Improve MV prediction accuracy to achieve performance gain

Add vp8_mv_pred() to better predict starting MV for NEWMV
mode in vp8_rd_pick_inter_mode(). Set different search
ranges according to MV prediction accuracy, which improves
encoder performance without hurting the quality. Also,
as Yaowu suggested, using diamond search result as full
search starting point and therefore adjusting(reducing)
full search range helps the performance.

Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6

8 files changed, 487 insertions, 42 deletions

diff --git a/vp8/encoder/firstpass.c b/vp8/encoder/firstpass.c
index 8a94fa369..539a28f4d 100644
--- a/vp8/encoder/firstpass.c
+++ b/vp8/encoder/firstpass.c
@@ -472,7 +472,7 @@ void vp8_first_pass_motion_search(VP8_COMP *cpi, MACROBLOCK *x, MV *ref_mv, MV *
     xd->pre.y_buffer = recon_buffer->y_buffer + recon_yoffset;
 
     // Initial step/diamond search centred on best mv
-    tmp_err = cpi->diamond_search_sad(x, b, d, ref_mv, &tmp_mv, step_param, x->errorperbit, &num00, &v_fn_ptr, x->mvsadcost, x->mvcost);
+    tmp_err = cpi->diamond_search_sad(x, b, d, ref_mv, &tmp_mv, step_param, x->errorperbit, &num00, &v_fn_ptr, x->mvsadcost, x->mvcost, ref_mv);
     if ( tmp_err < INT_MAX-new_mv_mode_penalty )
         tmp_err += new_mv_mode_penalty;
 
@@ -495,7 +495,7 @@ void vp8_first_pass_motion_search(VP8_COMP *cpi, MACROBLOCK *x, MV *ref_mv, MV *
             num00--;
         else
         {
-            tmp_err = cpi->diamond_search_sad(x, b, d, ref_mv, &tmp_mv, step_param + n, x->errorperbit, &num00, &v_fn_ptr, x->mvsadcost, x->mvcost);
+            tmp_err = cpi->diamond_search_sad(x, b, d, ref_mv, &tmp_mv, step_param + n, x->errorperbit, &num00, &v_fn_ptr, x->mvsadcost, x->mvcost, ref_mv);
             if ( tmp_err < INT_MAX-new_mv_mode_penalty )
                 tmp_err += new_mv_mode_penalty;
 
diff --git a/vp8/encoder/mcomp.c b/vp8/encoder/mcomp.c
index bb85afa6f..b9d86a20e 100644
--- a/vp8/encoder/mcomp.c
+++ b/vp8/encoder/mcomp.c
@@ -913,7 +913,8 @@ int vp8_diamond_search_sad
     int *num00,
     vp8_variance_fn_ptr_t *fn_ptr,
     int *mvsadcost[2],
-    int *mvcost[2]
+    int *mvcost[2],
+    MV *center_mv
 )
 {
     int i, j, step;
@@ -949,7 +950,7 @@ int vp8_diamond_search_sad
     (ref_row > x->mv_row_min) && (ref_row < x->mv_row_max))
     {
         // Check the starting position
-        bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff) + vp8_mv_err_cost(ref_mv, ref_mv, mvsadcost, error_per_bit);
+        bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff) + vp8_mv_err_cost(ref_mv, center_mv, mvsadcost, error_per_bit);
     }
 
     // search_param determines the length of the initial step and hence the number of iterations
@@ -982,7 +983,7 @@ int vp8_diamond_search_sad
                 {
                     this_mv.row = this_row_offset << 3;
                     this_mv.col = this_col_offset << 3;
-                    thissad += vp8_mv_err_cost(&this_mv, ref_mv, mvsadcost, error_per_bit);
+                    thissad += vp8_mv_err_cost(&this_mv, center_mv, mvsadcost, error_per_bit);
 
                     if (thissad < bestsad)
                     {
@@ -1013,7 +1014,7 @@ int vp8_diamond_search_sad
         return INT_MAX;
 
     return fn_ptr->vf(what, what_stride, best_address, in_what_stride, (unsigned int *)(&thissad))
-    + vp8_mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+    + vp8_mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
 }
 
 int vp8_diamond_search_sadx4
@@ -1028,7 +1029,8 @@ int vp8_diamond_search_sadx4
     int *num00,
     vp8_variance_fn_ptr_t *fn_ptr,
     int *mvsadcost[2],
-    int *mvcost[2]
+    int *mvcost[2],
+    MV *center_mv
 )
 {
     int i, j, step;
@@ -1064,7 +1066,7 @@ int vp8_diamond_search_sadx4
     (ref_row > x->mv_row_min) && (ref_row < x->mv_row_max))
     {
         // Check the starting position
-        bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff) + vp8_mv_err_cost(ref_mv, ref_mv, mvsadcost, error_per_bit);
+        bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff) + vp8_mv_err_cost(ref_mv, center_mv, mvsadcost, error_per_bit);
     }
 
     // search_param determines the length of the initial step and hence the number of iterations
@@ -1108,7 +1110,7 @@ int vp8_diamond_search_sadx4
                     {
                         this_mv.row = (best_mv->row + ss[i].mv.row) << 3;
                         this_mv.col = (best_mv->col + ss[i].mv.col) << 3;
-                        sad_array[t] += vp8_mv_err_cost(&this_mv, ref_mv, mvsadcost, error_per_bit);
+                        sad_array[t] += vp8_mv_err_cost(&this_mv, center_mv, mvsadcost, error_per_bit);
 
