libvpx: Pull from upstream

source/libvpx/vp9/encoder/vp9_rdopt.c

 /*
  *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
  *
  *  Use of this source code is governed by a BSD-style license
  *  that can be found in the LICENSE file in the root of the source
  *  tree. An additional intellectual property rights grant can be found
  *  in the file PATENTS.  All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
@@ skipping 23 matching lines @@
 #include "vp9/encoder/vp9_mcomp.h"
 #include "vp9/encoder/vp9_quantize.h"
 #include "vp9/encoder/vp9_ratectrl.h"
 #include "vp9/encoder/vp9_rd.h"
 #include "vp9/encoder/vp9_rdopt.h"
 #include "vp9/encoder/vp9_variance.h"
 
 #define RD_THRESH_MAX_FACT 64
 #define RD_THRESH_INC      1
 
-#define LAST_FRAME_MODE_MASK    0xFFEDCD60
-#define GOLDEN_FRAME_MODE_MASK  0xFFDA3BB0
-#define ALT_REF_MODE_MASK       0xFFC648D0
+#define LAST_FRAME_MODE_MASK    ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | \
+                                 (1 << INTRA_FRAME))
+#define GOLDEN_FRAME_MODE_MASK  ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | \
+                                 (1 << INTRA_FRAME))
+#define ALT_REF_MODE_MASK       ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | \
+                                 (1 << INTRA_FRAME))
+
+#define SECOND_REF_FRAME_MASK   ((1 << ALTREF_FRAME) | 0x01)
 
 #define MIN_EARLY_TERM_INDEX    3
 
 typedef struct {
   PREDICTION_MODE mode;
   MV_REFERENCE_FRAME ref_frame[2];
 } MODE_DEFINITION;
 
 typedef struct {
   MV_REFERENCE_FRAME ref_frame[2];
@@ skipping 22 matching lines @@
   {NEARESTMV, {GOLDEN_FRAME, NONE}},
 
   {DC_PRED,   {INTRA_FRAME,  NONE}},
 
   {NEWMV,     {LAST_FRAME,   NONE}},
   {NEWMV,     {ALTREF_FRAME, NONE}},
   {NEWMV,     {GOLDEN_FRAME, NONE}},
 
   {NEARMV,    {LAST_FRAME,   NONE}},
   {NEARMV,    {ALTREF_FRAME, NONE}},
+  {NEARMV,    {GOLDEN_FRAME, NONE}},
+
+  {ZEROMV,    {LAST_FRAME,   NONE}},
+  {ZEROMV,    {GOLDEN_FRAME, NONE}},
+  {ZEROMV,    {ALTREF_FRAME, NONE}},
+
   {NEARESTMV, {LAST_FRAME,   ALTREF_FRAME}},
   {NEARESTMV, {GOLDEN_FRAME, ALTREF_FRAME}},
 
   {TM_PRED,   {INTRA_FRAME,  NONE}},
 
   {NEARMV,    {LAST_FRAME,   ALTREF_FRAME}},
   {NEWMV,     {LAST_FRAME,   ALTREF_FRAME}},
-  {NEARMV,    {GOLDEN_FRAME, NONE}},
   {NEARMV,    {GOLDEN_FRAME, ALTREF_FRAME}},
   {NEWMV,     {GOLDEN_FRAME, ALTREF_FRAME}},
 
-  {ZEROMV,    {LAST_FRAME,   NONE}},
-  {ZEROMV,    {GOLDEN_FRAME, NONE}},
-  {ZEROMV,    {ALTREF_FRAME, NONE}},
   {ZEROMV,    {LAST_FRAME,   ALTREF_FRAME}},
   {ZEROMV,    {GOLDEN_FRAME, ALTREF_FRAME}},
 
   {H_PRED,    {INTRA_FRAME,  NONE}},
   {V_PRED,    {INTRA_FRAME,  NONE}},
   {D135_PRED, {INTRA_FRAME,  NONE}},
   {D207_PRED, {INTRA_FRAME,  NONE}},
   {D153_PRED, {INTRA_FRAME,  NONE}},
   {D63_PRED,  {INTRA_FRAME,  NONE}},
   {D117_PRED, {INTRA_FRAME,  NONE}},
@@ skipping 49 matching lines @@
 
 static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
                             MACROBLOCK *x, MACROBLOCKD *xd,
                             int *out_rate_sum, int64_t *out_dist_sum) {
   // 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.
   int i;
   int64_t rate_sum = 0;
   int64_t dist_sum = 0;
-  const int ref = xd->mi[0]->mbmi.ref_frame[0];
+  const int ref = xd->mi[0].src_mi->mbmi.ref_frame[0];
   unsigned int sse;
   unsigned int var = 0;
   unsigned int sum_sse = 0;
   const int shift = 8;
   int rate;
   int64_t dist;
 
   x->pred_sse[ref] = 0;
 
   for (i = 0; i < MAX_MB_PLANE; ++i) {
@@ skipping 54 matching lines @@
                                    pd->dequant[1] >> 3, &rate, &dist);
       rate_sum += rate;
       dist_sum += dist;
     }
   }
 
   *out_rate_sum = (int)rate_sum;
   *out_dist_sum = dist_sum << 4;
 }
 
-int64_t vp9_block_error_c(const int16_t *coeff, const int16_t *dqcoeff,
+int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
                           intptr_t block_size, int64_t *ssz) {
   int i;
   int64_t error = 0, sqcoeff = 0;
 
   for (i = 0; i < block_size; i++) {
     const int diff = coeff[i] - dqcoeff[i];
     error +=  diff * diff;
     sqcoeff += coeff[i] * coeff[i];
   }
 
@@ skipping 12 matching lines @@
   { 1, 2, 3, 4, 11,  256 - 21, 0 },
   { 1, 2, 3, 4, 11, 1024 - 21, 0 },
 };
 static INLINE int cost_coeffs(MACROBLOCK *x,
                               int plane, int block,
                               ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
                               TX_SIZE tx_size,
                               const int16_t *scan, const int16_t *nb,
                               int use_fast_coef_costing) {
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
   const struct macroblock_plane *p = &x->plane[plane];
   const struct macroblockd_plane *pd = &xd->plane[plane];
   const PLANE_TYPE type = pd->plane_type;
   const int16_t *band_count = &band_counts[tx_size][1];
   const int eob = p->eobs[block];
-  const int16_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
   unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
                    x->token_costs[tx_size][type][is_inter_block(mbmi)];
   uint8_t token_cache[32 * 32];
   int pt = combine_entropy_contexts(*A, *L);
   int c, cost;
   // Check for consistency of tx_size with mode info
   assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size
                               : get_uv_tx_size(mbmi, pd) == tx_size);
 
   if (eob == 0) {
@@ skipping 49 matching lines @@
 }
 static void dist_block(int plane, int block, TX_SIZE tx_size,
                        struct rdcost_block_args* args) {
   const int ss_txfrm_size = tx_size << 1;
   MACROBLOCK* const x = args->x;
   MACROBLOCKD* const xd = &x->e_mbd;
   const struct macroblock_plane *const p = &x->plane[plane];
   const struct macroblockd_plane *const pd = &xd->plane[plane];
   int64_t this_sse;
   int shift = tx_size == TX_32X32 ? 0 : 2;
-  int16_t *const coeff = BLOCK_OFFSET(p->coeff, block);
-  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
   args->dist = vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
                                &this_sse) >> shift;
   args->sse  = this_sse >> shift;
 
-  if (x->skip_encode && !is_inter_block(&xd->mi[0]->mbmi)) {
+  if (x->skip_encode && !is_inter_block(&xd->mi[0].src_mi->mbmi)) {
     // TODO(jingning): tune the model to better capture the distortion.
     int64_t p = (pd->dequant[1] * pd->dequant[1] *
                     (1 << ss_txfrm_size)) >> (shift + 2);
     args->dist += (p >> 4);
     args->sse  += p;
   }
 }
 
 static void rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
                        TX_SIZE tx_size, struct rdcost_block_args* args) {
   int x_idx, y_idx;
   txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x_idx, &y_idx);
 
   args->rate = cost_coeffs(args->x, plane, block, args->t_above + x_idx,
                            args->t_left + y_idx, tx_size,
                            args->so->scan, args->so->neighbors,
                            args->use_fast_coef_costing);
 }
 
 static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
                           TX_SIZE tx_size, void *arg) {
   struct rdcost_block_args *args = arg;
   MACROBLOCK *const x = args->x;
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
   int64_t rd1, rd2, rd;
 
   if (args->skip)
     return;
 
