text_src/16.04__vp8-bitstream__split-prediction.txt

#### 16.4 Split Prediction                                 {#h-16-04}


The remaining mode (`SPLITMV`) causes multiple vectors to be applied to
the Y subblocks.  It is immediately followed by a partition
specification that determines how many vectors will be specified and
how they will be assigned to the subblocks.  The possible partitions,
with indicated subdivisions and coding tree, are as follows.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
typedef enum
{
    mv_top_bottom,   /* two pieces {0...7} and {8...15} */
    mv_left_right,   /* {0,1,4,5,8,9,12,13} and
                        {2,3,6,7,10,11,14,15} */
    mv_quarters,    /* {0,1,4,5}, {2,3,6,7}, {8,9,12,13},
                       {10,11,14,15} */
    MV_16,          /* every subblock gets its own vector
                       {0} ... {15} */

    mv_num_partitions
}
MVpartition;

const tree_index mvpartition_tree [2 * (mvnum_partition - 1)] =
{
 -MV_16, 2,                         /* MV_16 = "0" */
  -mv_quarters, 4,                  /* mv_quarters = "10" */
   -mv_top_bottom, -mv_left_right   /* top_bottom = "110",
                                       left_right = "111" */
};
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
{:lang="c"}


The partition is decoded using a fixed, constant probability table:


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
const Prob mvpartition_probs [mvnum_partition - 1] =
  { 110, 111, 150};
part = (MVpartition) treed_read( d, mvpartition_tree,
  mvpartition_probs);
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
{:lang="c"}


After the partition come two (for `mv_top_bottom` or `mv_left_right`),
four (for `mv_quarters`), or sixteen (for `MV_16`) subblock inter-
prediction modes.  These modes occur in the order indicated by the
partition layouts (given as comments to the `MVpartition` enum) and are
coded as follows.  (As was done for the macroblock-level modes, we
offset the mode enumeration so that a single variable may
unambiguously hold either an intra- or inter-subblock mode.)

Prior to decoding each subblock, a decoding tree context is chosen as
illustrated in the code snippet below.  The context is based on the
immediate left and above subblock neighbors, and whether they are
equal, are zero, or a combination of those.


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
typedef enum
{
    LEFT4x4 = num_intra_bmodes,   /* use already-coded MV to
                                     my left */
    ABOVE4x4,             /* use already-coded MV above me */
    ZERO4x4,              /* use zero MV */
    NEW4x4,               /* explicit offset from "best" */

    num_sub_mv_ref
};
sub_mv_ref;

const tree_index sub_mv_ref_tree [2 * (num_sub_mv_ref - 1)] =
{
 -LEFT4X4, 2,           /* LEFT = "0" */
  -ABOVE4X4, 4,         /* ABOVE = "10" */
   -ZERO4X4, -NEW4X4    /* ZERO = "110", NEW = "111" */
};

/* Choose correct decoding tree context
 * Function parameters are left subblock neighbor MV and above
 * subblock neighbor MV */
int vp8_mvCont(MV *l, MV*a)
{
    int lez = (l->row == 0 && l->col == 0);   /* left neighbor
                                                 is zero */
    int aez = (a->row == 0 && a->col == 0);   /* above neighbor
                                                 is zero */
    int lea = (l->row == a->row && l->col == a->col);  /* left
                             neighbor equals above neighbor */

    if(lea && lez)
        return SUBMVREF_LEFT_ABOVE_ZED; /* =4 */

    if(lea)
        return SUBMVREF_LEFT_ABOVE_SAME; /* =3 */

    if(aez)
        return SUBMVREF_ABOVE_ZED; /* =2 */

    if(lez)
        return SUBMVREF_LEFT_ZED; /* =1*/

    return SUBMVREF_NORMAL; /* =0 */
}

/* Constant probabilities and decoding procedure. */

const Prob sub_mv_ref_prob [5][num_sub_mv_ref - 1] = {
    { 147,136,18 },
    { 106,145,1  },
    { 179,121,1  },
    { 223,1  ,34 },
    { 208,1  ,1  }
 };

    sub_ref = (sub_mv_ref) treed_read( d, sub_mv_ref_tree,
      sub_mv_ref_prob[context]);
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
{:lang="c"}


The first two sub-prediction modes simply copy the already-coded
motion vectors used by the blocks above and to the left of the
subblock at the upper left corner of the current subset (i.e.,
collection of subblocks being predicted).  These prediction blocks
need not lie in the current macroblock and, if the current subset
lies at the top or left edges of the frame, need not lie in the
frame.  In this latter case, their motion vectors are taken to be
zero, as are subblock motion vectors within an intra-predicted
macroblock.  Also, to ensure the correctness of prediction within
this macroblock, all subblocks lying in an already-decoded subset of
the current macroblock must have their motion vectors set.

`ZERO4x4` uses a zero motion vector and predicts the current subset
using the corresponding subset from the prediction frame.

`NEW4x4` is exactly like `NEWMV` except that `NEW4x4` is applied only to
the current subset.  It is followed by a two-dimensional motion
vector offset (described in the next section) that is added to the
best vector returned by the earlier call to `find_near_mvs` to form the
motion vector in effect for the subset.

Parsing of both inter-prediction modes and motion vectors (described
next) can be found in the reference decoder file `modemv.c`
(Section 20.11).