                         if (sad_array[t] < bestsad)
                         {
@@ -1137,7 +1139,7 @@ int vp8_diamond_search_sadx4
                     {
                         this_mv.row = this_row_offset << 3;
                         this_mv.col = this_col_offset << 3;
-                        thissad += vp8_mv_err_cost(&this_mv, ref_mv, mvsadcost, error_per_bit);
+                        thissad += vp8_mv_err_cost(&this_mv, center_mv, mvsadcost, error_per_bit);
 
                         if (thissad < bestsad)
                         {
@@ -1168,12 +1170,12 @@ int vp8_diamond_search_sadx4
         return INT_MAX;
 
     return fn_ptr->vf(what, what_stride, best_address, in_what_stride, (unsigned int *)(&thissad))
-    + vp8_mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+    + vp8_mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
 }
 
 
 #if !(CONFIG_REALTIME_ONLY)
-int vp8_full_search_sad(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int error_per_bit, int distance, vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2], int *mvsadcost[2])
+int vp8_full_search_sad(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int error_per_bit, int distance, vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2], int *mvsadcost[2], MV *center_mv)
 {
     unsigned char *what = (*(b->base_src) + b->src);
     int what_stride = b->src_stride;
@@ -1211,7 +1213,7 @@ int vp8_full_search_sad(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int erro
         // Baseline value at the centre
 
         //bestsad = fn_ptr->sf( what,what_stride,bestaddress,in_what_stride) + (int)sqrt(vp8_mv_err_cost(ref_mv,ref_mv, mvcost,error_per_bit*14));
-        bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride, 0x7fffffff) + vp8_mv_err_cost(ref_mv, ref_mv, mvsadcost, error_per_bit);
+        bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride, 0x7fffffff) + vp8_mv_err_cost(ref_mv, center_mv, mvsadcost, error_per_bit);
     }
 
     // Apply further limits to prevent us looking using vectors that stretch beyiond the UMV border
@@ -1239,7 +1241,7 @@ int vp8_full_search_sad(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int erro
             this_mv.col = c << 3;
             //thissad += (int)sqrt(vp8_mv_err_cost(&this_mv,ref_mv, mvcost,error_per_bit*14));
             //thissad  += error_per_bit * mv_bits_sadcost[mv_bits(&this_mv, ref_mv, mvcost)];
-            thissad  += vp8_mv_err_cost(&this_mv, ref_mv, mvsadcost, error_per_bit); //mv_bits(error_per_bit, &this_mv, ref_mv, mvsadcost);
+            thissad  += vp8_mv_err_cost(&this_mv, center_mv, mvsadcost, error_per_bit); //mv_bits(error_per_bit, &this_mv, ref_mv, mvsadcost);
 
             if (thissad < bestsad)
             {
@@ -1258,12 +1260,12 @@ int vp8_full_search_sad(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int erro
 
     if (bestsad < INT_MAX)
         return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, (unsigned int *)(&thissad))
-        + vp8_mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+        + vp8_mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
     else
         return INT_MAX;
 }
 
-int vp8_full_search_sadx3(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int error_per_bit, int distance, vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2], int *mvsadcost[2])
+int vp8_full_search_sadx3(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int error_per_bit, int distance, vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2], int *mvsadcost[2], MV *center_mv)
 {
     unsigned char *what = (*(b->base_src) + b->src);
     int what_stride = b->src_stride;
@@ -1301,7 +1303,7 @@ int vp8_full_search_sadx3(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int er
     (ref_row > x->mv_row_min) && (ref_row < x->mv_row_max))
     {
         // Baseline value at the centre
-        bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride, 0x7fffffff) + vp8_mv_err_cost(ref_mv, ref_mv, mvsadcost, error_per_bit);
+        bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride, 0x7fffffff) + vp8_mv_err_cost(ref_mv, center_mv, mvsadcost, error_per_bit);
     }
 
     // Apply further limits to prevent us looking using vectors that stretch beyiond the UMV border
@@ -1336,7 +1338,7 @@ int vp8_full_search_sadx3(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int er
                 if (thissad < bestsad)
                 {
                     this_mv.col = c << 3;
-                    thissad  += vp8_mv_err_cost(&this_mv, ref_mv, mvsadcost, error_per_bit);
+                    thissad  += vp8_mv_err_cost(&this_mv, center_mv, mvsadcost, error_per_bit);
 
                     if (thissad < bestsad)
                     {
@@ -1359,7 +1361,7 @@ int vp8_full_search_sadx3(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int er
             if (thissad < bestsad)
             {
                 this_mv.col = c << 3;
-                thissad  += vp8_mv_err_cost(&this_mv, ref_mv, mvsadcost, error_per_bit);
+                thissad  += vp8_mv_err_cost(&this_mv, center_mv, mvsadcost, error_per_bit);
 
                 if (thissad < bestsad)
                 {
@@ -1381,14 +1383,14 @@ int vp8_full_search_sadx3(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int er
 
     if (bestsad < INT_MAX)
         return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, (unsigned int *)(&thissad))
-        + vp8_mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+        + vp8_mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
     else
         return INT_MAX;
 }
 #endif
 
 
-int vp8_full_search_sadx8(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int error_per_bit, int distance, vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2], int *mvsadcost[2])
+int vp8_full_search_sadx8(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int error_per_bit, int distance, vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2], int *mvsadcost[2], MV *center_mv)
 {
     unsigned char *what = (*(b->base_src) + b->src);
     int what_stride = b->src_stride;
@@ -1427,7 +1429,7 @@ int vp8_full_search_sadx8(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int er
     (ref_row > x->mv_row_min) && (ref_row < x->mv_row_max))
     {
         // Baseline value at the centre
-        bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride, 0x7fffffff) + vp8_mv_err_cost(ref_mv, ref_mv, mvsadcost, error_per_bit);
+        bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride, 0x7fffffff) + vp8_mv_err_cost(ref_mv, center_mv, mvsadcost, error_per_bit);
     }
 
