Compound prediction mode for nonrd pickmode.
Allow for compound prediction mode in nonrd_pickmode for ZEROMV.
For real-time encoding, 1 pass with non-zero lag-in-frames.
Added speed feature to control the feature.
Enabled for speed >=6 for now, under VBR mode.
avgPSNR/SSIM metrics positive on ytlive set, for speed 6:
some clips up by ~3-5%, some clips neutral gain, average gain
across clips is ~1%.
Small/negligible decrease in speed.
Change-Id: I7a60c7596e69b9a928410c5ee2f9141eecd8613d
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index f60c0d3..b1587a8 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -4797,8 +4797,31 @@
}
}
} else {
+ FRAME_COUNTS *counts = cpi->td.counts;
cm->reference_mode = SINGLE_REFERENCE;
+ if (cpi->allow_comp_inter_inter && cpi->sf.use_compound_nonrd_pickmode &&
+ cpi->rc.alt_ref_gf_group && !cpi->rc.is_src_frame_alt_ref &&
+ cm->frame_type != KEY_FRAME)
+ cm->reference_mode = REFERENCE_MODE_SELECT;
+
encode_frame_internal(cpi);
+
+ if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+ int single_count_zero = 0;
+ int comp_count_zero = 0;
+ int i;
+ for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
+ single_count_zero += counts->comp_inter[i][0];
+ comp_count_zero += counts->comp_inter[i][1];
+ }
+ if (comp_count_zero == 0) {
+ cm->reference_mode = SINGLE_REFERENCE;
+ vp9_zero(counts->comp_inter);
+ } else if (single_count_zero == 0) {
+ cm->reference_mode = COMPOUND_REFERENCE;
+ vp9_zero(counts->comp_inter);
+ }
+ }
}
// If segmented AQ is enabled compute the average AQ weighting.
diff --git a/vp9/encoder/vp9_pickmode.c b/vp9/encoder/vp9_pickmode.c
index 870af84..cc95a34 100644
--- a/vp9/encoder/vp9_pickmode.c
+++ b/vp9/encoder/vp9_pickmode.c
@@ -1190,7 +1190,7 @@
int const_motion[MAX_REF_FRAMES], int *ref_frame_skip_mask,
const int flag_list[4], TileDataEnc *tile_data, int mi_row, int mi_col,
struct buf_2d yv12_mb[4][MAX_MB_PLANE], BLOCK_SIZE bsize,
- int force_skip_low_temp_var) {
+ int force_skip_low_temp_var, int comp_pred_allowed) {
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &x->e_mbd;
const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
@@ -1204,7 +1204,7 @@
int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
- if (cm->use_prev_frame_mvs) {
+ if (cm->use_prev_frame_mvs || comp_pred_allowed) {
vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
x->mbmi_ext->mode_context);
} else {
@@ -1426,7 +1426,7 @@
struct macroblockd_plane *const pd = &xd->plane[0];
PREDICTION_MODE best_mode = ZEROMV;
MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
- MV_REFERENCE_FRAME usable_ref_frame;
+ MV_REFERENCE_FRAME usable_ref_frame, second_ref_frame;
TX_SIZE best_tx_size = TX_SIZES;
INTERP_FILTER best_pred_filter = EIGHTTAP;
int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
@@ -1494,6 +1494,8 @@
int denoise_svc_pickmode = 1;
#endif
INTERP_FILTER filter_gf_svc = EIGHTTAP;
+ MV_REFERENCE_FRAME best_second_ref_frame = NONE;
+ int comp_modes = 0;
init_ref_frame_cost(cm, xd, ref_frame_cost);
@@ -1626,18 +1628,24 @@
x->last_sb_high_content > 40 || cpi->rc.frames_since_golden > 120))
usable_ref_frame = LAST_FRAME;
+ // Compound prediction modes: (0,0) on LAST/GOLDEN and ARF.
