libvpx: Pull from upstream

source/libvpx/vp9/encoder/vp9_pickmode.c

Index: source/libvpx/vp9/encoder/vp9_pickmode.c
diff --git a/source/libvpx/vp9/encoder/vp9_pickmode.c b/source/libvpx/vp9/encoder/vp9_pickmode.c
index 6dfb9eedde14e15b3bff64b5a32625fdfcb67156..9fb7cfba7bfc9e67024650502cb6a8164023f8b7 100644
--- a/source/libvpx/vp9/encoder/vp9_pickmode.c
+++ b/source/libvpx/vp9/encoder/vp9_pickmode.c
@@ -20,9 +20,11 @@
 #include "vp9/common/vp9_blockd.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_mvref_common.h"
+#include "vp9/common/vp9_pred_common.h"
 #include "vp9/common/vp9_reconinter.h"
 #include "vp9/common/vp9_reconintra.h"
 
+#include "vp9/encoder/vp9_cost.h"
 #include "vp9/encoder/vp9_encoder.h"
 #include "vp9/encoder/vp9_pickmode.h"
 #include "vp9/encoder/vp9_ratectrl.h"
@@ -49,7 +51,7 @@ static int mv_refs_rt(const VP9_COMMON *cm, const MACROBLOCKD *xd,
   int const_motion = 0;
 
   // Blank the reference vector list
-  vpx_memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
+  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,
@@ -58,7 +60,7 @@ static int mv_refs_rt(const VP9_COMMON *cm, const MACROBLOCKD *xd,
     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].src_mi;
+                                                   xd->mi_stride];
       const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
       // Keep counts for entropy encoding.
       context_counter += mode_2_counter[candidate->mode];
@@ -79,7 +81,7 @@ static int mv_refs_rt(const VP9_COMMON *cm, const MACROBLOCKD *xd,
     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].src_mi->mbmi;
+                                                    xd->mi_stride]->mbmi;
       different_ref_found = 1;
 
       if (candidate->ref_frame[0] == ref_frame)
@@ -95,7 +97,7 @@ static int mv_refs_rt(const VP9_COMMON *cm, const MACROBLOCKD *xd,
       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].src_mi->mbmi;
+                                              * xd->mi_stride]->mbmi;
 
         // If the candidate is INTRA we don't want to consider its mv.
         IF_DIFF_REF_FRAME_ADD_MV(candidate, ref_frame, ref_sign_bias,
@@ -120,7 +122,7 @@ static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                   int_mv *tmp_mv, int *rate_mv,
                                   int64_t best_rd_sofar) {
   MACROBLOCKD *xd = &x->e_mbd;
-  MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
   struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
   const int step_param = cpi->sf.mv.fullpel_search_step_param;
   const int sadpb = x->sadperbit16;
@@ -188,6 +190,8 @@ static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                  cond_cost_list(cpi, cost_list),
                                  x->nmvjointcost, x->mvcost,
                                  &dis, &x->pred_sse[ref], NULL, 0, 0);
+    *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
+                               x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
   }
 
   if (scaled_ref_frame) {
@@ -198,6 +202,250 @@ static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
   return rv;
 }
 