     // Apply further limits to prevent us looking using vectors that stretch beyiond the UMV border
@@ -1462,7 +1464,7 @@ int vp8_full_search_sadx8(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int er
                 if (thissad < bestsad)
                 {
                     this_mv.col = c << 3;
-                    thissad  += vp8_mv_err_cost(&this_mv, ref_mv, mvsadcost, error_per_bit);
+                    thissad  += vp8_mv_err_cost(&this_mv, center_mv, mvsadcost, error_per_bit);
 
                     if (thissad < bestsad)
                     {
@@ -1491,7 +1493,7 @@ int vp8_full_search_sadx8(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int er
                 if (thissad < bestsad)
                 {
                     this_mv.col = c << 3;
-                    thissad  += vp8_mv_err_cost(&this_mv, ref_mv, mvsadcost, error_per_bit);
+                    thissad  += vp8_mv_err_cost(&this_mv, center_mv, mvsadcost, error_per_bit);
 
                     if (thissad < bestsad)
                     {
@@ -1514,7 +1516,7 @@ int vp8_full_search_sadx8(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int er
             if (thissad < bestsad)
             {
                 this_mv.col = c << 3;
-                thissad  += vp8_mv_err_cost(&this_mv, ref_mv, mvsadcost, error_per_bit);
+                thissad  += vp8_mv_err_cost(&this_mv, center_mv, mvsadcost, error_per_bit);
 
                 if (thissad < bestsad)
                 {
@@ -1535,7 +1537,7 @@ int vp8_full_search_sadx8(MACROBLOCK *x, BLOCK *b, BLOCKD *d, MV *ref_mv, int er
 
     if (bestsad < INT_MAX)
         return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, (unsigned int *)(&thissad))
-        + vp8_mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
+        + vp8_mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
     else
         return INT_MAX;
 }
diff --git a/vp8/encoder/mcomp.h b/vp8/encoder/mcomp.h
index 7d6036248..b54fe7377 100644
--- a/vp8/encoder/mcomp.h
+++ b/vp8/encoder/mcomp.h
@@ -67,7 +67,8 @@ extern fractional_mv_step_fp vp8_skip_fractional_mv_step;
      int distance, \
      vp8_variance_fn_ptr_t *fn_ptr, \
      int *mvcost[2], \
-     int *mvsadcost[2] \
+     int *mvsadcost[2], \
+     MV *center_mv \
     )
 
 #define prototype_diamond_search_sad(sym)\
@@ -83,7 +84,8 @@ extern fractional_mv_step_fp vp8_skip_fractional_mv_step;
      int *num00, \
      vp8_variance_fn_ptr_t *fn_ptr, \
      int *mvsadcost[2], \
-     int *mvcost[2] \
+     int *mvcost[2], \
+     MV *center_mv \
     )
 
 #if ARCH_X86 || ARCH_X86_64
diff --git a/vp8/encoder/onyx_if.c b/vp8/encoder/onyx_if.c
index 5f02a5a02..fd7b5668a 100644
--- a/vp8/encoder/onyx_if.c
+++ b/vp8/encoder/onyx_if.c
@@ -283,6 +283,21 @@ static void setup_features(VP8_COMP *cpi)
 
 void vp8_dealloc_compressor_data(VP8_COMP *cpi)
 {
+    // Delete last frame MV storage buffers
+    if (cpi->lfmv != 0)
+        vpx_free(cpi->lfmv);
+
+    cpi->lfmv = 0;
+
+    if (cpi->lf_ref_frame_sign_bias != 0)
+        vpx_free(cpi->lf_ref_frame_sign_bias);
+
+    cpi->lf_ref_frame_sign_bias = 0;
+
+    if (cpi->lf_ref_frame != 0)
+        vpx_free(cpi->lf_ref_frame);
+
+    cpi->lf_ref_frame = 0;
 
     // Delete sementation map
     if (cpi->segmentation_map != 0)
@@ -2145,7 +2160,10 @@ VP8_PTR vp8_create_compressor(VP8_CONFIG *oxcf)
     cpi->alt_is_last  = 0 ;
     cpi->gold_is_alt  = 0 ;
 
-
+    // allocate memory for storing last frame's MVs for MV prediction.
+    CHECK_MEM_ERROR(cpi->lfmv, vpx_calloc((cpi->common.mb_rows+1) * (cpi->common.mb_cols+1), sizeof(int_mv)));
+    CHECK_MEM_ERROR(cpi->lf_ref_frame_sign_bias, vpx_calloc((cpi->common.mb_rows+1) * (cpi->common.mb_cols+1), sizeof(int)));
+    CHECK_MEM_ERROR(cpi->lf_ref_frame, vpx_calloc((cpi->common.mb_rows+1) * (cpi->common.mb_cols+1), sizeof(int)));
 
     // Create the encoder segmentation map and set all entries to 0
     CHECK_MEM_ERROR(cpi->segmentation_map, vpx_calloc(cpi->common.mb_rows * cpi->common.mb_cols, 1));
@@ -4165,6 +4183,60 @@ static void encode_frame_to_data_rate
     }
 #endif
 
