Merge "Merge upstream-master and enable ARM64" into emu-master-devemu-31-stable-releaseemu-31-release

1 files changed, 201 insertions, 73 deletions

diff --git a/third_party/libaom/source/libaom/av1/encoder/speed_features.c b/third_party/libaom/source/libaom/av1/encoder/speed_features.c
index 2244aaae91..916a818513 100644
--- a/third_party/libaom/source/libaom/av1/encoder/speed_features.c
+++ b/third_party/libaom/source/libaom/av1/encoder/speed_features.c
@@ -274,6 +274,20 @@ static void set_allintra_speed_feature_framesize_dependent(
 
     sf->part_sf.use_square_partition_only_threshold = BLOCK_16X16;
   }
+
+  if (speed >= 7) {
+    if (!is_480p_or_larger) {
+      sf->rt_sf.nonrd_check_partition_merge_mode = 2;
+    }
+  }
+
+  if (speed >= 8) {
+    // TODO(kyslov): add more speed features to control speed/quality
+  }
+
+  if (speed >= 9) {
+    // TODO(kyslov): add more speed features to control speed/quality
+  }
 }
 
 static void set_allintra_speed_features_framesize_independent(
@@ -289,8 +303,11 @@ static void set_allintra_speed_features_framesize_independent(
   sf->part_sf.prune_part4_search = 2;
   sf->part_sf.simple_motion_search_prune_rect = 1;
   sf->part_sf.ml_predict_breakout_level = use_hbd ? 1 : 3;
+  sf->part_sf.reuse_prev_rd_results_for_part_ab = 1;
+  sf->part_sf.use_best_rd_for_pruning = 1;
 
   sf->intra_sf.intra_pruning_with_hog = 1;
+  sf->intra_sf.dv_cost_upd_level = INTERNAL_COST_UPD_OFF;
 
   sf->tx_sf.adaptive_txb_search_level = 1;
   sf->tx_sf.intra_tx_size_search_init_depth_sqr = 1;
@@ -300,7 +317,7 @@ static void set_allintra_speed_features_framesize_independent(
   sf->rt_sf.use_nonrd_pick_mode = 0;
   sf->rt_sf.use_real_time_ref_set = 0;
 
-  if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION ||
+  if (cpi->ppi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION ||
       cpi->use_screen_content_tools) {
     sf->mv_sf.exhaustive_searches_thresh = (1 << 20);
   } else {
@@ -318,10 +335,12 @@ static void set_allintra_speed_features_framesize_independent(
     // speed feature accordingly
     sf->part_sf.simple_motion_search_split = allow_screen_content_tools ? 1 : 2;
     sf->part_sf.ml_predict_breakout_level = use_hbd ? 2 : 3;
+    sf->part_sf.reuse_best_prediction_for_part_ab = 1;
 
     sf->mv_sf.exhaustive_searches_thresh <<= 1;
 
     sf->intra_sf.prune_palette_search_level = 1;
+    sf->intra_sf.top_intra_model_count_allowed = 3;
 
     sf->tx_sf.adaptive_txb_search_level = 2;
     sf->tx_sf.inter_tx_size_search_init_depth_rect = 1;
@@ -348,6 +367,7 @@ static void set_allintra_speed_features_framesize_independent(
 
     sf->intra_sf.disable_smooth_intra = 1;
     sf->intra_sf.intra_pruning_with_hog = 2;
+    sf->intra_sf.prune_filter_intra_level = 1;
 
     sf->rd_sf.perform_coeff_opt = 3;
 
@@ -397,9 +417,6 @@ static void set_allintra_speed_features_framesize_independent(
     sf->intra_sf.intra_uv_mode_mask[TX_16X16] = UV_INTRA_DC_H_V_CFL;
     sf->intra_sf.intra_uv_mode_mask[TX_32X32] = UV_INTRA_DC_H_V_CFL;
     sf->intra_sf.intra_uv_mode_mask[TX_64X64] = UV_INTRA_DC_H_V_CFL;
-    sf->intra_sf.intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
-    sf->intra_sf.intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
-    sf->intra_sf.intra_y_mode_mask[TX_64X64] = INTRA_DC_H_V;
     sf->intra_sf.prune_chroma_modes_using_luma_winner = 1;
 
     sf->mv_sf.simple_motion_subpel_force_stop = HALF_PEL;
@@ -408,7 +425,7 @@ static void set_allintra_speed_features_framesize_independent(
     sf->tpl_sf.subpel_force_stop = HALF_PEL;
     sf->tpl_sf.search_method = FAST_BIGDIA;
 
-    sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 1;
+    sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 2;
     sf->tx_sf.tx_type_search.fast_intra_tx_type_search = 1;
     sf->tx_sf.tx_type_search.prune_2d_txfm_mode = TX_TYPE_PRUNE_3;
     sf->tx_sf.tx_type_search.prune_tx_type_est_rd = 1;
@@ -443,9 +460,10 @@ static void set_allintra_speed_features_framesize_independent(
   }
 
   if (speed >= 6) {
-    sf->intra_sf.disable_filter_intra = 1;
+    sf->intra_sf.prune_filter_intra_level = 2;
     sf->intra_sf.chroma_intra_pruning_with_hog = 4;
     sf->intra_sf.intra_pruning_with_hog = 4;
+    sf->intra_sf.cfl_search_range = 1;
 
     sf->part_sf.prune_rectangular_split_based_on_qidx =
         allow_screen_content_tools ? 0 : 1;
@@ -458,7 +476,7 @@ static void set_allintra_speed_features_framesize_independent(
 
     sf->mv_sf.use_bsize_dependent_search_method = 1;
 
