Refactor and reduce mfmv structure size by 3x
The DRL system would only access the first valid mfmv in stack per
8x8 block position. This commit refactors the mfmv access process
and reduces the needed data structure size by 3x. The cumulative
decoding time spent on av1_find_mv_refs() at decoder end is reduced
by 5%. (cumulative 43 ms -> 40.9 ms for pedestrian_1080p at 2000
kbps).
Change-Id: I2ceb9edd1fe4a1f85c2e01b7712bd8d9d83b3c63
diff --git a/av1/common/mvref_common.c b/av1/common/mvref_common.c
index a2bf92a..a720212 100644
--- a/av1/common/mvref_common.c
+++ b/av1/common/mvref_common.c
@@ -419,43 +419,40 @@
#endif
CANDIDATE_MV *ref_mv_stack = ref_mv_stacks[rf[0]];
- for (int i = 0; i < MFMV_STACK_SIZE; ++i) {
- if (prev_frame_mvs->mfmv0[i].as_int != INVALID_MV) {
- int_mv this_refmv;
+ if (prev_frame_mvs->mfmv0.as_int != INVALID_MV) {
+ int_mv this_refmv;
- get_mv_projection(&this_refmv.as_mv, prev_frame_mvs->mfmv0[i].as_mv,
- cur_offset_0, prev_frame_mvs->ref_frame_offset[i]);
+ get_mv_projection(&this_refmv.as_mv, prev_frame_mvs->mfmv0.as_mv,
+ cur_offset_0, prev_frame_mvs->ref_frame_offset);
#if CONFIG_AMVR
- lower_mv_precision(&this_refmv.as_mv, cm->allow_high_precision_mv,
- cm->cur_frame_force_integer_mv);
+ lower_mv_precision(&this_refmv.as_mv, cm->allow_high_precision_mv,
+ cm->cur_frame_force_integer_mv);
#else
- lower_mv_precision(&this_refmv.as_mv, cm->allow_high_precision_mv);
+ lower_mv_precision(&this_refmv.as_mv, cm->allow_high_precision_mv);
#endif
- if (blk_row == 0 && blk_col == 0)
- if (abs(this_refmv.as_mv.row - gm_mv_candidates[0].as_mv.row) >= 16 ||
- abs(this_refmv.as_mv.col - gm_mv_candidates[0].as_mv.col) >= 16)
- mode_context[ref_frame] |= (1 << GLOBALMV_OFFSET);
+ if (blk_row == 0 && blk_col == 0)
+ if (abs(this_refmv.as_mv.row - gm_mv_candidates[0].as_mv.row) >= 16 ||
+ abs(this_refmv.as_mv.col - gm_mv_candidates[0].as_mv.col) >= 16)
+ mode_context[ref_frame] |= (1 << GLOBALMV_OFFSET);
- for (idx = 0; idx < refmv_count[rf[0]]; ++idx)
- if (this_refmv.as_int == ref_mv_stack[idx].this_mv.as_int) break;
+ for (idx = 0; idx < refmv_count[rf[0]]; ++idx)
+ if (this_refmv.as_int == ref_mv_stack[idx].this_mv.as_int) break;
- if (idx < refmv_count[rf[0]])
- ref_mv_stack[idx].weight += 2 * weight_unit;
+ if (idx < refmv_count[rf[0]]) ref_mv_stack[idx].weight += 2 * weight_unit;
- if (idx == refmv_count[rf[0]] &&
- refmv_count[rf[0]] < MAX_REF_MV_STACK_SIZE) {
- ref_mv_stack[idx].this_mv.as_int = this_refmv.as_int;
- // TODO(jingning): Hard coded context number. Need to make it better
- // sense.
- ref_mv_stack[idx].pred_diff[0] = 1;
- ref_mv_stack[idx].weight = 2 * weight_unit;
- ++(refmv_count[rf[0]]);
- }
-
- ++coll_blk_count;
- return coll_blk_count;
+ if (idx == refmv_count[rf[0]] &&
+ refmv_count[rf[0]] < MAX_REF_MV_STACK_SIZE) {
+ ref_mv_stack[idx].this_mv.as_int = this_refmv.as_int;
+ // TODO(jingning): Hard coded context number. Need to make it better
+ // sense.