+
+
+
+////////////////////////////////
+////////////////////////////////
+    // This frame's MVs are saved and will be used in next frame's MV prediction.
+    if(cm->show_frame)   //do not save for altref frame
+    {
+      int mb_row;
+      int mb_col;
+      MODE_INFO *tmp = cm->mip; //point to beginning of allocated MODE_INFO arrays.
+      //static int last_video_frame = 0;
+
+      /*
+      if (cm->current_video_frame == 0)   //first frame: set to 0
+      {
+        for (mb_row = 0; mb_row < cm->mb_rows+1; mb_row ++)
+        {
+            for (mb_col = 0; mb_col < cm->mb_cols+1; mb_col ++)
+            {
+                cpi->lfmv[mb_col + mb_row*(cm->mode_info_stride)].as_int = 0;
+                cpi->lf_ref_frame_sign_bias[mb_col + mb_row*(cm->mode_info_stride)] = 0;
+                cpi->lf_ref_frame[mb_col + mb_row*(cm->mode_info_stride)] = 0;
+            }
+        }
+      }else
+      */
+
+      if(cm->frame_type != KEY_FRAME)
+      {
+        for (mb_row = 0; mb_row < cm->mb_rows+1; mb_row ++)
+        {
+          for (mb_col = 0; mb_col < cm->mb_cols+1; mb_col ++)
+          {
+              if(tmp->mbmi.ref_frame != INTRA_FRAME)
+                cpi->lfmv[mb_col + mb_row*(cm->mode_info_stride)].as_int = tmp->mbmi.mv.as_int;
+
+              cpi->lf_ref_frame_sign_bias[mb_col + mb_row*(cm->mode_info_stride)] = cm->ref_frame_sign_bias[tmp->mbmi.ref_frame];
+              cpi->lf_ref_frame[mb_col + mb_row*(cm->mode_info_stride)] = tmp->mbmi.ref_frame;
+              //printf("[%d, %d]  ", cpi->lfmv[mb_col + mb_row*(cm->mode_info_stride-1)].as_mv.row, cpi->lfmv[mb_col + mb_row*(cm->mode_info_stride-1)].as_mv.col);
+              tmp++;
+          }
+        }
+
+      //last_video_frame = cm->current_video_frame;
+      }
+    }
+
+//printf("after: %d   %d \n", cm->current_video_frame, cm->show_frame );
+
+
+
+
+
     // Update the GF useage maps.
     // This is done after completing the compression of a frame when all modes etc. are finalized but before loop filter
     vp8_update_gf_useage_maps(cpi, cm, &cpi->mb);
diff --git a/vp8/encoder/onyx_int.h b/vp8/encoder/onyx_int.h
index 81e32f031..31e627b5c 100644
--- a/vp8/encoder/onyx_int.h
+++ b/vp8/encoder/onyx_int.h
@@ -239,6 +239,12 @@ enum
     BLOCK_MAX_SEGMENTS
 };
 
+typedef union
+{
+    unsigned int as_int;
+    MV           as_mv;
+} int_mv;        /* facilitates rapid equality tests */
+
 typedef struct
 {
 
@@ -661,6 +667,10 @@ typedef struct
     unsigned char *gf_active_flags;   // Record of which MBs still refer to last golden frame either directly or through 0,0
     int gf_active_count;
 
+    //Store last frame's MV info for next frame MV prediction
+    int_mv *lfmv;
+    int *lf_ref_frame_sign_bias;
+    int *lf_ref_frame;
 
 } VP8_COMP;
 
diff --git a/vp8/encoder/pickinter.c b/vp8/encoder/pickinter.c
index 2f7dd9c7c..8dfca351c 100644
--- a/vp8/encoder/pickinter.c
+++ b/vp8/encoder/pickinter.c
@@ -685,7 +685,7 @@ int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int rec
 #if 0
 
             // Initial step Search
-            bestsme = vp8_diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param, x->errorperbit, &num00, &cpi->fn_ptr, cpi->mb.mvsadcost, cpi->mb.mvcost);
+            bestsme = vp8_diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param, x->errorperbit, &num00, &cpi->fn_ptr, cpi->mb.mvsadcost, cpi->mb.mvcost, &best_ref_mv1);
             mode_mv[NEWMV].row = d->bmi.mv.as_mv.row;
             mode_mv[NEWMV].col = d->bmi.mv.as_mv.col;
 
@@ -698,7 +698,7 @@ int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int rec
                     num00--;
                 else
                 {
-                    thissme = vp8_diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param + n, x->errorperbit, &num00, &cpi->fn_ptr, cpi->mb.mvsadcost, x->mvcost);
+                    thissme = vp8_diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param + n, x->errorperbit, &num00, &cpi->fn_ptr, cpi->mb.mvsadcost, x->mvcost, &best_ref_mv1);
 
                     if (thissme < bestsme)
                     {
@@ -724,7 +724,7 @@ int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int rec
             }
             else
             {
-                bestsme = cpi->diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param, sadpb / 2/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost); //sadpb < 9
+                bestsme = cpi->diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param, sadpb / 2/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost, &best_ref_mv1); //sadpb < 9
                 mode_mv[NEWMV].row = d->bmi.mv.as_mv.row;
                 mode_mv[NEWMV].col = d->bmi.mv.as_mv.col;
 
@@ -743,7 +743,7 @@ int vp8_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int rec
                         num00--;
                     else
                     {
-                        thissme = cpi->diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param + n, sadpb / 4/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost); //sadpb = 9
+                        thissme = cpi->diamond_search_sad(x, b, d, &best_ref_mv1, &d->bmi.mv.as_mv, step_param + n, sadpb / 4/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost, &best_ref_mv1); //sadpb = 9
 