-    sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 2;
+    sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 3;
     sf->tx_sf.tx_type_search.prune_tx_type_est_rd = 0;
     // Use largest txfm block size for square coding blocks.
     sf->tx_sf.intra_tx_size_search_init_depth_sqr = 2;
@@ -466,10 +484,39 @@ static void set_allintra_speed_features_framesize_independent(
 
     sf->rd_sf.perform_coeff_opt = 6;
     sf->lpf_sf.cdef_pick_method = CDEF_FAST_SEARCH_LVL4;
+    sf->lpf_sf.lpf_pick = LPF_PICK_FROM_Q;
 
     sf->winner_mode_sf.multi_winner_mode_type = MULTI_WINNER_MODE_OFF;
   }
 
+  if (speed >= 7) {
+    sf->part_sf.default_min_partition_size = BLOCK_8X8;
+    sf->part_sf.partition_search_type = VAR_BASED_PARTITION;
+
+    sf->lpf_sf.cdef_pick_method = CDEF_PICK_FROM_Q;
+
+    sf->rt_sf.mode_search_skip_flags |= FLAG_SKIP_INTRA_DIRMISMATCH;
+    sf->rt_sf.use_nonrd_pick_mode = 1;
+    sf->rt_sf.nonrd_check_partition_merge_mode = 1;
+    sf->rt_sf.nonrd_check_partition_split = 0;
+    sf->rt_sf.skip_intra_pred_if_tx_skip = 1;
+    // Set mask for intra modes.
+    for (int i = 0; i < BLOCK_SIZES; ++i)
+      if (i >= BLOCK_32X32)
+        sf->rt_sf.intra_y_mode_bsize_mask_nrd[i] = INTRA_DC;
+      else
+        // Use DC, H, V intra mode for block sizes < 32X32.
+        sf->rt_sf.intra_y_mode_bsize_mask_nrd[i] = INTRA_DC_H_V;
+  }
+
+  if (speed >= 8) {
+    // TODO(kyslov): add more speed features to control speed/quality
+  }
+
+  if (speed >= 9) {
+    // TODO(kyslov): add more speed features to control speed/quality
+  }
+
   // Intra txb hash is currently not compatible with multi-winner mode as the
   // hashes got reset during multi-winner mode processing.
   assert(IMPLIES(
@@ -480,6 +527,7 @@ static void set_allintra_speed_features_framesize_independent(
 static void set_good_speed_feature_framesize_dependent(
     const AV1_COMP *const cpi, SPEED_FEATURES *const sf, int speed) {
   const AV1_COMMON *const cm = &cpi->common;
+  const int is_480p_or_lesser = AOMMIN(cm->width, cm->height) <= 480;
   const int is_480p_or_larger = AOMMIN(cm->width, cm->height) >= 480;
   const int is_720p_or_larger = AOMMIN(cm->width, cm->height) >= 720;
   const int is_1080p_or_larger = AOMMIN(cm->width, cm->height) >= 1080;
@@ -518,7 +566,16 @@ static void set_good_speed_feature_framesize_dependent(
     sf->mv_sf.use_downsampled_sad = 1;
   }
 
+  if (!is_720p_or_larger) {
+    const RateControlCfg *const rc_cfg = &cpi->oxcf.rc_cfg;
+    const int rate_tolerance =
+        AOMMIN(rc_cfg->under_shoot_pct, rc_cfg->over_shoot_pct);
+    sf->hl_sf.recode_tolerance = 25 + (rate_tolerance >> 2);
+  }
+
   if (speed >= 1) {
+    if (is_480p_or_lesser) sf->inter_sf.skip_newmv_in_drl = 1;
+
     if (is_720p_or_larger) {
       sf->part_sf.use_square_partition_only_threshold = BLOCK_128X128;
     } else if (is_480p_or_larger) {
@@ -561,6 +618,12 @@ static void set_good_speed_feature_framesize_dependent(
     }
 
     if (is_480p_or_larger) {
+      sf->inter_sf.disable_interintra_wedge_var_thresh = 100;
+    } else {
+      sf->inter_sf.disable_interintra_wedge_var_thresh = UINT_MAX;
+    }
+
+    if (is_480p_or_larger) {
       sf->tx_sf.tx_type_search.prune_tx_type_using_stats = 1;
       if (use_hbd) sf->tx_sf.prune_tx_size_level = 2;
     } else {
@@ -573,6 +636,8 @@ static void set_good_speed_feature_framesize_dependent(
   }
 
   if (speed >= 3) {
+    sf->inter_sf.skip_newmv_in_drl = 2;
+
     sf->part_sf.ml_early_term_after_part_split_level = 0;
 
     if (is_720p_or_larger) {
@@ -584,6 +649,10 @@ static void set_good_speed_feature_framesize_dependent(
       sf->part_sf.partition_search_breakout_rate_thr = 120;
     }
     if (use_hbd) sf->tx_sf.prune_tx_size_level = 3;
+
+    if (is_480p_or_larger) sf->intra_sf.top_intra_model_count_allowed = 2;
+
+    sf->inter_sf.disable_interintra_wedge_var_thresh = UINT_MAX;
   }
 
   if (speed >= 4) {
@@ -598,11 +667,14 @@ static void set_good_speed_feature_framesize_dependent(
     }
 
     sf->inter_sf.prune_obmc_prob_thresh = INT_MAX;
+    if (is_480p_or_lesser) sf->inter_sf.skip_newmv_in_drl = 3;
 