   if (!is_inter_block(mbmi)) {
     vp9_encode_block_intra(x, plane, block, plane_bsize, tx_size, &mbmi->skip);
     dist_block(plane, block, tx_size, args);
   } else if (max_txsize_lookup[plane_bsize] == tx_size) {
     if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 0) {
       // full forward transform and quantization
       vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
       dist_block(plane, block, tx_size, args);
     } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 2) {
       // compute DC coefficient
-      int16_t *const coeff   = BLOCK_OFFSET(x->plane[plane].coeff, block);
-      int16_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
+      tran_low_t *const coeff   = BLOCK_OFFSET(x->plane[plane].coeff, block);
+      tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
       vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
       args->sse  = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
       args->dist = args->sse;
       if (!x->plane[plane].eobs[block])
         args->dist = args->sse - ((coeff[0] * coeff[0] -
             (coeff[0] - dqcoeff[0]) * (coeff[0] - dqcoeff[0])) >> 2);
     } else {
       // skip forward transform
       x->plane[plane].eobs[block] = 0;
       args->sse  = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
@@ skipping 34 matching lines @@
                              int use_fast_coef_casting) {
   MACROBLOCKD *const xd = &x->e_mbd;
   const struct macroblockd_plane *const pd = &xd->plane[plane];
   struct rdcost_block_args args;
   vp9_zero(args);
   args.x = x;
   args.best_rd = ref_best_rd;
   args.use_fast_coef_costing = use_fast_coef_casting;
 
   if (plane == 0)
-    xd->mi[0]->mbmi.tx_size = tx_size;
+    xd->mi[0].src_mi->mbmi.tx_size = tx_size;
 
   vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
 
   args.so = get_scan(xd, tx_size, pd->plane_type, 0);
 
   vp9_foreach_transformed_block_in_plane(xd, bsize, plane,
                                          block_rd_txfm, &args);
   if (args.skip) {
     *rate       = INT_MAX;
     *distortion = INT64_MAX;
     *sse        = INT64_MAX;
     *skippable  = 0;
   } else {
     *distortion = args.this_dist;
     *rate       = args.this_rate;
     *sse        = args.this_sse;
     *skippable  = vp9_is_skippable_in_plane(x, bsize, plane);
   }
 }
 
 static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x,
                                    int *rate, int64_t *distortion,
                                    int *skip, int64_t *sse,
                                    int64_t ref_best_rd,
                                    BLOCK_SIZE bs) {
   const TX_SIZE max_tx_size = max_txsize_lookup[bs];
   VP9_COMMON *const cm = &cpi->common;
   const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
 
   mbmi->tx_size = MIN(max_tx_size, largest_tx_size);
 
   txfm_rd_in_plane(x, rate, distortion, skip,
                    sse, ref_best_rd, 0, bs,
                    mbmi->tx_size, cpi->sf.use_fast_coef_costing);
 }
 
 static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
                                    int *rate,
                                    int64_t *distortion,
                                    int *skip,
                                    int64_t *psse,
                                    int64_t tx_cache[TX_MODES],
                                    int64_t ref_best_rd,
                                    BLOCK_SIZE bs) {
   const TX_SIZE max_tx_size = max_txsize_lookup[bs];
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
   vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
   int r[TX_SIZES][2], s[TX_SIZES];
   int64_t d[TX_SIZES], sse[TX_SIZES];
   int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
                              {INT64_MAX, INT64_MAX},
                              {INT64_MAX, INT64_MAX},
                              {INT64_MAX, INT64_MAX}};
   int n, m;
   int s0, s1;
   const TX_SIZE max_mode_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
@@ skipping 66 matching lines @@
 
 static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
                             int64_t *distortion, int *skip,
                             int64_t *psse, BLOCK_SIZE bs,
                             int64_t txfm_cache[TX_MODES],
                             int64_t ref_best_rd) {
   MACROBLOCKD *xd = &x->e_mbd;
   int64_t sse;
   int64_t *ret_sse = psse ? psse : &sse;
 
-  assert(bs == xd->mi[0]->mbmi.sb_type);
+  assert(bs == xd->mi[0].src_mi->mbmi.sb_type);
 
   if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
     vpx_memset(txfm_cache, 0, TX_MODES * sizeof(int64_t));
     choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
                            bs);
   } else {
     choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse,
                            txfm_cache, ref_best_rd, bs);
   }
 }
@@ skipping 43 matching lines @@
 
   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
   int idx, idy;
   uint8_t best_dst[8 * 8];
 
   assert(ib < 4);
 
   vpx_memcpy(ta, a, sizeof(ta));
   vpx_memcpy(tl, l, sizeof(tl));
-  xd->mi[0]->mbmi.tx_size = TX_4X4;
+  xd->mi[0].src_mi->mbmi.tx_size = TX_4X4;
 
   for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
     int64_t this_rd;
     int ratey = 0;
     int64_t distortion = 0;
     int rate = bmode_costs[mode];
 
     if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
       continue;
 
     // Only do the oblique modes if the best so far is
     // one of the neighboring directional modes
     if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
       if (conditional_skipintra(mode, *best_mode))
           continue;
     }
 
     vpx_memcpy(tempa, ta, sizeof(ta));
     vpx_memcpy(templ, tl, sizeof(tl));
 
     for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
       for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
         const int block = ib + idy * 2 + idx;
         const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
         uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
         int16_t *const src_diff = raster_block_offset_int16(BLOCK_8X8, block,
                                                             p->src_diff);
-        int16_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
-        xd->mi[0]->bmi[block].as_mode = mode;
+        tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
+        xd->mi[0].src_mi->bmi[block].as_mode = mode;
         vp9_predict_intra_block(xd, block, 1,
                                 TX_4X4, mode,
                                 x->skip_encode ? src : dst,
                                 x->skip_encode ? src_stride : dst_stride,
                                 dst, dst_stride, idx, idy, 0);
         vp9_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
 
         if (xd->lossless) {
           const scan_order *so = &vp9_default_scan_orders[TX_4X4];
           vp9_fwht4x4(src_diff, coeff, 8);
@@ skipping 52 matching lines @@
 
   return best_rd;
 }
 
 static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
                                             int *rate, int *rate_y,
                                             int64_t *distortion,
                                             int64_t best_rd) {
   int i, j;
   const MACROBLOCKD *const xd = &mb->e_mbd;
-  MODE_INFO *const mic = xd->mi[0];
-  const MODE_INFO *above_mi = xd->mi[-xd->mi_stride];
-  const MODE_INFO *left_mi = xd->left_available ? xd->mi[-1] : NULL;
-  const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  MODE_INFO *const mic = xd->mi[0].src_mi;
+  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 BLOCK_SIZE bsize = xd->mi[0].src_mi->mbmi.sb_type;
   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
   int idx, idy;
   int cost = 0;
   int64_t total_distortion = 0;
   int tot_rate_y = 0;
   int64_t total_rd = 0;
   ENTROPY_CONTEXT t_above[4], t_left[4];
   const int *bmode_costs = cpi->mbmode_cost;
 
@@ skipping 46 matching lines @@
 
 static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
                                       int *rate, int *rate_tokenonly,
                                       int64_t *distortion, int *skippable,
                                       BLOCK_SIZE bsize,
                                       int64_t tx_cache[TX_MODES],
                                       int64_t best_rd) {
   PREDICTION_MODE mode;
   PREDICTION_MODE mode_selected = DC_PRED;
   MACROBLOCKD *const xd = &x->e_mbd;
-  MODE_INFO *const mic = xd->mi[0];
+  MODE_INFO *const mic = xd->mi[0].src_mi;
   int this_rate, this_rate_tokenonly, s;
   int64_t this_distortion, this_rd;
   TX_SIZE best_tx = TX_4X4;
   int i;
   int *bmode_costs = cpi->mbmode_cost;
 
   if (cpi->sf.tx_size_search_method == USE_FULL_RD)
     for (i = 0; i < TX_MODES; i++)
       tx_cache[i] = INT64_MAX;
 
   /* Y Search for intra prediction mode */
   for (mode = DC_PRED; mode <= TM_PRED; mode++) {
     int64_t local_tx_cache[TX_MODES];
-    MODE_INFO *above_mi = xd->mi[-xd->mi_stride];
-    MODE_INFO *left_mi = xd->left_available ? xd->mi[-1] : NULL;
+    MODE_INFO *above_mi = xd->mi[-xd->mi_stride].src_mi;
+    MODE_INFO *left_mi = xd->left_available ? xd->mi[-1].src_mi : NULL;
 
     if (cpi->common.frame_type == KEY_FRAME) {
       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];
     }
     mic->mbmi.mode = mode;
 
     super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion,
@@ skipping 30 matching lines @@
   mic->mbmi.tx_size = best_tx;
 
   return best_rd;
 }
 
 static void super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x,
                              int *rate, int64_t *distortion, int *skippable,
                              int64_t *sse, BLOCK_SIZE bsize,
                              int64_t ref_best_rd) {
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
   const TX_SIZE uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
   int plane;
   int pnrate = 0, pnskip = 1;
   int64_t pndist = 0, pnsse = 0;
 
   if (ref_best_rd < 0)
     goto term;
 
   if (is_inter_block(mbmi)) {
     int plane;
@@ skipping 36 matching lines @@
   PREDICTION_MODE mode;
   PREDICTION_MODE mode_selected = DC_PRED;
   int64_t best_rd = INT64_MAX, this_rd;
   int this_rate_tokenonly, this_rate, s;
   int64_t this_distortion, this_sse;
 
   for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
     if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode)))
       continue;
 
