Remove secondary mv clamping from decode stage

This patch removes the secondary MV clamping from the MV decoder. This
behavior was consistent with limits placed on non-split MVs by the
reference encoder, but was inconsistent with the MVs generated in the
split case.

The purpose of this secondary clamping was only to prevent crashes on
invalid data. It was not intended to be a behaviour an encoder could or
should rely on. Instead of doing additional clamping in a way that
changes the entropy context, the secondary clamp is removed and the
border handling is made implmentation specific. With respect to the
spec, the border is treated as essentially infinite, limited only by
the clamping performed on the near/nearest reference and the maximum
encodable magnitude of the residual MV.

This does not affect any currently produced streams.

Change-Id: I68d35a2fbb51570d6569eab4ad233961405230a3
This commit is contained in:
John Koleszar 2010-06-09 11:29:20 -04:00
parent 4bb895e854
commit 3085025fa1
4 changed files with 117 additions and 39 deletions

View File

@ -174,9 +174,8 @@ typedef struct
int dc_diff;
unsigned char segment_id; // Which set of segmentation parameters should be used for this MB
int force_no_skip;
int need_to_clamp_mvs;
B_MODE_INFO partition_bmi[16];
} MB_MODE_INFO;

View File

@ -171,6 +171,7 @@ void vp8_decode_mode_mvs(VP8D_COMP *pbi)
VP8_COMMON *const pc = &pbi->common;
MACROBLOCKD *xd = &pbi->mb;
mbmi->need_to_clamp_mvs = 0;
vp8dx_bool_decoder_fill(bc);
// Distance of Mb to the various image edges.
@ -269,6 +270,17 @@ void vp8_decode_mode_mvs(VP8D_COMP *pbi)
break;
}
if (mv->col < xd->mb_to_left_edge
- LEFT_TOP_MARGIN
|| mv->col > xd->mb_to_right_edge
+ RIGHT_BOTTOM_MARGIN
|| mv->row < xd->mb_to_top_edge
- LEFT_TOP_MARGIN
|| mv->row > xd->mb_to_bottom_edge
+ RIGHT_BOTTOM_MARGIN
)
mbmi->need_to_clamp_mvs = 1;
/* Fill (uniform) modes, mvs of jth subset.
Must do it here because ensuing subsets can
refer back to us via "left" or "above". */
@ -325,27 +337,18 @@ void vp8_decode_mode_mvs(VP8D_COMP *pbi)
read_mv(bc, mv, (const MV_CONTEXT *) mvc);
mv->row += best_mv.row;
mv->col += best_mv.col;
/* Encoder should not produce invalid motion vectors, but since
* arbitrary length MVs can be parsed from the bitstream, we
* need to clamp them here in case we're reading bad data to
* avoid a crash.
/* Don't need to check this on NEARMV and NEARESTMV modes
* since those modes clamp the MV. The NEWMV mode does not,
* so signal to the prediction stage whether special
* handling may be required.
*/
#if CONFIG_DEBUG
assert(mv->col >= (xd->mb_to_left_edge - LEFT_TOP_MARGIN));
assert(mv->col <= (xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN));
assert(mv->row >= (xd->mb_to_top_edge - LEFT_TOP_MARGIN));
assert(mv->row <= (xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN));
#endif
if (mv->col < (xd->mb_to_left_edge - LEFT_TOP_MARGIN))
mv->col = xd->mb_to_left_edge - LEFT_TOP_MARGIN;
else if (mv->col > xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN)
mv->col = xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN;
if (mv->row < (xd->mb_to_top_edge - LEFT_TOP_MARGIN))
mv->row = xd->mb_to_top_edge - LEFT_TOP_MARGIN;
else if (mv->row > xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN)
mv->row = xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN;
if (mv->col < xd->mb_to_left_edge - LEFT_TOP_MARGIN
|| mv->col > xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN
|| mv->row < xd->mb_to_top_edge - LEFT_TOP_MARGIN
|| mv->row > xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN
)
mbmi->need_to_clamp_mvs = 1;
propagate_mv: /* same MV throughout */
{
@ -381,7 +384,6 @@ void vp8_decode_mode_mvs(VP8D_COMP *pbi)
assert(0);
#endif
}
}
else
{

