Improve ALL_ZERO flag handling with global motion
Change-Id: I9b2465807c2dd6e356807f3f3555394b382918ad
diff --git a/av1/common/mvref_common.c b/av1/common/mvref_common.c
index 36e916f..542ec5c 100644
--- a/av1/common/mvref_common.c
+++ b/av1/common/mvref_common.c
@@ -499,7 +499,8 @@
MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
int_mv *mv_ref_list, int block, int mi_row,
int mi_col, find_mv_refs_sync sync,
- void *const data, int16_t *mode_context) {
+ void *const data, int16_t *mode_context,
+ int_mv zeromv) {
const int *ref_sign_bias = cm->ref_frame_sign_bias;
int i, refmv_count = 0;
#if !CONFIG_REF_MV
@@ -661,7 +662,7 @@
if (mode_context)
mode_context[ref_frame] = counter_to_context[context_counter];
for (i = refmv_count; i < MAX_MV_REF_CANDIDATES; ++i)
- mv_ref_list[i].as_int = 0;
+ mv_ref_list[i].as_int = zeromv.as_int;
}
#if CONFIG_EXT_INTER
@@ -745,11 +746,12 @@
int_mv *mv_ref_list, int mi_row, int mi_col,
find_mv_refs_sync sync, void *const data,
int16_t *mode_context) {
+ int_mv zeromv[2];
#if CONFIG_REF_MV
-#if !CONFIG_GLOBAL_MOTION
- int idx;
+ int idx, all_zero = 1;
#endif
- int all_zero = 1;
+#if CONFIG_GLOBAL_MOTION
+ MV_REFERENCE_FRAME rf[2];
#endif
#if CONFIG_EXT_INTER
av1_update_mv_context(xd, mi, ref_frame, mv_ref_list, -1, mi_row, mi_col,
@@ -759,33 +761,51 @@
mode_context);
#endif // CONFIG_REF_MV
#endif // CONFIG_EXT_INTER
+
+#if CONFIG_GLOBAL_MOTION
+ av1_set_ref_frame(rf, ref_frame);
+ zeromv[0].as_int = gm_get_motion_vector(&cm->global_motion[rf[0]]).as_int;
+ zeromv[1].as_int =
+ (rf[1] != NONE) ? gm_get_motion_vector(&cm->global_motion[rf[1]]).as_int
+ : 0;
+#else
+ zeromv[0].as_int = zeromv[1].as_int = 0;
+#endif
+
#if CONFIG_REF_MV
if (ref_frame <= ALTREF_FRAME)
- find_mv_refs_idx(cm, xd, mi, ref_frame, mv_ref_list, -1, mi_row, mi_col,
- sync, data, mode_context);
-#else
- find_mv_refs_idx(cm, xd, mi, ref_frame, mv_ref_list, -1, mi_row, mi_col, sync,
- data, mode_context);
#endif // CONFIG_REF_MV
+ find_mv_refs_idx(cm, xd, mi, ref_frame, mv_ref_list, -1, mi_row, mi_col,
+ sync, data, mode_context, zeromv[0]);
#if CONFIG_REF_MV
setup_ref_mv_list(cm, xd, ref_frame, ref_mv_count, ref_mv_stack, mv_ref_list,
-1, mi_row, mi_col, mode_context);
-
-#if CONFIG_GLOBAL_MOTION
- if ((*ref_mv_count >= 2) || (ref_frame <= ALTREF_FRAME)) all_zero = 0;
-#else
+ /* Note: If global motion is enabled, then we want to set the ALL_ZERO flag
+ iff all of the MVs we could generate with NEARMV/NEARESTMV are equivalent
+ to the global motion vector.
+ Note: For the following to work properly, the encoder can't throw away
+ any global motion models after calling this function, even if they are
+ unused. Instead we rely on the recode loop: If any non-IDENTITY model
+ is unused, the whole frame will be re-encoded without it.
+ The problem is that, otherwise, we can end up in the following situation:
+ * Encoder has a global motion model with nonzero translational part,
+ and all candidate MVs are zero. So the ALL_ZERO flag is unset.
+ * Encoder throws away global motion because it is never used.
+ * Decoder sees that there is no global motion and all candidate MVs are
+ zero, so sets the ALL_ZERO flag.
+ * This leads to an encode/decode mismatch.
