Evaluate RT speed 4 features
Removed features that gave much worse than 10:1 speed/quality
tradeoff. Removed speed 4 for now since there weren't much speedup
to differentiate it from speed 0. We can always add more speeds
when it is needed. Here are borg tests results.
avg_psnr: ovr_psnr: ssim: speed:
speed 4: (Got over 1% quality back.)
rtc: -1.204 -1.359 -0.976 -3.600
rtc_derf: -1.481 -1.517 -1.175 -3.462
speed 5: (Made both quality & speed better.)
rtc: -0.492 -0.506 0.072 6.568
rtc_derf: -0.471 -0.503 0.217 6.756
speed 7: (This didn't affect speed 7.)
rtc: 0.000 0.000 0.000 0.000
rtc_derf: 0.000 0.000 0.000 0.000
STATS_CHANGED
Change-Id: I6ac27ac80675619775928d75e8bc93399616677f
diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index 438c12f..1c5111f 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -1242,6 +1242,8 @@
sf->inter_sf.disable_interintra_wedge_var_thresh = UINT_MAX;
sf->inter_sf.prune_comp_search_by_single_result = 2;
sf->inter_sf.selective_ref_frame = 4;
+ // TODO(any): need tuning for RT mode.
+ sf->inter_sf.alt_ref_search_fp = 1;
sf->interp_sf.use_fast_interpolation_filter_search = 1;
sf->interp_sf.use_interp_filter = 1;
@@ -1261,6 +1263,7 @@
sf->mv_sf.auto_mv_step_size = 1;
sf->mv_sf.subpel_iters_per_step = 1;
sf->mv_sf.use_accurate_subpel_search = USE_2_TAPS;
+ sf->mv_sf.subpel_search_method = SUBPEL_TREE_PRUNED;
sf->part_sf.ml_prune_partition = 1;
sf->part_sf.reuse_prev_rd_results_for_part_ab = 1;
@@ -1300,26 +1303,11 @@
sf->tx_sf.inter_tx_size_search_init_depth_sqr = 1;
sf->tx_sf.model_based_prune_tx_search_level = 0;
sf->tx_sf.tx_type_search.prune_2d_txfm_mode = TX_TYPE_PRUNE_2;
+ // TODO(any, yunqing): somehow this is needed by sf->rt_sf.use_nonrd_pick_mode
+ // at speed 7? Need more investigation.
+ sf->tx_sf.use_intra_txb_hash = 0;
- sf->winner_mode_sf.tx_size_search_level = boosted ? 0 : 2;
-
- // TODO(Yunqing, any): speed 4 overall speed/quality tradeoff isn't good. Need
- // evaluate it.
- if (speed >= 4) {
- sf->mv_sf.subpel_search_method = SUBPEL_TREE_PRUNED;
-
- sf->inter_sf.alt_ref_search_fp = 1;
-
- sf->interp_sf.skip_sharp_interp_filter_search = 1;
-
- sf->tx_sf.tx_type_search.fast_inter_tx_type_search = 2;
- sf->tx_sf.tx_type_search.fast_intra_tx_type_search = 1;
- sf->tx_sf.use_intra_txb_hash = 0;
-
- sf->rd_sf.use_mb_rd_hash = 0;
-
- sf->winner_mode_sf.tx_size_search_level = frame_is_intra_only(cm) ? 0 : 2;
- }
+ sf->winner_mode_sf.tx_size_search_level = frame_is_intra_only(cm) ? 0 : 2;
if (speed >= 5) {
sf->inter_sf.adaptive_rd_thresh = 4;
@@ -1360,6 +1348,7 @@
sf->tx_sf.tx_type_search.prune_2d_txfm_mode = TX_TYPE_PRUNE_3;
sf->tx_sf.use_inter_txb_hash = 0;
sf->tx_sf.refine_fast_tx_search_results = 0;
+ sf->tx_sf.tx_type_search.fast_intra_tx_type_search = 1;
sf->rd_sf.optimize_coefficients = NO_TRELLIS_OPT;
sf->rd_sf.simple_model_rd_from_var = 1;
