libvpx: Pull from upstream

source/libvpx/vp9/encoder/vp9_pickmode.c

Index: source/libvpx/vp9/encoder/vp9_pickmode.c
===================================================================
--- source/libvpx/vp9/encoder/vp9_pickmode.c	(revision 281795)
+++ source/libvpx/vp9/encoder/vp9_pickmode.c	(working copy)
@@ -23,9 +23,89 @@
 #include "vp9/common/vp9_reconintra.h"
 
 #include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_pickmode.h"
 #include "vp9/encoder/vp9_ratectrl.h"
-#include "vp9/encoder/vp9_rdopt.h"
+#include "vp9/encoder/vp9_rd.h"
 
+static int mv_refs_rt(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                       const TileInfo *const tile,
+                       MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                       int_mv *mv_ref_list,
+                       int mi_row, int mi_col) {
+  const int *ref_sign_bias = cm->ref_frame_sign_bias;
+  int i, refmv_count = 0;
+
+  const POSITION *const mv_ref_search = mv_ref_blocks[mi->mbmi.sb_type];
+
+  int different_ref_found = 0;
+  int context_counter = 0;
+  int const_motion = 0;
+
+  // Blank the reference vector list
+  vpx_memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
+
+  // The nearest 2 blocks are treated differently
+  // if the size < 8x8 we get the mv from the bmi substructure,
+  // and we also need to keep a mode count.
+  for (i = 0; i < 2; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MODE_INFO *const candidate_mi = xd->mi[mv_ref->col + mv_ref->row *
+                                                   xd->mi_stride];
+      const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
+      // Keep counts for entropy encoding.
+      context_counter += mode_2_counter[candidate->mode];
+      different_ref_found = 1;
+
+      if (candidate->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1));
+    }
+  }
+
+  const_motion = 1;
+
+  // Check the rest of the neighbors in much the same way
+  // as before except we don't need to keep track of sub blocks or
+  // mode counts.
+  for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row *
+                                                    xd->mi_stride]->mbmi;
+      different_ref_found = 1;
+
+      if (candidate->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(candidate->mv[0]);
+    }
+  }
+
+  // Since we couldn't find 2 mvs from the same reference frame
+  // go back through the neighbors and find motion vectors from
+  // different reference frames.
+  if (different_ref_found && !refmv_count) {
+    for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
+      const POSITION *mv_ref = &mv_ref_search[i];
+      if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+        const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row
+                                              * xd->mi_stride]->mbmi;
+
+        // If the candidate is INTRA we don't want to consider its mv.
+        IF_DIFF_REF_FRAME_ADD_MV(candidate);
+      }
+    }
+  }
+
+ Done:
+
+  mi->mbmi.mode_context[ref_frame] = counter_to_context[context_counter];
+
+  // Clamp vectors
+  for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
+    clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
+
+  return const_motion;
+}
+
 static void full_pixel_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                     BLOCK_SIZE bsize, int mi_row, int mi_col,
                                     int_mv *tmp_mv, int *rate_mv) {
@@ -61,7 +141,7 @@
 
   // TODO(jingning) exploiting adaptive motion search control in non-RD
   // mode decision too.
-  step_param = 6;
+  step_param = cpi->sf.mv.fullpel_search_step_param;
 
   for (i = LAST_FRAME; i <= LAST_FRAME && cpi->common.show_frame; ++i) {
     if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
@@ -172,6 +252,17 @@
   else
     x->skip_txfm = 0;
 
+  if (cpi->common.tx_mode == TX_MODE_SELECT) {
+    if (sse > (var << 2))
+      xd->mi[0]->mbmi.tx_size = MIN(max_txsize_lookup[bsize],
+                          tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+    else
+      xd->mi[0]->mbmi.tx_size = TX_8X8;
+  } else {
+    xd->mi[0]->mbmi.tx_size = MIN(max_txsize_lookup[bsize],
+                         tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+  }
+
   vp9_model_rd_from_var_lapndz(sse - var, 1 << num_pels_log2_lookup[bsize],
                                dc_quant >> 3, &rate, &dist);
   *out_rate_sum = rate >> 1;
@@ -183,6 +274,109 @@
   *out_dist_sum += dist << 4;
 }
 