-    xd->mi[0]->mbmi.uv_mode = mode;
+    xd->mi[0].src_mi->mbmi.uv_mode = mode;
 
     super_block_uvrd(cpi, x, &this_rate_tokenonly,
                      &this_distortion, &s, &this_sse, bsize, best_rd);
     if (this_rate_tokenonly == INT_MAX)
       continue;
     this_rate = this_rate_tokenonly +
                 cpi->intra_uv_mode_cost[cpi->common.frame_type][mode];
     this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
 
     if (this_rd < best_rd) {
       mode_selected   = mode;
       best_rd         = this_rd;
       *rate           = this_rate;
       *rate_tokenonly = this_rate_tokenonly;
       *distortion     = this_distortion;
       *skippable      = s;
       if (!x->select_tx_size)
         swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
     }
   }
 
-  xd->mi[0]->mbmi.uv_mode = mode_selected;
+  xd->mi[0].src_mi->mbmi.uv_mode = mode_selected;
   return best_rd;
 }
 
 static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x,
                               int *rate, int *rate_tokenonly,
                               int64_t *distortion, int *skippable,
                               BLOCK_SIZE bsize) {
   const VP9_COMMON *cm = &cpi->common;
   int64_t unused;
 
-  x->e_mbd.mi[0]->mbmi.uv_mode = DC_PRED;
+  x->e_mbd.mi[0].src_mi->mbmi.uv_mode = DC_PRED;
   super_block_uvrd(cpi, x, rate_tokenonly, distortion,
                    skippable, &unused, bsize, INT64_MAX);
   *rate = *rate_tokenonly + cpi->intra_uv_mode_cost[cm->frame_type][DC_PRED];
   return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
 }
 
 static void choose_intra_uv_mode(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
                                  BLOCK_SIZE bsize, TX_SIZE max_tx_size,
                                  int *rate_uv, int *rate_uv_tokenonly,
                                  int64_t *dist_uv, int *skip_uv,
                                  PREDICTION_MODE *mode_uv) {
   MACROBLOCK *const x = &cpi->mb;
 
   // Use an estimated rd for uv_intra based on DC_PRED if the
   // appropriate speed flag is set.
   if (cpi->sf.use_uv_intra_rd_estimate) {
     rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv,
                    skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
   // Else do a proper rd search for each possible transform size that may
   // be considered in the main rd loop.
   } else {
     rd_pick_intra_sbuv_mode(cpi, x, ctx,
                             rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
                             bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize, max_tx_size);
   }
-  *mode_uv = x->e_mbd.mi[0]->mbmi.uv_mode;
+  *mode_uv = x->e_mbd.mi[0].src_mi->mbmi.uv_mode;
 }
 
 static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
                        int mode_context) {
   assert(is_inter_mode(mode));
   return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
 }
 
 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                 BLOCK_SIZE bsize,
                                 int_mv *frame_mv,
                                 int mi_row, int mi_col,
                                 int_mv single_newmv[MAX_REF_FRAMES],
                                 int *rate_mv);
 
 static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCKD *xd, int i,
                                 PREDICTION_MODE mode, int_mv this_mv[2],
                                 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
                                 int_mv seg_mvs[MAX_REF_FRAMES],
                                 int_mv *best_ref_mv[2], const int *mvjcost,
                                 int *mvcost[2]) {
-  MODE_INFO *const mic = xd->mi[0];
+  MODE_INFO *const mic = xd->mi[0].src_mi;
   const MB_MODE_INFO *const mbmi = &mic->mbmi;
   int thismvcost = 0;
   int idx, idy;
   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
   const int is_compound = has_second_ref(mbmi);
 
   switch (mode) {
     case NEWMV:
       this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
@@ skipping 41 matching lines @@
                                        int i,
                                        int *labelyrate,
                                        int64_t *distortion, int64_t *sse,
                                        ENTROPY_CONTEXT *ta,
                                        ENTROPY_CONTEXT *tl,
                                        int mi_row, int mi_col) {
   int k;
   MACROBLOCKD *xd = &x->e_mbd;
   struct macroblockd_plane *const pd = &xd->plane[0];
   struct macroblock_plane *const p = &x->plane[0];
-  MODE_INFO *const mi = xd->mi[0];
+  MODE_INFO *const mi = xd->mi[0].src_mi;
   const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
   const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
   const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
   int idx, idy;
 
   const uint8_t *const src = &p->src.buf[raster_block_offset(BLOCK_8X8, i,
                                                              p->src.stride)];
   uint8_t *const dst = &pd->dst.buf[raster_block_offset(BLOCK_8X8, i,
                                                         pd->dst.stride)];
   int64_t thisdistortion = 0, thissse = 0;
@@ skipping 16 matching lines @@
 
   vp9_subtract_block(height, width,
                      raster_block_offset_int16(BLOCK_8X8, i, p->src_diff), 8,
                      src, p->src.stride,
                      dst, pd->dst.stride);
 
   k = i;
   for (idy = 0; idy < height / 4; ++idy) {
     for (idx = 0; idx < width / 4; ++idx) {
       int64_t ssz, rd, rd1, rd2;
-      int16_t* coeff;
+      tran_low_t* coeff;
 
       k += (idy * 2 + idx);
       coeff = BLOCK_OFFSET(p->coeff, k);
       x->fwd_txm4x4(raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
                     coeff, 8);
       vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
       thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
                                         16, &ssz);
       thissse += ssz;
       thisrate += cost_coeffs(x, 0, k, ta + (k & 1), tl + (k >> 1), TX_4X4,
@@ skipping 41 matching lines @@
 } BEST_SEG_INFO;
 
 static INLINE int mv_check_bounds(const MACROBLOCK *x, const MV *mv) {
   return (mv->row >> 3) < x->mv_row_min ||
          (mv->row >> 3) > x->mv_row_max ||
          (mv->col >> 3) < x->mv_col_min ||
          (mv->col >> 3) > x->mv_col_max;
 }
 
 static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
-  MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0]->mbmi;
+  MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0].src_mi->mbmi;
   struct macroblock_plane *const p = &x->plane[0];
   struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
 
   p->src.buf = &p->src.buf[raster_block_offset(BLOCK_8X8, i, p->src.stride)];
   assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
   pd->pre[0].buf = &pd->pre[0].buf[raster_block_offset(BLOCK_8X8, i,
                                                        pd->pre[0].stride)];
   if (has_second_ref(mbmi))
     pd->pre[1].buf = &pd->pre[1].buf[raster_block_offset(BLOCK_8X8, i,
                                                          pd->pre[1].stride)];
 }
 
 static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
                                   struct buf_2d orig_pre[2]) {
-  MB_MODE_INFO *mbmi = &x->e_mbd.mi[0]->mbmi;
+  MB_MODE_INFO *mbmi = &x->e_mbd.mi[0].src_mi->mbmi;
   x->plane[0].src = orig_src;
   x->e_mbd.plane[0].pre[0] = orig_pre[0];
   if (has_second_ref(mbmi))
     x->e_mbd.plane[0].pre[1] = orig_pre[1];
 }
 
 static INLINE int mv_has_subpel(const MV *mv) {
   return (mv->row & 0x0F) || (mv->col & 0x0F);
 }
 
 // Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
 // TODO(aconverse): Find out if this is still productive then clean up or remove
 static int check_best_zero_mv(
     const VP9_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES],
-    int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
-    int inter_mode_mask, int this_mode,
+    int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int this_mode,
     const MV_REFERENCE_FRAME ref_frames[2]) {
-  if ((inter_mode_mask & (1 << ZEROMV)) &&
-      (this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
+  if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
       frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
       (ref_frames[1] == NONE ||
        frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
     int rfc = mode_context[ref_frames[0]];
     int c1 = cost_mv_ref(cpi, NEARMV, rfc);
     int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
     int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
 
     if (this_mode == NEARMV) {
       if (c1 > c3) return 0;
@@ skipping 25 matching lines @@
                                         int *returnyrate,
                                         int64_t *returndistortion,
                                         int *skippable, int64_t *psse,
                                         int mvthresh,
                                         int_mv seg_mvs[4][MAX_REF_FRAMES],
                                         BEST_SEG_INFO *bsi_buf, int filter_idx,
                                         int mi_row, int mi_col) {
   int i;
   BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
   MACROBLOCKD *xd = &x->e_mbd;
-  MODE_INFO *mi = xd->mi[0];
+  MODE_INFO *mi = xd->mi[0].src_mi;
   MB_MODE_INFO *mbmi = &mi->mbmi;
   int mode_idx;
   int k, br = 0, idx, idy;
   int64_t bd = 0, block_sse = 0;
   PREDICTION_MODE this_mode;
   VP9_COMMON *cm = &cpi->common;
   struct macroblock_plane *const p = &x->plane[0];
   struct macroblockd_plane *const pd = &xd->plane[0];
   const int label_count = 4;
   int64_t this_segment_rd = 0;
@@ skipping 51 matching lines @@
       for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
         const struct buf_2d orig_src = x->plane[0].src;
         struct buf_2d orig_pre[2];
 
         mode_idx = INTER_OFFSET(this_mode);
         bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
         if (!(inter_mode_mask & (1 << this_mode)))
           continue;
 
         if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv,
-                                inter_mode_mask,
                                 this_mode, mbmi->ref_frame))
           continue;
 
         vpx_memcpy(orig_pre, pd->pre, sizeof(orig_pre));
         vpx_memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
                    sizeof(bsi->rdstat[i][mode_idx].ta));
         vpx_memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
                    sizeof(bsi->rdstat[i][mode_idx].tl));
 
