Extend prune_sgr_based_on_wiener for preset 2
Extended the speed feature 'prune_sgr_based_on_wiener' for speed preset
2. Pruning is based on the rdcost ratio of RESTORE_WIENER and
RESTORE_NONE.
The speed feature has been enabled for non-screen contents.
Instruction Count
cpu-used Reduction Quality Loss
2 1.36% 0.01%
STATS_CHANGED
Change-Id: I8ada2a62a039f4df33c0ba544be34add00dfae80
diff --git a/av1/encoder/pickrst.c b/av1/encoder/pickrst.c
index e371b0e..520649b 100644
--- a/av1/encoder/pickrst.c
+++ b/av1/encoder/pickrst.c
@@ -862,8 +862,7 @@
const int bit_depth = cm->seq_params.bit_depth;
const int64_t bits_none = x->sgrproj_restore_cost[0];
- // Prune evaluation of RESTORE_SGRPROJ if RESTORE_NONE was the winner (no loop
- // restoration)
+ // Prune evaluation of RESTORE_SGRPROJ if 'skip_sgr_eval' is set
if (rusi->skip_sgr_eval) {
rsc->bits += bits_none;
rsc->sse += rusi->sse[RESTORE_NONE];
@@ -1478,7 +1477,7 @@
rsc->sse += rusi->sse[RESTORE_NONE];
rusi->best_rtype[RESTORE_WIENER - 1] = RESTORE_NONE;
rusi->sse[RESTORE_WIENER] = INT64_MAX;
- if (rsc->sf->prune_sgr_based_on_wiener) rusi->skip_sgr_eval = 1;
+ if (rsc->sf->prune_sgr_based_on_wiener == 2) rusi->skip_sgr_eval = 1;
return;
}
@@ -1497,7 +1496,7 @@
rsc->sse += rusi->sse[RESTORE_NONE];
rusi->best_rtype[RESTORE_WIENER - 1] = RESTORE_NONE;
rusi->sse[RESTORE_WIENER] = INT64_MAX;
- if (rsc->sf->prune_sgr_based_on_wiener) rusi->skip_sgr_eval = 1;
+ if (rsc->sf->prune_sgr_based_on_wiener == 2) rusi->skip_sgr_eval = 1;
return;
}
@@ -1528,7 +1527,11 @@
(cost_wiener < cost_none) ? RESTORE_WIENER : RESTORE_NONE;
rusi->best_rtype[RESTORE_WIENER - 1] = rtype;
- if (rsc->sf->prune_sgr_based_on_wiener) {
+ // Set 'skip_sgr_eval' based on rdcost ratio of RESTORE_WIENER and
+ // RESTORE_NONE or based on best_rtype
+ if (rsc->sf->prune_sgr_based_on_wiener == 1) {
+ rusi->skip_sgr_eval = cost_wiener > (1.01 * cost_none);
+ } else if (rsc->sf->prune_sgr_based_on_wiener == 2) {
rusi->skip_sgr_eval = rusi->best_rtype[RESTORE_WIENER - 1] == RESTORE_NONE;
}
diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index be7b4e0..3babf45 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -349,6 +349,7 @@
sf->gm_erroradv_type = GM_ERRORADV_TR_2;
sf->selective_ref_frame = 3;
+ sf->prune_sgr_based_on_wiener = cm->allow_screen_content_tools ? 0 : 1;
// TODO(chiyotsai@google.com): We can get 10% speed up if we move
// adaptive_rd_thresh to speed 1. But currently it performs poorly on some
@@ -405,7 +406,7 @@
sf->enable_winner_mode_for_use_tx_domain_dist =
cm->allow_screen_content_tools ? 0 : 1;
sf->reduce_wiener_window_size = is_boosted_arf2_bwd_type ? 0 : 1;
- sf->prune_sgr_based_on_wiener = cm->allow_screen_content_tools ? 0 : 1;
+ sf->prune_sgr_based_on_wiener = cm->allow_screen_content_tools ? 0 : 2;
sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED;
sf->simple_motion_search_prune_agg = 1;
sf->disable_sb_level_mv_cost_upd = 1;
diff --git a/av1/encoder/speed_features.h b/av1/encoder/speed_features.h
index 9dc3226..88e1d03 100644
--- a/av1/encoder/speed_features.h
+++ b/av1/encoder/speed_features.h
@@ -804,6 +804,10 @@
int disable_loop_restoration_chroma;
// Prune self-guided loop restoration based on wiener search results
+ // 0 : no pruning
+ // 1 : pruning based on rdcost ratio of RESTORE_WIENER and RESTORE_NONE
+ // 2 : pruning based on winner restoration type among RESTORE_WIENER and
+ // RESTORE_NONE
int prune_sgr_based_on_wiener;
// Reduce the wiener filter win size for luma