+static void block_variance(const uint8_t *src, int src_stride,
+                           const uint8_t *ref, int ref_stride,
+                           int w, int h, unsigned int *sse, int *sum,
+                           int block_size, unsigned int *sse8x8,
+                           int *sum8x8, unsigned int *var8x8) {
+  int i, j, k = 0;
+
+  *sse = 0;
+  *sum = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      vp9_get8x8var(src + src_stride * i + j, src_stride,
+                    ref + ref_stride * i + j, ref_stride,
+                    &sse8x8[k], &sum8x8[k]);
+      *sse += sse8x8[k];
+      *sum += sum8x8[k];
+      var8x8[k] = sse8x8[k] - (((unsigned int)sum8x8[k] * sum8x8[k]) >> 6);
+      k++;
+    }
+  }
+}
+
+static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
+                               unsigned int *sse_i, int *sum_i,
+                               unsigned int *var_o, unsigned int *sse_o,
+                               int *sum_o) {
+  const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
+  const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
+  const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
+  int i, j, k = 0;
+
+  for (i = 0; i < nh; i += 2) {
+    for (j = 0; j < nw; j += 2) {
+      sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
+          sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
+      sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
+          sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
+      var_o[k] = sse_o[k] - (((unsigned int)sum_o[k] * sum_o[k]) >>
+          (b_width_log2_lookup[unit_size] +
+              b_height_log2_lookup[unit_size] + 6));
+      k++;
+    }
+  }
+}
+
+static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
+                                    MACROBLOCK *x, MACROBLOCKD *xd,
+                                    int *out_rate_sum, int64_t *out_dist_sum,
+                                    unsigned int *var_y, unsigned int *sse_y,
+                                    int mi_row, int mi_col, int *early_term) {
+  // Note our transform coeffs are 8 times an orthogonal transform.
+  // Hence quantizer step is also 8 times. To get effective quantizer
+  // we need to divide by 8 before sending to modeling function.
+  unsigned int sse;
+  int rate;
+  int64_t dist;
+  struct macroblock_plane *const p = &x->plane[0];
+  struct macroblockd_plane *const pd = &xd->plane[0];
+  const uint32_t dc_quant = pd->dequant[0];
+  const uint32_t ac_quant = pd->dequant[1];
+  const int64_t dc_thr = dc_quant * dc_quant >> 6;
+  const int64_t ac_thr = ac_quant * ac_quant >> 6;
+  unsigned int var;
+  int sum;
+  int skip_dc = 0;
+
+  const int bw = b_width_log2_lookup[bsize];
+  const int bh = b_height_log2_lookup[bsize];
+  const int num8x8 = 1 << (bw + bh - 2);
+  unsigned int sse8x8[64] = {0};
+  int sum8x8[64] = {0};
+  unsigned int var8x8[64] = {0};
+  TX_SIZE tx_size;
+  int i, k;
+
+  // Calculate variance for whole partition, and also save 8x8 blocks' variance
+  // to be used in following transform skipping test.
+  block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
+                 4 << bw, 4 << bh, &sse, &sum, 8, sse8x8, sum8x8, var8x8);
+  var = sse - (((int64_t)sum * sum) >> (bw + bh + 4));
+
+  *var_y = var;
+  *sse_y = sse;
+
+  if (cpi->common.tx_mode == TX_MODE_SELECT) {
+    if (sse > (var << 2))
+      tx_size = MIN(max_txsize_lookup[bsize],
+                    tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+    else
+      tx_size = TX_8X8;
+
+    if (cpi->sf.partition_search_type == VAR_BASED_PARTITION) {
+      if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
+          cyclic_refresh_segment_id_boosted(xd->mi[0]->mbmi.segment_id))
+        tx_size = TX_8X8;
+      else if (tx_size > TX_16X16)
+        tx_size = TX_16X16;
+    }
+  } else {
+    tx_size = MIN(max_txsize_lookup[bsize],
+                  tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+  }
+
+  assert(tx_size >= TX_8X8);
+  xd->mi[0]->mbmi.tx_size = tx_size;
+
+  // Evaluate if the partition block is a skippable block in Y plane.
+  {
+    unsigned int sse16x16[16] = {0};
+    int sum16x16[16] = {0};
+    unsigned int var16x16[16] = {0};
+    const int num16x16 = num8x8 >> 2;
+
+    unsigned int sse32x32[4] = {0};
+    int sum32x32[4] = {0};
+    unsigned int var32x32[4] = {0};
+    const int num32x32 = num8x8 >> 4;
+
+    int ac_test = 1;
+    int dc_test = 1;
+    const int num = (tx_size == TX_8X8) ? num8x8 :
+        ((tx_size == TX_16X16) ? num16x16 : num32x32);
+    const unsigned int *sse_tx = (tx_size == TX_8X8) ? sse8x8 :
+        ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
+    const unsigned int *var_tx = (tx_size == TX_8X8) ? var8x8 :
+        ((tx_size == TX_16X16) ? var16x16 : var32x32);
+
+    // Calculate variance if tx_size > TX_8X8
+    if (tx_size >= TX_16X16)
+      calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
+                         sum16x16);
+    if (tx_size == TX_32X32)
+      calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
+                         sse32x32, sum32x32);
+
+    // Skipping test
+    x->skip_txfm[0] = 0;
+    for (k = 0; k < num; k++)
+      // Check if all ac coefficients can be quantized to zero.
+      if (!(var_tx[k] < ac_thr || var == 0)) {
+        ac_test = 0;
+        break;
+      }
+
+    for (k = 0; k < num; k++)
+      // Check if dc coefficient can be quantized to zero.
+      if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
+        dc_test = 0;
+        break;
+      }
+
+    if (ac_test) {
+      x->skip_txfm[0] = 2;
+
+      if (dc_test)
+        x->skip_txfm[0] = 1;
+    } else if (dc_test) {
+      skip_dc = 1;
+    }
+  }
+
+  if (x->skip_txfm[0] == 1) {
+    int skip_uv[2] = {0};
+    unsigned int var_uv[2];
+    unsigned int sse_uv[2];
+
+    *out_rate_sum = 0;
+    *out_dist_sum = sse << 4;
+
+    // Transform skipping test in UV planes.
+    for (i = 1; i <= 2; i++) {
+      struct macroblock_plane *const p = &x->plane[i];
+      struct macroblockd_plane *const pd = &xd->plane[i];
+      const TX_SIZE uv_tx_size = get_uv_tx_size(&xd->mi[0]->mbmi, pd);
+      const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
+      const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
+      const int uv_bw = b_width_log2_lookup[uv_bsize];
+      const int uv_bh = b_height_log2_lookup[uv_bsize];
+      const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
+          (uv_bh - b_height_log2_lookup[unit_size]);
+      const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
+      const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
+      int j = i - 1;
+
+      vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
+      var_uv[j] = cpi->fn_ptr[uv_bsize].vf(p->src.buf, p->src.stride,
+          pd->dst.buf, pd->dst.stride, &sse_uv[j]);
+
+      if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
+          (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
+        skip_uv[j] = 1;
+      else
+        break;
+    }
+
+    // If the transform in YUV planes are skippable, the mode search checks
+    // fewer inter modes and doesn't check intra modes.
+    if (skip_uv[0] & skip_uv[1]) {
+      *early_term = 1;
+    }
+
+    return;
+  }
+
+  if (!skip_dc) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
+                                   dc_quant >> (xd->bd - 5), &rate, &dist);
+    } else {
+      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
+                                   dc_quant >> 3, &rate, &dist);
+    }
+#else
+    vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
+                                 dc_quant >> 3, &rate, &dist);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+
+  if (!skip_dc) {
+    *out_rate_sum = rate >> 1;
+    *out_dist_sum = dist << 3;
+  } else {
+    *out_rate_sum = 0;
+    *out_dist_sum = (sse - var) << 4;
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
+                                 ac_quant >> (xd->bd - 5), &rate, &dist);
+  } else {
+    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
+                                 ac_quant >> 3, &rate, &dist);
+  }
+#else
+  vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
+                               ac_quant >> 3, &rate, &dist);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  *out_rate_sum += rate;
+  *out_dist_sum += dist << 4;
+}
 