+ if (cm->reference_mode == REFERENCE_MODE_SELECT &&
+ cpi->sf.use_compound_nonrd_pickmode && usable_ref_frame == ALTREF_FRAME)
+ comp_modes = 2;
+
for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
if (!skip_ref_find_pred[ref_frame]) {
find_predictors(cpi, x, ref_frame, frame_mv, const_motion,
&ref_frame_skip_mask, flag_list, tile_data, mi_row,
- mi_col, yv12_mb, bsize, force_skip_low_temp_var);
+ mi_col, yv12_mb, bsize, force_skip_low_temp_var,
+ comp_modes > 0);
}
}
if (cpi->use_svc || cpi->oxcf.speed <= 7 || bsize < BLOCK_32X32)
x->sb_use_mv_part = 0;
- for (idx = 0; idx < RT_INTER_MODES; ++idx) {
+ for (idx = 0; idx < RT_INTER_MODES + comp_modes; ++idx) {
int rate_mv = 0;
int mode_rd_thresh;
int mode_index;
@@ -1648,15 +1656,37 @@
int rd_computed = 0;
int inter_mv_mode = 0;
int skip_this_mv = 0;
+ int comp_pred = 0;
+ PREDICTION_MODE this_mode;
+ second_ref_frame = NONE;
- PREDICTION_MODE this_mode = ref_mode_set[idx].pred_mode;
+ if (idx < RT_INTER_MODES) {
+ this_mode = ref_mode_set[idx].pred_mode;
+ ref_frame = ref_mode_set[idx].ref_frame;
- ref_frame = ref_mode_set[idx].ref_frame;
-
- if (cpi->use_svc) {
- this_mode = ref_mode_set_svc[idx].pred_mode;
- ref_frame = ref_mode_set_svc[idx].ref_frame;
+ if (cpi->use_svc) {
+ this_mode = ref_mode_set_svc[idx].pred_mode;
+ ref_frame = ref_mode_set_svc[idx].ref_frame;
+ }
+ } else {
+ // Add (0,0) compound modes.
+ this_mode = ZEROMV;
+ ref_frame = LAST_FRAME;
+ if (idx == RT_INTER_MODES + comp_modes - 1) ref_frame = GOLDEN_FRAME;
+ second_ref_frame = ALTREF_FRAME;
+ comp_pred = 1;
}
+
+ if (comp_pred) {
+ const struct segmentation *const seg = &cm->seg;
+ if (!cpi->allow_comp_inter_inter) continue;
+ // Skip compound inter modes if ARF is not available.
+ if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
+ // Do not allow compound prediction if the segment level reference frame
+ // feature is in use as in this case there can only be one reference.
+ if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME)) continue;
+ }
+
if (ref_frame > usable_ref_frame) continue;
if (skip_ref_find_pred[ref_frame]) continue;
@@ -1748,11 +1778,14 @@
if (ref_frame_skip_mask & (1 << ref_frame)) continue;
// Select prediction reference frames.
- for (i = 0; i < MAX_MB_PLANE; i++)
+ for (i = 0; i < MAX_MB_PLANE; i++) {
xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
+ if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
+ }
mi->ref_frame[0] = ref_frame;
- set_ref_ptrs(cm, xd, ref_frame, NONE);
+ mi->ref_frame[1] = second_ref_frame;
+ set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
mode_rd_thresh = best_mode_skip_txfm ? rd_threshes[mode_index] << 1
@@ -1862,7 +1895,7 @@
// causes some regression, leave it for duplicate zero-mv for now, until
// regression issue is resolved.
for (inter_mv_mode = NEARESTMV; inter_mv_mode <= NEWMV; inter_mv_mode++) {
- if (inter_mv_mode == this_mode) continue;
+ if (inter_mv_mode == this_mode || comp_pred) continue;
if (mode_checked[inter_mv_mode][ref_frame] &&
frame_mv[this_mode][ref_frame].as_int ==
frame_mv[inter_mv_mode][ref_frame].as_int &&
@@ -1888,13 +1921,14 @@
x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
}
- if (this_mode != NEARESTMV &&
+ if (this_mode != NEARESTMV && !comp_pred &&
frame_mv[this_mode][ref_frame].as_int ==
frame_mv[NEARESTMV][ref_frame].as_int)
continue;
mi->mode = this_mode;
mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
+ mi->mv[1].as_int = 0;
// Search for the best prediction filter type, when the resulting
// motion vector is at sub-pixel accuracy level for luma component, i.e.,
@@ -2050,6 +2084,7 @@
this_rdc.rate += rate_mv;
this_rdc.rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
[INTER_OFFSET(this_mode)];
+ // TODO(marpan): Add costing for compound mode.
