commit 3418cad83167e61a447d3f5745688d5fc21abbbd
author chiyotsai <chiyotsai@google.com> Fri Sep 13 11:06:58 2019 -0700
committer Chi Yo Tsai <chiyotsai@google.com> Mon Sep 16 21:00:06 2019 +0000
tree 1c31b2d4d31781bae12760ee8a8d5534991d920c
parent bcacb3260abb3f482cfe322e42aa7066416a0636

Move high performing speed feature from spd2 to spd1

The adaptive_rd_thresh also gives significant speed up on hdres (~10%),
but has some issues on certain clips (e.g. dinner_1080p30.y4m), so it is
not moved in this commit.

Performance:
   RES   | AVG_PSNR | OVR_PSNR |   SSIM  |  AVG_SPD |  OVR_SPD
  LOWRES |  +0.157% |  +0.155% | +0.100% | +10.934% | +12.628%
  MIDRES |  +0.106% |  +0.110% | +0.061% | +13.779% | +18.624%
   HDRES |  +0.078% |  +0.082% | +0.048% | +14.174% | +17.341%

STATS_CHANGED

Change-Id: Id333a501a5247f3620d8354fcb8ad7d19f8c1010

av1/encoder/speed_features.c

diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index 1bfae8b..065d963 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -288,18 +288,27 @@
             ? 0
             : (boosted ? 1 : 2);
     sf->intra_cnn_split = (speed == 1);
+
+    sf->enable_sgr_ep_pruning = 1;
+    sf->inter_tx_size_search_init_depth_rect = 1;
+    sf->inter_tx_size_search_init_depth_sqr = 1;
+    sf->cdef_pick_method = CDEF_FAST_SEARCH;
+    sf->cb_pred_filter_search = 0;
+    sf->model_based_prune_tx_search_level = 0;
+    sf->reduce_inter_modes = boosted ? 1 : 2;
+    sf->tx_type_search.prune_mode = PRUNE_2D_FAST;
+    sf->prune_comp_type_by_model_rd = boosted ? 0 : 1;
   }
 
   if (speed >= 2) {
     sf->gm_erroradv_type = GM_ERRORADV_TR_2;
 
-    sf->enable_sgr_ep_pruning = 1;
     sf->selective_ref_frame = 3;
-    sf->inter_tx_size_search_init_depth_rect = 1;
-    sf->inter_tx_size_search_init_depth_sqr = 1;
 
-    sf->cdef_pick_method = CDEF_FAST_SEARCH;
-
+    // 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
+    // clips (e.g. 5% loss on dinner_1080p). We need to examine the sequence a
+    // bit more closely to figure out why.
     sf->adaptive_rd_thresh = 1;
     sf->mv.auto_mv_step_size = 1;
     sf->mv.subpel_iters_per_step = 1;
@@ -315,15 +324,10 @@
     sf->use_dist_wtd_comp_flag = DIST_WTD_COMP_DISABLED;
     sf->prune_comp_type_by_comp_avg = 2;
     // TODO(Sachin): Enable/Enhance this speed feature for speed 2 & 3
-    sf->cb_pred_filter_search = 0;
     sf->adaptive_interp_filter_search = 1;
     sf->perform_coeff_opt = is_boosted_arf2_bwd_type ? 2 : 3;
-    sf->model_based_prune_tx_search_level = 0;
 
-    sf->reduce_inter_modes = boosted ? 1 : 2;
-    sf->tx_type_search.prune_mode = PRUNE_2D_FAST;
     sf->prune_warp_using_wmtype = 1;
-    sf->prune_comp_type_by_model_rd = boosted ? 0 : 1;
     sf->disable_smooth_intra =
         !frame_is_intra_only(&cpi->common) || (cpi->rc.frames_to_key != 1);
   }