Merge "Inter-frame non-RD mode decision"

vp9/encoder/vp9_block.h

@@ -116,6 +116,7 @@ struct macroblock {
   unsigned int source_variance;
   unsigned int pred_sse[MAX_REF_FRAMES];
   int pred_mv_sad[MAX_REF_FRAMES];
+  int mode_sad[MAX_REF_FRAMES][INTER_MODES + 1];
 
   int nmvjointcost[MV_JOINTS];
   int nmvcosts[2][MV_VALS];

vp9/encoder/vp9_encodeframe.c

@@ -34,6 +34,7 @@
 #include "vp9/encoder/vp9_encodemv.h"
 #include "vp9/encoder/vp9_extend.h"
 #include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/encoder/vp9_pickmode.h"
 #include "vp9/encoder/vp9_rdopt.h"
 #include "vp9/encoder/vp9_segmentation.h"
 #include "vp9/encoder/vp9_tokenize.h"
@@ -1038,6 +1039,132 @@ static int sb_has_motion(const VP9_COMMON *cm, MODE_INFO **prev_mi_8x8) {
   return 0;
 }
 
+// TODO(jingning) This currently serves as a test framework for non-RD mode
+// decision. To be continued on optimizing the partition type decisions.
+static void pick_partition_type(VP9_COMP *cpi,
+                                const TileInfo *const tile,
+                                MODE_INFO **mi_8x8, TOKENEXTRA **tp,
+                                int mi_row, int mi_col,
+                                BLOCK_SIZE bsize, int *rate, int64_t *dist,
+                                int do_recon) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->mb;
+  const int mi_stride = cm->mode_info_stride;
+  const int num_8x8_subsize = (num_8x8_blocks_wide_lookup[bsize] >> 1);
+  int i;
+  PARTITION_TYPE partition = PARTITION_NONE;
+  BLOCK_SIZE subsize;
+  BLOCK_SIZE bs_type = mi_8x8[0]->mbmi.sb_type;
+  int sub_rate[4] = {0};
+  int64_t sub_dist[4] = {0};
+  int mi_offset;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  partition = partition_lookup[b_width_log2(bsize)][bs_type];
+  subsize = get_subsize(bsize, partition);
+
+  if (bsize < BLOCK_8X8) {
+    // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
+    // there is nothing to be done.
+    if (x->ab_index != 0) {
+      *rate = 0;
+      *dist = 0;
+      return;
+    }
+  } else {
+    *(get_sb_partitioning(x, bsize)) = subsize;
+  }
+
+  switch (partition) {
+    case PARTITION_NONE:
+      pick_sb_modes(cpi, tile, mi_row, mi_col, rate, dist,
+                    bsize, get_block_context(x, bsize), INT64_MAX);
+      break;
+    case PARTITION_HORZ:
+      *get_sb_index(x, subsize) = 0;
+      pick_sb_modes(cpi, tile, mi_row, mi_col, &sub_rate[0], &sub_dist[0],
+                    subsize, get_block_context(x, subsize), INT64_MAX);
+      if (bsize >= BLOCK_8X8 && mi_row + num_8x8_subsize < cm->mi_rows) {
+        update_state(cpi, get_block_context(x, subsize), subsize, 0);
+        encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
+        *get_sb_index(x, subsize) = 1;
+        pick_sb_modes(cpi, tile, mi_row + num_8x8_subsize, mi_col,
+                      &sub_rate[1], &sub_dist[1], subsize,
+                      get_block_context(x, subsize), INT64_MAX);
+      }
+      *rate = sub_rate[0] + sub_rate[1];
+      *dist = sub_dist[0] + sub_dist[1];
+      break;
+    case PARTITION_VERT:
+      *get_sb_index(x, subsize) = 0;
+      pick_sb_modes(cpi, tile, mi_row, mi_col, &sub_rate[0], &sub_dist[0],
+                    subsize, get_block_context(x, subsize), INT64_MAX);
+      if (bsize >= BLOCK_8X8 && mi_col + num_8x8_subsize < cm->mi_cols) {
+        update_state(cpi, get_block_context(x, subsize), subsize, 0);
+        encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
+        *get_sb_index(x, subsize) = 1;
+        pick_sb_modes(cpi, tile, mi_row, mi_col + num_8x8_subsize,
+                      &sub_rate[1], &sub_dist[1], subsize,
+                      get_block_context(x, subsize), INT64_MAX);
+      }
+      *rate = sub_rate[0] + sub_rate[1];
+      *dist = sub_dist[1] + sub_dist[1];
+      break;
+    case PARTITION_SPLIT:
+      *get_sb_index(x, subsize) = 0;
+      pick_partition_type(cpi, tile, mi_8x8, tp, mi_row, mi_col, subsize,
+                          &sub_rate[0], &sub_dist[0], 0);
+
+      if ((mi_col + num_8x8_subsize) < cm->mi_cols) {
+        *get_sb_index(x, subsize) = 1;
+        pick_partition_type(cpi, tile, mi_8x8 + num_8x8_subsize, tp,
+                            mi_row, mi_col + num_8x8_subsize, subsize,
+                            &sub_rate[1], &sub_dist[1], 0);
+      }
+
+      if ((mi_row + num_8x8_subsize) < cm->mi_rows) {
+        *get_sb_index(x, subsize) = 2;
+        pick_partition_type(cpi, tile, mi_8x8 + num_8x8_subsize * mi_stride, tp,
+                            mi_row + num_8x8_subsize, mi_col, subsize,
+                            &sub_rate[2], &sub_dist[2], 0);
+      }
+
+      if ((mi_col + num_8x8_subsize) < cm->mi_cols &&
+          (mi_row + num_8x8_subsize) < cm->mi_rows) {
+        *get_sb_index(x, subsize) = 3;
+        mi_offset = num_8x8_subsize * mi_stride + num_8x8_subsize;
+        pick_partition_type(cpi, tile, mi_8x8 + mi_offset, tp,
+                            mi_row + num_8x8_subsize, mi_col + num_8x8_subsize,
+                            subsize, &sub_rate[3], &sub_dist[3], 0);
+      }
+
+      for (i = 0; i < 4; ++i) {
+        *rate += sub_rate[i];
+        *dist += sub_dist[i];
+      }
+
+      break;
+    default:
+      assert(0);
+  }
+
+  if (do_recon) {
+    int output_enabled = (bsize == BLOCK_64X64);
+
+    // Check the projected output rate for this SB against it's target
+    // and and if necessary apply a Q delta using segmentation to get
+    // closer to the target.
+    if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) {
+      select_in_frame_q_segment(cpi, mi_row, mi_col,
+                                output_enabled, *rate);
+    }
+
+    encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize);
+  }
+}
+
 static void rd_use_partition(VP9_COMP *cpi,
                              const TileInfo *const tile,
                              MODE_INFO **mi_8x8,
@@ -1875,6 +2002,34 @@ static void rd_pick_reference_frame(VP9_COMP *cpi, const TileInfo *const tile,
   restore_context(cpi, mi_row, mi_col, a, l, sa, sl, BLOCK_64X64);
 }
 
