Merge changes Ib9d16a4b,I6061f38c

* changes:
  Refactor av1_nb_mvs_inconsistency()
  Recompute mv inconsistency after mv search is done

3 files changed, 31 insertions, 15 deletions

diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 35f5aa324..549430ab7 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -52,6 +52,9 @@
 #include "vp9/encoder/vp9_extend.h"
 #include "vp9/encoder/vp9_firstpass.h"
 #include "vp9/encoder/vp9_mbgraph.h"
+#if CONFIG_NON_GREEDY_MV
+#include "vp9/encoder/vp9_mcomp.h"
+#endif
 #include "vp9/encoder/vp9_multi_thread.h"
 #include "vp9/encoder/vp9_noise_estimate.h"
 #include "vp9/encoder/vp9_picklpf.h"
@@ -6089,6 +6092,7 @@ static void do_motion_search(VP9_COMP *cpi, ThreadData *td, int frame_idx,
 
 #define CHANGE_MV_SEARCH_ORDER 1
 #define USE_PQSORT 1
+#define RE_COMPUTE_MV_INCONSISTENCY 1
 
 #if CHANGE_MV_SEARCH_ORDER
 #if USE_PQSORT
@@ -6327,11 +6331,21 @@ void mc_flow_dispenser(VP9_COMP *cpi, GF_PICTURE *gf_picture, int frame_idx,
         TplDepStats *this_tpl_stats =
             &tpl_frame->tpl_stats_ptr[mi_row * tpl_frame->stride + mi_col];
         for (rf_idx = 0; rf_idx < 3; ++rf_idx) {
+#if RE_COMPUTE_MV_INCONSISTENCY
+          MV full_mv;
+          int_mv nb_full_mvs[NB_MVS_NUM];
+          prepare_nb_full_mvs(tpl_frame, mi_row, mi_col, rf_idx, bsize,
+                              nb_full_mvs);
+          full_mv.row = this_tpl_stats->mv_arr[rf_idx].as_mv.row >> 3;
+          full_mv.col = this_tpl_stats->mv_arr[rf_idx].as_mv.col >> 3;
+          this_tpl_stats->mv_cost[rf_idx] =
+              av1_nb_mvs_inconsistency(&full_mv, nb_full_mvs);
+#endif  // RE_COMPUTE_MV_INCONSISTENCY
           tpl_frame->mv_dist_sum[rf_idx] += this_tpl_stats->mv_dist[rf_idx];
           tpl_frame->mv_cost_sum[rf_idx] += this_tpl_stats->mv_cost[rf_idx];
         }
       }
-#endif
+#endif  // CONFIG_NON_GREEDY_MV
     }
   }
 }
diff --git a/vp9/encoder/vp9_mcomp.c b/vp9/encoder/vp9_mcomp.c
index 496251622..a2543035c 100644
--- a/vp9/encoder/vp9_mcomp.c
+++ b/vp9/encoder/vp9_mcomp.c
@@ -1733,9 +1733,10 @@ static int exhuastive_mesh_search(const MACROBLOCK *x, MV *ref_mv, MV *best_mv,
 }
 
 #if CONFIG_NON_GREEDY_MV
-static double nb_mvs_inconsistency(const MV *mv, const int_mv *nb_mvs) {
+double av1_nb_mvs_inconsistency(const MV *mv, const int_mv *nb_mvs) {
   int i;
-  double best_cost = -1;
+  int update = 0;
+  double best_cost = 0;
   vpx_clear_system_state();
   for (i = 0; i < NB_MVS_NUM; ++i) {
     if (nb_mvs[i].as_int != INVALID_MV) {
@@ -1744,18 +1745,15 @@ static double nb_mvs_inconsistency(const MV *mv, const int_mv *nb_mvs) {
       const double col_diff = mv->col - nb_mv.col;
       double cost = row_diff * row_diff + col_diff * col_diff;
       cost = log2(1 + cost);
-      if (best_cost < 0) {
+      if (update == 0) {
         best_cost = cost;
+        update = 1;
       } else {
         best_cost = cost < best_cost ? cost : best_cost;
       }
     }
   }
-  if (best_cost < 0) {
-    return 0;
-  } else {
-    return best_cost;
-  }
+  return best_cost;
 }
 
 double vp9_diamond_search_sad_new(const MACROBLOCK *x,
@@ -1801,7 +1799,7 @@ double vp9_diamond_search_sad_new(const MACROBLOCK *x,
 
