libvpx: Pull from upstream

source/libvpx/vp9/encoder/vp9_encodeframe.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.
  */
 
 #include <limits.h>
 #include <math.h>
 #include <stdio.h>
 
 #include "./vp9_rtcd.h"
 #include "./vpx_config.h"
 
 #include "vpx_ports/vpx_timer.h"
 
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_entropy.h"
 #include "vp9/common/vp9_entropymode.h"
-#include "vp9/common/vp9_extend.h"
 #include "vp9/common/vp9_findnearmv.h"
 #include "vp9/common/vp9_idct.h"
 #include "vp9/common/vp9_mvref_common.h"
 #include "vp9/common/vp9_pred_common.h"
 #include "vp9/common/vp9_quant_common.h"
 #include "vp9/common/vp9_reconintra.h"
 #include "vp9/common/vp9_reconinter.h"
 #include "vp9/common/vp9_seg_common.h"
 #include "vp9/common/vp9_tile_common.h"
 #include "vp9/encoder/vp9_encodeframe.h"
-#include "vp9/encoder/vp9_encodeintra.h"
 #include "vp9/encoder/vp9_encodemb.h"
 #include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_extend.h"
 #include "vp9/encoder/vp9_onyx_int.h"
 #include "vp9/encoder/vp9_rdopt.h"
 #include "vp9/encoder/vp9_segmentation.h"
 #include "vp9/common/vp9_systemdependent.h"
 #include "vp9/encoder/vp9_tokenize.h"
 #include "vp9/encoder/vp9_vaq.h"
 
 
 #define DBG_PRNT_SEGMAP 0
 
 
 // #define ENC_DEBUG
 #ifdef ENC_DEBUG
 int enc_debug = 0;
 #endif
 
-static INLINE uint8_t *get_sb_index(MACROBLOCKD *xd, BLOCK_SIZE subsize) {
+static INLINE uint8_t *get_sb_index(MACROBLOCK *x, BLOCK_SIZE subsize) {
   switch (subsize) {
     case BLOCK_64X64:
     case BLOCK_64X32:
     case BLOCK_32X64:
     case BLOCK_32X32:
-      return &xd->sb_index;
+      return &x->sb_index;
     case BLOCK_32X16:
     case BLOCK_16X32:
     case BLOCK_16X16:
-      return &xd->mb_index;
+      return &x->mb_index;
     case BLOCK_16X8:
     case BLOCK_8X16:
     case BLOCK_8X8:
-      return &xd->b_index;
+      return &x->b_index;
     case BLOCK_8X4:
     case BLOCK_4X8:
     case BLOCK_4X4:
-      return &xd->ab_index;
+      return &x->ab_index;
     default:
       assert(0);
       return NULL;
   }
 }
 
 static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
                               int mi_row, int mi_col, BLOCK_SIZE bsize);
 
 static void adjust_act_zbin(VP9_COMP *cpi, MACROBLOCK *x);
@@ skipping 272 matching lines @@
 
   x->rdmult = (unsigned int) (((int64_t) x->rdmult * b + (a >> 1)) / a);
   x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
   x->errorperbit += (x->errorperbit == 0);
 #endif
 