                         if (thissme < bestsme)
                         {
diff --git a/vp8/encoder/rdopt.c b/vp8/encoder/rdopt.c
index 8a753fd44..99129ace0 100644
--- a/vp8/encoder/rdopt.c
+++ b/vp8/encoder/rdopt.c
@@ -1236,7 +1236,7 @@ static int vp8_rd_pick_best_mbsegmentation(VP8_COMP *cpi, MACROBLOCK *x, MV *bes
                             bestsme = vp8_hex_search(x, c, e, best_ref_mv, &mode_mv[NEW4X4], step_param, sadpb/*x->errorperbit*/, &num00, v_fn_ptr, x->mvsadcost, mvcost);
                         else
                         {
-                            bestsme = cpi->diamond_search_sad(x, c, e, best_ref_mv, &mode_mv[NEW4X4], step_param, sadpb / 2/*x->errorperbit*/, &num00, v_fn_ptr, x->mvsadcost, mvcost);
+                            bestsme = cpi->diamond_search_sad(x, c, e, best_ref_mv, &mode_mv[NEW4X4], step_param, sadpb / 2/*x->errorperbit*/, &num00, v_fn_ptr, x->mvsadcost, mvcost, best_ref_mv);
 
                             n = num00;
                             num00 = 0;
@@ -1249,7 +1249,7 @@ static int vp8_rd_pick_best_mbsegmentation(VP8_COMP *cpi, MACROBLOCK *x, MV *bes
                                     num00--;
                                 else
                                 {
-                                    thissme = cpi->diamond_search_sad(x, c, e, best_ref_mv, &temp_mv, step_param + n, sadpb / 2/*x->errorperbit*/, &num00, v_fn_ptr, x->mvsadcost, mvcost);
+                                    thissme = cpi->diamond_search_sad(x, c, e, best_ref_mv, &temp_mv, step_param + n, sadpb / 2/*x->errorperbit*/, &num00, v_fn_ptr, x->mvsadcost, mvcost, best_ref_mv);
 
                                     if (thissme < bestsme)
                                     {
@@ -1264,7 +1264,7 @@ static int vp8_rd_pick_best_mbsegmentation(VP8_COMP *cpi, MACROBLOCK *x, MV *bes
                         // Should we do a full search (best quality only)
                         if ((compressor_speed == 0) && (bestsme >> sseshift) > 4000)
                         {
-                            thissme = cpi->full_search_sad(x, c, e, best_ref_mv, sadpb / 4, 16, v_fn_ptr, x->mvcost, x->mvsadcost);
+                            thissme = cpi->full_search_sad(x, c, e, best_ref_mv, sadpb / 4, 16, v_fn_ptr, x->mvcost, x->mvsadcost, best_ref_mv);
 
                             if (thissme < bestsme)
                             {
@@ -1385,6 +1385,273 @@ static int vp8_rd_pick_best_mbsegmentation(VP8_COMP *cpi, MACROBLOCK *x, MV *bes
 }
 