 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize,
                               MACROBLOCK *x, MACROBLOCKD *xd,
@@ -224,21 +472,21 @@ static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize,
 
   if (cpi->common.tx_mode == TX_MODE_SELECT) {
     if (sse > (var << 2))
-      xd->mi[0].src_mi->mbmi.tx_size =
+      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].src_mi->mbmi.tx_size = TX_8X8;
+      xd->mi[0]->mbmi.tx_size = TX_8X8;
 
     if (cpi->sf.partition_search_type == VAR_BASED_PARTITION) {
       if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
-          cyclic_refresh_segment_id_boosted(xd->mi[0].src_mi->mbmi.segment_id))
-        xd->mi[0].src_mi->mbmi.tx_size = TX_8X8;
-      else if (xd->mi[0].src_mi->mbmi.tx_size > TX_16X16)
-        xd->mi[0].src_mi->mbmi.tx_size = TX_16X16;
+          cyclic_refresh_segment_id_boosted(xd->mi[0]->mbmi.segment_id))
+        xd->mi[0]->mbmi.tx_size = TX_8X8;
+      else if (xd->mi[0]->mbmi.tx_size > TX_16X16)
+        xd->mi[0]->mbmi.tx_size = TX_16X16;
     }
   } else {
-    xd->mi[0].src_mi->mbmi.tx_size =
+    xd->mi[0]->mbmi.tx_size =
         MIN(max_txsize_lookup[bsize],
             tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
   }
@@ -246,7 +494,7 @@ static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize,
   // Evaluate if the partition block is a skippable block in Y plane.
   {
     const BLOCK_SIZE unit_size =
-        txsize_to_bsize[xd->mi[0].src_mi->mbmi.tx_size];
+        txsize_to_bsize[xd->mi[0]->mbmi.tx_size];
     const unsigned int num_blk_log2 =
         (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
         (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
@@ -312,6 +560,132 @@ static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize,
   *out_dist_sum += dist << 4;
 }
 