     if (is_720p_or_larger)
       sf->hl_sf.recode_tolerance = 32;
     else
       sf->hl_sf.recode_tolerance = 55;
+
+    sf->intra_sf.top_intra_model_count_allowed = 2;
   }
 
   if (speed >= 5) {
@@ -612,6 +684,8 @@ static void set_good_speed_feature_framesize_dependent(
       sf->inter_sf.prune_warped_prob_thresh = 8;
     }
     if (is_720p_or_larger) sf->hl_sf.recode_tolerance = 40;
+
+    sf->inter_sf.skip_newmv_in_drl = 4;
   }
 
   if (speed >= 6) {
@@ -630,7 +704,9 @@ static void set_good_speed_feature_framesize_dependent(
     }
 
     if (!is_720p_or_larger) {
-      sf->inter_sf.mv_cost_upd_level = 2;
+      sf->inter_sf.mv_cost_upd_level = INTERNAL_COST_UPD_SBROW_SET;
+      sf->inter_sf.coeff_cost_upd_level = INTERNAL_COST_UPD_SBROW;
+      sf->inter_sf.mode_cost_upd_level = INTERNAL_COST_UPD_SBROW;
     }
 
     if (is_720p_or_larger) {
@@ -650,10 +726,10 @@ static void set_good_speed_feature_framesize_dependent(
 static void set_good_speed_features_framesize_independent(
     const AV1_COMP *const cpi, SPEED_FEATURES *const sf, int speed) {
   const AV1_COMMON *const cm = &cpi->common;
-  const GF_GROUP *const gf_group = &cpi->gf_group;
+  const GF_GROUP *const gf_group = &cpi->ppi->gf_group;
   const int boosted = frame_is_boosted(cpi);
   const int is_boosted_arf2_bwd_type =
-      boosted || gf_group->update_type[gf_group->index] == INTNL_ARF_UPDATE;
+      boosted || gf_group->update_type[cpi->gf_frame_index] == INTNL_ARF_UPDATE;
   const int allow_screen_content_tools =
       cm->features.allow_screen_content_tools;
   const int use_hbd = cpi->oxcf.use_highbitdepth;
@@ -670,6 +746,8 @@ static void set_good_speed_features_framesize_independent(
   sf->part_sf.prune_part4_search = 2;
   sf->part_sf.simple_motion_search_prune_rect = 1;
   sf->part_sf.ml_predict_breakout_level = use_hbd ? 1 : 3;
+  sf->part_sf.reuse_prev_rd_results_for_part_ab = 1;
+  sf->part_sf.use_best_rd_for_pruning = 1;
 
   // TODO(debargha): Test, tweak and turn on either 1 or 2
   sf->inter_sf.inter_mode_rd_model_estimation = 1;
@@ -698,7 +776,7 @@ static void set_good_speed_features_framesize_independent(
   sf->rt_sf.use_nonrd_pick_mode = 0;
   sf->rt_sf.use_real_time_ref_set = 0;
 
-  if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION ||
+  if (cpi->ppi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION ||
       cpi->use_screen_content_tools) {
     sf->mv_sf.exhaustive_searches_thresh = (1 << 20);
   } else {
@@ -725,7 +803,6 @@ static void set_good_speed_features_framesize_independent(
     sf->mv_sf.use_accurate_subpel_search = USE_4_TAPS;
     sf->mv_sf.disable_extensive_joint_motion_search = 1;
 
-    sf->inter_sf.disable_interinter_wedge_newmv_search = boosted ? 0 : 1;
     sf->inter_sf.prune_comp_search_by_single_result = boosted ? 2 : 1;
     sf->inter_sf.prune_comp_type_by_comp_avg = 1;
     sf->inter_sf.prune_comp_type_by_model_rd = boosted ? 0 : 1;
@@ -736,7 +813,6 @@ static void set_good_speed_features_framesize_independent(
     sf->inter_sf.reduce_inter_modes = boosted ? 1 : 3;
     sf->inter_sf.reuse_inter_intra_mode = 1;
     sf->inter_sf.selective_ref_frame = 2;
-    sf->inter_sf.skip_repeated_newmv = 1;
 
     sf->interp_sf.use_interp_filter = 1;
 
@@ -766,7 +842,11 @@ static void set_good_speed_features_framesize_independent(
   if (speed >= 2) {
     sf->hl_sf.recode_loop = ALLOW_RECODE_KFARFGF;
 
+    sf->fp_sf.skip_motion_search_threshold = 25;
+
     sf->part_sf.allow_partition_search_skip = 1;
+    sf->part_sf.reuse_best_prediction_for_part_ab =
+        !frame_is_intra_only(&cpi->common);
 
     sf->mv_sf.auto_mv_step_size = 1;
     sf->mv_sf.simple_motion_subpel_force_stop = QUARTER_PEL;
@@ -778,20 +858,21 @@ static void set_good_speed_features_framesize_independent(
     // bit more closely to figure out why.
     sf->inter_sf.adaptive_rd_thresh = 1;
     sf->inter_sf.comp_inter_joint_search_thresh = BLOCK_SIZES_ALL;
-    sf->inter_sf.disable_interintra_wedge_var_thresh = 100;
     sf->inter_sf.disable_interinter_wedge_var_thresh = 100;
     sf->inter_sf.fast_interintra_wedge_search = 1;
     sf->inter_sf.prune_comp_search_by_single_result = boosted ? 4 : 1;
-    sf->inter_sf.prune_compound_using_neighbors = 1;
+    sf->inter_sf.prune_ext_comp_using_neighbors = 1;
     sf->inter_sf.prune_comp_using_best_single_mode_ref = 2;
     sf->inter_sf.prune_comp_type_by_comp_avg = 2;
-    sf->inter_sf.reuse_best_prediction_for_part_ab = 1;
     sf->inter_sf.selective_ref_frame = 3;
     sf->inter_sf.use_dist_wtd_comp_flag = DIST_WTD_COMP_DISABLED;
     // Enable fast search only for COMPOUND_DIFFWTD type.
     sf->inter_sf.enable_fast_compound_mode_search = 1;
     sf->inter_sf.reuse_mask_search_results = 1;
     sf->inter_sf.txfm_rd_gate_level = boosted ? 0 : 1;
+    sf->inter_sf.disable_interinter_wedge_newmv_search =
+        is_boosted_arf2_bwd_type ? 0 : 1;
+    sf->inter_sf.inter_mode_txfm_breakout = boosted ? 0 : 1;
 