+static int get_pred_buffer(PRED_BUFFER *p, int len) {
+  int i;
+
+  for (i = 0; i < len; i++) {
+    if (!p[i].in_use) {
+      p[i].in_use = 1;
+      return i;
+    }
+  }
+  return -1;
+}
+
+static void free_pred_buffer(PRED_BUFFER *p) {
+  p->in_use = 0;
+}
+
+static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x,
+                                 BLOCK_SIZE bsize, int mi_row, int mi_col,
+                                 MV_REFERENCE_FRAME ref_frame,
+                                 PREDICTION_MODE this_mode,
+                                 unsigned int var_y, unsigned int sse_y,
+                                 struct buf_2d yv12_mb[][MAX_MB_PLANE],
+                                 int *rate, int64_t *dist) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+
+  const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
+  unsigned int var = var_y, sse = sse_y;
+  // Skipping threshold for ac.
+  unsigned int thresh_ac;
+  // Skipping threshold for dc.
+  unsigned int thresh_dc;
+  if (x->encode_breakout > 0) {
+    // Set a maximum for threshold to avoid big PSNR loss in low bit rate
+    // case. Use extreme low threshold for static frames to limit
+    // skipping.
+    const unsigned int max_thresh = 36000;
+    // The encode_breakout input
+    const unsigned int min_thresh =
+        MIN(((unsigned int)x->encode_breakout << 4), max_thresh);
+
+    // Calculate threshold according to dequant value.
+    thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
+    thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
+
+    // Adjust ac threshold according to partition size.
+    thresh_ac >>=
+        8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
+
+    thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
+  } else {
+    thresh_ac = 0;
+    thresh_dc = 0;
+  }
+
+  // Y skipping condition checking for ac and dc.
+  if (var <= thresh_ac && (sse - var) <= thresh_dc) {
+    unsigned int sse_u, sse_v;
+    unsigned int var_u, var_v;
+
+    // Skip UV prediction unless breakout is zero (lossless) to save
+    // computation with low impact on the result
+    if (x->encode_breakout == 0) {
+      xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
+      xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
+      vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
+    }
+
+    var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
+                                    x->plane[1].src.stride,
+                                    xd->plane[1].dst.buf,
+                                    xd->plane[1].dst.stride, &sse_u);
+
+    // U skipping condition checking
+    if ((var_u * 4 <= thresh_ac) && (sse_u - var_u <= thresh_dc)) {
+      var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
+                                      x->plane[2].src.stride,
+                                      xd->plane[2].dst.buf,
+                                      xd->plane[2].dst.stride, &sse_v);
+
+      // V skipping condition checking
+      if ((var_v * 4 <= thresh_ac) && (sse_v - var_v <= thresh_dc)) {
+        x->skip = 1;
+
+        // The cost of skip bit needs to be added.
+        *rate = cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
+                                     [INTER_OFFSET(this_mode)];
+
+        // More on this part of rate
+        // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+
+        // Scaling factor for SSE from spatial domain to frequency
+        // domain is 16. Adjust distortion accordingly.
+        // TODO(yunqingwang): In this function, only y-plane dist is
+        // calculated.
+        *dist = (sse << 4);  // + ((sse_u + sse_v) << 4);
+
+        // *disable_skip = 1;
+      }
+    }
+  }
+}
+
 // TODO(jingning) placeholder for inter-frame non-RD mode decision.
 // this needs various further optimizations. to be continued..
 int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
@@ -197,6 +391,8 @@
   struct macroblockd_plane *const pd = &xd->plane[0];
   PREDICTION_MODE this_mode, best_mode = ZEROMV;
   MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
+  TX_SIZE best_tx_size = MIN(max_txsize_lookup[bsize],
+                             tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
   INTERP_FILTER best_pred_filter = EIGHTTAP;
   int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
@@ -224,11 +420,43 @@
   const int *const rd_thresh_freq_fact = cpi->rd.thresh_freq_fact[bsize];
   // Mode index conversion form THR_MODES to PREDICTION_MODE for a ref frame.
   int mode_idx[MB_MODE_COUNT] = {0};
-  INTERP_FILTER filter_ref = SWITCHABLE;
+  INTERP_FILTER filter_ref = cm->interp_filter;
   int bsl = mi_width_log2_lookup[bsize];
-  const int pred_filter_search = (((mi_row + mi_col) >> bsl) +
-                                      get_chessboard_index(cm)) % 2;
+  const int pred_filter_search = cm->interp_filter == SWITCHABLE ?
+      (((mi_row + mi_col) >> bsl) + get_chessboard_index(cm)) % 2 : 0;
+  int const_motion[MAX_REF_FRAMES] = { 0 };
 
