Apply fast motion search to golden reference frame

This commit enables the rtc coding mode to run integral projection
based motion search for golden reference frame. It improves the
speed -6 compression performance by 1.1% on average, 3.46% for
jimred_vga, 6.46% for tacomascmvvga, and 0.5% for vidyo clips. The
speed -6 is about 6% slower.

Change-Id: I0fe402ad2edf0149d0349ad304ab9b2abdf0c804

2 files changed, 40 insertions, 12 deletions

diff --git a/vp9/encoder/vp9_pickmode.c b/vp9/encoder/vp9_pickmode.c
index 6c2576add..23a2569c8 100644
--- a/vp9/encoder/vp9_pickmode.c
+++ b/vp9/encoder/vp9_pickmode.c
@@ -784,15 +784,43 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
         continue;
 
       if (this_mode == NEWMV) {
-        if (ref_frame > LAST_FRAME)
-          continue;
         if (cpi->sf.partition_search_type != VAR_BASED_PARTITION &&
             best_rdc.rdcost < (int64_t)(1 << num_pels_log2_lookup[bsize]))
           continue;
-        if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
-                                    &frame_mv[NEWMV][ref_frame],
-                                    &rate_mv, best_rdc.rdcost))
+
+        if (ref_frame > LAST_FRAME) {
+          int tmp_sad;
+          int dis, cost_list[5];
+
+          if (bsize < BLOCK_16X16)
+            continue;
+
+          tmp_sad = vp9_int_pro_motion_estimation(cpi, x, bsize);
+          if (tmp_sad > x->pred_mv_sad[LAST_FRAME])
+            continue;
+
+          frame_mv[NEWMV][ref_frame].as_int = mbmi->mv[0].as_int;
+          rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
+                                    &mbmi->ref_mvs[ref_frame][0].as_mv,
+                                    x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+          frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
+          frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
+
+          cpi->find_fractional_mv_step(x, &frame_mv[NEWMV][ref_frame].as_mv,
+                                       &mbmi->ref_mvs[ref_frame][0].as_mv,
+                                       cpi->common.allow_high_precision_mv,
+                                       x->errorperbit,
+                                       &cpi->fn_ptr[bsize],
+                                       cpi->sf.mv.subpel_force_stop,
+                                       cpi->sf.mv.subpel_iters_per_step,
+                                       cond_cost_list(cpi, cost_list),
+                                       x->nmvjointcost, x->mvcost, &dis,
+                                       &x->pred_sse[ref_frame], NULL, 0, 0);
+        } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
+                                           &frame_mv[NEWMV][ref_frame],
+                                           &rate_mv, best_rdc.rdcost)) {
           continue;
+        }
       }
 
       if (this_mode != NEARESTMV &&
@@ -817,7 +845,7 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
       }
 
       if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
-          pred_filter_search &&
+          pred_filter_search && (ref_frame == LAST_FRAME) &&
           ((mbmi->mv[0].as_mv.row & 0x07) != 0 ||
            (mbmi->mv[0].as_mv.col & 0x07) != 0)) {
         int pf_rate[3];
diff --git a/vp9/encoder/x86/vp9_avg_intrin_sse2.c b/vp9/encoder/x86/vp9_avg_intrin_sse2.c
index f49949940..618b5f73d 100644
--- a/vp9/encoder/x86/vp9_avg_intrin_sse2.c
+++ b/vp9/encoder/x86/vp9_avg_intrin_sse2.c
@@ -61,7 +61,7 @@ void vp9_int_pro_row_sse2(int16_t *hbuf, uint8_t const*ref,
                           const int ref_stride, const int height) {
   int idx;
   __m128i zero = _mm_setzero_si128();
-  __m128i src_line = _mm_load_si128((const __m128i *)ref);
+  __m128i src_line = _mm_loadu_si128((const __m128i *)ref);
   __m128i s0 = _mm_unpacklo_epi8(src_line, zero);
   __m128i s1 = _mm_unpackhi_epi8(src_line, zero);
   __m128i t0, t1;
@@ -69,14 +69,14 @@ void vp9_int_pro_row_sse2(int16_t *hbuf, uint8_t const*ref,
   ref += ref_stride;
 
   for (idx = 1; idx < height_1; idx += 2) {
-    src_line = _mm_load_si128((const __m128i *)ref);
+    src_line = _mm_loadu_si128((const __m128i *)ref);
     t0 = _mm_unpacklo_epi8(src_line, zero);
     t1 = _mm_unpackhi_epi8(src_line, zero);
     s0 = _mm_adds_epu16(s0, t0);
     s1 = _mm_adds_epu16(s1, t1);
     ref += ref_stride;
 
-    src_line = _mm_load_si128((const __m128i *)ref);
+    src_line = _mm_loadu_si128((const __m128i *)ref);
     t0 = _mm_unpacklo_epi8(src_line, zero);
     t1 = _mm_unpackhi_epi8(src_line, zero);
     s0 = _mm_adds_epu16(s0, t0);
@@ -84,7 +84,7 @@ void vp9_int_pro_row_sse2(int16_t *hbuf, uint8_t const*ref,
     ref += ref_stride;
   }
 
-  src_line = _mm_load_si128((const __m128i *)ref);
+  src_line = _mm_loadu_si128((const __m128i *)ref);
   t0 = _mm_unpacklo_epi8(src_line, zero);
   t1 = _mm_unpackhi_epi8(src_line, zero);
   s0 = _mm_adds_epu16(s0, t0);
@@ -101,9 +101,9 @@ void vp9_int_pro_row_sse2(int16_t *hbuf, uint8_t const*ref,
     s1 = _mm_srai_epi16(s1, 3);
   }
 
-  _mm_store_si128((__m128i *)hbuf, s0);
+  _mm_storeu_si128((__m128i *)hbuf, s0);
   hbuf += 8;
-  _mm_store_si128((__m128i *)hbuf, s1);
+  _mm_storeu_si128((__m128i *)hbuf, s1);
 }
 
 int16_t vp9_int_pro_col_sse2(uint8_t const *ref, const int width) {