     // TODO(Sachin): Enable/Enhance this speed feature for speed 2 & 3
     sf->interp_sf.adaptive_interp_filter_search = 1;
@@ -831,7 +912,8 @@ static void set_good_speed_features_framesize_independent(
     sf->mv_sf.search_method = DIAMOND;
     sf->mv_sf.disable_second_mv = 2;
 
-    sf->inter_sf.mv_cost_upd_level = 1;
+    sf->inter_sf.disable_interinter_wedge_newmv_search = boosted ? 0 : 1;
+    sf->inter_sf.mv_cost_upd_level = INTERNAL_COST_UPD_SBROW;
     sf->inter_sf.disable_onesided_comp = 1;
     // TODO(yunqing): evaluate this speed feature for speed 1 & 2, and combine
     // it with cpi->sf.disable_wedge_search_var_thresh.
@@ -843,10 +925,11 @@ static void set_good_speed_features_framesize_independent(
     sf->inter_sf.prune_comp_search_by_single_result = boosted ? 4 : 2;
     sf->inter_sf.selective_ref_frame = 5;
     sf->inter_sf.skip_repeated_ref_mv = 1;
-    sf->inter_sf.skip_repeated_full_newmv = 1;
     sf->inter_sf.reuse_compound_type_decision = 1;
     sf->inter_sf.txfm_rd_gate_level =
         boosted ? 0 : (is_boosted_arf2_bwd_type ? 1 : 2);
+    sf->inter_sf.enable_fast_wedge_mask_search = 1;
+    sf->inter_sf.inter_mode_txfm_breakout = boosted ? 0 : 2;
 
     sf->interp_sf.adaptive_interp_filter_search = 2;
 
@@ -865,6 +948,8 @@ static void set_good_speed_features_framesize_independent(
     sf->tx_sf.adaptive_txb_search_level = boosted ? 2 : 3;
     sf->tx_sf.tx_type_search.use_skip_flag_prediction = 2;
     sf->tx_sf.use_intra_txb_hash = 1;
+    sf->tx_sf.tx_type_search.prune_2d_txfm_mode = TX_TYPE_PRUNE_3;
+    sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 1;
 
     // TODO(any): Refactor the code related to following winner mode speed
     // features
@@ -874,10 +959,10 @@ static void set_good_speed_features_framesize_independent(
         frame_is_intra_only(&cpi->common) ? 0 : 1;
     sf->winner_mode_sf.enable_winner_mode_for_use_tx_domain_dist = 1;
     sf->winner_mode_sf.motion_mode_for_winner_cand =
-        boosted
-            ? 0
-            : gf_group->update_type[gf_group->index] == INTNL_ARF_UPDATE ? 1
-                                                                         : 2;
+        boosted ? 0
+                : gf_group->update_type[cpi->gf_frame_index] == INTNL_ARF_UPDATE
+                      ? 1
+                      : 2;
 
     // TODO(any): evaluate if these lpf features can be moved to speed 2.
     // For screen content, "prune_sgr_based_on_wiener = 2" cause large quality
@@ -889,6 +974,8 @@ static void set_good_speed_features_framesize_independent(
   }
 
   if (speed >= 4) {
+    sf->gm_sf.prune_zero_mv_with_sse = 1;
+
     sf->mv_sf.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
 
     sf->part_sf.simple_motion_search_prune_agg = 2;
@@ -901,7 +988,7 @@ static void set_good_speed_features_framesize_independent(
     sf->inter_sf.txfm_rd_gate_level = boosted ? 0 : 3;
 
     sf->inter_sf.prune_inter_modes_based_on_tpl = boosted ? 0 : 2;
-    sf->inter_sf.prune_compound_using_neighbors = 2;
+    sf->inter_sf.prune_ext_comp_using_neighbors = 2;
     sf->inter_sf.prune_obmc_prob_thresh = INT_MAX;
 
     sf->interp_sf.cb_pred_filter_search = 1;
@@ -911,9 +998,10 @@ static void set_good_speed_features_framesize_independent(
     sf->intra_sf.intra_uv_mode_mask[TX_16X16] = UV_INTRA_DC_H_V_CFL;
     sf->intra_sf.intra_uv_mode_mask[TX_32X32] = UV_INTRA_DC_H_V_CFL;
     sf->intra_sf.intra_uv_mode_mask[TX_64X64] = UV_INTRA_DC_H_V_CFL;
-    sf->intra_sf.intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
-    sf->intra_sf.intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
-    sf->intra_sf.intra_y_mode_mask[TX_64X64] = INTRA_DC_H_V;
+    // TODO(any): "intra_y_mode_mask" doesn't help much at speed 4.
+    // sf->intra_sf.intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
+    // sf->intra_sf.intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    // sf->intra_sf.intra_y_mode_mask[TX_64X64] = INTRA_DC_H_V;
     // TODO(any): Experiment with this speed feature set to 2 for higher quality
     // presets as well
     sf->intra_sf.skip_intra_in_interframe = 2;
@@ -923,10 +1011,10 @@ static void set_good_speed_features_framesize_independent(
     sf->tpl_sf.prune_starting_mv = 2;
     sf->tpl_sf.subpel_force_stop = HALF_PEL;
     sf->tpl_sf.search_method = FAST_BIGDIA;
+    sf->tpl_sf.gop_length_decision_method = 1;
 