+  // For speed 6, the result of interp filter is reused later in actual encoding
+  // process.
+  int bh = num_4x4_blocks_high_lookup[bsize] << 2;
+  int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
+  int pixels_in_block = bh * bw;
+  // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
+  PRED_BUFFER tmp[4];
+  DECLARE_ALIGNED_ARRAY(16, uint8_t, pred_buf, 3 * 64 * 64);
+  struct buf_2d orig_dst = pd->dst;
+  PRED_BUFFER *best_pred = NULL;
+  PRED_BUFFER *this_mode_pred = NULL;
+  int i;
+
+#if CONFIG_DENOISING
+  if (cpi->oxcf.noise_sensitivity > 0) {
+    vp9_denoiser_reset_frame_stats(&cpi->denoiser);
+  }
+#endif
+
+  if (cpi->sf.reuse_inter_pred_sby) {
+    for (i = 0; i < 3; i++) {
+      tmp[i].data = &pred_buf[pixels_in_block * i];
+      tmp[i].stride = bw;
+      tmp[i].in_use = 0;
+    }
+
+    tmp[3].data = pd->dst.buf;
+    tmp[3].stride = pd->dst.stride;
+    tmp[3].in_use = 0;
+  }
+
   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
 
   x->skip = 0;
@@ -241,18 +469,36 @@
   mbmi->ref_frame[0] = NONE;
   mbmi->ref_frame[1] = NONE;
   mbmi->tx_size = MIN(max_txsize_lookup[bsize],
-                      tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
-  mbmi->interp_filter = cpi->common.interp_filter == SWITCHABLE ?
-                        EIGHTTAP : cpi->common.interp_filter;
+                      tx_mode_to_biggest_tx_size[cm->tx_mode]);
+  mbmi->interp_filter = cm->interp_filter == SWITCHABLE ?
+                        EIGHTTAP : cm->interp_filter;
   mbmi->skip = 0;
   mbmi->segment_id = segment_id;
 
   for (ref_frame = LAST_FRAME; ref_frame <= LAST_FRAME ; ++ref_frame) {
     x->pred_mv_sad[ref_frame] = INT_MAX;
     if (cpi->ref_frame_flags & flag_list[ref_frame]) {
-      vp9_setup_buffer_inter(cpi, x, tile,
-                             ref_frame, bsize, mi_row, mi_col,
-                             frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
+      const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
+      int_mv *const candidates = mbmi->ref_mvs[ref_frame];
+      const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
+      vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
+                           sf, sf);
+
+      if (cm->coding_use_prev_mi)
+        vp9_find_mv_refs(cm, xd, tile, xd->mi[0], ref_frame,
+                         candidates, mi_row, mi_col);
+      else
+        const_motion[ref_frame] = mv_refs_rt(cm, xd, tile, xd->mi[0],
+                                             ref_frame, candidates,
+                                             mi_row, mi_col);
+
+      vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
+                            &frame_mv[NEARESTMV][ref_frame],
+                            &frame_mv[NEARMV][ref_frame]);
+
+      if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8)
+        vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
+                    ref_frame, bsize);
     }
     frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
     frame_mv[ZEROMV][ref_frame].as_int = 0;
@@ -286,6 +532,10 @@
     for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
       int rate_mv = 0;
 
+      if (const_motion[ref_frame] &&
+          (this_mode == NEARMV || this_mode == ZEROMV))
+        continue;
+
       if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode)))
         continue;
 
@@ -324,6 +574,16 @@
       // Search for the best prediction filter type, when the resulting
       // motion vector is at sub-pixel accuracy level for luma component, i.e.,
       // the last three bits are all zeros.
+      if (cpi->sf.reuse_inter_pred_sby) {
+        if (this_mode == NEARESTMV) {
+          this_mode_pred = &tmp[3];
+        } else {
+          this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
+          pd->dst.buf = this_mode_pred->data;
+          pd->dst.stride = bw;
+        }
+      }
+
       if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
           pred_filter_search &&
           ((mbmi->mv[0].as_mv.row & 0x07) != 0 ||
@@ -332,8 +592,10 @@
         int64_t pf_dist[3];
         unsigned int pf_var[3];
         unsigned int pf_sse[3];
+        TX_SIZE pf_tx_size[3];
         int64_t best_cost = INT64_MAX;
         INTERP_FILTER best_filter = SWITCHABLE, filter;
+        PRED_BUFFER *current_pred = this_mode_pred;
 