   // Check the starting position
   *best_mv_dist = fn_ptr->sdf(what, what_stride, in_what, in_what_stride);
-  *best_mv_cost = nb_mvs_inconsistency(best_full_mv, nb_full_mvs);
+  *best_mv_cost = av1_nb_mvs_inconsistency(best_full_mv, nb_full_mvs);
   bestsad = (*best_mv_dist) + lambda * (*best_mv_cost);
 
   i = 0;
@@ -1834,7 +1832,8 @@ double vp9_diamond_search_sad_new(const MACROBLOCK *x,
           const MV this_mv = { best_full_mv->row + ss_mv[i].row,
                                best_full_mv->col + ss_mv[i].col };
           const double mv_dist = sad_array[t];
-          const double mv_cost = nb_mvs_inconsistency(&this_mv, nb_full_mvs);
+          const double mv_cost =
+              av1_nb_mvs_inconsistency(&this_mv, nb_full_mvs);
           double thissad = mv_dist + lambda * mv_cost;
           if (thissad < bestsad) {
             bestsad = thissad;
@@ -1854,7 +1853,8 @@ double vp9_diamond_search_sad_new(const MACROBLOCK *x,
           const uint8_t *const check_here = ss_os[i] + best_address;
           const double mv_dist =
               fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
-          const double mv_cost = nb_mvs_inconsistency(&this_mv, nb_full_mvs);
+          const double mv_cost =
+              av1_nb_mvs_inconsistency(&this_mv, nb_full_mvs);
           double thissad = mv_dist + lambda * mv_cost;
           if (thissad < bestsad) {
             bestsad = thissad;
@@ -2440,7 +2440,7 @@ double vp9_refining_search_sad_new(const MACROBLOCK *x, MV *best_full_mv,
   vpx_clear_system_state();
   *best_mv_dist =
       fn_ptr->sdf(what->buf, what->stride, best_address, in_what->stride);
-  *best_mv_cost = nb_mvs_inconsistency(best_full_mv, nb_full_mvs);
+  *best_mv_cost = av1_nb_mvs_inconsistency(best_full_mv, nb_full_mvs);
   best_sad = (*best_mv_dist) + lambda * (*best_mv_cost);
 
   for (i = 0; i < search_range; i++) {
@@ -2462,7 +2462,7 @@ double vp9_refining_search_sad_new(const MACROBLOCK *x, MV *best_full_mv,
         const MV mv = { best_full_mv->row + neighbors[j].row,
                         best_full_mv->col + neighbors[j].col };
         const double mv_dist = sads[j];
-        const double mv_cost = nb_mvs_inconsistency(&mv, nb_full_mvs);
+        const double mv_cost = av1_nb_mvs_inconsistency(&mv, nb_full_mvs);
         const double thissad = mv_dist + lambda * mv_cost;
         if (thissad < best_sad) {
           best_sad = thissad;
@@ -2480,7 +2480,7 @@ double vp9_refining_search_sad_new(const MACROBLOCK *x, MV *best_full_mv,
           const double mv_dist =
               fn_ptr->sdf(what->buf, what->stride,
                           get_buf_from_mv(in_what, &mv), in_what->stride);
-          const double mv_cost = nb_mvs_inconsistency(&mv, nb_full_mvs);
+          const double mv_cost = av1_nb_mvs_inconsistency(&mv, nb_full_mvs);
           const double thissad = mv_dist + lambda * mv_cost;
           if (thissad < best_sad) {
             best_sad = thissad;
diff --git a/vp9/encoder/vp9_mcomp.h b/vp9/encoder/vp9_mcomp.h
index 558ecbcd8..a159cb288 100644
--- a/vp9/encoder/vp9_mcomp.h
+++ b/vp9/encoder/vp9_mcomp.h
@@ -134,6 +134,8 @@ double vp9_full_pixel_diamond_new(const struct VP9_COMP *cpi, MACROBLOCK *x,
                                   const vp9_variance_fn_ptr_t *fn_ptr,
                                   const int_mv *nb_full_mvs,
                                   struct TplDepStats *tpl_stats, int rf_idx);
+
+double av1_nb_mvs_inconsistency(const MV *mv, const int_mv *nb_mvs);
 #endif  // CONFIG_NON_GREEDY_MV
 #ifdef __cplusplus
 }  // extern "C"