 
+
+/////////////////////////
+static void mv_bias(const MODE_INFO *x, int refframe, int_mv *mvp, const int *ref_frame_sign_bias)
+{
+    MV xmv;
+    xmv = x->mbmi.mv.as_mv;
+
+    if (ref_frame_sign_bias[x->mbmi.ref_frame] != ref_frame_sign_bias[refframe])
+    {
+        xmv.row *= -1;
+        xmv.col *= -1;
+    }
+
+    mvp->as_mv = xmv;
+}
+
+static void lf_mv_bias(const int lf_ref_frame_sign_bias, int refframe, int_mv *mvp, const int *ref_frame_sign_bias)
+{
+    MV xmv;
+    xmv = mvp->as_mv;
+
+    if (lf_ref_frame_sign_bias != ref_frame_sign_bias[refframe])
+    {
+        xmv.row *= -1;
+        xmv.col *= -1;
+    }
+
+    mvp->as_mv = xmv;
+}
+
+static void vp8_clamp_mv(MV *mv, const MACROBLOCKD *xd)
+{
+    if (mv->col < (xd->mb_to_left_edge - LEFT_TOP_MARGIN))
+        mv->col = xd->mb_to_left_edge - LEFT_TOP_MARGIN;
+    else if (mv->col > xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN)
+        mv->col = xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN;
+
+    if (mv->row < (xd->mb_to_top_edge - LEFT_TOP_MARGIN))
+        mv->row = xd->mb_to_top_edge - LEFT_TOP_MARGIN;
+    else if (mv->row > xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN)
+        mv->row = xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN;
+}
+
+static void swap(int *x,int *y)
+{
+   int tmp;
+
+   tmp = *x;
+   *x = *y;
+   *y = tmp;
+}
+
+static void quicksortmv(int arr[],int left, int right)
+{
+   int lidx,ridx,pivot;
+
+   lidx = left;
+   ridx = right;
+
+   if( left < right)
+   {
+      pivot = (left + right)/2;
+
+      while(lidx <=pivot && ridx >=pivot)
+      {
+          while(arr[lidx] < arr[pivot] && lidx <= pivot)
+              lidx++;
+          while(arr[ridx] > arr[pivot] && ridx >= pivot)
+              ridx--;
+          swap(&arr[lidx], &arr[ridx]);
+          lidx++;
+          ridx--;
+          if(lidx-1 == pivot)
+          {
+              ridx++;
+              pivot = ridx;
+          }
+          else if(ridx+1 == pivot)
+          {
+              lidx--;
+              pivot = lidx;
+          }
+      }
+      quicksortmv(arr, left, pivot - 1);
+      quicksortmv(arr, pivot + 1, right);
+   }
+}
+
+static void quicksortsad(int arr[],int idx[], int left, int right)
+{
+   int lidx,ridx,pivot;
+
+   lidx = left;
+   ridx = right;
+
+   if( left < right)
+   {
+      pivot = (left + right)/2;
+
+      while(lidx <=pivot && ridx >=pivot)
+      {
+          while(arr[lidx] < arr[pivot] && lidx <= pivot)
+              lidx++;
+          while(arr[ridx] > arr[pivot] && ridx >= pivot)
+              ridx--;
+          swap(&arr[lidx], &arr[ridx]);
+          swap(&idx[lidx], &idx[ridx]);
+          lidx++;
+          ridx--;
+          if(lidx-1 == pivot)
+          {
+              ridx++;
+              pivot = ridx;
+          }
+          else if(ridx+1 == pivot)
+          {
+              lidx--;
+              pivot = lidx;
+          }
+      }
+      quicksortsad(arr, idx, left, pivot - 1);
+      quicksortsad(arr, idx, pivot + 1, right);
+   }
+}
+
+//The improved MV prediction
+static void vp8_mv_pred
+(
+    VP8_COMP *cpi,
+    MACROBLOCKD *xd,
+    const MODE_INFO *here,
+    MV *mvp,
+    int refframe,
+    int *ref_frame_sign_bias,
+    int *sr,
+    int near_sadidx[]
+)
+{
+    const MODE_INFO *above = here - xd->mode_info_stride;
+    const MODE_INFO *left = here - 1;
+    const MODE_INFO *aboveleft = above - 1;
+    int_mv           near_mvs[7];
+    int              near_ref[7];
+    int_mv           mv;
+    int              vcnt=0;
+    int              find=0;
+    int              mb_offset;
+
+    int              mvx[7];
+    int              mvy[7];
+    int              i;
+
+    mv.as_int = 0;
+
+    if(here->mbmi.ref_frame != INTRA_FRAME)
+    {
+        near_mvs[0].as_int = near_mvs[1].as_int = near_mvs[2].as_int = near_mvs[3].as_int = near_mvs[4].as_int = near_mvs[5].as_int = near_mvs[6].as_int = 0;
+        near_ref[0] = near_ref[1] = near_ref[2] = near_ref[3] = near_ref[4] = near_ref[5] = near_ref[6] = 0;
+
+        // read in 3 nearby block's MVs from current frame as prediction candidates.
+        if (above->mbmi.ref_frame != INTRA_FRAME)
+        {
+            near_mvs[vcnt].as_int = above->mbmi.mv.as_int;
+            mv_bias(above, refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+            near_ref[vcnt] =  above->mbmi.ref_frame;
+        }
+        vcnt++;
+        if (left->mbmi.ref_frame != INTRA_FRAME)
+        {
+            near_mvs[vcnt].as_int = left->mbmi.mv.as_int;
+            mv_bias(left, refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+            near_ref[vcnt] =  left->mbmi.ref_frame;
+        }
+        vcnt++;
+        if (aboveleft->mbmi.ref_frame != INTRA_FRAME)
+        {
+            near_mvs[vcnt].as_int = aboveleft->mbmi.mv.as_int;
+            mv_bias(aboveleft, refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+            near_ref[vcnt] =  aboveleft->mbmi.ref_frame;
+        }
+        vcnt++;
+
+        // read in 4 nearby block's MVs from last frame.
+        if(cpi->common.last_frame_type != KEY_FRAME)
+        {
+            mb_offset = (-xd->mb_to_top_edge/128 + 1) * (xd->mode_info_stride) + (-xd->mb_to_left_edge/128 +1) ;
+
+            // current in last frame
+            if (cpi->lf_ref_frame[mb_offset] != INTRA_FRAME)
+            {
+                near_mvs[vcnt].as_int = cpi->lfmv[mb_offset].as_int;
+                lf_mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+                near_ref[vcnt] =  cpi->lf_ref_frame[mb_offset];
+            }
+            vcnt++;
+
+            // above in last frame
+            if (cpi->lf_ref_frame[mb_offset - xd->mode_info_stride] != INTRA_FRAME)
+            {
+                near_mvs[vcnt].as_int = cpi->lfmv[mb_offset - xd->mode_info_stride].as_int;
+                lf_mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset - xd->mode_info_stride], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+                near_ref[vcnt] =  cpi->lf_ref_frame[mb_offset - xd->mode_info_stride];
+            }
+            vcnt++;
+
+            // left in last frame
+            if (cpi->lf_ref_frame[mb_offset-1] != INTRA_FRAME)
+            {
+                near_mvs[vcnt].as_int = cpi->lfmv[mb_offset -1].as_int;
+                lf_mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset -1], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+                near_ref[vcnt] =  cpi->lf_ref_frame[mb_offset - 1];
+            }
+            vcnt++;
+
+            // aboveleft in last frame
+            if (cpi->lf_ref_frame[mb_offset - xd->mode_info_stride -1] != INTRA_FRAME)
+            {
+                near_mvs[vcnt].as_int = cpi->lfmv[mb_offset - xd->mode_info_stride -1].as_int;
+                lf_mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset - xd->mode_info_stride -1], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
+                near_ref[vcnt] =  cpi->lf_ref_frame[mb_offset - xd->mode_info_stride -1];
+            }
+            vcnt++;
+        }
+
+        for(i=0; i< vcnt; i++)
+        {
+            if(near_ref[near_sadidx[i]] != INTRA_FRAME)
+            {
+                if(here->mbmi.ref_frame == near_ref[near_sadidx[i]])
+                {
+                    mv.as_int = near_mvs[near_sadidx[i]].as_int;
+                    find = 1;
+                    if(vcnt<2)
+                        *sr = 4;
+                    else if (vcnt<4)
+                        *sr = 3;
+                    else
+                        *sr = 2;
+                    break;
+                }
+            }
+        }
+
+        if(!find)
+        {
+            for(i=0; i<vcnt; i++)
+            {
+                mvx[i] = near_mvs[i].as_mv.row;
+                mvy[i] = near_mvs[i].as_mv.col;
+            }
+
+            quicksortmv (mvx, 0, vcnt-1);
+            quicksortmv (mvy, 0, vcnt-1);
+            mv.as_mv.row = mvx[vcnt/2];
+            mv.as_mv.col = mvy[vcnt/2];
+
+            find = 1;
+            //sr is set to 0 to allow calling function to decide the search range.
+            *sr = 0;
+        }
+    }
+
+    /* Set up return values */
+    *mvp = mv.as_mv;
+    vp8_clamp_mv(mvp, xd);
+}
+
 int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset, int *returnrate, int *returndistortion, int *returnintra)
 {
     BLOCK *b = &x->block[0];
@@ -1422,6 +1689,12 @@ int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int
     int tteob = 0;
     int force_no_skip = 0;
 