-    sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 1;
+    sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 2;
     sf->tx_sf.tx_type_search.fast_intra_tx_type_search = 1;
-    sf->tx_sf.tx_type_search.prune_2d_txfm_mode = TX_TYPE_PRUNE_3;
     sf->tx_sf.tx_type_search.prune_tx_type_est_rd = 1;
     // TODO(any): Experiment with enabling of this speed feature as hash state
     // is reset during winner mode processing
@@ -948,9 +1036,14 @@ static void set_good_speed_features_framesize_independent(
   }
 
   if (speed >= 5) {
+    sf->fp_sf.reduce_mv_step_param = 4;
+
     sf->part_sf.simple_motion_search_prune_agg = 3;
     sf->part_sf.ext_partition_eval_thresh =
         allow_screen_content_tools ? BLOCK_8X8 : BLOCK_16X16;
+    sf->part_sf.prune_sub_8x8_partition_level =
+        (allow_screen_content_tools || frame_is_intra_only(&cpi->common)) ? 0
+                                                                          : 2;
 
     sf->inter_sf.disable_interinter_wedge_var_thresh = UINT_MAX;
     sf->inter_sf.prune_inter_modes_if_skippable = 1;
@@ -974,8 +1067,11 @@ static void set_good_speed_features_framesize_independent(
     sf->tpl_sf.prune_starting_mv = 3;
     sf->tpl_sf.use_y_only_rate_distortion = 1;
     sf->tpl_sf.subpel_force_stop = FULL_PEL;
+    sf->tpl_sf.gop_length_decision_method = 2;
 
     sf->winner_mode_sf.dc_blk_pred_level = 1;
+
+    sf->fp_sf.disable_recon = 1;
   }
 
   if (speed >= 6) {
@@ -986,9 +1082,14 @@ static void set_good_speed_features_framesize_independent(
     sf->inter_sf.prune_inter_modes_based_on_tpl = boosted ? 0 : 3;
     sf->inter_sf.prune_nearmv_using_neighbors = 1;
     sf->inter_sf.selective_ref_frame = 6;
+    sf->inter_sf.prune_ext_comp_using_neighbors = 3;
 
     sf->intra_sf.chroma_intra_pruning_with_hog = 4;
     sf->intra_sf.intra_pruning_with_hog = 4;
+    sf->intra_sf.intra_uv_mode_mask[TX_32X32] = UV_INTRA_DC;
+    sf->intra_sf.intra_uv_mode_mask[TX_64X64] = UV_INTRA_DC;
+    sf->intra_sf.intra_y_mode_mask[TX_32X32] = INTRA_DC;
+    sf->intra_sf.intra_y_mode_mask[TX_64X64] = INTRA_DC;
 
     sf->part_sf.prune_rectangular_split_based_on_qidx =
         boosted || allow_screen_content_tools ? 0 : 1;
@@ -1000,10 +1101,10 @@ static void set_good_speed_features_framesize_independent(
     sf->mv_sf.simple_motion_subpel_force_stop = FULL_PEL;
     sf->mv_sf.use_bsize_dependent_search_method = 1;
 
-    sf->tpl_sf.disable_gop_length_decision = 1;
+    sf->tpl_sf.gop_length_decision_method = 3;
     sf->tpl_sf.disable_filtered_key_tpl = 1;
 
-    sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 2;
+    sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning = 4;
     sf->tx_sf.use_intra_txb_hash = 1;
     sf->tx_sf.tx_type_search.prune_tx_type_est_rd = 0;
 
@@ -1052,10 +1153,13 @@ static void set_rt_speed_feature_framesize_dependent(const AV1_COMP *const cpi,
 #endif
     }
   } else {
-    if (speed == 8 && !cpi->use_svc) {
+    if (speed == 8 && !cpi->ppi->use_svc) {
       sf->rt_sf.short_circuit_low_temp_var = 0;
       sf->rt_sf.use_nonrd_altref_frame = 1;
     }
+    if (speed >= 9) {
+      sf->rt_sf.skip_cdef_sb = 1;
+    }
   }
   if (!is_480p_or_larger) {
     if (speed == 7) {
@@ -1088,6 +1192,8 @@ static void set_rt_speed_features_framesize_independent(AV1_COMP *cpi,
   sf->part_sf.less_rectangular_check_level = 1;
   sf->part_sf.ml_prune_partition = 1;
   sf->part_sf.prune_ext_partition_types_search_level = 1;
+  sf->part_sf.reuse_prev_rd_results_for_part_ab = 1;
+  sf->part_sf.use_best_rd_for_pruning = 1;
 
   // TODO(debargha): Test, tweak and turn on either 1 or 2
   sf->inter_sf.inter_mode_rd_model_estimation = 0;
@@ -1103,6 +1209,7 @@ static void set_rt_speed_features_framesize_independent(AV1_COMP *cpi,
 
   sf->interp_sf.use_fast_interpolation_filter_search = 1;
 