+ */
if (*ref_mv_count >= 2) {
for (idx = 0; idx < AOMMIN(3, *ref_mv_count); ++idx) {
- if (ref_mv_stack[idx].this_mv.as_int != 0) all_zero = 0;
+ if (ref_mv_stack[idx].this_mv.as_int != zeromv[0].as_int) all_zero = 0;
if (ref_frame > ALTREF_FRAME)
- if (ref_mv_stack[idx].comp_mv.as_int != 0) all_zero = 0;
+ if (ref_mv_stack[idx].comp_mv.as_int != zeromv[1].as_int) all_zero = 0;
}
} else if (ref_frame <= ALTREF_FRAME) {
for (idx = 0; idx < MAX_MV_REF_CANDIDATES; ++idx)
- if (mv_ref_list[idx].as_int != 0) all_zero = 0;
+ if (mv_ref_list[idx].as_int != zeromv[0].as_int) all_zero = 0;
}
-#endif
if (all_zero) mode_context[ref_frame] |= (1 << ALL_ZERO_FLAG_OFFSET);
#endif
@@ -818,6 +838,7 @@
MODE_INFO *const mi = xd->mi[0];
b_mode_info *bmi = mi->bmi;
int n;
+ int_mv zeromv;
#if CONFIG_REF_MV
CANDIDATE_MV tmp_mv;
uint8_t idx;
@@ -828,8 +849,13 @@
assert(MAX_MV_REF_CANDIDATES == 2);
+#if CONFIG_GLOBAL_MOTION
+ zeromv.as_int = gm_get_motion_vector(&cm->global_motion[ref]).as_int;
+#else
+ zeromv.as_int = 0;
+#endif
find_mv_refs_idx(cm, xd, mi, mi->mbmi.ref_frame[ref], mv_list, block, mi_row,
- mi_col, NULL, NULL, NULL);
+ mi_col, NULL, NULL, NULL, zeromv);
#if CONFIG_REF_MV
scan_blk_mbmi(cm, xd, mi_row, mi_col, block, rf, -1, 0, ref_mv_stack,
diff --git a/av1/decoder/decodemv.c b/av1/decoder/decodemv.c
index 101ed3e..826d30d 100644
--- a/av1/decoder/decodemv.c
+++ b/av1/decoder/decodemv.c
@@ -1479,14 +1479,26 @@
if (xd->ref_mv_count[ref_frame] < 2) {
MV_REFERENCE_FRAME rf[2];
+ int_mv zeromv[2];
av1_set_ref_frame(rf, ref_frame);
+#if CONFIG_GLOBAL_MOTION
+ zeromv[0].as_int = gm_get_motion_vector(&cm->global_motion[rf[0]]).as_int;
+ zeromv[1].as_int =
+ (rf[1] != NONE)
+ ? gm_get_motion_vector(&cm->global_motion[rf[1]]).as_int
+ : 0;
+#else
+ zeromv[0].as_int = zeromv[1].as_int = 0;
+#endif
for (ref = 0; ref < 2; ++ref) {
lower_mv_precision(&ref_mvs[rf[ref]][0].as_mv, allow_hp);
lower_mv_precision(&ref_mvs[rf[ref]][1].as_mv, allow_hp);
}
- if (ref_mvs[rf[0]][0].as_int != 0 || ref_mvs[rf[0]][1].as_int != 0 ||
- ref_mvs[rf[1]][0].as_int != 0 || ref_mvs[rf[1]][1].as_int != 0)
+ if (ref_mvs[rf[0]][0].as_int != zeromv[0].as_int ||
+ ref_mvs[rf[0]][1].as_int != zeromv[0].as_int ||
+ ref_mvs[rf[1]][0].as_int != zeromv[1].as_int ||
+ ref_mvs[rf[1]][1].as_int != zeromv[1].as_int)
inter_mode_ctx[ref_frame] &= ~(1 << ALL_ZERO_FLAG_OFFSET);
}
}
diff --git a/av1/encoder/bitstream.c b/av1/encoder/bitstream.c
index 4915f78..acce688 100644
--- a/av1/encoder/bitstream.c
+++ b/av1/encoder/bitstream.c
@@ -4094,9 +4094,14 @@
AV1_COMMON *const cm = &cpi->common;
int frame;
for (frame = LAST_FRAME; frame <= ALTREF_FRAME; ++frame) {
+#if !CONFIG_REF_MV
+ // With ref-mv, clearing unused global motion models here is
+ // unsafe, and we need to rely on the recode loop to do it
+ // instead. See av1_find_mv_refs for details.