+#if CONFIG_VP9_HIGHBITDEPTH
+static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate, int64_t *dist,
+                      int *skippable, int64_t *sse, int plane,
+                      BLOCK_SIZE bsize, TX_SIZE tx_size) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  unsigned int var_y, sse_y;
+  (void)plane;
+  (void)tx_size;
+  model_rd_for_sb_y(cpi, bsize, x, xd, rate, dist, &var_y, &sse_y);
+  *sse = INT_MAX;
+  *skippable = 0;
+  return;
+}
+#else
+static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate, int64_t *dist,
+                      int *skippable, int64_t *sse, int plane,
+                      BLOCK_SIZE bsize, TX_SIZE tx_size) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  const struct macroblockd_plane *pd = &xd->plane[plane];
+  const struct macroblock_plane *const p = &x->plane[plane];
+  const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+  const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+  const int step = 1 << (tx_size << 1);
+  const int block_step = (1 << tx_size);
+  int block = 0, r, c;
+  int shift = tx_size == TX_32X32 ? 0 : 2;
+  const int max_blocks_wide = num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 :
+      xd->mb_to_right_edge >> (5 + pd->subsampling_x));
+  const int max_blocks_high = num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 :
+      xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
+  int eob_cost = 0;
+
+  (void)cpi;
+  vp9_subtract_plane(x, bsize, plane);
+  *skippable = 1;
+  // Keep track of the row and column of the blocks we use so that we know
+  // if we are in the unrestricted motion border.
+  for (r = 0; r < max_blocks_high; r += block_step) {
+    for (c = 0; c < num_4x4_w; c += block_step) {
+      if (c < max_blocks_wide) {
+        const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
+        tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+        tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+        tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+        uint16_t *const eob = &p->eobs[block];
+        const int diff_stride = 4 * num_4x4_blocks_wide_lookup[bsize];
+        const int16_t *src_diff;
+        src_diff = &p->src_diff[(r * diff_stride + c) << 2];
+
+        switch (tx_size) {
+          case TX_32X32:
+            vp9_fdct32x32_rd(src_diff, coeff, diff_stride);
+            vp9_quantize_fp_32x32(coeff, 1024, x->skip_block, p->zbin,
+                                  p->round_fp, p->quant_fp, p->quant_shift,
+                                  qcoeff, dqcoeff, pd->dequant, eob,
+                                  scan_order->scan, scan_order->iscan);
+            break;
+          case TX_16X16:
+            vp9_hadamard_16x16(src_diff, diff_stride, (int16_t *)coeff);
+            vp9_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp,
+                            p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                            pd->dequant, eob,
+                            scan_order->scan, scan_order->iscan);
+            break;
+          case TX_8X8:
+            vp9_hadamard_8x8(src_diff, diff_stride, (int16_t *)coeff);
+            vp9_quantize_fp(coeff, 64, x->skip_block, p->zbin, p->round_fp,
+                            p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                            pd->dequant, eob,
+                            scan_order->scan, scan_order->iscan);
+            break;
+          case TX_4X4:
+            x->fwd_txm4x4(src_diff, coeff, diff_stride);
+            vp9_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp,
+                            p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
+                            pd->dequant, eob,
+                            scan_order->scan, scan_order->iscan);
+            break;
+          default:
+            assert(0);
+            break;
+        }
+        *skippable &= (*eob == 0);
+        eob_cost += 1;
+      }
+      block += step;
+    }
+  }
+
+  if (*skippable && *sse < INT64_MAX) {
+    *rate = 0;
+    *dist = (*sse << 6) >> shift;
+    *sse = *dist;
+    return;
+  }
+
+  block = 0;
+  *rate = 0;
+  *dist = 0;
+  *sse = (*sse << 6) >> shift;
+  for (r = 0; r < max_blocks_high; r += block_step) {
+    for (c = 0; c < num_4x4_w; c += block_step) {
+      if (c < max_blocks_wide) {
+        tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+        tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+        tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+        uint16_t *const eob = &p->eobs[block];
+
+        if (*eob == 1)
+          *rate += (int)abs(qcoeff[0]);
+        else if (*eob > 1)
+          *rate += (int)vp9_satd((const int16_t *)qcoeff, step << 4);
+
+        *dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> shift;
+      }
+      block += step;
+    }
+  }
+
+  if (*skippable == 0) {
+    *rate <<= 10;
+    *rate += (eob_cost << 8);
+  }
+}
+#endif
+
 static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE bsize,
                                MACROBLOCK *x, MACROBLOCKD *xd,
                                int *out_rate_sum, int64_t *out_dist_sum,
@@ -402,7 +776,7 @@ static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x,
                                  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].src_mi->mbmi;
+  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;
@@ -518,7 +892,9 @@ static void estimate_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
   int i, j;
   int rate;
   int64_t dist;
-  unsigned int var_y, sse_y;
+  int64_t this_sse = INT64_MAX;
+  int is_skippable;
+
   txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
   assert(plane == 0);
   (void) plane;
@@ -533,8 +909,13 @@ static void estimate_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
                           x->skip_encode ? src_stride : dst_stride,
                           pd->dst.buf, dst_stride,
                           i, j, 0);
-  // This procedure assumes zero offset from p->src.buf and pd->dst.buf.
-  model_rd_for_sb_y(cpi, bsize_tx, x, xd, &rate, &dist, &var_y, &sse_y);
+
+  // TODO(jingning): This needs further refactoring.
+  block_yrd(cpi, x, &rate, &dist, &is_skippable, &this_sse, 0,
+            bsize_tx, MIN(tx_size, TX_16X16));
+  x->skip_txfm[0] = is_skippable;
+  rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), is_skippable);
+
   p->src.buf = src_buf_base;
   pd->dst.buf = dst_buf_base;
   args->rate += rate;
@@ -542,7 +923,7 @@ static void estimate_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
 }
 
 static const THR_MODES mode_idx[MAX_REF_FRAMES - 1][4] = {
-  {THR_DC, THR_H_PRED, THR_V_PRED, THR_TM},
+  {THR_DC, THR_V_PRED, THR_H_PRED, THR_TM},
   {THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV},
   {THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG},
 };
@@ -551,20 +932,54 @@ static const PREDICTION_MODE intra_mode_list[] = {
   DC_PRED, V_PRED, H_PRED, TM_PRED
 };
 