+  sf->intra_sf.dv_cost_upd_level = INTERNAL_COST_UPD_OFF;
   sf->intra_sf.intra_pruning_with_hog = 1;
 
   sf->mv_sf.full_pixel_search_level = 1;
@@ -1140,7 +1247,6 @@ static void set_rt_speed_features_framesize_independent(AV1_COMP *cpi,
     sf->inter_sf.prune_comp_search_by_single_result = 1;
     sf->inter_sf.reuse_inter_intra_mode = 1;
     sf->inter_sf.selective_ref_frame = 2;
-    sf->inter_sf.skip_repeated_newmv = 1;
     sf->inter_sf.disable_interintra_wedge_var_thresh = 0;
     sf->inter_sf.disable_interinter_wedge_var_thresh = 0;
     sf->inter_sf.prune_comp_type_by_comp_avg = 1;
@@ -1191,7 +1297,7 @@ static void set_rt_speed_features_framesize_independent(AV1_COMP *cpi,
   if (speed >= 3) {
     sf->hl_sf.recode_loop = ALLOW_RECODE_KFARFGF;
 
-    sf->gm_sf.gm_search_type = GM_DISABLE_SEARCH;
+    sf->gm_sf.gm_search_type = GM_REDUCED_REF_SEARCH_SKIP_L2_L3_ARF2;
 
     sf->part_sf.less_rectangular_check_level = 2;
 
@@ -1202,7 +1308,7 @@ static void set_rt_speed_features_framesize_independent(AV1_COMP *cpi,
     // sf->mv_sf.adaptive_motion_search = 1;
 
     sf->inter_sf.adaptive_rd_thresh = 2;
-    sf->inter_sf.mv_cost_upd_level = 1;
+    sf->inter_sf.mv_cost_upd_level = INTERNAL_COST_UPD_SBROW;
     // TODO(yunqing): evaluate this speed feature for speed 1 & 2, and combine
     // it with cpi->sf.disable_wedge_search_var_thresh.
     sf->inter_sf.disable_interintra_wedge_var_thresh = UINT_MAX;
@@ -1306,12 +1412,20 @@ static void set_rt_speed_features_framesize_independent(AV1_COMP *cpi,
     sf->part_sf.default_min_partition_size = BLOCK_8X8;
     sf->part_sf.partition_search_type = VAR_BASED_PARTITION;
 
+    sf->gm_sf.gm_search_type = GM_DISABLE_SEARCH;
+
     sf->mv_sf.search_method = FAST_DIAMOND;
     sf->mv_sf.subpel_force_stop = QUARTER_PEL;
     sf->mv_sf.subpel_search_method = SUBPEL_TREE_PRUNED;
 
     sf->inter_sf.inter_mode_rd_model_estimation = 2;
 
+    // Disable intra_y_mode_mask pruning since the performance at speed 7 isn't
+    // good. May need more study.
+    for (int i = 0; i < TX_SIZES; ++i) {
+      sf->intra_sf.intra_y_mode_mask[i] = INTRA_ALL;
+    }
+
     sf->lpf_sf.lpf_pick = LPF_PICK_FROM_Q;
 
     sf->rt_sf.mode_search_skip_flags |= FLAG_SKIP_INTRA_DIRMISMATCH;
@@ -1348,7 +1462,7 @@ static void set_rt_speed_features_framesize_independent(AV1_COMP *cpi,
     // TODO(marpan): Look into why enabling skip_loopfilter_non_reference is
     // not bitexact on rtc testset, its very close (< ~0.01 bdrate), but not
     // always bitexact.
-    if (cpi->use_svc && cpi->svc.non_reference_frame &&
+    if (cpi->ppi->use_svc && cpi->svc.non_reference_frame &&
         sf->lpf_sf.cdef_pick_method == CDEF_PICK_FROM_Q &&
         sf->lpf_sf.lpf_pick == LPF_PICK_FROM_Q)
       sf->rt_sf.skip_loopfilter_non_reference = 1;
@@ -1398,8 +1512,14 @@ static AOM_INLINE void init_hl_sf(HIGH_LEVEL_SPEED_FEATURES *hl_sf) {
   hl_sf->second_alt_ref_filtering = 1;
 }
 
+static AOM_INLINE void init_fp_sf(FIRST_PASS_SPEED_FEATURES *fp_sf) {
+  fp_sf->reduce_mv_step_param = 3;
+  fp_sf->skip_motion_search_threshold = 0;
+  fp_sf->disable_recon = 0;
+}
+
 static AOM_INLINE void init_tpl_sf(TPL_SPEED_FEATURES *tpl_sf) {
-  tpl_sf->disable_gop_length_decision = 0;
+  tpl_sf->gop_length_decision_method = 0;
   tpl_sf->prune_intra_modes = 0;
   tpl_sf->prune_starting_mv = 0;
   tpl_sf->reduce_first_step_size = 0;
@@ -1415,6 +1535,7 @@ static AOM_INLINE void init_tpl_sf(TPL_SPEED_FEATURES *tpl_sf) {
 static AOM_INLINE void init_gm_sf(GLOBAL_MOTION_SPEED_FEATURES *gm_sf) {
   gm_sf->gm_search_type = GM_FULL_SEARCH;
   gm_sf->prune_ref_frame_for_gm_search = 0;
+  gm_sf->prune_zero_mv_with_sse = 0;
 }
 
 static AOM_INLINE void init_part_sf(PARTITION_SPEED_FEATURES *part_sf) {
@@ -1454,6 +1575,9 @@ static AOM_INLINE void init_part_sf(PARTITION_SPEED_FEATURES *part_sf) {
   part_sf->ml_predict_breakout_level = 0;
   part_sf->prune_sub_8x8_partition_level = 0;
   part_sf->simple_motion_search_rect_split = 0;
+  part_sf->reuse_prev_rd_results_for_part_ab = 0;
+  part_sf->reuse_best_prediction_for_part_ab = 0;
+  part_sf->use_best_rd_for_pruning = 0;
 }
 
