_idx = CNT_ZERO;
    int cur_sign_bias = s->sign_bias[mb->ref_frame];
    int *sign_bias = s->sign_bias;

    /* Process MB on top, left and top-left */
    #define MV_EDGE_CHECK(n)\
    {\
        VP8Macroblock *edge = mb_edge[n];\
        int edge_ref = edge->ref_frame;\
        if (edge_ref != VP56_FRAME_CURRENT) {\
            uint32_t mv = AV_RN32A(&edge->mv);\
            if (mv) {\
                if (cur_sign_bias != sign_bias[edge_ref]) {\
                    /* SWAR negate of the values in mv. */\
                    mv = ((mv&0x80008000) + 0x00010001) ^ (mv&0x7fff7fff);\
                }\
                if (!n || mv != AV_RN32A(&near_mv[idx]))\
                    AV_WN32A(&near_mv[++idx], mv);\
                cnt[idx]      += 1 + (n != 2);\
            } else\
                cnt[CNT_ZERO] += 1 + (n != 2);\
        }\
    }
    MV_EDGE_CHECK(0)
    MV_EDGE_CHECK(1)
    MV_EDGE_CHECK(2)

    /* If we have three distinct MVs, merge first and last if they're the same */
    if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT]))
        cnt[CNT_NEAREST] += 1;

    cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode   == VP8_MVMODE_SPLIT) +
                        (mb_edge[EDGE_TOP]->mode    == VP8_MVMODE_SPLIT)) * 2 +
                       (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);

    /* Swap near and nearest if necessary */
    if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {
        FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);
        FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);
    }

    /* Choose the best mv out of 0,0 and the nearest mv */
    if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])
        best_idx = CNT_NEAREST;

static void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
                           uint8_t *intra4x4)
{
    VP56RangeCoder *c = &s->c;

    if (s->segmentation.update_map)
        mb->segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);

    mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;

        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
        mb->ref_frame = VP56_FRAME_CURRENT;
    } else if (vp56_rac_get_prob(c, s->prob->intra)) {
        VP56mv near[2], best;
        uint8_t cnt[4] = { 0 };
        uint8_t p[4];

        // inter MB, 16.2
        if (vp56_rac_get_prob(c, s->prob->last))
            mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?

        else
            mb->ref_frame = VP56_FRAME_PREVIOUS;

        // motion vectors, 16.3
        find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt);
        p[0] = vp8_mode_contexts[cnt[0]][0];
        p[1] = vp8_mode_contexts[cnt[1]][1];
        p[2] = vp8_mode_contexts[cnt[2]][2];
        p[3] = vp8_mode_contexts[cnt[3]][3];
        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_mvinter, p);
        switch (mb->mode) {
        case VP8_MVMODE_SPLIT:
            mb->mv = mb->bmv[decode_splitmvs(s, c, mb, &best) - 1];
            break;