+    MV mvp;
+    int near_sad[7]; // 0-cf above, 1-cf left, 2-cf aboveleft, 3-lf current, 4-lf above, 5-lf left, 6-lf aboveleft
+    int near_sadidx[7] = {0, 1, 2, 3, 4, 5, 6};
+    int saddone=0;
+    int sr=0;    //search range got from mv_pred(). It uses step_param levels. (0-7)
+
     *returnintra = INT_MAX;
 
     vpx_memset(&best_mbmode, 0, sizeof(best_mbmode)); // clean
@@ -1589,6 +1862,72 @@ int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int
                           &mode_mv[NEARESTMV], &mode_mv[NEARMV], &best_ref_mv,
                           mdcounts, x->e_mbd.mode_info_context->mbmi.ref_frame, cpi->common.ref_frame_sign_bias);
 
+        if(x->e_mbd.mode_info_context->mbmi.mode == NEWMV)
+        {
+            if(!saddone)
+            {
+                //calculate sad for current frame 3 nearby MBs.
+                if( xd->mb_to_top_edge==0 && xd->mb_to_left_edge ==0)
+                {
+                    near_sad[0] = near_sad[1] = near_sad[2] = INT_MAX;
+                }else if(xd->mb_to_top_edge==0)
+                {   //only has left MB for sad calculation.
+                    near_sad[0] = near_sad[2] = INT_MAX;
+                    near_sad[1] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, xd->dst.y_buffer - 16,xd->dst.y_stride, 0x7fffffff);
+                }else if(xd->mb_to_left_edge ==0)
+                {   //only has left MB for sad calculation.
+                    near_sad[1] = near_sad[2] = INT_MAX;
+                    near_sad[0] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, xd->dst.y_buffer - xd->dst.y_stride *16,xd->dst.y_stride, 0x7fffffff);
+                }else
+                {
+                    near_sad[0] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, xd->dst.y_buffer - xd->dst.y_stride *16,xd->dst.y_stride, 0x7fffffff);
+                    near_sad[1] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, xd->dst.y_buffer - 16,xd->dst.y_stride, 0x7fffffff);
+                    near_sad[2] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, xd->dst.y_buffer - xd->dst.y_stride *16 -16,xd->dst.y_stride, 0x7fffffff);
+                }
+
+                if(cpi->common.last_frame_type != KEY_FRAME)
+                {
+                    //calculate sad for last frame 4 nearby MBs.
+                    unsigned char *pre_y_buffer = cpi->common.yv12_fb[cpi->common.lst_fb_idx].y_buffer + recon_yoffset;
+                    int pre_y_stride = cpi->common.yv12_fb[cpi->common.lst_fb_idx].y_stride;
+
+                    if( xd->mb_to_top_edge==0 && xd->mb_to_left_edge ==0)
+                    {
+                        near_sad[4] = near_sad[5] = near_sad[6] = INT_MAX;
+                        near_sad[3] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, pre_y_buffer, pre_y_stride, 0x7fffffff);
+                    }else if(xd->mb_to_top_edge==0)
+                    {   //only has left MB for sad calculation.
+                        near_sad[4] = near_sad[6] = INT_MAX;
+                        near_sad[3] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, pre_y_buffer, pre_y_stride, 0x7fffffff);
+                        near_sad[5] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, pre_y_buffer - 16, pre_y_stride, 0x7fffffff);
+                    }else if(xd->mb_to_left_edge ==0)
+                    {   //only has left MB for sad calculation.
+                        near_sad[5] = near_sad[6] = INT_MAX;
+                        near_sad[3] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, pre_y_buffer, pre_y_stride, 0x7fffffff);
+                        near_sad[4] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, pre_y_buffer - pre_y_stride *16, pre_y_stride, 0x7fffffff);
+                    }else
+                    {
+                        near_sad[3] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, pre_y_buffer, pre_y_stride, 0x7fffffff);
+                        near_sad[4] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, pre_y_buffer - pre_y_stride *16, pre_y_stride, 0x7fffffff);
+                        near_sad[5] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, pre_y_buffer - 16, pre_y_stride, 0x7fffffff);
+                        near_sad[6] = cpi->fn_ptr[BLOCK_16X16].sdf(x->src.y_buffer, x->src.y_stride, pre_y_buffer - pre_y_stride *16 -16, pre_y_stride, 0x7fffffff);
+                    }
+                }
+
+                if(cpi->common.last_frame_type != KEY_FRAME)
+                {
+                    quicksortsad(near_sad, near_sadidx, 0, 6);
+                }else
+                {
+                    quicksortsad(near_sad, near_sadidx, 0, 2);
+                }
+
+                saddone = 1;
+            }
+
+            vp8_mv_pred(cpi, &x->e_mbd, x->e_mbd.mode_info_context, &mvp,
+                        x->e_mbd.mode_info_context->mbmi.ref_frame, cpi->common.ref_frame_sign_bias, &sr, &near_sadidx[0]);
+        }
 