+static void encode_sb_row_rt(VP9_COMP *cpi, const TileInfo *const tile,
+                             int mi_row, TOKENEXTRA **tp) {
+  VP9_COMMON *const cm = &cpi->common;
+  int mi_col;
+
+  cpi->sf.always_this_block_size = BLOCK_8X8;
+
+  // Initialize the left context for the new SB row
+  vpx_memset(&cpi->left_context, 0, sizeof(cpi->left_context));
+  vpx_memset(cpi->left_seg_context, 0, sizeof(cpi->left_seg_context));
+
+  // Code each SB in the row
+  for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
+       mi_col += MI_BLOCK_SIZE) {
+    int dummy_rate;
+    int64_t dummy_dist;
+    const int idx_str = cm->mode_info_stride * mi_row + mi_col;
+    MODE_INFO **mi_8x8 = cm->mi_grid_visible + idx_str;
+
+    vp9_zero(cpi->mb.pred_mv);
+
+    set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
+    set_partitioning(cpi, tile, mi_8x8, mi_row, mi_col);
+    pick_partition_type(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
+                        &dummy_rate, &dummy_dist, 1);
+  }
+}
+
 static void encode_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
                           int mi_row, TOKENEXTRA **tp) {
   VP9_COMMON *const cm = &cpi->common;
@@ -2101,7 +2256,11 @@ static void encode_frame_internal(VP9_COMP *cpi) {
           vp9_tile_init(&tile, cm, tile_row, tile_col);
           for (mi_row = tile.mi_row_start;
                mi_row < tile.mi_row_end; mi_row += 8)
+#if 1
             encode_sb_row(cpi, &tile, mi_row, &tp);
+#else
+            encode_sb_row_rt(cpi, &tile, mi_row, &tp);
+#endif
 
           cpi->tok_count[tile_row][tile_col] = (unsigned int)(tp - tp_old);
           assert(tp - cpi->tok <= get_token_alloc(cm->mb_rows, cm->mb_cols));