   // Activity based Zbin adjustment
   adjust_act_zbin(cpi, x);
 }
 
+// Select a segment for the current SB64
+static void select_in_frame_q_segment(VP9_COMP *cpi,
+                                      int mi_row, int mi_col,
+                                      int output_enabled, int projected_rate) {
+  VP9_COMMON * const cm = &cpi->common;
+  int target_rate = cpi->rc.sb64_target_rate << 8;   // convert to bits << 8
+
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = 1 << mi_width_log2(BLOCK_64X64);
+  const int bh = 1 << mi_height_log2(BLOCK_64X64);
+  const int xmis = MIN(cm->mi_cols - mi_col, bw);
+  const int ymis = MIN(cm->mi_rows - mi_row, bh);
+  int complexity_metric = 64;
+  int x, y;
+
+  unsigned char segment;
+
+  if (!output_enabled) {
+    segment = 0;
+  } else {
+    // Rate depends on fraction of a SB64 in frame (xmis * ymis / bw * bh).
+    // It is converted to bits * 256 units
+    target_rate = (cpi->rc.sb64_target_rate * xmis * ymis * 256) / (bw * bh);
+
+    if (projected_rate < (target_rate / 4)) {
+      segment = 2;
+    } else if (projected_rate < (target_rate / 2)) {
+      segment = 1;
+    } else {
+      segment = 0;
+    }
+
+    complexity_metric =
+      clamp((int)((projected_rate * 64) / target_rate), 16, 255);
+  }
+
+  // Fill in the entires in the segment map corresponding to this SB64
+  for (y = 0; y < ymis; y++) {
+    for (x = 0; x < xmis; x++) {
+      cpi->segmentation_map[mi_offset + y * cm->mi_cols + x] = segment;
+      cpi->complexity_map[mi_offset + y * cm->mi_cols + x] =
+        (unsigned char)complexity_metric;
+    }
+  }
+}
+
 static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
                          BLOCK_SIZE bsize, int output_enabled) {
   int i, x_idx, y;
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *const p = x->plane;
+  struct macroblockd_plane *const pd = xd->plane;
   MODE_INFO *mi = &ctx->mic;
   MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
   MODE_INFO *mi_addr = xd->mi_8x8[0];
 
   int mb_mode_index = ctx->best_mode_index;
   const int mis = cm->mode_info_stride;
   const int mi_width = num_8x8_blocks_wide_lookup[bsize];
   const int mi_height = num_8x8_blocks_high_lookup[bsize];
+  int max_plane;
 
   assert(mi->mbmi.mode < MB_MODE_COUNT);
   assert(mi->mbmi.ref_frame[0] < MAX_REF_FRAMES);
   assert(mi->mbmi.ref_frame[1] < MAX_REF_FRAMES);
   assert(mi->mbmi.sb_type == bsize);
 
+  // For in frame adaptive Q copy over the chosen segment id into the
+  // mode innfo context for the chosen mode / partition.
+  if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && output_enabled)
+    mi->mbmi.segment_id = xd->mi_8x8[0]->mbmi.segment_id;
+
   *mi_addr = *mi;
 
+  max_plane = is_inter_block(mbmi) ? MAX_MB_PLANE : 1;
+  for (i = 0; i < max_plane; ++i) {
+    p[i].coeff = ctx->coeff_pbuf[i][1];
+    p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
+    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
+    p[i].eobs = ctx->eobs_pbuf[i][1];
+  }
+
+  for (i = max_plane; i < MAX_MB_PLANE; ++i) {
+    p[i].coeff = ctx->coeff_pbuf[i][2];
+    p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
+    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
+    p[i].eobs = ctx->eobs_pbuf[i][2];
+  }
+
   // Restore the coding context of the MB to that that was in place
   // when the mode was picked for it
   for (y = 0; y < mi_height; y++)
     for (x_idx = 0; x_idx < mi_width; x_idx++)
       if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx
-          && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y)
+        && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
         xd->mi_8x8[x_idx + y * mis] = mi_addr;
+      }
 
-  if (cpi->sf.variance_adaptive_quantization) {
+    if ((cpi->oxcf.aq_mode == VARIANCE_AQ) ||
+        (cpi->oxcf.aq_mode == COMPLEXITY_AQ)) {
     vp9_mb_init_quantizer(cpi, x);
   }
 