 static AOM_INLINE void init_mv_sf(MV_SPEED_FEATURES *mv_sf) {
@@ -1487,16 +1611,17 @@ static AOM_INLINE void init_inter_sf(INTER_MODE_SPEED_FEATURES *inter_sf) {
   inter_sf->fast_wedge_sign_estimate = 0;
   inter_sf->use_dist_wtd_comp_flag = DIST_WTD_COMP_ENABLED;
   inter_sf->reuse_inter_intra_mode = 0;
-  inter_sf->mv_cost_upd_level = 0;
+  inter_sf->mv_cost_upd_level = INTERNAL_COST_UPD_SB;
+  inter_sf->coeff_cost_upd_level = INTERNAL_COST_UPD_SB;
+  inter_sf->mode_cost_upd_level = INTERNAL_COST_UPD_SB;
   inter_sf->prune_inter_modes_based_on_tpl = 0;
   inter_sf->prune_nearmv_using_neighbors = 0;
   inter_sf->prune_comp_search_by_single_result = 0;
   inter_sf->skip_repeated_ref_mv = 0;
-  inter_sf->skip_repeated_newmv = 0;
-  inter_sf->skip_repeated_full_newmv = 0;
+  inter_sf->skip_newmv_in_drl = 0;
   inter_sf->inter_mode_rd_model_estimation = 0;
   inter_sf->prune_compound_using_single_ref = 0;
-  inter_sf->prune_compound_using_neighbors = 0;
+  inter_sf->prune_ext_comp_using_neighbors = 0;
   inter_sf->prune_comp_using_best_single_mode_ref = 0;
   inter_sf->disable_onesided_comp = 0;
   inter_sf->prune_mode_search_simple_translation = 0;
@@ -1514,9 +1639,10 @@ static AOM_INLINE void init_inter_sf(INTER_MODE_SPEED_FEATURES *inter_sf) {
   inter_sf->txfm_rd_gate_level = 0;
   inter_sf->prune_inter_modes_if_skippable = 0;
   inter_sf->disable_masked_comp = 0;
-  inter_sf->reuse_best_prediction_for_part_ab = 0;
   inter_sf->enable_fast_compound_mode_search = 0;
   inter_sf->reuse_mask_search_results = 0;
+  inter_sf->enable_fast_wedge_mask_search = 0;
+  inter_sf->inter_mode_txfm_breakout = 0;
 }
 
 static AOM_INLINE void init_interp_sf(INTERP_FILTER_SPEED_FEATURES *interp_sf) {
@@ -1529,6 +1655,7 @@ static AOM_INLINE void init_interp_sf(INTERP_FILTER_SPEED_FEATURES *interp_sf) {
 }
 
 static AOM_INLINE void init_intra_sf(INTRA_MODE_SPEED_FEATURES *intra_sf) {
+  intra_sf->dv_cost_upd_level = INTERNAL_COST_UPD_SB;
   intra_sf->skip_intra_in_interframe = 1;
   intra_sf->intra_pruning_with_hog = 0;
   intra_sf->chroma_intra_pruning_with_hog = 0;
@@ -1539,8 +1666,10 @@ static AOM_INLINE void init_intra_sf(INTRA_MODE_SPEED_FEATURES *intra_sf) {
     intra_sf->intra_uv_mode_mask[i] = UV_INTRA_ALL;
   }
   intra_sf->disable_smooth_intra = 0;
-  intra_sf->disable_filter_intra = 0;
+  intra_sf->prune_filter_intra_level = 0;
   intra_sf->prune_chroma_modes_using_luma_winner = 0;
+  intra_sf->cfl_search_range = 3;
+  intra_sf->top_intra_model_count_allowed = TOP_INTRA_MODEL_COUNT;
 }
 
 static AOM_INLINE void init_tx_sf(TX_SPEED_FEATURES *tx_sf) {
@@ -1650,9 +1779,11 @@ void av1_set_speed_features_framesize_dependent(AV1_COMP *cpi, int speed) {
       break;
   }
 
-  if (!cpi->seq_params_locked) {
-    cpi->common.seq_params.enable_masked_compound &=
+  if (!cpi->ppi->seq_params_locked) {
+    cpi->common.seq_params->enable_masked_compound &=
         !sf->inter_sf.disable_masked_comp;
+    cpi->common.seq_params->enable_interintra_compound &=
+        (sf->inter_sf.disable_interintra_wedge_var_thresh != UINT_MAX);
   }
 
   // This is only used in motion vector unit test.
@@ -1662,9 +1793,9 @@ void av1_set_speed_features_framesize_dependent(AV1_COMP *cpi, int speed) {
     cpi->mv_search_params.find_fractional_mv_step = av1_return_min_sub_pixel_mv;
 
   if ((cpi->oxcf.row_mt == 1) && (cpi->oxcf.max_threads > 1)) {
-    if (sf->inter_sf.mv_cost_upd_level > 1) {
+    if (sf->inter_sf.mv_cost_upd_level < INTERNAL_COST_UPD_SBROW) {
       // Set mv_cost_upd_level to use row level update.
-      sf->inter_sf.mv_cost_upd_level = 1;
+      sf->inter_sf.mv_cost_upd_level = INTERNAL_COST_UPD_SBROW;
     }
   }
 }
@@ -1676,6 +1807,7 @@ void av1_set_speed_features_framesize_independent(AV1_COMP *cpi, int speed) {
   int i;
 
   init_hl_sf(&sf->hl_sf);
+  init_fp_sf(&sf->fp_sf);
   init_tpl_sf(&sf->tpl_sf);
   init_gm_sf(&sf->gm_sf);
   init_part_sf(&sf->part_sf);
@@ -1701,12 +1833,12 @@ void av1_set_speed_features_framesize_independent(AV1_COMP *cpi, int speed) {
       break;
   }
 