vp9/encoder/vp9_pickmode.c

@@ -0,0 +1,230 @@
+/*
+ *  Copyright (c) 2014 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.
+ */
+
+#include <stdio.h>
+#include <math.h>
+#include <limits.h>
+#include <assert.h>
+
+#include "vp9/common/vp9_pragmas.h"
+#include "vp9/encoder/vp9_tokenize.h"
+#include "vp9/encoder/vp9_treewriter.h"
+#include "vp9/encoder/vp9_onyx_int.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/encoder/vp9_encodemb.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_variance.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_rdopt.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp9/common/vp9_systemdependent.h"
+#include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_entropy.h"
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_mvref_common.h"
+#include "vp9/common/vp9_common.h"
+
+static int full_pixel_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
+                                     const TileInfo *const tile,
+                                     BLOCK_SIZE bsize, int mi_row, int mi_col,
+                                     int_mv *tmp_mv, int *rate_mv) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}};
+  int bestsme = INT_MAX;
+  int further_steps, step_param;
+  int sadpb = x->sadperbit16;
+  MV mvp_full;
+  int ref = mbmi->ref_frame[0];
+  int_mv ref_mv = mbmi->ref_mvs[ref][0];
+  int i;
+
+  int tmp_col_min = x->mv_col_min;
+  int tmp_col_max = x->mv_col_max;
+  int tmp_row_min = x->mv_row_min;
+  int tmp_row_max = x->mv_row_max;
+
+  int buf_offset;
+  int stride = xd->plane[0].pre[0].stride;
+
+  YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi, ref);
+
+  if (scaled_ref_frame) {
+    int i;
+    // Swap out the reference frame for a version that's been scaled to
+    // match the resolution of the current frame, allowing the existing
+    // motion search code to be used without additional modifications.
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      backup_yv12[i] = xd->plane[i].pre[0];
+
+    setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
+  }
+
+  vp9_set_mv_search_range(x, &ref_mv.as_mv);
+
+  // TODO(jingning) exploiting adaptive motion search control in non-RD
+  // mode decision too.
+  step_param = 6;
+  further_steps = (cpi->sf.max_step_search_steps - 1) - step_param;
+
+  for (i = LAST_FRAME; i <= ALTREF_FRAME && cpi->common.show_frame; ++i) {
+    if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
+      tmp_mv->as_int = INVALID_MV;
+
+      if (scaled_ref_frame) {
+        int i;
+        for (i = 0; i < MAX_MB_PLANE; i++)
+          xd->plane[i].pre[0] = backup_yv12[i];
+      }
+      return INT_MAX;
+    }
+  }
+
+  mvp_full = mbmi->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
+
+  mvp_full.col >>= 3;
+  mvp_full.row >>= 3;
+
+  bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param,
+                                   sadpb, further_steps, 1,
+                                   &cpi->fn_ptr[bsize],
+                                   &ref_mv.as_mv, tmp_mv);
+
+  x->mv_col_min = tmp_col_min;
+  x->mv_col_max = tmp_col_max;
+  x->mv_row_min = tmp_row_min;
+  x->mv_row_max = tmp_row_max;
+
+  if (scaled_ref_frame) {
+    int i;
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      xd->plane[i].pre[0] = backup_yv12[i];
+  }
+
+  // TODO(jingning) This step can be merged into full pixel search step in the
+  // re-designed log-diamond search
+  buf_offset = tmp_mv->as_mv.row * stride + tmp_mv->as_mv.col;
+
+  // Find sad for current vector.
