blob: 936c11064f5ba892e67b362aba2dd8667e2b8842 [file] [log] [blame]
#### 16.2 Inter-Predicted Macroblocks {#h-16-02}
Otherwise (when the above bool is true), we are using inter-
prediction (which of course only happens for interframes), to which
we now restrict our attention.
The next datum is then another bool, `B( prob_last)`, selecting the
reference frame. If 0, the reference frame is the previous frame
(the last frame); if 1, another bool (`prob_gf`) selects the reference
frame between the golden frame (0) and the altref frame (1). The
probabilities `prob_last` and `prob_gf` are set in field J of the frame
Together with setting the reference frame, the purpose of inter-mode
decoding is to set a motion vector for each of the sixteen Y
subblocks of the current macroblock. These settings then define the
calculation of the inter-prediction buffer (detailed in Section 18).
While the net effect of inter-mode decoding is straightforward, the
implementation is somewhat complex; the (lossless) compression
achieved by this method justifies the complexity.
After the reference frame selector comes the mode (or motion vector
reference) applied to the macroblock as a whole, coded using the
following enumeration and tree. Setting `mv_nearest = num_ymodes` is a
convenience that allows a single variable to unambiguously hold an
inter- or intra-prediction mode.
typedef enum
mv_nearest = num_ymodes, /* use "nearest" motion vector
for entire MB */
mv_near, /* use "next nearest" "" */
mv_zero, /* use zero "" */
mv_new, /* use explicit offset from
implicit "" */
mv_split, /* use multiple motion vectors */
num_mv_refs = mv_split + 1 - mv_nearest
const tree_index mv_ref_tree [2 * (num_mv_refs - 1)] =
-mv_zero, 2, /* zero = "0" */
-mv_nearest, 4, /* nearest = "10" */
-mv_near, 6, /* near = "110" */
-mv_new, -mv_split /* new = "1110", split = "1111" */