         // motion search for newmv (single predictor case only)
@@ skipping 349 matching lines @@
                          const int64_t tx_size_diff[TX_MODES],
                          int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS],
                          int skippable) {
   MACROBLOCKD *const xd = &x->e_mbd;
 
   // Take a snapshot of the coding context so it can be
   // restored if we decide to encode this way
   ctx->skip = x->skip;
   ctx->skippable = skippable;
   ctx->best_mode_index = mode_index;
-  ctx->mic = *xd->mi[0];
+  ctx->mic = *xd->mi[0].src_mi;
   ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
   ctx->comp_pred_diff   = (int)comp_pred_diff[COMPOUND_REFERENCE];
   ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
 
   vpx_memcpy(ctx->tx_rd_diff, tx_size_diff, sizeof(ctx->tx_rd_diff));
   vpx_memcpy(ctx->best_filter_diff, best_filter_diff,
              sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
 }
 
 static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
                                const TileInfo *const tile,
                                MV_REFERENCE_FRAME ref_frame,
                                BLOCK_SIZE block_size,
                                int mi_row, int mi_col,
                                int_mv frame_nearest_mv[MAX_REF_FRAMES],
                                int_mv frame_near_mv[MAX_REF_FRAMES],
                                struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
   const VP9_COMMON *cm = &cpi->common;
   const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
   MACROBLOCKD *const xd = &x->e_mbd;
-  MODE_INFO *const mi = xd->mi[0];
+  MODE_INFO *const mi = xd->mi[0].src_mi;
   int_mv *const candidates = mi->mbmi.ref_mvs[ref_frame];
   const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
 
   // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
   // use the UV scaling factors.
   vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
 
   // Gets an initial list of candidate vectors from neighbours and orders them
   vp9_find_mv_refs(cm, xd, tile, mi, ref_frame, candidates, mi_row, mi_col);
 
   // Candidate refinement carried out at encoder and decoder
   vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
                         &frame_nearest_mv[ref_frame],
                         &frame_near_mv[ref_frame]);
 
   // Further refinement that is encode side only to test the top few candidates
   // in full and choose the best as the centre point for subsequent searches.
   // The current implementation doesn't support scaling.
   if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
     vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
                 ref_frame, block_size);
 }
 
 static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                  BLOCK_SIZE bsize,
                                  int mi_row, int mi_col,
                                  int_mv *tmp_mv, int *rate_mv) {
   MACROBLOCKD *xd = &x->e_mbd;
   const VP9_COMMON *cm = &cpi->common;
-  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
   struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
   int bestsme = INT_MAX;
   int step_param;
   int sadpb = x->sadperbit16;
   MV mvp_full;
   int ref = mbmi->ref_frame[0];
   MV ref_mv = mbmi->ref_mvs[ref][0].as_mv;
 
   int tmp_col_min = x->mv_col_min;
   int tmp_col_max = x->mv_col_max;
@@ skipping 106 matching lines @@
 
 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                 BLOCK_SIZE bsize,
                                 int_mv *frame_mv,
                                 int mi_row, int mi_col,
                                 int_mv single_newmv[MAX_REF_FRAMES],
                                 int *rate_mv) {
   const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
   const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
   MACROBLOCKD *xd = &x->e_mbd;
-  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
   const int refs[2] = { mbmi->ref_frame[0],
                         mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1] };
   int_mv ref_mv[2];
   int ite, ref;
   // Prediction buffer from second frame.
   uint8_t *second_pred = vpx_memalign(16, pw * ph * sizeof(uint8_t));
   const InterpKernel *kernel = vp9_get_interp_kernel(mbmi->interp_filter);
 
   // Do joint motion search in compound mode to get more accurate mv.
   struct buf_2d backup_yv12[2][MAX_MB_PLANE];
@@ skipping 221 matching lines @@
                                  int_mv (*mode_mv)[MAX_REF_FRAMES],
                                  int mi_row, int mi_col,
                                  int_mv single_newmv[MAX_REF_FRAMES],
                                  INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
                                  int (*single_skippable)[MAX_REF_FRAMES],
                                  int64_t *psse,
                                  const int64_t ref_best_rd) {
   VP9_COMMON *cm = &cpi->common;
   RD_OPT *rd_opt = &cpi->rd;
   MACROBLOCKD *xd = &x->e_mbd;
-  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
   const int is_comp_pred = has_second_ref(mbmi);
   const int this_mode = mbmi->mode;
   int_mv *frame_mv = mode_mv[this_mode];
   int i;
   int refs[2] = { mbmi->ref_frame[0],
     (mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) };
   int_mv cur_mv[2];
   int64_t this_rd = 0;
   DECLARE_ALIGNED_ARRAY(16, uint8_t, tmp_buf, MAX_MB_PLANE * 64 * 64);
   int pred_exists = 0;
   int intpel_mv;
   int64_t rd, tmp_rd, best_rd = INT64_MAX;
   int best_needs_copy = 0;
   uint8_t *orig_dst[MAX_MB_PLANE];
   int orig_dst_stride[MAX_MB_PLANE];
   int rs = 0;
   INTERP_FILTER best_filter = SWITCHABLE;
   uint8_t skip_txfm[MAX_MB_PLANE << 2] = {0};
   int64_t bsse[MAX_MB_PLANE << 2] = {0};
 
   int bsl = mi_width_log2_lookup[bsize];
   int pred_filter_search = cpi->sf.cb_pred_filter_search ?
       (((mi_row + mi_col) >> bsl) +
        get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
 
   if (pred_filter_search) {
     INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
     if (xd->up_available)
-      af = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
+      af = xd->mi[-xd->mi_stride].src_mi->mbmi.interp_filter;
     if (xd->left_available)
-      lf = xd->mi[-1]->mbmi.interp_filter;
+      lf = xd->mi[-1].src_mi->mbmi.interp_filter;
 
     if ((this_mode != NEWMV) || (af == lf))
       best_filter = af;
   }
 
   if (is_comp_pred) {
     if (frame_mv[refs[0]].as_int == INVALID_MV ||
         frame_mv[refs[1]].as_int == INVALID_MV)
       return INT64_MAX;
 
@@ skipping 24 matching lines @@
       }
       *rate2 += rate_mv;
     } else {
       int_mv tmp_mv;
       single_motion_search(cpi, x, bsize, mi_row, mi_col,
                            &tmp_mv, &rate_mv);
       if (tmp_mv.as_int == INVALID_MV)
         return INT64_MAX;
       *rate2 += rate_mv;
       frame_mv[refs[0]].as_int =
-          xd->mi[0]->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
+          xd->mi[0].src_mi->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
       single_newmv[refs[0]].as_int = tmp_mv.as_int;
     }
   }
 
   for (i = 0; i < is_comp_pred + 1; ++i) {
     cur_mv[i] = frame_mv[refs[i]];
     // Clip "next_nearest" so that it does not extend to far out of image
     if (this_mode != NEWMV)
       clamp_mv2(&cur_mv[i].as_mv, xd);
 
@@ skipping 231 matching lines @@
                                PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCKD *const xd = &x->e_mbd;
   struct macroblockd_plane *const pd = xd->plane;
   int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
   int y_skip = 0, uv_skip = 0;
   int64_t dist_y = 0, dist_uv = 0, tx_cache[TX_MODES] = { 0 };
   TX_SIZE max_uv_tx_size;
   x->skip_encode = 0;
   ctx->skip = 0;
-  xd->mi[0]->mbmi.ref_frame[0] = INTRA_FRAME;
+  xd->mi[0].src_mi->mbmi.ref_frame[0] = INTRA_FRAME;
 
   if (bsize >= BLOCK_8X8) {
     if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly,
                                &dist_y, &y_skip, bsize, tx_cache,
                                best_rd) >= best_rd) {
       *returnrate = INT_MAX;
       return;
     }
-    max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->mbmi.tx_size, bsize,
+    max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0].src_mi->mbmi.tx_size, bsize,
                                          pd[1].subsampling_x,
                                          pd[1].subsampling_y);
     rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
                             &dist_uv, &uv_skip, bsize, max_uv_tx_size);
   } else {
     y_skip = 0;
     if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
                                      &dist_y, best_rd) >= best_rd) {
       *returnrate = INT_MAX;
       return;
     }
-    max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->mbmi.tx_size, bsize,
+    max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0].src_mi->mbmi.tx_size, bsize,
                                          pd[1].subsampling_x,
                                          pd[1].subsampling_y);
     rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
                             &dist_uv, &uv_skip, BLOCK_8X8, max_uv_tx_size);
   }
 
   if (y_skip && uv_skip) {
     *returnrate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
                   vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
     *returndist = dist_y + dist_uv;
     vp9_zero(ctx->tx_rd_diff);
   } else {
     int i;
     *returnrate = rate_y + rate_uv + vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
     *returndist = dist_y + dist_uv;
     if (cpi->sf.tx_size_search_method == USE_FULL_RD)
       for (i = 0; i < TX_MODES; i++) {
         if (tx_cache[i] < INT64_MAX && tx_cache[cm->tx_mode] < INT64_MAX)
           ctx->tx_rd_diff[i] = tx_cache[i] - tx_cache[cm->tx_mode];
         else
           ctx->tx_rd_diff[i] = 0;
       }
   }
 