+ ref_mv_stack[idx].pred_diff[0] = 1;
+ ref_mv_stack[idx].weight = 2 * weight_unit;
+ ++(refmv_count[rf[0]]);
}
+
+ ++coll_blk_count;
+ return coll_blk_count;
}
} else {
// Process compound inter mode
@@ -477,55 +474,53 @@
#endif
CANDIDATE_MV *ref_mv_stack = ref_mv_stacks[ref_frame];
- for (int i = 0; i < MFMV_STACK_SIZE; ++i) {
- if (prev_frame_mvs->mfmv0[i].as_int != INVALID_MV) {
- int_mv this_refmv;
- int_mv comp_refmv;
- get_mv_projection(&this_refmv.as_mv, prev_frame_mvs->mfmv0[i].as_mv,
- cur_offset_0, prev_frame_mvs->ref_frame_offset[i]);
- get_mv_projection(&comp_refmv.as_mv, prev_frame_mvs->mfmv0[i].as_mv,
- cur_offset_1, prev_frame_mvs->ref_frame_offset[i]);
+ if (prev_frame_mvs->mfmv0.as_int != INVALID_MV) {
+ int_mv this_refmv;
+ int_mv comp_refmv;
+ get_mv_projection(&this_refmv.as_mv, prev_frame_mvs->mfmv0.as_mv,
+ cur_offset_0, prev_frame_mvs->ref_frame_offset);
+ get_mv_projection(&comp_refmv.as_mv, prev_frame_mvs->mfmv0.as_mv,
+ cur_offset_1, prev_frame_mvs->ref_frame_offset);
#if CONFIG_AMVR
- lower_mv_precision(&this_refmv.as_mv, cm->allow_high_precision_mv,
- cm->cur_frame_force_integer_mv);
- lower_mv_precision(&comp_refmv.as_mv, cm->allow_high_precision_mv,
- cm->cur_frame_force_integer_mv);
+ lower_mv_precision(&this_refmv.as_mv, cm->allow_high_precision_mv,
+ cm->cur_frame_force_integer_mv);
+ lower_mv_precision(&comp_refmv.as_mv, cm->allow_high_precision_mv,
+ cm->cur_frame_force_integer_mv);
#else
- lower_mv_precision(&this_refmv.as_mv, cm->allow_high_precision_mv);
- lower_mv_precision(&comp_refmv.as_mv, cm->allow_high_precision_mv);
+ lower_mv_precision(&this_refmv.as_mv, cm->allow_high_precision_mv);
+ lower_mv_precision(&comp_refmv.as_mv, cm->allow_high_precision_mv);
#endif
- if (blk_row == 0 && blk_col == 0)
- if (abs(this_refmv.as_mv.row - gm_mv_candidates[0].as_mv.row) >= 16 ||
- abs(this_refmv.as_mv.col - gm_mv_candidates[0].as_mv.col) >= 16 ||
- abs(comp_refmv.as_mv.row - gm_mv_candidates[1].as_mv.row) >= 16 ||
- abs(comp_refmv.as_mv.col - gm_mv_candidates[1].as_mv.col) >= 16)
- mode_context[ref_frame] |= (1 << GLOBALMV_OFFSET);
+ if (blk_row == 0 && blk_col == 0)
+ if (abs(this_refmv.as_mv.row - gm_mv_candidates[0].as_mv.row) >= 16 ||
+ abs(this_refmv.as_mv.col - gm_mv_candidates[0].as_mv.col) >= 16 ||
+ abs(comp_refmv.as_mv.row - gm_mv_candidates[1].as_mv.row) >= 16 ||
+ abs(comp_refmv.as_mv.col - gm_mv_candidates[1].as_mv.col) >= 16)
+ mode_context[ref_frame] |= (1 << GLOBALMV_OFFSET);
- for (idx = 0; idx < refmv_count[ref_frame]; ++idx)
- if (this_refmv.as_int == ref_mv_stack[idx].this_mv.as_int &&
- comp_refmv.as_int == ref_mv_stack[idx].comp_mv.as_int)
- break;
+ for (idx = 0; idx < refmv_count[ref_frame]; ++idx)
+ if (this_refmv.as_int == ref_mv_stack[idx].this_mv.as_int &&
+ comp_refmv.as_int == ref_mv_stack[idx].comp_mv.as_int)
+ break;
- if (idx < refmv_count[ref_frame])
- ref_mv_stack[idx].weight += 2 * weight_unit;
+ if (idx < refmv_count[ref_frame])
+ ref_mv_stack[idx].weight += 2 * weight_unit;
- if (idx == refmv_count[ref_frame] &&
- refmv_count[ref_frame] < MAX_REF_MV_STACK_SIZE) {
- ref_mv_stack[idx].this_mv.as_int = this_refmv.as_int;
- ref_mv_stack[idx].comp_mv.as_int = comp_refmv.as_int;
- // TODO(jingning): Hard coded context number. Need to make it better
- // sense.