+static int mode_offset(const PREDICTION_MODE mode) {
+  if (mode >= NEARESTMV) {
+    return INTER_OFFSET(mode);
+  } else {
+    switch (mode) {
+      case DC_PRED:
+        return 0;
+      case V_PRED:
+        return 1;
+      case H_PRED:
+        return 2;
+      case TM_PRED:
+        return 3;
+      default:
+        return -1;
+    }
+  }
+}
+
+static INLINE void update_thresh_freq_fact(VP9_COMP *cpi,
+                                           TileDataEnc *tile_data,
+                                           BLOCK_SIZE bsize,
+                                           MV_REFERENCE_FRAME ref_frame,
+                                           THR_MODES best_mode_idx,
+                                           PREDICTION_MODE mode) {
+  THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
+  int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
+  if (thr_mode_idx == best_mode_idx)
+    *freq_fact -= (*freq_fact >> 4);
+  else
+    *freq_fact = MIN(*freq_fact + RD_THRESH_INC,
+        cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
+}
+
 void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
                          BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
   RD_COST this_rdc, best_rdc;
   PREDICTION_MODE this_mode;
   struct estimate_block_intra_args args = { cpi, x, DC_PRED, 0, 0 };
   const TX_SIZE intra_tx_size =
       MIN(max_txsize_lookup[bsize],
           tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
-  MODE_INFO *const mic = xd->mi[0].src_mi;
+  MODE_INFO *const mic = xd->mi[0];
   int *bmode_costs;
-  const MODE_INFO *above_mi = xd->mi[-xd->mi_stride].src_mi;
-  const MODE_INFO *left_mi = xd->left_available ? xd->mi[-1].src_mi : NULL;
+  const MODE_INFO *above_mi = xd->mi[-xd->mi_stride];
+  const MODE_INFO *left_mi = xd->left_available ? xd->mi[-1] : NULL;
   const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
   const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
   bmode_costs = cpi->y_mode_costs[A][L];
@@ -576,7 +991,7 @@ void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
   mbmi->ref_frame[0] = INTRA_FRAME;
   mbmi->mv[0].as_int = INVALID_MV;
   mbmi->uv_mode = DC_PRED;
-  vpx_memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
+  memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
 
   // Change the limit of this loop to add other intra prediction
   // mode tests.
@@ -602,9 +1017,23 @@ void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
   *rd_cost = best_rdc;
 }
 
-static const int ref_frame_cost[MAX_REF_FRAMES] = {
-    1235, 229, 530, 615,
-};
+static void init_ref_frame_cost(VP9_COMMON *const cm,
+                                MACROBLOCKD *const xd,
+                                int ref_frame_cost[MAX_REF_FRAMES]) {
+  vp9_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
+  vp9_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
+  vp9_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
+
+  ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
+  ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
+    ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
+
+  ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
+  ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
+  ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
+  ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
+  ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
+}
 
 typedef struct {
   MV_REFERENCE_FRAME ref_frame;
@@ -632,7 +1061,7 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   VP9_COMMON *const cm = &cpi->common;
   TileInfo *const tile_info = &tile_data->tile_info;
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
   struct macroblockd_plane *const pd = &xd->plane[0];
   PREDICTION_MODE best_mode = ZEROMV;
   MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
@@ -651,7 +1080,7 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   // Reduce the intra cost penalty for small blocks (<=16x16).
   const int reduction_fac =
       (cpi->sf.partition_search_type == VAR_BASED_PARTITION &&
-       bsize <= BLOCK_16X16) ? 2 : 0;
+       bsize <= BLOCK_16X16) ? ((bsize <= BLOCK_8X8) ? 4 : 2) : 0;
   const int intra_cost_penalty = vp9_get_intra_cost_penalty(
       cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth) >> reduction_fac;
   const int64_t inter_mode_thresh = RDCOST(x->rdmult, x->rddiv,
@@ -670,9 +1099,9 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   // process.
   // 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);
+  DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]);
 #if CONFIG_VP9_HIGHBITDEPTH
-  DECLARE_ALIGNED_ARRAY(16, uint16_t, pred_buf_16, 3 * 64 * 64);
+  DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]);
 #endif
   struct buf_2d orig_dst = pd->dst;
   PRED_BUFFER *best_pred = NULL;
@@ -682,6 +1111,10 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   int ref_frame_skip_mask = 0;
   int idx;
   int best_pred_sad = INT_MAX;
+  int best_early_term = 0;
+  int ref_frame_cost[MAX_REF_FRAMES];
+
+  init_ref_frame_cost(cm, xd, ref_frame_cost);
 
   if (reuse_inter_pred) {
     int i;
@@ -706,9 +1139,9 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   x->skip = 0;
 
   if (xd->up_available)
-    filter_ref = xd->mi[-xd->mi_stride].src_mi->mbmi.interp_filter;
+    filter_ref = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
   else if (xd->left_available)
-    filter_ref = xd->mi[-1].src_mi->mbmi.interp_filter;
+    filter_ref = xd->mi[-1]->mbmi.interp_filter;
   else
     filter_ref = cm->interp_filter;
 