   // FIXME(rbultje) I'm pretty sure this should go to the end of this block
   // (i.e. after the output_enabled)
   if (bsize < BLOCK_32X32) {
     if (bsize < BLOCK_16X16)
       ctx->tx_rd_diff[ALLOW_16X16] = ctx->tx_rd_diff[ALLOW_8X8];
     ctx->tx_rd_diff[ALLOW_32X32] = ctx->tx_rd_diff[ALLOW_16X16];
   }
@@ skipping 49 matching lines @@
       mbmi->best_mv[0].as_int = best_mv[0].as_int;
       mbmi->best_mv[1].as_int = best_mv[1].as_int;
       vp9_update_mv_count(cpi, x, best_mv);
     }
 
     if (cm->mcomp_filter_type == SWITCHABLE && is_inter_mode(mbmi->mode)) {
       const int ctx = vp9_get_pred_context_switchable_interp(xd);
       ++cm->counts.switchable_interp[ctx][mbmi->interp_filter];
     }
 
-    cpi->rd_comp_pred_diff[SINGLE_PREDICTION_ONLY] += ctx->single_pred_diff;
-    cpi->rd_comp_pred_diff[COMP_PREDICTION_ONLY] += ctx->comp_pred_diff;
-    cpi->rd_comp_pred_diff[HYBRID_PREDICTION] += ctx->hybrid_pred_diff;
+    cpi->rd_comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
+    cpi->rd_comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
+    cpi->rd_comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
 
     for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
       cpi->rd_filter_diff[i] += ctx->best_filter_diff[i];
   }
 }
 
 void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
                           int mi_row, int mi_col) {
   uint8_t *const buffers[4] = {src->y_buffer, src->u_buffer, src->v_buffer,
                                src->alpha_buffer};
@@ skipping 56 matching lines @@
 
   /* set up source buffers */
   vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
 
   /* R/D setup */
   x->rddiv = cpi->RDDIV;
   x->rdmult = cpi->RDMULT;
 
   /* segment ID */
   if (seg->enabled) {
-    if (!cpi->sf.variance_adaptive_quantization) {
+    if (cpi->oxcf.aq_mode != VARIANCE_AQ) {
       uint8_t *map = seg->update_map ? cpi->segmentation_map
           : cm->last_frame_seg_map;
       mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
     }
     vp9_mb_init_quantizer(cpi, x);
 
     if (seg->enabled && cpi->seg0_cnt > 0
         && !vp9_segfeature_active(seg, 0, SEG_LVL_REF_FRAME)
         && vp9_segfeature_active(seg, 1, SEG_LVL_REF_FRAME)) {
       cpi->seg0_progress = (cpi->seg0_idx << 16) / cpi->seg0_cnt;
@@ skipping 17 matching lines @@
 }
 
 static void pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile,
                           int mi_row, int mi_col,
                           int *totalrate, int64_t *totaldist,
                           BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
                           int64_t best_rd) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *const p = x->plane;
+  struct macroblockd_plane *const pd = xd->plane;
+  int i;
   int orig_rdmult = x->rdmult;
   double rdmult_ratio;
 
   vp9_clear_system_state();  // __asm emms;
   rdmult_ratio = 1.0;  // avoid uninitialized warnings
 
   // Use the lower precision, but faster, 32x32 fdct for mode selection.
   x->use_lp32x32fdct = 1;
 
   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 (xd->ab_index != 0) {
+    if (x->ab_index != 0) {
       *totalrate = 0;
       *totaldist = 0;
       return;
     }
   }
 
   set_offsets(cpi, tile, mi_row, mi_col, bsize);
   xd->mi_8x8[0]->mbmi.sb_type = bsize;
 
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    p[i].coeff = ctx->coeff_pbuf[i][0];
+    p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
+    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
+    p[i].eobs = ctx->eobs_pbuf[i][0];
+  }
+  ctx->is_coded = 0;
+  x->skip_recode = 0;
+
   // Set to zero to make sure we do not use the previous encoded frame stats
   xd->mi_8x8[0]->mbmi.skip_coeff = 0;
 
   x->source_variance = get_sby_perpixel_variance(cpi, x, bsize);
 
-  if (cpi->sf.variance_adaptive_quantization) {
+  if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
     int energy;
     if (bsize <= BLOCK_16X16) {
       energy = x->mb_energy;
     } else {
       energy = vp9_block_energy(cpi, x, bsize);
     }
 
     xd->mi_8x8[0]->mbmi.segment_id = vp9_vaq_segment_id(energy);
     rdmult_ratio = vp9_vaq_rdmult_ratio(energy);
     vp9_mb_init_quantizer(cpi, x);
   }
 
   if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
     vp9_activity_masking(cpi, x);
 