-  if (!cpi->seq_params_locked) {
-    cpi->common.seq_params.enable_dual_filter &=
+  if (!cpi->ppi->seq_params_locked) {
+    cpi->common.seq_params->enable_dual_filter &=
         !sf->interp_sf.disable_dual_filter;
-    cpi->common.seq_params.enable_restoration &= !sf->lpf_sf.disable_lr_filter;
+    cpi->common.seq_params->enable_restoration &= !sf->lpf_sf.disable_lr_filter;
 
-    cpi->common.seq_params.enable_interintra_compound &=
+    cpi->common.seq_params->enable_interintra_compound &=
         (sf->inter_sf.disable_interintra_wedge_var_thresh != UINT_MAX);
   }
 
@@ -1821,10 +1953,11 @@ void av1_set_speed_features_qindex_dependent(AV1_COMP *cpi, int speed) {
   SPEED_FEATURES *const sf = &cpi->sf;
   WinnerModeParams *const winner_mode_params = &cpi->winner_mode_params;
   const int boosted = frame_is_boosted(cpi);
+  const int is_480p_or_larger = AOMMIN(cm->width, cm->height) >= 480;
   const int is_720p_or_larger = AOMMIN(cm->width, cm->height) >= 720;
   const int is_1080p_or_larger = AOMMIN(cm->width, cm->height) >= 1080;
   const int is_arf2_bwd_type =
-      cpi->gf_group.update_type[cpi->gf_group.index] == INTNL_ARF_UPDATE;
+      cpi->ppi->gf_group.update_type[cpi->gf_frame_index] == INTNL_ARF_UPDATE;
 
   if (cpi->oxcf.mode == REALTIME) return;
 
@@ -1832,7 +1965,6 @@ void av1_set_speed_features_qindex_dependent(AV1_COMP *cpi, int speed) {
     // qindex_thresh for resolution < 720p
     const int qindex_thresh = boosted ? 70 : (is_arf2_bwd_type ? 110 : 140);
     if (!is_720p_or_larger && cm->quant_params.base_qindex <= qindex_thresh) {
-      sf->inter_sf.skip_repeated_newmv = 1;
       sf->part_sf.simple_motion_search_split =
           cm->features.allow_screen_content_tools ? 1 : 2;
       sf->part_sf.simple_motion_search_early_term_none = 1;
@@ -1849,7 +1981,6 @@ void av1_set_speed_features_qindex_dependent(AV1_COMP *cpi, int speed) {
       sf->tx_sf.inter_tx_size_search_init_depth_rect = 1;
       sf->tx_sf.inter_tx_size_search_init_depth_sqr = 1;
       sf->tx_sf.intra_tx_size_search_init_depth_rect = 1;
-      sf->inter_sf.skip_repeated_newmv = 1;
       sf->tx_sf.model_based_prune_tx_search_level = 0;
 
       if (is_1080p_or_larger && cm->quant_params.base_qindex <= 108) {
@@ -1866,28 +1997,25 @@ void av1_set_speed_features_qindex_dependent(AV1_COMP *cpi, int speed) {
     }
   }
 
-  if (speed >= 3) {
-    // Disable extended partitions for lower quantizers
-    const int qindex_thresh =
-        cm->features.allow_screen_content_tools ? 50 : 100;
-    if (cm->quant_params.base_qindex <= qindex_thresh && !boosted) {
-      sf->part_sf.ext_partition_eval_thresh = BLOCK_128X128;
-    }
-  }
-
-  if (speed >= 4) {
+  if (speed >= 2) {
     // Disable extended partitions for lower quantizers
-    const int qindex_thresh = boosted ? 80 : 120;
-    if (cm->quant_params.base_qindex <= qindex_thresh &&
-        !frame_is_intra_only(&cpi->common)) {
-      sf->part_sf.ext_partition_eval_thresh = BLOCK_128X128;
+    const int aggr = AOMMIN(3, speed - 2);
+    const int qindex_thresh1[4] = { 50, 50, 80, 100 };
+    const int qindex_thresh2[4] = { 80, 100, 120, 160 };
+    int qindex_thresh;
+    int disable_ext_part;
+    if (aggr <= 1) {
+      const int qthresh2 =
+          (!aggr && !is_480p_or_larger) ? 70 : qindex_thresh2[aggr];
+      qindex_thresh = cm->features.allow_screen_content_tools
+                          ? qindex_thresh1[aggr]
+                          : qthresh2;
+      disable_ext_part = !boosted;
+    } else {
+      qindex_thresh = boosted ? qindex_thresh1[aggr] : qindex_thresh2[aggr];
+      disable_ext_part = !frame_is_intra_only(cm);
     }
-  }
-
-  if (speed >= 5) {
-    const int qindex_thresh = boosted ? 100 : 160;
-    if (cm->quant_params.base_qindex <= qindex_thresh &&
-        !frame_is_intra_only(&cpi->common)) {
+    if (cm->quant_params.base_qindex <= qindex_thresh && disable_ext_part) {
       sf->part_sf.ext_partition_eval_thresh = BLOCK_128X128;
     }
   }