-  ctx->mic = *xd->mi[0];
+  ctx->mic = *xd->mi[0].src_mi;
 }
 
 // Updating rd_thresh_freq_fact[] here means that the different
 // partition/block sizes are handled independently based on the best
 // choice for the current partition. It may well be better to keep a scaled
 // best rd so far value and update rd_thresh_freq_fact based on the mode/size
 // combination that wins out.
 static void update_rd_thresh_fact(VP9_COMP *cpi, int bsize,
                                   int best_mode_index) {
   if (cpi->sf.adaptive_rd_thresh > 0) {
@@ skipping 16 matching lines @@
                                   const TileInfo *const tile,
                                   int mi_row, int mi_col,
                                   int *returnrate,
                                   int64_t *returndistortion,
                                   BLOCK_SIZE bsize,
                                   PICK_MODE_CONTEXT *ctx,
                                   int64_t best_rd_so_far) {
   VP9_COMMON *const cm = &cpi->common;
   RD_OPT *const rd_opt = &cpi->rd;
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
   const struct segmentation *const seg = &cm->seg;
   struct macroblockd_plane *const pd = xd->plane;
   PREDICTION_MODE this_mode;
   MV_REFERENCE_FRAME ref_frame, second_ref_frame;
   unsigned char segment_id = mbmi->segment_id;
   int comp_pred, i, k;
   int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
   int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
   INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
   int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
   static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
                                     VP9_ALT_FLAG };
   int64_t best_rd = best_rd_so_far;
   int64_t best_tx_rd[TX_MODES];
   int64_t best_tx_diff[TX_MODES];
   int64_t best_pred_diff[REFERENCE_MODES];
   int64_t best_pred_rd[REFERENCE_MODES];
   int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
   MB_MODE_INFO best_mbmode;
   int best_mode_skippable = 0;
-  int mode_index, best_mode_index = -1;
+  int midx, best_mode_index = -1;
   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
   vp9_prob comp_mode_p;
   int64_t best_intra_rd = INT64_MAX;
-  int64_t best_inter_rd = INT64_MAX;
+  unsigned int best_pred_sse = UINT_MAX;
   PREDICTION_MODE best_intra_mode = DC_PRED;
-  MV_REFERENCE_FRAME best_inter_ref_frame = LAST_FRAME;
   int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
   int64_t dist_uv[TX_SIZES];
   int skip_uv[TX_SIZES];
   PREDICTION_MODE mode_uv[TX_SIZES];
-  int intra_cost_penalty = 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q);
+  const int intra_cost_penalty =
+      20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
   int best_skip2 = 0;
-  int mode_skip_mask = 0;
+  uint8_t ref_frame_skip_mask[2] = { 0 };
+  uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
   int mode_skip_start = cpi->sf.mode_skip_start + 1;
   const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
   const int *const rd_thresh_freq_fact = rd_opt->thresh_freq_fact[bsize];
+  int mode_threshold[MAX_MODES];
+  int *mode_map = rd_opt->mode_map[bsize];
   const int mode_search_skip_flags = cpi->sf.mode_search_skip_flags;
-  const int intra_y_mode_mask =
-      cpi->sf.intra_y_mode_mask[max_txsize_lookup[bsize]];
-  int inter_mode_mask = cpi->sf.inter_mode_mask[bsize];
   vp9_zero(best_mbmode);
+
   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
 
   estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
                            &comp_mode_p);
 
   for (i = 0; i < REFERENCE_MODES; ++i)
     best_pred_rd[i] = INT64_MAX;
   for (i = 0; i < TX_MODES; i++)
     best_tx_rd[i] = INT64_MAX;
   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
@@ skipping 15 matching lines @@
     x->pred_mv_sad[ref_frame] = INT_MAX;
     if (cpi->ref_frame_flags & flag_list[ref_frame]) {
       setup_buffer_inter(cpi, x, tile, ref_frame, bsize, mi_row, mi_col,
                          frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
     }
     frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
     frame_mv[ZEROMV][ref_frame].as_int = 0;
   }
 
   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
-    // All modes from vp9_mode_order that use this frame as any ref
-    static const int ref_frame_mask_all[] = {
-        0x0, 0x123291, 0x25c444, 0x39b722
-    };
-    // Fixed mv modes (NEARESTMV, NEARMV, ZEROMV) from vp9_mode_order that use
-    // this frame as their primary ref
-    static const int ref_frame_mask_fixedmv[] = {
-        0x0, 0x121281, 0x24c404, 0x080102
-    };
     if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
-      // Skip modes for missing references
-      mode_skip_mask |= ref_frame_mask_all[ref_frame];
+      // Skip checking missing references in both single and compound reference
+      // modes. Note that a mode will be skipped iff both reference frames
+      // are masked out.
+      ref_frame_skip_mask[0] |= (1 << ref_frame);
+      ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
     } else if (cpi->sf.reference_masking) {
       for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
         // Skip fixed mv modes for poor references
         if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
-          mode_skip_mask |= ref_frame_mask_fixedmv[ref_frame];
+          mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
           break;
         }
       }
     }
     // If the segment reference frame feature is enabled....
     // then do nothing if the current ref frame is not allowed..
     if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
         vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
-      mode_skip_mask |= ref_frame_mask_all[ref_frame];
+      ref_frame_skip_mask[0] |= (1 << ref_frame);
+      ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
     }
   }
 
   // Disable this drop out case if the ref frame
   // segment level feature is enabled for this segment. This is to
   // prevent the possibility that we end up unable to pick any mode.
   if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
     // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
     // unless ARNR filtering is enabled in which case we want
     // an unfiltered alternative. We allow near/nearest as well
     // because they may result in zero-zero MVs but be cheaper.
     if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
-      mode_skip_mask =
-          ~((1 << THR_NEARESTA) | (1 << THR_NEARA) | (1 << THR_ZEROA));
+      ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
+      ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
+      mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
       if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
-        mode_skip_mask |= (1 << THR_NEARA);
+        mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
       if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
-        mode_skip_mask |= (1 << THR_NEARESTA);
+        mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
+    }
+  }
+
+  if (cpi->rc.is_src_frame_alt_ref) {
+    if (cpi->sf.alt_ref_search_fp) {
+      mode_skip_mask[ALTREF_FRAME] = 0;
+      ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
+      ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
     }
   }
 
   if (bsize > cpi->sf.max_intra_bsize) {
-    const int all_intra_modes = (1 << THR_DC) | (1 << THR_TM) |
-        (1 << THR_H_PRED) | (1 << THR_V_PRED) | (1 << THR_D135_PRED) |
-        (1 << THR_D207_PRED) | (1 << THR_D153_PRED) | (1 << THR_D63_PRED) |
-        (1 << THR_D117_PRED) | (1 << THR_D45_PRED);
-    mode_skip_mask |= all_intra_modes;
+    ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
+    ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
   }
 
-  for (mode_index = 0; mode_index < MAX_MODES; ++mode_index) {
+  mode_skip_mask[INTRA_FRAME] |=
+      ~(cpi->sf.intra_y_mode_mask[max_txsize_lookup[bsize]]);
+
+  for (i = 0; i < MAX_MODES; ++i)
+    mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
+
+  midx =  cpi->sf.schedule_mode_search ? mode_skip_start : 0;
+  while (midx > 4) {
+    uint8_t end_pos = 0;
+    for (i = 5; i < midx; ++i) {
+      if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) {
+        uint8_t tmp = mode_map[i];
+        mode_map[i] = mode_map[i - 1];
+        mode_map[i - 1] = tmp;
+        end_pos = i;
+      }
+    }
+    midx = end_pos;
+  }
+
+  for (midx = 0; midx < MAX_MODES; ++midx) {
+    int mode_index = mode_map[midx];
     int mode_excluded = 0;
     int64_t this_rd = INT64_MAX;
     int disable_skip = 0;
     int compmode_cost = 0;
     int rate2 = 0, rate_y = 0, rate_uv = 0;
     int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
     int skippable = 0;
     int64_t tx_cache[TX_MODES];
-    int i;
     int this_skip2 = 0;
     int64_t total_sse = INT64_MAX;
     int early_term = 0;
 
     this_mode = vp9_mode_order[mode_index].mode;
     ref_frame = vp9_mode_order[mode_index].ref_frame[0];
-    if (ref_frame != INTRA_FRAME && !(inter_mode_mask & (1 << this_mode)))
-      continue;
     second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
 
