Retune global motion related speed features
* Enable global motion for "good" mode speeds 3 and 4
* Use disflow-based global motion by default, as it is much faster
and gives better results than feature matching
* Add a speed feature, num_refinement_steps, which determines
how many refinement iterations to do in av1_refine_integerized_param
* Fix interaction between the prune_zero_mv_with_sse speed feature
and global motion
* Fix assertions in prune_zero_mv_with_sse code, to check the correct
conditions for compound blocks
* Retune the following speed features:
* Global motion usage
* Number of global motion refinement steps
* prune_zero_mv_with_sse
Results:
| lowres2 | midres2 | hdres2
Speed | BDRATE | Enc time | BDRATE | Enc Time | BDRATE | Enc Time
------+---------+----------+---------+----------+---------+---------
1 | -0.029% | -2.289% | -0.057% | -2.201% | -0.058% | -0.192%
2 | +0.010% | -5.912% | -0.036% | -4.855% | -0.023% | -1.333%
3 | -1.249% | +3.028% | -0.808% | +3.212% | -0.379% | +3.384%
4 | -1.304% | +4.202% | -0.846% | +4.760% | -0.429% | +4.960%
No change for speed 5+, or for realtime mode
STATS_CHANGED
Change-Id: Ie054a79890ecf6ef8b57f0dbffd2ebca5aecbda1
diff --git a/av1/encoder/global_motion_facade.c b/av1/encoder/global_motion_facade.c
index 1c31148..941118b 100644
--- a/av1/encoder/global_motion_facade.c
+++ b/av1/encoder/global_motion_facade.c
@@ -103,6 +103,7 @@
TransformationType model;
int bit_depth = cpi->common.seq_params->bit_depth;
GlobalMotionMethod global_motion_method = cpi->oxcf.global_motion_method;
+ int num_refinements = cpi->sf.gm_sf.num_refinement_steps;
for (model = ROTZOOM; model < GLOBAL_TRANS_TYPES_ENC; ++model) {
int64_t best_warp_error = INT64_MAX;
@@ -151,7 +152,7 @@
ref_buf[frame]->y_buffer, ref_buf[frame]->y_crop_width,
ref_buf[frame]->y_crop_height, ref_buf[frame]->y_stride,
cpi->source->y_buffer, src_width, src_height, src_stride,
- GM_REFINEMENT_COUNT, best_warp_error, segment_map, segment_map_w,
+ num_refinements, best_warp_error, segment_map, segment_map_w,
erroradv_threshold);
// av1_refine_integerized_param() can return a TRANSLATION type model