-  if (cpi->sf.variance_adaptive_quantization) {
+  if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
     vp9_clear_system_state();  // __asm emms;
     x->rdmult = round(x->rdmult * rdmult_ratio);
+  } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
+    const int mi_offset = mi_row * cm->mi_cols + mi_col;
+    unsigned char complexity = cpi->complexity_map[mi_offset];
+    const int is_edge = (mi_row == 0) || (mi_row == (cm->mi_rows - 1)) ||
+                        (mi_col == 0) || (mi_col == (cm->mi_cols - 1));
+
+    if (!is_edge && (complexity > 128))
+      x->rdmult = x->rdmult  + ((x->rdmult * (complexity - 128)) / 256);
   }
 
   // Find best coding mode & reconstruct the MB so it is available
   // as a predictor for MBs that follow in the SB
   if (frame_is_intra_only(cm)) {
     vp9_rd_pick_intra_mode_sb(cpi, x, totalrate, totaldist, bsize, ctx,
                               best_rd);
   } else {
     if (bsize >= BLOCK_8X8)
       vp9_rd_pick_inter_mode_sb(cpi, x, tile, mi_row, mi_col,
                                 totalrate, totaldist, bsize, ctx, best_rd);
     else
       vp9_rd_pick_inter_mode_sub8x8(cpi, x, tile, mi_row, mi_col, totalrate,
                                     totaldist, bsize, ctx, best_rd);
   }
 
-  if (cpi->sf.variance_adaptive_quantization) {
+  if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
     x->rdmult = orig_rdmult;
     if (*totalrate != INT_MAX) {
       vp9_clear_system_state();  // __asm emms;
       *totalrate = round(*totalrate * rdmult_ratio);
     }
   }
 }
 
 static void update_stats(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
   MODE_INFO *mi = xd->mi_8x8[0];
   MB_MODE_INFO *const mbmi = &mi->mbmi;
 
   if (!frame_is_intra_only(cm)) {
     const int seg_ref_active = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
                                                      SEG_LVL_REF_FRAME);
 
     if (!seg_ref_active)
-      cpi->intra_inter_count[vp9_get_pred_context_intra_inter(xd)]
+      cpi->intra_inter_count[vp9_get_intra_inter_context(xd)]
                             [is_inter_block(mbmi)]++;
 
     // If the segment reference feature is enabled we have only a single
     // reference frame allowed for the segment so exclude it from
     // the reference frame counts used to work out probabilities.
     if (is_inter_block(mbmi) && !seg_ref_active) {
-      if (cm->comp_pred_mode == HYBRID_PREDICTION)
-        cpi->comp_inter_count[vp9_get_pred_context_comp_inter_inter(cm, xd)]
+      if (cm->reference_mode == REFERENCE_MODE_SELECT)
+        cpi->comp_inter_count[vp9_get_reference_mode_context(cm, xd)]
                              [has_second_ref(mbmi)]++;
 
       if (has_second_ref(mbmi)) {
         cpi->comp_ref_count[vp9_get_pred_context_comp_ref_p(cm, xd)]
                            [mbmi->ref_frame[0] == GOLDEN_FRAME]++;
       } else {
         cpi->single_ref_count[vp9_get_pred_context_single_ref_p1(xd)][0]
                              [mbmi->ref_frame[0] != LAST_FRAME]++;
         if (mbmi->ref_frame[0] != LAST_FRAME)
           cpi->single_ref_count[vp9_get_pred_context_single_ref_p2(xd)][1]
                                [mbmi->ref_frame[0] != GOLDEN_FRAME]++;
       }
     }
-
-    // Count of last ref frame 0,0 usage
-    if (mbmi->mode == ZEROMV && mbmi->ref_frame[0] == LAST_FRAME)
-      cpi->inter_zz_count++;
   }
 }
 