+  bestsme = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf, x->plane[0].src.stride,
+                                   xd->plane[0].pre[0].buf + buf_offset,
+                                   stride, 0x7fffffff);
+
+  // scale to 1/8 pixel resolution
+  tmp_mv->as_mv.row = tmp_mv->as_mv.row << 3;
+  tmp_mv->as_mv.col = tmp_mv->as_mv.col << 3;
+
+  // calculate the bit cost on motion vector
+  *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv.as_mv,
+                             x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+
+
+  return bestsme;
+}
+
+// TODO(jingning) placeholder for inter-frame non-RD mode decision.
+// this needs various further optimizations. to be continued..
+int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
+                            const TileInfo *const tile,
+                            int mi_row, int mi_col,
+                            int *returnrate,
+                            int64_t *returndistortion,
+                            BLOCK_SIZE bsize,
+                            PICK_MODE_CONTEXT *ctx) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
+  const BLOCK_SIZE block_size = get_plane_block_size(bsize, &xd->plane[0]);
+  MB_PREDICTION_MODE this_mode;
+  MV_REFERENCE_FRAME ref_frame;
+  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 = INT64_MAX;
+  int64_t this_rd;
+
+  x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
+
+  x->skip = 0;
+  if (cpi->active_map_enabled && x->active_ptr[0] == 0)
+    x->skip = 1;
+
+  // initialize mode decisions
+  *returnrate = INT_MAX;
+  vpx_memset(mbmi, 0, sizeof(MB_MODE_INFO));
+  mbmi->sb_type = bsize;
+  mbmi->ref_frame[0] = NONE;
+  mbmi->ref_frame[1] = NONE;
+  mbmi->tx_size = MIN(max_txsize_lookup[bsize],
+                      tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    x->pred_mv_sad[ref_frame] = INT_MAX;
+    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
+      vp9_setup_buffer_inter(cpi, x, tile, get_ref_frame_idx(cpi, ref_frame),
+                             ref_frame, block_size, 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) {
+    int rate_mv = 0;
+
+    if (!(cpi->ref_frame_flags & flag_list[ref_frame]))
+      continue;
+
+    // Select prediction reference frames.
+    xd->plane[0].pre[0] = yv12_mb[ref_frame][0];
+
+
+    x->mode_sad[ref_frame][INTER_OFFSET(NEWMV)] =
+        full_pixel_motion_search(cpi, x, tile, bsize, mi_row, mi_col,
+                                 &frame_mv[NEWMV][ref_frame], &rate_mv);
+
+    if (frame_mv[NEWMV][ref_frame].as_int == INVALID_MV)
+      continue;
+
+    clamp_mv2(&frame_mv[NEARESTMV][ref_frame].as_mv, xd);
+    clamp_mv2(&frame_mv[NEARMV][ref_frame].as_mv, xd);
+
+    for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
+      int rate = x->inter_mode_cost[mbmi->mode_context[ref_frame]]
+                                   [INTER_OFFSET(this_mode)];
+      int64_t dist = x->mode_sad[ref_frame][INTER_OFFSET(this_mode)] *
+                      x->mode_sad[ref_frame][INTER_OFFSET(this_mode)];
+      this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
+
+      if (this_rd < best_rd) {
+        best_rd = this_rd;
+        mbmi->mode = this_mode;
+        mbmi->ref_frame[0] = ref_frame;
+        mbmi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
+      }
+    }
+  }
+
+  // TODO(jingning) sub-pixel motion search, if NEWMV is chosen
+
+  // TODO(jingning) intra prediction search, if the best SAD is above a certain
+  // threshold.
+
+  // store mode decisions
+  ctx->mic = *xd->mi_8x8[0];
+
+  return INT64_MAX;
+}

