comparison h264_cabac.c @ 11004:1c3b1752d6d6 libavcodec

Optimize decode_cabac_field_decoding_flag(). ~4 cpu cycles faster

11003:72c026446d67

    }
}

static int decode_cabac_field_decoding_flag(H264Context *h) {
    MpegEncContext * const s = &h->s;
    const int mb_x = s->mb_x;
    const int mb_y = s->mb_y & ~1;
    const int mba_xy = mb_x - 1 +  mb_y   *s->mb_stride;
    const int mbb_xy = mb_x     + (mb_y-2)*s->mb_stride;

    unsigned int ctx = 0;

    if( h->slice_table[mba_xy] == h->slice_num && IS_INTERLACED( s->current_picture.mb_type[mba_xy] ) ) {
        ctx += 1;
    }
    if( h->slice_table[mbb_xy] == h->slice_num && IS_INTERLACED( s->current_picture.mb_type[mbb_xy] ) ) {
        ctx += 1;
    }

    return get_cabac_noinline( &h->cabac, &h->cabac_state[70 + ctx] );
}

static int decode_cabac_intra_mb_type(H264Context *h, int ctx_base, int intra_slice) {
    uint8_t *state= &h->cabac_state[ctx_base];
    int mb_type;

11004:1c3b1752d6d6

    }
}

static int decode_cabac_field_decoding_flag(H264Context *h) {
    MpegEncContext * const s = &h->s;
    const long mba_xy = h->mb_xy - 1L;
    const long mbb_xy = h->mb_xy - 2L*s->mb_stride;

    unsigned long ctx = 0;

    ctx += (s->current_picture.mb_type[mba_xy]>>7)&(h->slice_table[mba_xy] == h->slice_num);
    ctx += (s->current_picture.mb_type[mbb_xy]>>7)&(h->slice_table[mbb_xy] == h->slice_num);

    return get_cabac_noinline( &h->cabac, &(h->cabac_state+70)[ctx] );
}

static int decode_cabac_intra_mb_type(H264Context *h, int ctx_base, int intra_slice) {
    uint8_t *state= &h->cabac_state[ctx_base];
    int mb_type;