         // Estimate the reference frame signaling cost and add it to the rolling cost variable.
         frame_cost = ref_frame_cost[x->e_mbd.mode_info_context->mbmi.ref_frame];
@@ -1764,6 +2103,10 @@ int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int
                 int further_steps;
                 int n;
 
+                //adjust search range according to sr from mv prediction
+                if(sr > step_param)
+                    step_param = sr;
+
                 // Work out how long a search we should do
                 search_range = MAXF(abs(best_ref_mv.col), abs(best_ref_mv.row)) >> 3;
 
@@ -1784,7 +2127,7 @@ int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int
                     }
                     else
                     {
-                        bestsme = cpi->diamond_search_sad(x, b, d, &best_ref_mv, &d->bmi.mv.as_mv, step_param, sadpb / 2/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost); //sadpb < 9
+                        bestsme = cpi->diamond_search_sad(x, b, d, &mvp, &d->bmi.mv.as_mv, step_param, sadpb / 2/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost, &best_ref_mv); //sadpb < 9
                         mode_mv[NEWMV].row = d->bmi.mv.as_mv.row;
                         mode_mv[NEWMV].col = d->bmi.mv.as_mv.col;
 
@@ -1803,7 +2146,7 @@ int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int
                                 num00--;
                             else
                             {
-                                thissme = cpi->diamond_search_sad(x, b, d, &best_ref_mv, &d->bmi.mv.as_mv, step_param + n, sadpb / 4/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost); //sadpb = 9
+                                thissme = cpi->diamond_search_sad(x, b, d, &mvp, &d->bmi.mv.as_mv, step_param + n, sadpb / 4/*x->errorperbit*/, &num00, &cpi->fn_ptr[BLOCK_16X16], x->mvsadcost, x->mvcost, &best_ref_mv); //sadpb = 9
 
                                 if (thissme < bestsme)
                                 {
@@ -1827,9 +2170,14 @@ int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int
                 {
                     int thissme;
                     int full_flag_thresh = 0;
+                    MV full_mvp;
+
+                    full_mvp.row = d->bmi.mv.as_mv.row <<3;    // use diamond search result as full search staring point
+                    full_mvp.col = d->bmi.mv.as_mv.col <<3;
 
                     // Update x->vector_range based on best vector found in step search
-                    search_range = MAXF(abs(d->bmi.mv.as_mv.row), abs(d->bmi.mv.as_mv.col));
+                    search_range = MAXF(abs((mvp.row>>3) - d->bmi.mv.as_mv.row), abs((mvp.col>>3) - d->bmi.mv.as_mv.col));
+                    //search_range *= 1.4;  //didn't improve PSNR
 
                     if (search_range > x->vector_range)
                         x->vector_range = search_range;
@@ -1838,9 +2186,20 @@ int vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int
 
                     // Apply limits
                     search_range = (search_range > cpi->sf.max_fs_radius) ? cpi->sf.max_fs_radius : search_range;
+
+                    //add this to reduce full search range.
+                    if(sr<=3 && search_range > 8) search_range = 8;
+
                     {
                         int sadpb = x->sadperbit16 >> 2;
-                        thissme = cpi->full_search_sad(x, b, d, &best_ref_mv, sadpb, search_range, &cpi->fn_ptr[BLOCK_16X16], x->mvcost, x->mvsadcost);
+                        thissme = cpi->full_search_sad(x, b, d, &full_mvp, sadpb, search_range, &cpi->fn_ptr[BLOCK_16X16], x->mvcost, x->mvsadcost,&best_ref_mv);
+                        /*
+                        MV dia_ref_mv;
+                        dia_ref_mv.row = d->bmi.mv.as_mv.row << 3;
+                        dia_ref_mv.col = d->bmi.mv.as_mv.col << 3;
+                        thissme = cpi->full_search_sad(x, b, d, &dia_ref_mv, sadpb, search_range, &cpi->fn_ptr[BLOCK_16X16], x->mvcost, x->mvsadcost,&best_ref_mv);
+                        */
+
                     }
 
                     // Barrier threshold to initiating full search
diff --git a/vp8/encoder/temporal_filter.c b/vp8/encoder/temporal_filter.c
index fd5dd7ede..745dc31fe 100644
--- a/vp8/encoder/temporal_filter.c
+++ b/vp8/encoder/temporal_filter.c
@@ -246,7 +246,7 @@ static int find_matching_mb
             step_param,
             sadpb / 2/*x->errorperbit*/,
             &num00, &cpi->fn_ptr[BLOCK_16X16],
-            mvsadcost, mvcost); //sadpb < 9
+            mvsadcost, mvcost, &best_ref_mv1); //sadpb < 9
 
         // Further step/diamond searches as necessary
         n = 0;
@@ -268,7 +268,7 @@ static int find_matching_mb
                     step_param + n,
                     sadpb / 4/*x->errorperbit*/,
                     &num00, &cpi->fn_ptr[BLOCK_16X16],
-                    mvsadcost, mvcost); //sadpb = 9
+                    mvsadcost, mvcost, &best_ref_mv1); //sadpb = 9
 
                 if (thissme < bestsme)
                 {