From fec5988657cfb810072d657dd545c20036ba9fa1 Mon Sep 17 00:00:00 2001
From: Jingning Han <jingning@google.com>
Date: Fri, 19 Feb 2016 11:53:24 -0800
Subject: [PATCH] Unify motion vector cost system

This commit unifies the motion vector cost buffers for full pixel
and sub-pixel motion search. The new motion vector coding system
provides 0.5% coding gains for 720p and above sequences and 0.2%
for lower resolution sets.

Change-Id: I927ec81eadc39d11a3c12b375221a1ddd2e8bf24
---
 vp10/encoder/block.h           |  4 ++--
 vp10/encoder/encoder.c         |  4 +++-
 vp10/encoder/mcomp.c           | 10 ++++++++++
 vp10/encoder/rd.c              |  4 ++++
 vp10/encoder/temporal_filter.c |  7 +++++++
 5 files changed, 26 insertions(+), 3 deletions(-)

diff --git a/vp10/encoder/block.h b/vp10/encoder/block.h
index aa9139398..0c3e48fea 100644
--- a/vp10/encoder/block.h
+++ b/vp10/encoder/block.h
@@ -113,15 +113,15 @@ struct macroblock {
   int *nmvcost[NMV_CONTEXTS][2];
   int *nmvcost_hp[NMV_CONTEXTS][2];
   int **mv_cost_stack[NMV_CONTEXTS];
+  int *nmvjointsadcost;
 #else
   int nmvjointcost[MV_JOINTS];
   int *nmvcost[2];
   int *nmvcost_hp[2];
+  int nmvjointsadcost[MV_JOINTS];
 #endif
 
   int **mvcost;
-
-  int nmvjointsadcost[MV_JOINTS];
   int *nmvsadcost[2];
   int *nmvsadcost_hp[2];
   int **mvsadcost;
diff --git a/vp10/encoder/encoder.c b/vp10/encoder/encoder.c
index 1ecd668ee..cadd53b53 100644
--- a/vp10/encoder/encoder.c
+++ b/vp10/encoder/encoder.c
@@ -1627,12 +1627,14 @@ void vp10_change_config(struct VP10_COMP *cpi, const VP10EncoderConfig *oxcf) {
 #endif
 #define log2f(x) (log (x) / (float) M_LOG2_E)
 
+#if !CONFIG_REF_MV
 static void cal_nmvjointsadcost(int *mvjointsadcost) {
   mvjointsadcost[0] = 600;
   mvjointsadcost[1] = 300;
   mvjointsadcost[2] = 300;
   mvjointsadcost[3] = 300;
 }
+#endif
 
 static void cal_nmvsadcosts(int *mvsadcost[2]) {
   int i = 1;
@@ -1794,7 +1796,6 @@ VP10_COMP *vp10_create_compressor(VP10EncoderConfig *oxcf,
 
   cpi->first_time_stamp_ever = INT64_MAX;
 
-  cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost);
 #if CONFIG_REF_MV
   for (i = 0; i < NMV_CONTEXTS; ++i) {
     cpi->td.mb.nmvcost[i][0] = &cpi->nmv_costs[i][0][MV_MAX];
@@ -1803,6 +1804,7 @@ VP10_COMP *vp10_create_compressor(VP10EncoderConfig *oxcf,
     cpi->td.mb.nmvcost_hp[i][1] = &cpi->nmv_costs_hp[i][1][MV_MAX];
   }
 #else
+  cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost);
   cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX];
   cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX];
   cpi->td.mb.nmvcost_hp[0] = &cpi->nmvcosts_hp[0][MV_MAX];
diff --git a/vp10/encoder/mcomp.c b/vp10/encoder/mcomp.c
index dd19e029f..6e3b06ab9 100644
--- a/vp10/encoder/mcomp.c
+++ b/vp10/encoder/mcomp.c
@@ -97,12 +97,22 @@ static int mv_err_cost(const MV *mv, const MV *ref, const int *mvjcost,
 
 static int mvsad_err_cost(const MACROBLOCK *x, const MV *mv, const MV *ref,
                           int sad_per_bit) {
+#if CONFIG_REF_MV
+  const MV diff = { (mv->row - ref->row) << 3,
+                    (mv->col - ref->col) << 3 };
+  return ROUND_POWER_OF_TWO(
+      (unsigned)mv_cost(&diff, x->nmvjointsadcost, x->mvsadcost) *
+          sad_per_bit,
+      VP9_PROB_COST_SHIFT);
+#else
   const MV diff = { mv->row - ref->row,
                     mv->col - ref->col };
+
   return ROUND_POWER_OF_TWO(
       (unsigned)mv_cost(&diff, x->nmvjointsadcost, x->nmvsadcost) *
           sad_per_bit,
       VP9_PROB_COST_SHIFT);
+#endif
 }
 
 void vp10_init_dsmotion_compensation(search_site_config *cfg, int stride) {
diff --git a/vp10/encoder/rd.c b/vp10/encoder/rd.c
index 77fc16228..299b7612b 100644
--- a/vp10/encoder/rd.c
+++ b/vp10/encoder/rd.c
@@ -338,6 +338,8 @@ void vp10_set_mvcost(MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame) {
                              mbmi_ext->ref_mv_stack[ref_frame]);
   x->mvcost = x->mv_cost_stack[nmv_ctx];
   x->nmvjointcost = x->nmv_vec_cost[nmv_ctx];
+  x->mvsadcost = x->mvcost;
+  x->nmvjointsadcost = x->nmvjointcost;
 }
 #endif
 
@@ -382,6 +384,8 @@ void vp10_initialize_rd_consts(VP10_COMP *cpi) {
     }
     x->mvcost = x->mv_cost_stack[0];
     x->nmvjointcost = x->nmv_vec_cost[0];
+    x->mvsadcost = x->mvcost;
+    x->nmvjointsadcost = x->nmvjointcost;
 #else
     vp10_build_nmv_cost_table(x->nmvjointcost,
                              cm->allow_high_precision_mv ? x->nmvcost_hp
diff --git a/vp10/encoder/temporal_filter.c b/vp10/encoder/temporal_filter.c
index 035b66abf..afe555d5a 100644
--- a/vp10/encoder/temporal_filter.c
+++ b/vp10/encoder/temporal_filter.c
@@ -293,6 +293,13 @@ static int temporal_filter_find_matching_mb_c(VP10_COMP *cpi,
   step_param = mv_sf->reduce_first_step_size;
   step_param = VPXMIN(step_param, MAX_MVSEARCH_STEPS - 2);
 
+#if CONFIG_REF_MV
+  x->mvcost = x->mv_cost_stack[0];
+  x->nmvjointcost = x->nmv_vec_cost[0];
+  x->mvsadcost = x->mvcost;
+  x->nmvjointsadcost = x->nmvjointcost;
+#endif
+
   // Ignore mv costing by sending NULL pointer instead of cost arrays
   vp10_hex_search(x, &best_ref_mv1_full, step_param, sadpb, 1,
                  cond_cost_list(cpi, cost_list),
-- 
2.40.0