 static BLOCK_SIZE *get_sb_partitioning(MACROBLOCK *x, BLOCK_SIZE bsize) {
-  MACROBLOCKD *const xd = &x->e_mbd;
   switch (bsize) {
     case BLOCK_64X64:
       return &x->sb64_partitioning;
     case BLOCK_32X32:
-      return &x->sb_partitioning[xd->sb_index];
+      return &x->sb_partitioning[x->sb_index];
     case BLOCK_16X16:
-      return &x->mb_partitioning[xd->sb_index][xd->mb_index];
+      return &x->mb_partitioning[x->sb_index][x->mb_index];
     case BLOCK_8X8:
-      return &x->b_partitioning[xd->sb_index][xd->mb_index][xd->b_index];
+      return &x->b_partitioning[x->sb_index][x->mb_index][x->b_index];
     default:
       assert(0);
       return NULL;
   }
 }
 
 static void restore_context(VP9_COMP *cpi, int mi_row, int mi_col,
                             ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
                             ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
                             PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
@@ skipping 52 matching lines @@
   }
   vpx_memcpy(sa, cpi->above_seg_context + mi_col,
              sizeof(*cpi->above_seg_context) * mi_width);
   vpx_memcpy(sl, cpi->left_seg_context + (mi_row & MI_MASK),
              sizeof(cpi->left_seg_context[0]) * mi_height);
 }
 
 static void encode_b(VP9_COMP *cpi, const TileInfo *const tile,
                      TOKENEXTRA **tp, int mi_row, int mi_col,
                      int output_enabled, BLOCK_SIZE bsize, int sub_index) {
-  VP9_COMMON * const cm = &cpi->common;
-  MACROBLOCK * const x = &cpi->mb;
-  MACROBLOCKD * const xd = &x->e_mbd;
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->mb;
 
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
   if (sub_index != -1)
-    *get_sb_index(xd, bsize) = sub_index;
+    *get_sb_index(x, bsize) = sub_index;
 
   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 (xd->ab_index > 0)
+    if (x->ab_index > 0)
       return;
   }
   set_offsets(cpi, tile, mi_row, mi_col, bsize);
   update_state(cpi, get_block_context(x, bsize), bsize, output_enabled);
   encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize);
 
   if (output_enabled) {
     update_stats(cpi);
 
     (*tp)->token = EOSB_TOKEN;
     (*tp)++;
   }
 }
 
 static void encode_sb(VP9_COMP *cpi, const TileInfo *const tile,
                       TOKENEXTRA **tp, int mi_row, int mi_col,
                       int output_enabled, BLOCK_SIZE bsize) {
-  VP9_COMMON * const cm = &cpi->common;
-  MACROBLOCK * const x = &cpi->mb;
-  MACROBLOCKD * const xd = &x->e_mbd;
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->mb;
   BLOCK_SIZE c1 = BLOCK_8X8;
   const int bsl = b_width_log2(bsize), bs = (1 << bsl) / 4;
   int pl = 0;
   PARTITION_TYPE partition;
   BLOCK_SIZE subsize;
   int i;
 
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
   c1 = BLOCK_4X4;
   if (bsize >= BLOCK_8X8) {
     pl = partition_plane_context(cpi->above_seg_context, cpi->left_seg_context,
                                  mi_row, mi_col, bsize);
     c1 = *(get_sb_partitioning(x, bsize));
   }
   partition = partition_lookup[bsl][c1];
 