View File

@ -126,6 +126,47 @@ static void skip_recon_mb(VP8D_COMP *pbi, MACROBLOCKD *xd)
}
}
static void clamp_mv_to_umv_border(MV *mv, const MACROBLOCKD *xd)
{
/* If the MV points so far into the UMV border that no visible pixels
* are used for reconstruction, the subpel part of the MV can be
* discarded and the MV limited to 16 pixels with equivalent results.
*
* This limit kicks in at 19 pixels for the top and left edges, for
* the 16 pixels plus 3 taps right of the central pixel when subpel
* filtering. The bottom and right edges use 16 pixels plus 2 pixels
* left of the central pixel when filtering.
*/
if (mv->col < (xd->mb_to_left_edge - (19 << 3)))
mv->col = xd->mb_to_left_edge - (16 << 3);
else if (mv->col > xd->mb_to_right_edge + (18 << 3))
mv->col = xd->mb_to_right_edge + (16 << 3);
if (mv->row < (xd->mb_to_top_edge - (19 << 3)))
mv->row = xd->mb_to_top_edge - (16 << 3);
else if (mv->row > xd->mb_to_bottom_edge + (18 << 3))
mv->row = xd->mb_to_bottom_edge + (16 << 3);
}
static void clamp_mvs(MACROBLOCKD *xd)
{
if (xd->mbmi.mode == SPLITMV)
{
int i;
for (i=0; i<16; i++)
clamp_mv_to_umv_border(&xd->block[i].bmi.mv.as_mv, xd);
}
else
{
clamp_mv_to_umv_border(&xd->mbmi.mv.as_mv, xd);
clamp_mv_to_umv_border(&xd->block[16].bmi.mv.as_mv, xd);
}
}
static void reconstruct_mb(VP8D_COMP *pbi, MACROBLOCKD *xd)
{
if (xd->frame_type == KEY_FRAME || xd->mbmi.ref_frame == INTRA_FRAME)
@ -233,6 +274,8 @@ static void de_quantand_idct(VP8D_COMP *pbi, MACROBLOCKD *xd)
void vp8_decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd)
{
int eobtotal = 0;
MV orig_mvs[24];
int i, do_clamp = xd->mbmi.need_to_clamp_mvs;
if (xd->mbmi.mb_skip_coeff)
{
@ -243,13 +286,28 @@ void vp8_decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd)
eobtotal = vp8_decode_mb_tokens(pbi, xd);
}
/* Perform temporary clamping of the MV to be used for prediction */
if (do_clamp)
{
if (xd->mbmi.mode == SPLITMV)
for (i=0; i<24; i++)
orig_mvs[i] = xd->block[i].bmi.mv.as_mv;
else
{
orig_mvs[0] = xd->mbmi.mv.as_mv;
orig_mvs[1] = xd->block[16].bmi.mv.as_mv;
}
clamp_mvs(xd);
}
xd->mode_info_context->mbmi.dc_diff = 1;
do {
if (xd->mbmi.mode != B_PRED && xd->mbmi.mode != SPLITMV && eobtotal == 0)
{
xd->mode_info_context->mbmi.dc_diff = 0;
skip_recon_mb(pbi, xd);
return;
break;
}
if (xd->segmentation_enabled)
@ -257,6 +315,21 @@ void vp8_decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd)
de_quantand_idct(pbi, xd);
reconstruct_mb(pbi, xd);
} while(0);
/* Restore the original MV so as not to affect the entropy context. */
if (do_clamp)
{
if (xd->mbmi.mode == SPLITMV)
for (i=0; i<24; i++)
xd->block[i].bmi.mv.as_mv = orig_mvs[i];
else
{
xd->mbmi.mv.as_mv = orig_mvs[0];
xd->block[16].bmi.mv.as_mv = orig_mvs[1];
}
}
}
static int get_delta_q(vp8_reader *bc, int prev, int *q_update)
@ -314,7 +387,9 @@ void vp8_decode_mb_row(VP8D_COMP *pbi,
for (mb_col = 0; mb_col < pc->mb_cols; mb_col++)
{
// Take a copy of the mode and Mv information for this macroblock into the xd->mbmi
vpx_memcpy(&xd->mbmi, &xd->mode_info_context->mbmi, 32); //sizeof(MB_MODE_INFO) );
// the partition_bmi array is unused in the decoder, so don't copy it.
vpx_memcpy(&xd->mbmi, &xd->mode_info_context->mbmi,
sizeof(MB_MODE_INFO) - sizeof(xd->mbmi.partition_bmi));
if (xd->mbmi.mode == SPLITMV || xd->mbmi.mode == B_PRED)
{

View File

@ -154,7 +154,9 @@ THREAD_FUNCTION vp8_thread_decoding_proc(void *p_data)
}
// Take a copy of the mode and Mv information for this macroblock into the xd->mbmi
vpx_memcpy(&xd->mbmi, &xd->mode_info_context->mbmi, 32); //sizeof(MB_MODE_INFO) );
// the partition_bmi array is unused in the decoder, so don't copy it.
vpx_memcpy(&xd->mbmi, &xd->mode_info_context->mbmi,
sizeof(MB_MODE_INFO) - sizeof(xd->mbmi.partition_bmi));
if (xd->mbmi.mode == SPLITMV || xd->mbmi.mode == B_PRED)
{