vp9/encoder/vp9_pickmode.h

@@ -0,0 +1,19 @@
+/*
+ *  Copyright (c) 2014 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.
+ */
+
+#include "vp9/encoder/vp9_onyx_int.h"
+
+int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
+                            const struct TileInfo *const tile,
+                            int mi_row, int mi_col,
+                            int *returnrate,
+                            int64_t *returndistortion,
+                            BLOCK_SIZE bsize,
+                            PICK_MODE_CONTEXT *ctx);

vp9/encoder/vp9_rdopt.c

@@ -37,8 +37,6 @@
 #include "vp9/common/vp9_mvref_common.h"
 #include "vp9/common/vp9_common.h"
 
-#define INVALID_MV 0x80008000
-
 /* Factor to weigh the rate for switchable interp filters */
 #define SWITCHABLE_INTERP_RATE_FACTOR 1
 
@@ -113,14 +111,6 @@ const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
 static int rd_thresh_block_size_factor[BLOCK_SIZES] =
   {2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32};
 
-#define RD_THRESH_MAX_FACT 64
-#define RD_THRESH_INC      1
-#define RD_THRESH_POW      1.25
-#define RD_MULT_EPB_RATIO  64
-
-#define MV_COST_WEIGHT      108
-#define MV_COST_WEIGHT_SUB  120
-
 static int raster_block_offset(BLOCK_SIZE plane_bsize,
                                int raster_block, int stride) {
   const int bw = b_width_log2(plane_bsize);
@@ -2133,8 +2123,10 @@ static void mv_pred(VP9_COMP *cpi, MACROBLOCK *x,
     max_mv = MAX(max_mv,
                  MAX(abs(this_mv.as_mv.row), abs(this_mv.as_mv.col)) >> 3);
     // only need to check zero mv once
-    if (!this_mv.as_int && zero_seen)
+    if (!this_mv.as_int && zero_seen) {
+      x->mode_sad[ref_frame][i] = x->mode_sad[ref_frame][INTER_OFFSET(ZEROMV)];
       continue;
+    }
     zero_seen = zero_seen || !this_mv.as_int;
 
     row_offset = this_mv.as_mv.row >> 3;
@@ -2145,6 +2137,9 @@ static void mv_pred(VP9_COMP *cpi, MACROBLOCK *x,
     this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride,
                                            ref_y_ptr, ref_y_stride,
                                            0x7fffffff);
+    x->mode_sad[ref_frame][i] = this_sad;
+    if (this_mv.as_int == 0)
+      x->mode_sad[ref_frame][INTER_OFFSET(ZEROMV)] = this_sad;
 
     // Note if it is the best so far.
     if (this_sad < best_sad) {
@@ -2153,6 +2148,12 @@ static void mv_pred(VP9_COMP *cpi, MACROBLOCK *x,
     }
   }
 
+  if (!zero_seen)
+    x->mode_sad[ref_frame][INTER_OFFSET(ZEROMV)] =
+        cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride,
+                                    ref_y_buffer, ref_y_stride,
+                                    0x7fffffff);
+
   // Note the index of the mv that worked best in the reference list.
   x->mv_best_ref_index[ref_frame] = best_index;
   x->max_mv_context[ref_frame] = max_mv;
@@ -2312,7 +2313,7 @@ void vp9_setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
             frame_type, block_size);
 }
 