   switch (partition) {
     case PARTITION_NONE:
       if (output_enabled && bsize >= BLOCK_8X8)
-        cpi->partition_count[pl][PARTITION_NONE]++;
+        cm->counts.partition[pl][PARTITION_NONE]++;
       encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, c1, -1);
       break;
     case PARTITION_VERT:
       if (output_enabled)
-        cpi->partition_count[pl][PARTITION_VERT]++;
+        cm->counts.partition[pl][PARTITION_VERT]++;
       encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, c1, 0);
       encode_b(cpi, tile, tp, mi_row, mi_col + bs, output_enabled, c1, 1);
       break;
     case PARTITION_HORZ:
       if (output_enabled)
-        cpi->partition_count[pl][PARTITION_HORZ]++;
+        cm->counts.partition[pl][PARTITION_HORZ]++;
       encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, c1, 0);
       encode_b(cpi, tile, tp, mi_row + bs, mi_col, output_enabled, c1, 1);
       break;
     case PARTITION_SPLIT:
       subsize = get_subsize(bsize, PARTITION_SPLIT);
 
       if (output_enabled)
-        cpi->partition_count[pl][PARTITION_SPLIT]++;
+        cm->counts.partition[pl][PARTITION_SPLIT]++;
 
       for (i = 0; i < 4; i++) {
         const int x_idx = i & 1, y_idx = i >> 1;
 
-        *get_sb_index(xd, subsize) = i;
+        *get_sb_index(x, subsize) = i;
         encode_sb(cpi, tile, tp, mi_row + y_idx * bs, mi_col + x_idx * bs,
                   output_enabled, subsize);
       }
       break;
     default:
       assert(0);
       break;
   }
 
   if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
@@ skipping 106 matching lines @@
   }
   return 0;
 }
 
 static void rd_use_partition(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;
-  MACROBLOCKD *xd = &cpi->mb.e_mbd;
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->mb;
   const int mis = cm->mode_info_stride;
   int bsl = b_width_log2(bsize);
   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 ms = num_4x4_blocks_wide / 2;
   int mh = num_4x4_blocks_high / 2;
   int bss = (1 << bsl) / 4;
   int i, pl;
   PARTITION_TYPE partition = PARTITION_NONE;
   BLOCK_SIZE subsize;
@@ skipping 14 matching lines @@
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
   partition = partition_lookup[bsl][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 (xd->ab_index != 0) {
+    if (x->ab_index != 0) {
       *rate = 0;
       *dist = 0;
       return;
     }
   } else {
     *(get_sb_partitioning(x, bsize)) = subsize;
   }
   save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
 
   if (bsize == BLOCK_16X16) {
@@ skipping 37 matching lines @@
       *(get_sb_partitioning(x, bsize)) = subsize;
     }
   }
 
   switch (partition) {
     case PARTITION_NONE:
       pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate, &last_part_dist,
                     bsize, get_block_context(x, bsize), INT64_MAX);
       break;
     case PARTITION_HORZ:
-      *get_sb_index(xd, subsize) = 0;
+      *get_sb_index(x, subsize) = 0;
       pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate, &last_part_dist,
                     subsize, get_block_context(x, subsize), INT64_MAX);
       if (last_part_rate != INT_MAX &&
           bsize >= BLOCK_8X8 && mi_row + (mh >> 1) < cm->mi_rows) {
         int rt = 0;
         int64_t dt = 0;
         update_state(cpi, get_block_context(x, subsize), subsize, 0);
         encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
-        *get_sb_index(xd, subsize) = 1;
+        *get_sb_index(x, subsize) = 1;
         pick_sb_modes(cpi, tile, mi_row + (ms >> 1), mi_col, &rt, &dt, subsize,
                       get_block_context(x, subsize), INT64_MAX);
         if (rt == INT_MAX || dt == INT_MAX) {
           last_part_rate = INT_MAX;
           last_part_dist = INT_MAX;
           break;
         }
 