- ref_mv_stack[idx].pred_diff[0] = 1;
- ref_mv_stack[idx].pred_diff[1] = 1;
- ref_mv_stack[idx].weight = 2 * weight_unit;
- ++(refmv_count[ref_frame]);
- }
-
- ++coll_blk_count;
- return coll_blk_count;
+ if (idx == refmv_count[ref_frame] &&
+ refmv_count[ref_frame] < MAX_REF_MV_STACK_SIZE) {
+ ref_mv_stack[idx].this_mv.as_int = this_refmv.as_int;
+ ref_mv_stack[idx].comp_mv.as_int = comp_refmv.as_int;
+ // TODO(jingning): Hard coded context number. Need to make it better
+ // sense.
+ ref_mv_stack[idx].pred_diff[0] = 1;
+ ref_mv_stack[idx].pred_diff[1] = 1;
+ ref_mv_stack[idx].weight = 2 * weight_unit;
+ ++(refmv_count[ref_frame]);
}
+
+ ++coll_blk_count;
+ return coll_blk_count;
}
}
@@ -1111,7 +1106,7 @@
}
static int motion_field_projection(AV1_COMMON *cm, MV_REFERENCE_FRAME ref_frame,
- int ref_stamp, int dir) {
+ int dir) {
TPL_MV_REF *tpl_mvs_base = cm->tpl_mvs;
int cur_rf_index[TOTAL_REFS_PER_FRAME] = { 0 };
int cur_offset[TOTAL_REFS_PER_FRAME] = { 0 };
@@ -1190,12 +1185,11 @@
if (pos_valid) {
int mi_offset = mi_r * (cm->mi_stride >> 1) + mi_c;
- tpl_mvs_base[mi_offset].mfmv0[ref_stamp].as_mv.row =
+ tpl_mvs_base[mi_offset].mfmv0.as_mv.row =
(dir == 1) ? -fwd_mv.row : fwd_mv.row;
- tpl_mvs_base[mi_offset].mfmv0[ref_stamp].as_mv.col =
+ tpl_mvs_base[mi_offset].mfmv0.as_mv.col =
(dir == 1) ? -fwd_mv.col : fwd_mv.col;
- tpl_mvs_base[mi_offset].ref_frame_offset[ref_stamp] =
- ref_frame_offset;
+ tpl_mvs_base[mi_offset].ref_frame_offset = ref_frame_offset;
}
}
}
@@ -1211,10 +1205,8 @@
TPL_MV_REF *tpl_mvs_base = cm->tpl_mvs;
int size = ((cm->mi_rows + MAX_MIB_SIZE) >> 1) * (cm->mi_stride >> 1);
for (int idx = 0; idx < size; ++idx) {
- for (int i = 0; i < MFMV_STACK_SIZE; ++i) {
- tpl_mvs_base[idx].mfmv0[i].as_int = INVALID_MV;
- tpl_mvs_base[idx].ref_frame_offset[i] = 0;
- }
+ tpl_mvs_base[idx].mfmv0.as_int = INVALID_MV;
+ tpl_mvs_base[idx].ref_frame_offset = 0;
}
const int cur_order_hint = cm->cur_frame->cur_frame_offset;
@@ -1248,24 +1240,24 @@
const int is_lst_overlay =
(alt_of_lst_order_hint == ref_order_hint[GOLDEN_FRAME - LAST_FRAME]);
- if (!is_lst_overlay) motion_field_projection(cm, LAST_FRAME, ref_stamp, 2);
+ if (!is_lst_overlay) motion_field_projection(cm, LAST_FRAME, 2);
--ref_stamp;
}
if (ref_order_hint[BWDREF_FRAME - LAST_FRAME] > cur_order_hint) {
- if (motion_field_projection(cm, BWDREF_FRAME, ref_stamp, 0)) --ref_stamp;
+ if (motion_field_projection(cm, BWDREF_FRAME, 0)) --ref_stamp;
}
if (ref_order_hint[ALTREF2_FRAME - LAST_FRAME] > cur_order_hint) {
- if (motion_field_projection(cm, ALTREF2_FRAME, ref_stamp, 0)) --ref_stamp;
+ if (motion_field_projection(cm, ALTREF2_FRAME, 0)) --ref_stamp;
}
if (ref_order_hint[ALTREF_FRAME - LAST_FRAME] > cur_order_hint &&
ref_stamp >= 0)
- if (motion_field_projection(cm, ALTREF_FRAME, ref_stamp, 0)) --ref_stamp;
+ if (motion_field_projection(cm, ALTREF_FRAME, 0)) --ref_stamp;
if (ref_stamp >= 0 && ref_buf_idx[LAST2_FRAME - LAST_FRAME] >= 0)
- if (motion_field_projection(cm, LAST2_FRAME, ref_stamp, 2)) --ref_stamp;
+ if (motion_field_projection(cm, LAST2_FRAME, 2)) --ref_stamp;
}
#if CONFIG_EXT_WARPED_MOTION