     // Look at the reference frame of the best mode so far and set the
     // skip mask to look at a subset of the remaining modes.
-    if (mode_index == mode_skip_start && best_mode_index >= 0) {
-      switch (vp9_mode_order[best_mode_index].ref_frame[0]) {
+    if (midx == mode_skip_start && best_mode_index >= 0) {
+      switch (best_mbmode.ref_frame[0]) {
         case INTRA_FRAME:
           break;
         case LAST_FRAME:
-          mode_skip_mask |= LAST_FRAME_MODE_MASK;
+          ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
+          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
           break;
         case GOLDEN_FRAME:
-          mode_skip_mask |= GOLDEN_FRAME_MODE_MASK;
+          ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
+          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
           break;
         case ALTREF_FRAME:
-          mode_skip_mask |= ALT_REF_MODE_MASK;
+          ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK;
           break;
         case NONE:
         case MAX_REF_FRAMES:
           assert(0 && "Invalid Reference frame");
           break;
       }
     }
 
-    if (cpi->sf.alt_ref_search_fp && cpi->rc.is_src_frame_alt_ref) {
-      mode_skip_mask = 0;
-      if (!(ref_frame == ALTREF_FRAME && second_ref_frame == NONE))
-        continue;
-    }
+    if (ref_frame_skip_mask[0] & (1 << ref_frame) &&
+        ref_frame_skip_mask[1] & (1 << MAX(0, second_ref_frame)))
+      continue;
 
-    if (mode_skip_mask & (1 << mode_index))
+    if (mode_skip_mask[ref_frame] & (1 << this_mode))
       continue;
 
     // Test best rd so far against threshold for trying this mode.
-    if (rd_less_than_thresh(best_rd, rd_threshes[mode_index],
-                            rd_thresh_freq_fact[mode_index]))
+    if (best_mode_skippable && cpi->sf.schedule_mode_search)
+      mode_threshold[mode_index] <<= 1;
+
+    if (best_rd < mode_threshold[mode_index])
       continue;
 
     if (cpi->sf.motion_field_mode_search) {
       const int mi_width  = MIN(num_8x8_blocks_wide_lookup[bsize],
                                 tile->mi_col_end - mi_col);
       const int mi_height = MIN(num_8x8_blocks_high_lookup[bsize],
                                 tile->mi_row_end - mi_row);
       const int bsl = mi_width_log2(bsize);
       int cb_partition_search_ctrl = (((mi_row + mi_col) >> bsl)
           + get_chessboard_index(cm->current_video_frame)) & 0x1;
       MB_MODE_INFO *ref_mbmi;
       int const_motion = 1;
       int skip_ref_frame = !cb_partition_search_ctrl;
       MV_REFERENCE_FRAME rf = NONE;
       int_mv ref_mv;
       ref_mv.as_int = INVALID_MV;
 
       if ((mi_row - 1) >= tile->mi_row_start) {
-        ref_mv = xd->mi[-xd->mi_stride]->mbmi.mv[0];
-        rf = xd->mi[-xd->mi_stride]->mbmi.ref_frame[0];
+        ref_mv = xd->mi[-xd->mi_stride].src_mi->mbmi.mv[0];
+        rf = xd->mi[-xd->mi_stride].src_mi->mbmi.ref_frame[0];
         for (i = 0; i < mi_width; ++i) {
-          ref_mbmi = &xd->mi[-xd->mi_stride + i]->mbmi;
+          ref_mbmi = &xd->mi[-xd->mi_stride + i].src_mi->mbmi;
           const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
                           (ref_frame == ref_mbmi->ref_frame[0]);
           skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
         }
       }
 
       if ((mi_col - 1) >= tile->mi_col_start) {
         if (ref_mv.as_int == INVALID_MV)
-          ref_mv = xd->mi[-1]->mbmi.mv[0];
+          ref_mv = xd->mi[-1].src_mi->mbmi.mv[0];
         if (rf == NONE)
-          rf = xd->mi[-1]->mbmi.ref_frame[0];
+          rf = xd->mi[-1].src_mi->mbmi.ref_frame[0];
         for (i = 0; i < mi_height; ++i) {
-          ref_mbmi = &xd->mi[i * xd->mi_stride - 1]->mbmi;
+          ref_mbmi = &xd->mi[i * xd->mi_stride - 1].src_mi->mbmi;
           const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
                           (ref_frame == ref_mbmi->ref_frame[0]);
           skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
         }
       }
 
       if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
         if (rf > INTRA_FRAME)
           if (ref_frame != rf)
             continue;
 
       if (const_motion)
         if (this_mode == NEARMV || this_mode == ZEROMV)
           continue;
     }
 
     comp_pred = second_ref_frame > INTRA_FRAME;
     if (comp_pred) {
       if (!cm->allow_comp_inter_inter)
         continue;
 
+      // Skip compound inter modes if ARF is not available.
+      if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
+        continue;
+
+      // Do not allow compound prediction if the segment level reference frame
+      // feature is in use as in this case there can only be one reference.
+      if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
+        continue;
+
       if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
-          best_mode_index >=0 &&
-          vp9_mode_order[best_mode_index].ref_frame[0] == INTRA_FRAME)
+          best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
         continue;
-      if ((mode_search_skip_flags & FLAG_SKIP_COMP_REFMISMATCH) &&
-          ref_frame != best_inter_ref_frame &&
-          second_ref_frame != best_inter_ref_frame)
-        continue;
+
       mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
     } else {
       if (ref_frame != INTRA_FRAME)
         mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
     }
 
     if (ref_frame == INTRA_FRAME) {
       if (cpi->sf.adaptive_mode_search)
-        if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_intra_rd)
+        if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
           continue;
 
-      if (!(intra_y_mode_mask & (1 << this_mode)))
-        continue;
       if (this_mode != DC_PRED) {
         // Disable intra modes other than DC_PRED for blocks with low variance
         // Threshold for intra skipping based on source variance
         // TODO(debargha): Specialize the threshold for super block sizes
         const unsigned int skip_intra_var_thresh = 64;
         if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
             x->source_variance < skip_intra_var_thresh)
           continue;
         // Only search the oblique modes if the best so far is
         // one of the neighboring directional modes
         if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
             (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
           if (best_mode_index >= 0 &&
-              vp9_mode_order[best_mode_index].ref_frame[0] > INTRA_FRAME)
+              best_mbmode.ref_frame[0] > INTRA_FRAME)
             continue;
         }
         if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
           if (conditional_skipintra(this_mode, best_intra_mode))
               continue;
         }
       }
     } else {
       const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame};
       if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv,
-                              inter_mode_mask, this_mode, ref_frames))
+                              this_mode, ref_frames))
         continue;
     }
 
     mbmi->mode = this_mode;
     mbmi->uv_mode = DC_PRED;
     mbmi->ref_frame[0] = ref_frame;
     mbmi->ref_frame[1] = second_ref_frame;
     // Evaluate all sub-pel filters irrespective of whether we can use
     // them for this frame.
     mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
@@ skipping 60 matching lines @@
     // Estimate the reference frame signaling cost and add it
     // to the rolling cost variable.
     if (comp_pred) {
       rate2 += ref_costs_comp[ref_frame];
     } else {
       rate2 += ref_costs_single[ref_frame];
     }
 
     if (!disable_skip) {
       if (skippable) {
-        vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
-
         // Back out the coefficient coding costs
         rate2 -= (rate_y + rate_uv);
-        // for best yrd calculation
-        rate_uv = 0;
 
         // Cost the skip mb case
-        if (skip_prob) {
-          int prob_skip_cost = vp9_cost_bit(skip_prob, 1);
-          rate2 += prob_skip_cost;
-        }
+        rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
       } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
         if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
             RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
           // Add in the cost of the no skip flag.
           rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
         } else {
           // FIXME(rbultje) make this work for splitmv also
           rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
           distortion2 = total_sse;
           assert(total_sse >= 0);
           rate2 -= (rate_y + rate_uv);
-          rate_y = 0;
-          rate_uv = 0;
           this_skip2 = 1;
         }
       } else {
         // Add in the cost of the no skip flag.
         rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
       }
 
       // Calculate the final RD estimate for this mode.
       this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
     }
 
     if (ref_frame == INTRA_FRAME) {
     // Keep record of best intra rd
       if (this_rd < best_intra_rd) {
         best_intra_rd = this_rd;
         best_intra_mode = mbmi->mode;
       }
-    } else {
-      // Keep record of best inter rd with single reference
-      if (!comp_pred && !mode_excluded && this_rd < best_inter_rd) {
-        best_inter_rd = this_rd;
-        best_inter_ref_frame = ref_frame;
-      }
     }
 
     if (!disable_skip && ref_frame == INTRA_FRAME) {
       for (i = 0; i < REFERENCE_MODES; ++i)
         best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
         best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
     }
 
     // Did this mode help.. i.e. is it the new best mode
     if (this_rd < best_rd || x->skip) {
       int max_plane = MAX_MB_PLANE;
       if (!mode_excluded) {
         // Note index of best mode so far
         best_mode_index = mode_index;
 
         if (ref_frame == INTRA_FRAME) {
           /* required for left and above block mv */
           mbmi->mv[0].as_int = 0;
           max_plane = 1;
         } else {
-          best_intra_rd = x->pred_sse[ref_frame];
+          best_pred_sse = x->pred_sse[ref_frame];
         }
 