if (!cpi->global_motion_used[frame]) {
set_default_gmparams(&cm->global_motion[frame]);
}
+#endif
write_global_motion_params(&cm->global_motion[frame],
cm->fc->global_motion_types_prob, w);
/*
diff --git a/av1/encoder/encoder.c b/av1/encoder/encoder.c
index dd65ee6..81122c7 100644
--- a/av1/encoder/encoder.c
+++ b/av1/encoder/encoder.c
@@ -2889,7 +2889,11 @@
if (cm->global_motion[i].wmtype != IDENTITY &&
cpi->global_motion_used[i] < MIN_GLOBAL_MOTION_BLKS) {
set_default_gmparams(&cm->global_motion[i]);
+#if CONFIG_REF_MV
+ recode = 1;
+#else
recode |= (cpi->global_motion_used[i] > 0);
+#endif
}
}
return recode;
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index 0f5db59..98c5455 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -8611,10 +8611,13 @@
if (mbmi_ext->ref_mv_count[ref_frame] < 2) {
MV_REFERENCE_FRAME rf[2];
av1_set_ref_frame(rf, ref_frame);
- if (mbmi_ext->ref_mvs[rf[0]][0].as_int != 0 ||
- mbmi_ext->ref_mvs[rf[0]][1].as_int != 0 ||
- mbmi_ext->ref_mvs[rf[1]][0].as_int != 0 ||
- mbmi_ext->ref_mvs[rf[1]][1].as_int != 0)
+ if (mbmi_ext->ref_mvs[rf[0]][0].as_int !=
+ frame_mv[ZEROMV][rf[0]].as_int ||
+ mbmi_ext->ref_mvs[rf[0]][1].as_int !=
+ frame_mv[ZEROMV][rf[0]].as_int ||
+ mbmi_ext->ref_mvs[rf[1]][0].as_int !=
+ frame_mv[ZEROMV][rf[1]].as_int ||
+ mbmi_ext->ref_mvs[rf[1]][1].as_int != frame_mv[ZEROMV][rf[1]].as_int)
mbmi_ext->mode_context[ref_frame] &= ~(1 << ALL_ZERO_FLAG_OFFSET);
}
}
@@ -9969,27 +9972,34 @@
}
#if CONFIG_REF_MV
- if (best_mbmode.ref_frame[0] > INTRA_FRAME && best_mbmode.mv[0].as_int == 0 &&
-#if CONFIG_EXT_INTER
- (best_mbmode.ref_frame[1] <= INTRA_FRAME)
-#else
- (best_mbmode.ref_frame[1] == NONE || best_mbmode.mv[1].as_int == 0)
-#endif // CONFIG_EXT_INTER
- ) {
+ {
int8_t ref_frame_type = av1_ref_frame_type(best_mbmode.ref_frame);
int16_t mode_ctx = mbmi_ext->mode_context[ref_frame_type];
-
if (mode_ctx & (1 << ALL_ZERO_FLAG_OFFSET)) {
- best_mbmode.mode = ZEROMV;
+ int_mv zeromv[2];
#if CONFIG_GLOBAL_MOTION
- best_mbmode.mv[0].as_int =
- gm_get_motion_vector(&cm->global_motion[best_mbmode.ref_frame[0]])
- .as_int;
- if (best_mbmode.ref_frame[1] != NONE)
- best_mbmode.mv[1].as_int =
- gm_get_motion_vector(&cm->global_motion[best_mbmode.ref_frame[1]])
- .as_int;
-#endif
+ const MV_REFERENCE_FRAME refs[2] = { best_mbmode.ref_frame[0],
+ best_mbmode.ref_frame[1] };
+ zeromv[0].as_int =
+ gm_get_motion_vector(&cm->global_motion[refs[0]]).as_int;
+ zeromv[1].as_int =
+ gm_get_motion_vector(&cm->global_motion[refs[1]]).as_int;
+ lower_mv_precision(&zeromv[0].as_mv, cm->allow_high_precision_mv);
+ lower_mv_precision(&zeromv[1].as_mv, cm->allow_high_precision_mv);
+#else
+ zeromv[0].as_int = zeromv[1].as_int = 0;
+#endif // CONFIG_GLOBAL_MOTION
+ if (best_mbmode.ref_frame[0] > INTRA_FRAME &&
+ best_mbmode.mv[0].as_int == zeromv[0].as_int &&
+#if CONFIG_EXT_INTER
+ (best_mbmode.ref_frame[1] <= INTRA_FRAME)
+#else
+ (best_mbmode.ref_frame[1] == NONE ||
+ best_mbmode.mv[1].as_int == zeromv[1].as_int)
+#endif // CONFIG_EXT_INTER
+ ) {
+ best_mbmode.mode = ZEROMV;
+ }
}
}
#endif