         for (filter = EIGHTTAP; filter <= EIGHTTAP_SHARP; ++filter) {
           int64_t cost;
@@ -344,14 +606,32 @@
           cost = RDCOST(x->rdmult, x->rddiv,
                         vp9_get_switchable_rate(cpi) + pf_rate[filter],
                         pf_dist[filter]);
+          pf_tx_size[filter] = mbmi->tx_size;
           if (cost < best_cost) {
-              best_filter = filter;
-              best_cost = cost;
-              skip_txfm = x->skip_txfm;
+            best_filter = filter;
+            best_cost = cost;
+            skip_txfm = x->skip_txfm;
+
+            if (cpi->sf.reuse_inter_pred_sby) {
+              if (this_mode_pred != current_pred) {
+                free_pred_buffer(this_mode_pred);
+                this_mode_pred = current_pred;
+              }
+
+              if (filter < EIGHTTAP_SHARP) {
+                current_pred = &tmp[get_pred_buffer(tmp, 3)];
+                pd->dst.buf = current_pred->data;
+                pd->dst.stride = bw;
+              }
+            }
           }
         }
 
+        if (cpi->sf.reuse_inter_pred_sby && this_mode_pred != current_pred)
+          free_pred_buffer(current_pred);
+
         mbmi->interp_filter = best_filter;
+        mbmi->tx_size = pf_tx_size[mbmi->interp_filter];
         rate = pf_rate[mbmi->interp_filter];
         dist = pf_dist[mbmi->interp_filter];
         var_y = pf_var[mbmi->interp_filter];
@@ -370,77 +650,19 @@
 
       // Skipping checking: test to see if this block can be reconstructed by
       // prediction only.
-      if (!x->in_active_map) {
-        x->skip = 1;
-      } else if (cpi->allow_encode_breakout && x->encode_breakout) {
-        const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
-        unsigned int var = var_y, sse = sse_y;
-        // Skipping threshold for ac.
-        unsigned int thresh_ac;
-        // Skipping threshold for dc.
-        unsigned int thresh_dc;
-        // Set a maximum for threshold to avoid big PSNR loss in low bit rate
-        // case. Use extreme low threshold for static frames to limit skipping.
-        const unsigned int max_thresh = 36000;
-        // The encode_breakout input
-        const unsigned int min_thresh =
-            MIN(((unsigned int)x->encode_breakout << 4), max_thresh);
-
-        // Calculate threshold according to dequant value.
-        thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
-        thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
-
-        // Adjust ac threshold according to partition size.
-        thresh_ac >>= 8 - (b_width_log2_lookup[bsize] +
-            b_height_log2_lookup[bsize]);
-
-        thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
-
-        // Y skipping condition checking for ac and dc.
-        if (var <= thresh_ac && (sse - var) <= thresh_dc) {
-          unsigned int sse_u, sse_v;
-          unsigned int var_u, var_v;
-
-          // Skip u v prediction for less calculation, that won't affect
-          // result much.
-          var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
-                                          x->plane[1].src.stride,
-                                          xd->plane[1].dst.buf,
-                                          xd->plane[1].dst.stride, &sse_u);
-
-          // U skipping condition checking
-          if ((var_u * 4 <= thresh_ac) && (sse_u - var_u <= thresh_dc)) {
-            var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
-                                            x->plane[2].src.stride,
-                                            xd->plane[2].dst.buf,
-                                            xd->plane[2].dst.stride, &sse_v);
-
-            // V skipping condition checking
-            if ((var_v * 4 <= thresh_ac) && (sse_v - var_v <= thresh_dc)) {
-              x->skip = 1;
-
-              // The cost of skip bit needs to be added.
-              rate = rate_mv;
-              rate += cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
-                                           [INTER_OFFSET(this_mode)];
-
-              // More on this part of rate
-              // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
-
-              // Scaling factor for SSE from spatial domain to frequency
-              // domain is 16. Adjust distortion accordingly.
-              // TODO(yunqingwang): In this function, only y-plane dist is
-              // calculated.
-              dist = (sse << 4);  // + ((sse_u + sse_v) << 4);
-              this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
-              // *disable_skip = 1;
-            }
-          }
+      if (cpi->allow_encode_breakout) {
+        encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame,
+                             this_mode, var_y, sse_y, yv12_mb, &rate, &dist);
+        if (x->skip) {
+          rate += rate_mv;
+          this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
         }
       }
 
