Speed feature: skip obmc or wm if mv field is identical
Check if the neighbors' motions used by obmc have same parameters
as for the current block. If all the parameters are identical, we
skip trying obmc or warped motion modes.
Two speed features below are added but not turned on by default.
sf.skip_obmc_in_uniform_mv_field
sf.skip_wm_in_uniform_mv_field
If both features are turned on, at speed 0
Compression performance loss: 0.059% midres, 0.021% lowres
Speed up: ~1%
Change-Id: Ie4abfe50c0eb5b785f75b27101dcbfc3ddbb94a5
diff --git a/av1/common/reconinter.c b/av1/common/reconinter.c
index 29ed929..16c4b3c 100644
--- a/av1/common/reconinter.c
+++ b/av1/common/reconinter.c
@@ -803,6 +803,53 @@
return;
}
+struct obmc_check_mv_field_ctxt {
+ MB_MODE_INFO *current_mi;
+ int mv_field_check_result;
+};
+
+static INLINE void obmc_check_identical_mv(MACROBLOCKD *xd, int rel_mi_col,
+ uint8_t nb_mi_width,
+ MB_MODE_INFO *nb_mi, void *fun_ctxt,
+ const int num_planes) {
+ (void)xd;
+ (void)rel_mi_col;
+ (void)nb_mi_width;
+ (void)num_planes;
+ struct obmc_check_mv_field_ctxt *ctxt =
+ (struct obmc_check_mv_field_ctxt *)fun_ctxt;
+ const MB_MODE_INFO *current_mi = ctxt->current_mi;
+
+ if (ctxt->mv_field_check_result == 0) return;
+
+ if (nb_mi->ref_frame[0] != current_mi->ref_frame[0] ||
+ nb_mi->mv[0].as_int != current_mi->mv[0].as_int ||
+ nb_mi->interp_filters != current_mi->interp_filters) {
+ ctxt->mv_field_check_result = 0;
+ }
+ return;
+}
+
+// Check if the neighbors' motions used by obmc have same parameters as for
+// the current block. If all the parameters are identical, obmc will produce
+// the same prediction as from regular bmc, therefore we can skip the
+// overlapping operations for less complexity. The parameters checked include
+// reference frame, motion vector, and interpolation filter.
+int av1_check_identical_obmc_mv_field(const AV1_COMMON *cm, MACROBLOCKD *xd,
+ int mi_row, int mi_col) {
+ const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
+ struct obmc_check_mv_field_ctxt mv_field_check_ctxt = { xd->mi[0], 1 };
+
+ foreach_overlappable_nb_above(cm, xd, mi_col,
+ max_neighbor_obmc[mi_size_wide_log2[bsize]],
+ obmc_check_identical_mv, &mv_field_check_ctxt);
+ foreach_overlappable_nb_left(cm, xd, mi_row,
+ max_neighbor_obmc[mi_size_high_log2[bsize]],
+ obmc_check_identical_mv, &mv_field_check_ctxt);
+
+ return mv_field_check_ctxt.mv_field_check_result;
+}
+
struct obmc_inter_pred_ctxt {
uint8_t **adjacent;
int *adjacent_stride;
diff --git a/av1/common/reconinter.h b/av1/common/reconinter.h
index 5550cf2..56645d0 100644
--- a/av1/common/reconinter.h
+++ b/av1/common/reconinter.h
@@ -161,6 +161,8 @@
void av1_modify_neighbor_predictor_for_obmc(MB_MODE_INFO *mbmi);
int av1_skip_u4x4_pred_in_obmc(BLOCK_SIZE bsize,
const struct macroblockd_plane *pd, int dir);
+int av1_check_identical_obmc_mv_field(const AV1_COMMON *cm, MACROBLOCKD *xd,
+ int mi_row, int mi_col);
static INLINE int is_interinter_compound_used(COMPOUND_TYPE type,
BLOCK_SIZE sb_type) {
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index f9e8284..371aa90 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -8923,6 +8923,11 @@
av1_is_interp_needed(xd) ? av1_get_switchable_rate(cm, x, xd) : 0;
int64_t best_rd = INT64_MAX;
int best_rate_mv = rate_mv0;
+ int identical_obmc_mv_field_detected =
+ (cpi->sf.skip_obmc_in_uniform_mv_field ||
+ cpi->sf.skip_wm_in_uniform_mv_field)
+ ? av1_check_identical_obmc_mv_field(cm, xd, mi_row, mi_col)
+ : 0;
for (int mode_index = (int)SIMPLE_TRANSLATION;
mode_index <= (int)last_motion_mode_allowed + interintra_allowed;
mode_index++) {
@@ -8944,6 +8949,15 @@
assert(mbmi->ref_frame[1] != INTRA_FRAME);
}
+ if (identical_obmc_mv_field_detected) {
+ if (cpi->sf.skip_obmc_in_uniform_mv_field &&
+ mbmi->motion_mode == OBMC_CAUSAL)
+ continue;
+ if (cpi->sf.skip_wm_in_uniform_mv_field &&
+ mbmi->motion_mode == WARPED_CAUSAL)
+ continue;
+ }
+
if (mbmi->motion_mode == SIMPLE_TRANSLATION && !is_interintra_mode) {
// SIMPLE_TRANSLATION mode: no need to recalculate.
// The prediction is calculated before motion_mode_rd() is called in
diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index ef1c858..7990710 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -464,6 +464,8 @@
sf->use_jnt_comp_flag = JNT_COMP_ENABLED;
sf->reuse_inter_intra_mode = 0;
sf->intra_angle_estimation = 0;
+ sf->skip_obmc_in_uniform_mv_field = 0;
+ sf->skip_wm_in_uniform_mv_field = 0;
for (i = 0; i < TX_SIZES; i++) {
sf->intra_y_mode_mask[i] = INTRA_ALL;
diff --git a/av1/encoder/speed_features.h b/av1/encoder/speed_features.h
index 070d199..a46a912 100644
--- a/av1/encoder/speed_features.h
+++ b/av1/encoder/speed_features.h
@@ -580,6 +580,11 @@
// Performs full pixel motion search before none_partition to decide if we
// want to split directly without trying other partition types.
int full_pixel_motion_search_based_split;
+
+ // Skip obmc or warped motion mode when neighborhood motion field is
+ // identical
+ int skip_obmc_in_uniform_mv_field;
+ int skip_wm_in_uniform_mv_field;
} SPEED_FEATURES;
struct AV1_COMP;