         *returnrate = rate2;
         *returndistortion = distortion2;
         best_rd = this_rd;
         best_mbmode = *mbmi;
         best_skip2 = this_skip2;
         best_mode_skippable = skippable;
 
         if (!x->select_tx_size)
@@ skipping 29 matching lines @@
         hybrid_rate = rate2;
       } else {
         single_rate = rate2;
         hybrid_rate = rate2 + compmode_cost;
       }
 
       single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
       hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
 
       if (!comp_pred) {
-        if (single_rd < best_pred_rd[SINGLE_REFERENCE]) {
+        if (single_rd < best_pred_rd[SINGLE_REFERENCE])
           best_pred_rd[SINGLE_REFERENCE] = single_rd;
-        }
       } else {
-        if (single_rd < best_pred_rd[COMPOUND_REFERENCE]) {
+        if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
           best_pred_rd[COMPOUND_REFERENCE] = single_rd;
-        }
       }
       if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
         best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
 
       /* keep record of best filter type */
       if (!mode_excluded && cm->interp_filter != BILINEAR) {
         int64_t ref = rd_opt->filter_cache[cm->interp_filter == SWITCHABLE ?
                               SWITCHABLE_FILTERS : cm->interp_filter];
 
         for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
@@ skipping 60 matching lines @@
         ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) || !comp_pred_mode))
       best_mbmode.mode = ZEROMV;
   }
 
   if (best_mode_index < 0 || best_rd >= best_rd_so_far)
     return INT64_MAX;
 
   // If we used an estimate for the uv intra rd in the loop above...
   if (cpi->sf.use_uv_intra_rd_estimate) {
     // Do Intra UV best rd mode selection if best mode choice above was intra.
-    if (vp9_mode_order[best_mode_index].ref_frame[0] == INTRA_FRAME) {
+    if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
       TX_SIZE uv_tx_size;
       *mbmi = best_mbmode;
       uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
       rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
                               &rate_uv_tokenonly[uv_tx_size],
                               &dist_uv[uv_tx_size],
                               &skip_uv[uv_tx_size],
                               bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
                               uv_tx_size);
     }
   }
 
   assert((cm->interp_filter == SWITCHABLE) ||
          (cm->interp_filter == best_mbmode.interp_filter) ||
          !is_inter_block(&best_mbmode));
 
-  update_rd_thresh_fact(cpi, bsize, best_mode_index);
+  if (!cpi->rc.is_src_frame_alt_ref)
+    update_rd_thresh_fact(cpi, bsize, best_mode_index);
 
   // macroblock modes
   *mbmi = best_mbmode;
   x->skip |= best_skip2;
 
   for (i = 0; i < REFERENCE_MODES; ++i) {
     if (best_pred_rd[i] == INT64_MAX)
       best_pred_diff[i] = INT_MIN;
     else
       best_pred_diff[i] = best_rd - best_pred_rd[i];
@@ skipping 12 matching lines @@
       if (best_tx_rd[i] == INT64_MAX)
         best_tx_diff[i] = 0;
       else
         best_tx_diff[i] = best_rd - best_tx_rd[i];
     }
   } else {
     vp9_zero(best_filter_diff);
     vp9_zero(best_tx_diff);
   }
 
-  set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
+  // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
+  // updating code causes PSNR loss. Need to figure out the confliction.
+  x->skip |= best_mode_skippable;
+
   store_coding_context(x, ctx, best_mode_index, best_pred_diff,
                        best_tx_diff, best_filter_diff, best_mode_skippable);
 
   return best_rd;
 }
 
 int64_t vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi, MACROBLOCK *x,
                                            int *returnrate,
                                            int64_t *returndistortion,
                                            BLOCK_SIZE bsize,
                                            PICK_MODE_CONTEXT *ctx,
                                            int64_t best_rd_so_far) {
   VP9_COMMON *const cm = &cpi->common;
   RD_OPT *const rd_opt = &cpi->rd;
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
   unsigned char segment_id = mbmi->segment_id;
   const int comp_pred = 0;
   int i;
   int64_t best_tx_diff[TX_MODES];
   int64_t best_pred_diff[REFERENCE_MODES];
   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
   vp9_prob comp_mode_p;
   INTERP_FILTER best_filter = SWITCHABLE;
   int64_t this_rd = INT64_MAX;
@@ skipping 86 matching lines @@
                                       const TileInfo *const tile,
                                       int mi_row, int mi_col,
                                       int *returnrate,
                                       int64_t *returndistortion,
                                       BLOCK_SIZE bsize,
                                       PICK_MODE_CONTEXT *ctx,
                                       int64_t best_rd_so_far) {
   VP9_COMMON *const cm = &cpi->common;
   RD_OPT *const rd_opt = &cpi->rd;
   MACROBLOCKD *const xd = &x->e_mbd;
-  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
   const struct segmentation *const seg = &cm->seg;
   MV_REFERENCE_FRAME ref_frame, second_ref_frame;
   unsigned char segment_id = mbmi->segment_id;
   int comp_pred, i;
   int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
   static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
                                     VP9_ALT_FLAG };
   int64_t best_rd = best_rd_so_far;
   int64_t best_yrd = best_rd_so_far;  // FIXME(rbultje) more precise
   static const int64_t best_tx_diff[TX_MODES] = { 0 };
   int64_t best_pred_diff[REFERENCE_MODES];
   int64_t best_pred_rd[REFERENCE_MODES];
   int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
   MB_MODE_INFO best_mbmode;
   int ref_index, best_ref_index = 0;
   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
   vp9_prob comp_mode_p;
-  int64_t best_inter_rd = INT64_MAX;
-  MV_REFERENCE_FRAME best_inter_ref_frame = LAST_FRAME;
   INTERP_FILTER tmp_best_filter = SWITCHABLE;
   int rate_uv_intra, rate_uv_tokenonly;
   int64_t dist_uv;
   int skip_uv;
   PREDICTION_MODE mode_uv = DC_PRED;
-  int intra_cost_penalty = 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q);
+  const int intra_cost_penalty =
+      20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
   int_mv seg_mvs[4][MAX_REF_FRAMES];
   b_mode_info best_bmodes[4];
   int best_skip2 = 0;
-  int mode_skip_mask = 0;
+  int ref_frame_skip_mask[2] = { 0 };
 
   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
   vpx_memset(x->zcoeff_blk[TX_4X4], 0, 4);
   vp9_zero(best_mbmode);
 
   for (i = 0; i < 4; i++) {
     int j;
     for (j = 0; j < MAX_REF_FRAMES; j++)
       seg_mvs[i][j].as_int = INVALID_MV;
   }
 
   estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
                            &comp_mode_p);
 
   for (i = 0; i < REFERENCE_MODES; ++i)
     best_pred_rd[i] = INT64_MAX;
   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
     best_filter_rd[i] = INT64_MAX;
   rate_uv_intra = INT_MAX;
 
   *returnrate = INT_MAX;
 
   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
     if (cpi->ref_frame_flags & flag_list[ref_frame]) {
       setup_buffer_inter(cpi, x, tile,
                              ref_frame, bsize, mi_row, mi_col,
                              frame_mv[NEARESTMV], frame_mv[NEARMV],
                              yv12_mb);
+    } else {
+      ref_frame_skip_mask[0] |= (1 << ref_frame);
+      ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
     }
     frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
     frame_mv[ZEROMV][ref_frame].as_int = 0;
   }
 
   for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
     int mode_excluded = 0;
     int64_t this_rd = INT64_MAX;
     int disable_skip = 0;
     int compmode_cost = 0;
     int rate2 = 0, rate_y = 0, rate_uv = 0;
     int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
     int skippable = 0;
     int i;
     int this_skip2 = 0;
     int64_t total_sse = INT_MAX;
     int early_term = 0;
 
     ref_frame = vp9_ref_order[ref_index].ref_frame[0];
     second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
 
     // Look at the reference frame of the best mode so far and set the
     // skip mask to look at a subset of the remaining modes.
     if (ref_index > 2 && cpi->sf.mode_skip_start < MAX_MODES) {
       if (ref_index == 3) {
-        switch (vp9_ref_order[best_ref_index].ref_frame[0]) {
+        switch (best_mbmode.ref_frame[0]) {
           case INTRA_FRAME:
-            mode_skip_mask = 0;
             break;
           case LAST_FRAME:
-            mode_skip_mask = 0x0010;
+            ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
+            ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
             break;
           case GOLDEN_FRAME:
-            mode_skip_mask = 0x0008;
+            ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
+            ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
             break;
           case ALTREF_FRAME:
-            mode_skip_mask = 0x0000;
+            ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
             break;
           case NONE:
           case MAX_REF_FRAMES:
             assert(0 && "Invalid Reference frame");
             break;
         }
       }
-      if (mode_skip_mask & (1 << ref_index))
-        continue;
     }
 
+    if (ref_frame_skip_mask[0] & (1 << ref_frame) &&
+        ref_frame_skip_mask[1] & (1 << MAX(0, second_ref_frame)))
+      continue;
+
     // Test best rd so far against threshold for trying this mode.
     if (rd_less_than_thresh(best_rd,
                             rd_opt->threshes[segment_id][bsize][ref_index],
                             rd_opt->thresh_freq_fact[bsize][ref_index]))
       continue;
 