@@ -726,7 +1159,6 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
 #endif
 
   if (cpi->rc.frames_since_golden == 0) {
-    cpi->ref_frame_flags &= (~VP9_GOLD_FLAG);
     usable_ref_frame = LAST_FRAME;
   } else {
     usable_ref_frame = GOLDEN_FRAME;
@@ -747,11 +1179,11 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
                            sf, sf);
 
       if (cm->use_prev_frame_mvs)
-        vp9_find_mv_refs(cm, xd, tile_info, xd->mi[0].src_mi, ref_frame,
+        vp9_find_mv_refs(cm, xd, tile_info, xd->mi[0], ref_frame,
                          candidates, mi_row, mi_col, NULL, NULL);
       else
         const_motion[ref_frame] = mv_refs_rt(cm, xd, tile_info,
-                                             xd->mi[0].src_mi,
+                                             xd->mi[0],
                                              ref_frame, candidates,
                                              mi_row, mi_col);
 
@@ -773,6 +1205,10 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
     int mode_index;
     int i;
     PREDICTION_MODE this_mode = ref_mode_set[idx].pred_mode;
+    int64_t this_sse;
+    int is_skippable;
+    int this_early_term = 0;
+
     if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode)))
       continue;
 
@@ -850,6 +1286,7 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
       best_pred_sad = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
                                    x->plane[0].src.stride,
                                    pre_buf, pre_stride);
+      x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
     }
 
     if (this_mode != NEARESTMV &&
@@ -885,7 +1322,7 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
       INTERP_FILTER best_filter = SWITCHABLE, filter;
       PRED_BUFFER *current_pred = this_mode_pred;
 
-      for (filter = EIGHTTAP; filter <= EIGHTTAP_SHARP; ++filter) {
+      for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) {
         int64_t cost;
         mbmi->interp_filter = filter;
         vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
@@ -924,17 +1361,55 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
       var_y = pf_var[best_filter];
       sse_y = pf_sse[best_filter];
       x->skip_txfm[0] = skip_txfm;
+      if (reuse_inter_pred) {
+        pd->dst.buf = this_mode_pred->data;
+        pd->dst.stride = this_mode_pred->stride;
+      }
     } else {
       mbmi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
       vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
-      model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
-                        &var_y, &sse_y);
-      this_rdc.rate +=
-          cm->interp_filter == SWITCHABLE ?
-              vp9_get_switchable_rate(cpi, xd) : 0;
+
+      // For large partition blocks, extra testing is done.
+      if (bsize > BLOCK_32X32 &&
+        !cyclic_refresh_segment_id_boosted(xd->mi[0]->mbmi.segment_id) &&
+        cm->base_qindex) {
+        model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
+                                &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
+                                &this_early_term);
+      } else {
+        model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
+                          &var_y, &sse_y);
+      }
+    }
+
+    if (!this_early_term) {
+      this_sse = (int64_t)sse_y;
+      block_yrd(cpi, x, &this_rdc.rate, &this_rdc.dist, &is_skippable,
+                &this_sse, 0, bsize, MIN(mbmi->tx_size, TX_16X16));
+      x->skip_txfm[0] = is_skippable;
+      if (is_skippable) {
+        this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+      } else {
+        if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
+            RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
+          this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
+        } else {
+          this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
+          this_rdc.dist = this_sse;
+          x->skip_txfm[0] = 1;
+        }
+      }
+
+      if (cm->interp_filter == SWITCHABLE) {
+        if ((mbmi->mv[0].as_mv.row | mbmi->mv[0].as_mv.col) & 0x07)
+          this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
+      }
+    } else {
+      this_rdc.rate += cm->interp_filter == SWITCHABLE ?
+          vp9_get_switchable_rate(cpi, xd) : 0;
+      this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
     }
 
-    // chroma component rate-distortion cost modeling
     if (x->color_sensitivity[0] || x->color_sensitivity[1]) {
       int uv_rate = 0;
       int64_t uv_dist = 0;
@@ -942,7 +1417,8 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
         vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
       if (x->color_sensitivity[1])
         vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
-      model_rd_for_sb_uv(cpi, bsize, x, xd, &uv_rate, &uv_dist, &var_y, &sse_y);
+      model_rd_for_sb_uv(cpi, bsize, x, xd, &uv_rate, &uv_dist,
+                         &var_y, &sse_y);
       this_rdc.rate += uv_rate;
       this_rdc.dist += uv_dist;
     }
@@ -981,6 +1457,7 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
       best_tx_size = mbmi->tx_size;
       best_ref_frame = ref_frame;
       best_mode_skip_txfm = x->skip_txfm[0];
+      best_early_term = this_early_term;
 