-static YV12_BUFFER_CONFIG *get_scaled_ref_frame(VP9_COMP *cpi, int ref_frame) {
+YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(VP9_COMP *cpi, int ref_frame) {
   YV12_BUFFER_CONFIG *scaled_ref_frame = NULL;
   int fb = get_ref_frame_idx(cpi, ref_frame);
   int fb_scale = get_scale_ref_frame_idx(cpi, ref_frame);
@@ -2350,7 +2351,7 @@ static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
   int tmp_row_min = x->mv_row_min;
   int tmp_row_max = x->mv_row_max;
 
-  YV12_BUFFER_CONFIG *scaled_ref_frame = get_scaled_ref_frame(cpi, ref);
+  YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi, ref);
 
   int_mv pred_mv[3];
   pred_mv[0] = mbmi->ref_mvs[ref][0];
@@ -2498,8 +2499,8 @@ static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
   struct buf_2d scaled_first_yv12 = xd->plane[0].pre[0];
   int last_besterr[2] = {INT_MAX, INT_MAX};
   YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
-    get_scaled_ref_frame(cpi, mbmi->ref_frame[0]),
+    vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[0]),
-    get_scaled_ref_frame(cpi, mbmi->ref_frame[1])
+    vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[1])
   };
 
   for (ref = 0; ref < 2; ++ref) {

vp9/encoder/vp9_rdopt.h

@@ -19,6 +19,16 @@
   (((128 + ((int64_t)R) * (RM)) >> 8) + (D << DM))
 #define QIDX_SKIP_THRESH     115
 
+#define RD_THRESH_MAX_FACT 64
+#define RD_THRESH_INC      1
+#define RD_THRESH_POW      1.25
+#define RD_MULT_EPB_RATIO  64
+
+#define MV_COST_WEIGHT      108
+#define MV_COST_WEIGHT_SUB  120
+
+#define INVALID_MV 0x80008000
+
 struct TileInfo;
 
 int vp9_compute_rd_mult(VP9_COMP *cpi, int qindex);
@@ -36,6 +46,8 @@ void vp9_setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
                             int_mv frame_near_mv[MAX_REF_FRAMES],
                             struct buf_2d yv12_mb[4][MAX_MB_PLANE]);
 
+YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(VP9_COMP *cpi, int ref_frame);
+
 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
                                int *r, int64_t *d, BLOCK_SIZE bsize,
                                PICK_MODE_CONTEXT *ctx, int64_t best_rd);

vp9/vp9cx.mk

@@ -43,6 +43,7 @@ VP9_CX_SRCS-yes += encoder/vp9_psnr.h
 VP9_CX_SRCS-yes += encoder/vp9_quantize.h
 VP9_CX_SRCS-yes += encoder/vp9_ratectrl.h
 VP9_CX_SRCS-yes += encoder/vp9_rdopt.h
+VP9_CX_SRCS-yes += encoder/vp9_pickmode.h
 VP9_CX_SRCS-yes += encoder/vp9_sadmxn.h
 VP9_CX_SRCS-yes += encoder/vp9_tokenize.h
 VP9_CX_SRCS-yes += encoder/vp9_treewriter.h
@@ -55,6 +56,7 @@ VP9_CX_SRCS-yes += encoder/vp9_psnr.c
 VP9_CX_SRCS-yes += encoder/vp9_quantize.c
 VP9_CX_SRCS-yes += encoder/vp9_ratectrl.c
 VP9_CX_SRCS-yes += encoder/vp9_rdopt.c
+VP9_CX_SRCS-yes += encoder/vp9_pickmode.c
 VP9_CX_SRCS-yes += encoder/vp9_sad_c.c
 VP9_CX_SRCS-yes += encoder/vp9_segmentation.c
 VP9_CX_SRCS-yes += encoder/vp9_segmentation.h