-    if (ref_frame > INTRA_FRAME &&
-        !(cpi->ref_frame_flags & flag_list[ref_frame])) {
-      continue;
-    }
-
     comp_pred = second_ref_frame > INTRA_FRAME;
     if (comp_pred) {
       if (!cm->allow_comp_inter_inter)
         continue;
-
       if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
         continue;
       // Do not allow compound prediction if the segment level reference frame
       // feature is in use as in this case there can only be one reference.
       if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
         continue;
+
       if ((cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
-          vp9_ref_order[best_ref_index].ref_frame[0] == INTRA_FRAME)
-        continue;
-      if ((cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_REFMISMATCH) &&
-          ref_frame != best_inter_ref_frame &&
-          second_ref_frame != best_inter_ref_frame)
+          best_mbmode.ref_frame[0] == INTRA_FRAME)
         continue;
     }
 
     // TODO(jingning, jkoleszar): scaling reference frame not supported for
     // sub8x8 blocks.
     if (ref_frame > INTRA_FRAME &&
         vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
       continue;
 
     if (second_ref_frame > INTRA_FRAME &&
         vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf))
       continue;
 
     if (comp_pred)
       mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
     else if (ref_frame != INTRA_FRAME)
       mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
 
     // If the segment reference frame feature is enabled....
     // then do nothing if the current ref frame is not allowed..
     if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
-        vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) !=
-            (int)ref_frame) {
+        vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
       continue;
     // Disable this drop out case if the ref frame
     // segment level feature is enabled for this segment. This is to
     // prevent the possibility that we end up unable to pick any mode.
-    } else if (!vp9_segfeature_active(seg, segment_id,
-                                      SEG_LVL_REF_FRAME)) {
+    } else if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
       // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
       // unless ARNR filtering is enabled in which case we want
       // an unfiltered alternative. We allow near/nearest as well
       // because they may result in zero-zero MVs but be cheaper.
       if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
         continue;
     }
 
     mbmi->tx_size = TX_4X4;
     mbmi->uv_mode = DC_PRED;
@@ skipping 108 matching lines @@
                 (mbmi->interp_filter == cm->interp_filter &&
                  cm->interp_filter != SWITCHABLE)) {
               tmp_best_rdu = tmp_rd;
               tmp_best_rate = rate;
               tmp_best_ratey = rate_y;
               tmp_best_distortion = distortion;
               tmp_best_sse = total_sse;
               tmp_best_skippable = skippable;
               tmp_best_mbmode = *mbmi;
               for (i = 0; i < 4; i++) {
-                tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
+                tmp_best_bmodes[i] = xd->mi[0].src_mi->bmi[i];
                 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
               }
               pred_exists = 1;
               if (switchable_filter_index == 0 &&
                   cpi->sf.use_rd_breakout &&
                   best_rd < INT64_MAX) {
                 if (tmp_best_rdu / 2 > best_rd) {
                   // skip searching the other filters if the first is
                   // already substantially larger than the best so far
                   tmp_best_filter = mbmi->interp_filter;
@@ skipping 23 matching lines @@
         if (tmp_rd == INT64_MAX)
           continue;
       } else {
         total_sse = tmp_best_sse;
         rate = tmp_best_rate;
         rate_y = tmp_best_ratey;
         distortion = tmp_best_distortion;
         skippable = tmp_best_skippable;
         *mbmi = tmp_best_mbmode;
         for (i = 0; i < 4; i++)
-          xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
+          xd->mi[0].src_mi->bmi[i] = tmp_best_bmodes[i];
       }
 
       rate2 += rate;
       distortion2 += distortion;
 
       if (cm->interp_filter == SWITCHABLE)
         rate2 += vp9_get_switchable_rate(cpi);
 
       if (!mode_excluded)
         mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
@@ skipping 53 matching lines @@
         }
       } else {
         // Add in the cost of the no skip flag.
         rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
       }
 
       // Calculate the final RD estimate for this mode.
       this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
     }
 
-    // Keep record of best inter rd with single reference
-    if (is_inter_block(mbmi) &&
-        !has_second_ref(mbmi) &&
-        !mode_excluded &&
-        this_rd < best_inter_rd) {
-      best_inter_rd = this_rd;
-      best_inter_ref_frame = ref_frame;
-    }
-
     if (!disable_skip && ref_frame == INTRA_FRAME) {
       for (i = 0; i < REFERENCE_MODES; ++i)
         best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
         best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
     }
 
     // Did this mode help.. i.e. is it the new best mode
     if (this_rd < best_rd || x->skip) {
       if (!mode_excluded) {
@@ skipping 13 matching lines @@
         best_yrd = best_rd -
                    RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
         best_mbmode = *mbmi;
         best_skip2 = this_skip2;
         if (!x->select_tx_size)
           swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
         vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
                    sizeof(uint8_t) * ctx->num_4x4_blk);
 
         for (i = 0; i < 4; i++)
-          best_bmodes[i] = xd->mi[0]->bmi[i];
+          best_bmodes[i] = xd->mi[0].src_mi->bmi[i];
 
         // TODO(debargha): enhance this test with a better distortion prediction
         // based on qp, activity mask and history
         if ((cpi->sf.mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
             (ref_index > MIN_EARLY_TERM_INDEX)) {
           const int qstep = xd->plane[0].dequant[1];
           // TODO(debargha): Enhance this by specializing for each mode_index
           int scale = 4;
           if (x->source_variance < UINT_MAX) {
             const int var_adjust = (x->source_variance < 16);
@@ skipping 15 matching lines @@
         single_rate = rate2 - compmode_cost;
         hybrid_rate = rate2;
       } else {
         single_rate = rate2;
         hybrid_rate = rate2 + compmode_cost;
       }
 
       single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
       hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
 
-      if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE]) {
+      if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
         best_pred_rd[SINGLE_REFERENCE] = single_rd;
-      } else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE]) {
+      else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
         best_pred_rd[COMPOUND_REFERENCE] = single_rd;
-      }
+
       if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
         best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
     }
 
     /* keep record of best filter type */
     if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
         cm->interp_filter != BILINEAR) {
       int64_t ref = rd_opt->filter_cache[cm->interp_filter == SWITCHABLE ?
                               SWITCHABLE_FILTERS : cm->interp_filter];
       int64_t adj_rd;
@@ skipping 20 matching lines @@
     if (x->skip && !comp_pred)
       break;
   }
 
   if (best_rd >= best_rd_so_far)
     return INT64_MAX;
 
   // If we used an estimate for the uv intra rd in the loop above...
   if (cpi->sf.use_uv_intra_rd_estimate) {
     // Do Intra UV best rd mode selection if best mode choice above was intra.
-    if (vp9_ref_order[best_ref_index].ref_frame[0] == INTRA_FRAME) {
+    if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
       *mbmi = best_mbmode;
       rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra,
                               &rate_uv_tokenonly,
                               &dist_uv,
                               &skip_uv,
                               BLOCK_8X8, TX_4X4);
     }
   }
 
   if (best_rd == INT64_MAX) {
     *returnrate = INT_MAX;
     *returndistortion = INT64_MAX;
     return best_rd;
   }
 
   assert((cm->interp_filter == SWITCHABLE) ||
          (cm->interp_filter == best_mbmode.interp_filter) ||
          !is_inter_block(&best_mbmode));
 
   update_rd_thresh_fact(cpi, bsize, best_ref_index);
 
   // macroblock modes
   *mbmi = best_mbmode;
   x->skip |= best_skip2;
   if (!is_inter_block(&best_mbmode)) {
     for (i = 0; i < 4; i++)
-      xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
+      xd->mi[0].src_mi->bmi[i].as_mode = best_bmodes[i].as_mode;
   } else {
     for (i = 0; i < 4; ++i)
-      vpx_memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
+      vpx_memcpy(&xd->mi[0].src_mi->bmi[i], &best_bmodes[i],
+                 sizeof(b_mode_info));
 
-    mbmi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
-    mbmi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
+    mbmi->mv[0].as_int = xd->mi[0].src_mi->bmi[3].as_mv[0].as_int;
+    mbmi->mv[1].as_int = xd->mi[0].src_mi->bmi[3].as_mv[1].as_int;
   }
 
   for (i = 0; i < REFERENCE_MODES; ++i) {
     if (best_pred_rd[i] == INT64_MAX)
       best_pred_diff[i] = INT_MIN;
     else
       best_pred_diff[i] = best_rd - best_pred_rd[i];
   }
 
   if (!x->skip) {
     for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
       if (best_filter_rd[i] == INT64_MAX)
         best_filter_diff[i] = 0;
       else
         best_filter_diff[i] = best_rd - best_filter_rd[i];
     }
     if (cm->interp_filter == SWITCHABLE)
       assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
   } else {
     vp9_zero(best_filter_diff);
   }
 
-  set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
   store_coding_context(x, ctx, best_ref_index,
                        best_pred_diff, best_tx_diff, best_filter_diff, 0);
 
   return best_rd;
 }
+