 #if CONFIG_DENOISING
-    vp9_denoiser_update_frame_stats();
+      if (cpi->oxcf.noise_sensitivity > 0) {
+        vp9_denoiser_update_frame_stats(&cpi->denoiser, mbmi, sse_y, this_mode);
+      }
 #endif
 
       if (this_rd < best_rd || x->skip) {
@@ -449,8 +671,19 @@
         *returndistortion = dist;
         best_mode = this_mode;
         best_pred_filter = mbmi->interp_filter;
+        best_tx_size = mbmi->tx_size;
         best_ref_frame = ref_frame;
         skip_txfm = x->skip_txfm;
+
+        if (cpi->sf.reuse_inter_pred_sby) {
+          if (best_pred != NULL)
+            free_pred_buffer(best_pred);
+
+          best_pred = this_mode_pred;
+        }
+      } else {
+        if (cpi->sf.reuse_inter_pred_sby)
+          free_pred_buffer(this_mode_pred);
       }
 
       if (x->skip)
@@ -458,11 +691,25 @@
     }
   }
 
+  // If best prediction is not in dst buf, then copy the prediction block from
+  // temp buf to dst buf.
+  if (cpi->sf.reuse_inter_pred_sby && best_pred->data != orig_dst.buf) {
+    uint8_t *copy_from, *copy_to;
 
-  mbmi->mode = best_mode;
+    pd->dst = orig_dst;
+    copy_to = pd->dst.buf;
+
+    copy_from = best_pred->data;
+
+    vp9_convolve_copy(copy_from, bw, copy_to, pd->dst.stride, NULL, 0, NULL, 0,
+                      bw, bh);
+  }
+
+  mbmi->mode          = best_mode;
   mbmi->interp_filter = best_pred_filter;
-  mbmi->ref_frame[0] = best_ref_frame;
-  mbmi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int;
+  mbmi->tx_size       = best_tx_size;
+  mbmi->ref_frame[0]  = best_ref_frame;
+  mbmi->mv[0].as_int  = frame_mv[best_mode][best_ref_frame].as_int;
   xd->mi[0]->bmi[0].as_mv[0].as_int = mbmi->mv[0].as_int;
   x->skip_txfm = skip_txfm;
 
@@ -470,22 +717,57 @@
   // threshold.
   if (!x->skip && best_rd > inter_mode_thresh &&
       bsize <= cpi->sf.max_intra_bsize) {
+    int i, j;
+    const int width  = num_4x4_blocks_wide_lookup[bsize];
+    const int height = num_4x4_blocks_high_lookup[bsize];
+
+    int rate2 = 0;
+    int64_t dist2 = 0;
+    const int dst_stride = pd->dst.stride;
+    const int src_stride = p->src.stride;
+    int block_idx = 0;
+
+    TX_SIZE tmp_tx_size = MIN(max_txsize_lookup[bsize],
+                              tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+    const int step = 1 << tmp_tx_size;
+
     for (this_mode = DC_PRED; this_mode <= DC_PRED; ++this_mode) {
-      vp9_predict_intra_block(xd, 0, b_width_log2(bsize),
-                              mbmi->tx_size, this_mode,
-                              &p->src.buf[0], p->src.stride,
-                              &pd->dst.buf[0], pd->dst.stride, 0, 0, 0);
+      if (cpi->sf.reuse_inter_pred_sby) {
+        pd->dst.buf = tmp[0].data;
+        pd->dst.stride = bw;
+      }
 
-      model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
+      for (j = 0; j < height; j += step) {
+        for (i = 0; i < width; i += step) {
+          vp9_predict_intra_block(xd, block_idx, b_width_log2(bsize),
+                                  tmp_tx_size, this_mode,
+                                  &p->src.buf[4 * (j * dst_stride + i)],
+                                  src_stride,
+                                  &pd->dst.buf[4 * (j * dst_stride + i)],
+                                  dst_stride, i, j, 0);
+          model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
+          rate2 += rate;
+          dist2 += dist;
+          ++block_idx;
+        }
+      }
+
+      rate = rate2;
+      dist = dist2;
+
       rate += cpi->mbmode_cost[this_mode];
       rate += intra_cost_penalty;
       this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
 
+      if (cpi->sf.reuse_inter_pred_sby)
+        pd->dst = orig_dst;
+
       if (this_rd + intra_mode_cost < best_rd) {
         best_rd = this_rd;
         *returnrate = rate;
         *returndistortion = dist;
         mbmi->mode = this_mode;
+        mbmi->tx_size = tmp_tx_size;
         mbmi->ref_frame[0] = INTRA_FRAME;
         mbmi->uv_mode = this_mode;
         mbmi->mv[0].as_int = INVALID_MV;
@@ -494,8 +776,11 @@
       }
     }
   }
+
 #if CONFIG_DENOISING
-  vp9_denoiser_denoise(&cpi->denoiser, x, mi_row, mi_col, bsize);
+  if (cpi->oxcf.noise_sensitivity > 0) {
+    vp9_denoiser_denoise(&cpi->denoiser, x, mi_row, mi_col, bsize);
+  }
 #endif
 
   return INT64_MAX;