       if (reuse_inter_pred) {
         free_pred_buffer(best_pred);
@@ -993,6 +1470,13 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
 
     if (x->skip)
       break;
+
+    // If early termination flag is 1 and at least 2 modes are checked,
+    // the mode search is terminated.
+    if (best_early_term && idx > 0) {
+      x->skip = 1;
+      break;
+    }
   }
 
   mbmi->mode          = best_mode;
@@ -1000,7 +1484,7 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   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].src_mi->bmi[0].as_mv[0].as_int = mbmi->mv[0].as_int;
+  xd->mi[0]->bmi[0].as_mv[0].as_int = mbmi->mv[0].as_int;
   x->skip_txfm[0] = best_mode_skip_txfm;
 
   // Perform intra prediction search, if the best SAD is above a certain
@@ -1039,8 +1523,18 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
 
     for (i = 0; i < 4; ++i) {
       const PREDICTION_MODE this_mode = intra_mode_list[i];
-      if (!((1 << this_mode) & cpi->sf.intra_y_mode_mask[intra_tx_size]))
+      THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
+      int mode_rd_thresh = rd_threshes[mode_index];
+
+      if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
+        continue;
+
+      if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
+                              rd_thresh_freq_fact[mode_index]))
         continue;
+
+      mbmi->mode = this_mode;
+      mbmi->ref_frame[0] = INTRA_FRAME;
       args.mode = this_mode;
       args.rate = 0;
       args.dist = 0;
@@ -1057,17 +1551,17 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
 
       if (this_rdc.rdcost < best_rdc.rdcost) {
         best_rdc = this_rdc;
-        mbmi->mode = this_mode;
+        best_mode = this_mode;
         best_intra_tx_size = mbmi->tx_size;
-        mbmi->ref_frame[0] = INTRA_FRAME;
+        best_ref_frame = INTRA_FRAME;
         mbmi->uv_mode = this_mode;
         mbmi->mv[0].as_int = INVALID_MV;
+        best_mode_skip_txfm = x->skip_txfm[0];
       }
     }
 
     // Reset mb_mode_info to the best inter mode.
-    if (mbmi->ref_frame[0] != INTRA_FRAME) {
-      x->skip_txfm[0] = best_mode_skip_txfm;
+    if (best_ref_frame != INTRA_FRAME) {
       mbmi->tx_size = best_tx_size;
     } else {
       mbmi->tx_size = best_intra_tx_size;
@@ -1075,6 +1569,9 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   }
 
   pd->dst = orig_dst;
+  mbmi->mode = best_mode;
+  mbmi->ref_frame[0] = best_ref_frame;
+  x->skip_txfm[0] = best_mode_skip_txfm;
 
   if (reuse_inter_pred && best_pred != NULL) {
     if (best_pred->data != orig_dst.buf && is_inter_mode(mbmi->mode)) {
@@ -1096,20 +1593,28 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   }
 
   if (cpi->sf.adaptive_rd_thresh) {
-    THR_MODES best_mode_idx = is_inter_block(mbmi) ?
-        mode_idx[best_ref_frame][INTER_OFFSET(mbmi->mode)] :
-        mode_idx[INTRA_FRAME][mbmi->mode];
+    THR_MODES best_mode_idx = mode_idx[best_ref_frame][mode_offset(mbmi->mode)];
     PREDICTION_MODE this_mode;
-    for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
-      if (best_ref_frame != ref_frame) continue;
-      for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
-        THR_MODES thr_mode_idx = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
-        int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
-        if (thr_mode_idx == best_mode_idx)
-          *freq_fact -= (*freq_fact >> 4);
-        else
-          *freq_fact = MIN(*freq_fact + RD_THRESH_INC,
-                           cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
+
+    if (best_ref_frame == INTRA_FRAME) {
+      // Only consider the modes that are included in the intra_mode_list.
+      int intra_modes = sizeof(intra_mode_list)/sizeof(PREDICTION_MODE);
+      int i;
+
+      // TODO(yunqingwang): Check intra mode mask and only update freq_fact
+      // for those valid modes.
+      for (i = 0; i < intra_modes; i++) {
+        PREDICTION_MODE this_mode = intra_mode_list[i];
+        update_thresh_freq_fact(cpi, tile_data, bsize, INTRA_FRAME,
+                                best_mode_idx, this_mode);
+      }
+    } else {
+      for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
+        if (best_ref_frame != ref_frame) continue;
+        for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
+          update_thresh_freq_fact(cpi, tile_data, bsize, ref_frame,
+                                  best_mode_idx, this_mode);
+        }
       }
     }
   }
@@ -1125,7 +1630,7 @@ void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
   TileInfo *const tile_info = &tile_data->tile_info;
   SPEED_FEATURES *const sf = &cpi->sf;
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
   const struct segmentation *const seg = &cm->seg;
   MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
   MV_REFERENCE_FRAME best_ref_frame = NONE;
@@ -1154,7 +1659,7 @@ void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
                              &cm->frame_refs[ref_frame - 1].sf;
       vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
                            sf, sf);
-      vp9_find_mv_refs(cm, xd, tile_info, xd->mi[0].src_mi, ref_frame,
+      vp9_find_mv_refs(cm, xd, tile_info, xd->mi[0], ref_frame,
                        candidates, mi_row, mi_col, NULL, NULL);
 
       vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
@@ -1214,7 +1719,7 @@ void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
         const struct buf_2d orig_src = p->src;
         const struct buf_2d orig_dst = pd->dst;
         struct buf_2d orig_pre[2];
-        vpx_memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
+        memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
 
         // set buffer pointers for sub8x8 motion search.
         p->src.buf =
@@ -1233,7 +1738,7 @@ void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
 
         for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
           int b_rate = 0;
-          xd->mi[0].bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
+          xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
 
           if (this_mode == NEWMV) {
             const int step_param = cpi->sf.mv.fullpel_search_step_param;
@@ -1250,8 +1755,8 @@ void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
               mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
               mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
             } else {
-              mvp_full.row = xd->mi[0].bmi[0].as_mv[0].as_mv.row >> 3;
-              mvp_full.col = xd->mi[0].bmi[0].as_mv[0].as_mv.col >> 3;
+              mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
+              mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
             }
 
             vp9_set_mv_search_range(x, &mbmi->ref_mvs[0]->as_mv);
@@ -1293,7 +1798,7 @@ void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
                                          &dummy_dist,
                                          &x->pred_sse[ref_frame], NULL, 0, 0);
 
-            xd->mi[0].bmi[i].as_mv[0].as_mv = tmp_mv;
+            xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
           } else {
             b_rate += cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
                                           [INTER_OFFSET(this_mode)];
@@ -1303,7 +1808,7 @@ void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
           if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
             vp9_highbd_build_inter_predictor(pd->pre[0].buf, pd->pre[0].stride,
                                     pd->dst.buf, pd->dst.stride,
-                                    &xd->mi[0].bmi[i].as_mv[0].as_mv,
+                                    &xd->mi[0]->bmi[i].as_mv[0].as_mv,
                                     &xd->block_refs[0]->sf,
                                     4 * num_4x4_blocks_wide,
                                     4 * num_4x4_blocks_high, 0,
@@ -1315,7 +1820,7 @@ void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
 #endif
             vp9_build_inter_predictor(pd->pre[0].buf, pd->pre[0].stride,
                                      pd->dst.buf, pd->dst.stride,
-                                     &xd->mi[0].bmi[i].as_mv[0].as_mv,
+                                     &xd->mi[0]->bmi[i].as_mv[0].as_mv,
                                      &xd->block_refs[0]->sf,
                                      4 * num_4x4_blocks_wide,
                                      4 * num_4x4_blocks_high, 0,
@@ -1337,7 +1842,7 @@ void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
           if (this_rdc.rdcost < b_best_rd) {
             b_best_rd = this_rdc.rdcost;
             bsi[ref_frame][i].as_mode = this_mode;
-            bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0].bmi[i].as_mv[0].as_mv;
+            bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
           }
         }  // mode search
 
@@ -1347,11 +1852,11 @@ void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
         pd->dst = orig_dst;
         this_rd += b_best_rd;
 
-        xd->mi[0].bmi[i] = bsi[ref_frame][i];
+        xd->mi[0]->bmi[i] = bsi[ref_frame][i];
         if (num_4x4_blocks_wide > 1)
-          xd->mi[0].bmi[i + 1] = xd->mi[0].bmi[i];
+          xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
         if (num_4x4_blocks_high > 1)
-          xd->mi[0].bmi[i + 2] = xd->mi[0].bmi[i];
+          xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
       }
     }  // loop through sub8x8 blocks
 
@@ -1366,15 +1871,15 @@ void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
   for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
     for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
       const int block = idy * 2 + idx;
-      xd->mi[0].bmi[block] = bsi[best_ref_frame][block];
+      xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
       if (num_4x4_blocks_wide > 1)
-        xd->mi[0].bmi[block + 1] = bsi[best_ref_frame][block];
+        xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
       if (num_4x4_blocks_high > 1)
-        xd->mi[0].bmi[block + 2] = bsi[best_ref_frame][block];
+        xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
     }
   }
-  mbmi->mode = xd->mi[0].bmi[3].as_mode;
-  ctx->mic = *(xd->mi[0].src_mi);
+  mbmi->mode = xd->mi[0]->bmi[3].as_mode;
+  ctx->mic = *(xd->mi[0]);
   ctx->skip_txfm[0] = 0;
   ctx->skip = 0;
   // Dummy assignment for speed -5. No effect in speed -6.