this_rdc.rate += ref_frame_cost[ref_frame];
this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
@@ -2099,6 +2134,7 @@
best_ref_frame = ref_frame;
best_mode_skip_txfm = x->skip_txfm[0];
best_early_term = this_early_term;
+ best_second_ref_frame = second_ref_frame;
if (reuse_inter_pred) {
free_pred_buffer(best_pred);
@@ -2125,6 +2161,7 @@
mi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int;
xd->mi[0]->bmi[0].as_mv[0].as_int = mi->mv[0].as_int;
x->skip_txfm[0] = best_mode_skip_txfm;
+ mi->ref_frame[1] = best_second_ref_frame;
// For spatial enhancemanent layer: perform intra prediction only if base
// layer is chosen as the reference. Always perform intra prediction if
@@ -2238,8 +2275,10 @@
best_mode = this_mode;
best_intra_tx_size = mi->tx_size;
best_ref_frame = INTRA_FRAME;
+ best_second_ref_frame = NONE;
mi->uv_mode = this_mode;
mi->mv[0].as_int = INVALID_MV;
+ mi->mv[1].as_int = INVALID_MV;
best_mode_skip_txfm = x->skip_txfm[0];
}
}
@@ -2255,6 +2294,7 @@
pd->dst = orig_dst;
mi->mode = best_mode;
mi->ref_frame[0] = best_ref_frame;
+ mi->ref_frame[1] = best_second_ref_frame;
x->skip_txfm[0] = best_mode_skip_txfm;
if (!is_inter_block(mi)) {
@@ -2300,7 +2340,7 @@
}
#endif
- if (best_ref_frame == ALTREF_FRAME)
+ if (best_ref_frame == ALTREF_FRAME || best_second_ref_frame == ALTREF_FRAME)
x->arf_frame_usage++;
else if (best_ref_frame != INTRA_FRAME)
x->lastgolden_frame_usage++;
diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c
index cebaca7..178861f 100644
--- a/vp9/encoder/vp9_speed_features.c
+++ b/vp9/encoder/vp9_speed_features.c
@@ -370,6 +370,7 @@
sf->use_simple_block_yrd = 0;
sf->adapt_partition_source_sad = 0;
sf->use_altref_onepass = 0;
+ sf->use_compound_nonrd_pickmode = 0;
sf->nonrd_keyframe = 0;
if (speed >= 1) {
@@ -531,8 +532,10 @@
}
if (speed >= 6) {
- if (cpi->oxcf.rc_mode == VPX_VBR && cpi->oxcf.lag_in_frames > 0)
+ if (cpi->oxcf.rc_mode == VPX_VBR && cpi->oxcf.lag_in_frames > 0) {
sf->use_altref_onepass = 1;
+ sf->use_compound_nonrd_pickmode = 1;
+ }
sf->partition_search_type = VAR_BASED_PARTITION;
// Turn on this to use non-RD key frame coding mode.
sf->use_nonrd_pick_mode = 1;
diff --git a/vp9/encoder/vp9_speed_features.h b/vp9/encoder/vp9_speed_features.h
index 9e5bf9a..56c0b8c 100644
--- a/vp9/encoder/vp9_speed_features.h
+++ b/vp9/encoder/vp9_speed_features.h
@@ -500,6 +500,9 @@
// Enable use of alt-refs in 1 pass VBR.
int use_altref_onepass;
+ // Enable use of compound prediction, for nonrd_pickmode with nonzero lag.
+ int use_compound_nonrd_pickmode;
+
// Always use nonrd_pick_intra for all block sizes on keyframes.
int nonrd_keyframe;
} SPEED_FEATURES;