         last_part_rate += rt;
         last_part_dist += dt;
       }
       break;
     case PARTITION_VERT:
-      *get_sb_index(xd, subsize) = 0;
+      *get_sb_index(x, subsize) = 0;
       pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate, &last_part_dist,
                     subsize, get_block_context(x, subsize), INT64_MAX);
       if (last_part_rate != INT_MAX &&
           bsize >= BLOCK_8X8 && mi_col + (ms >> 1) < cm->mi_cols) {
         int rt = 0;
         int64_t dt = 0;
         update_state(cpi, get_block_context(x, subsize), subsize, 0);
         encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
-        *get_sb_index(xd, subsize) = 1;
+        *get_sb_index(x, subsize) = 1;
         pick_sb_modes(cpi, tile, mi_row, mi_col + (ms >> 1), &rt, &dt, subsize,
                       get_block_context(x, subsize), INT64_MAX);
         if (rt == INT_MAX || dt == INT_MAX) {
           last_part_rate = INT_MAX;
           last_part_dist = INT_MAX;
           break;
         }
         last_part_rate += rt;
         last_part_dist += dt;
       }
       break;
     case PARTITION_SPLIT:
       // Split partition.
       last_part_rate = 0;
       last_part_dist = 0;
       for (i = 0; i < 4; i++) {
         int x_idx = (i & 1) * (ms >> 1);
         int y_idx = (i >> 1) * (ms >> 1);
         int jj = i >> 1, ii = i & 0x01;
         int rt;
         int64_t dt;
 
         if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
           continue;
 
-        *get_sb_index(xd, subsize) = i;
+        *get_sb_index(x, subsize) = i;
 
         rd_use_partition(cpi, tile, mi_8x8 + jj * bss * mis + ii * bss, tp,
                          mi_row + y_idx, mi_col + x_idx, subsize, &rt, &dt,
                          i != 3);
         if (rt == INT_MAX || dt == INT_MAX) {
           last_part_rate = INT_MAX;
           last_part_dist = INT_MAX;
           break;
         }
         last_part_rate += rt;
@@ skipping 20 matching lines @@
 
     // Split partition.
     for (i = 0; i < 4; i++) {
       int x_idx = (i & 1) * (num_4x4_blocks_wide >> 2);
       int y_idx = (i >> 1) * (num_4x4_blocks_wide >> 2);
       int rt = 0;
       int64_t dt = 0;
       ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
       PARTITION_CONTEXT sl[8], sa[8];
 
-      if ((mi_row + y_idx >= cm->mi_rows)
-          || (mi_col + x_idx >= cm->mi_cols))
+      if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
         continue;
 
-      *get_sb_index(xd, split_subsize) = i;
+      *get_sb_index(x, split_subsize) = i;
       *get_sb_partitioning(x, bsize) = split_subsize;
       *get_sb_partitioning(x, split_subsize) = split_subsize;
 
       save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
 
       pick_sb_modes(cpi, tile, mi_row + y_idx, mi_col + x_idx, &rt, &dt,
                     split_subsize, get_block_context(x, split_subsize),
                     INT64_MAX);
 
       restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
 
       if (rt == INT_MAX || dt == INT_MAX) {
         split_rate = INT_MAX;
         split_dist = INT_MAX;
         break;
       }
 
       if (i != 3)
         encode_sb(cpi, tile, tp,  mi_row + y_idx, mi_col + x_idx, 0,
                   split_subsize);
 
       split_rate += rt;
       split_dist += dt;
       pl = partition_plane_context(cpi->above_seg_context,
                                    cpi->left_seg_context,
-                                   mi_row + y_idx, mi_col + x_idx, bsize);
+                                   mi_row + y_idx, mi_col + x_idx,
+                                   split_subsize);
       split_rate += x->partition_cost[pl][PARTITION_NONE];
     }
     pl = partition_plane_context(cpi->above_seg_context, cpi->left_seg_context,
                                  mi_row, mi_col, bsize);
     if (split_rate < INT_MAX) {
       split_rate += x->partition_cost[pl][PARTITION_SPLIT];
 
       chosen_rate = split_rate;
       chosen_dist = split_dist;
     }
@@ skipping 17 matching lines @@
     chosen_dist = none_dist;
   }
 
   restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
 
   // We must have chosen a partitioning and encoding or we'll fail later on.
   // No other opportunities for success.
   if ( bsize == BLOCK_64X64)
     assert(chosen_rate < INT_MAX && chosen_dist < INT_MAX);
 
-  if (do_recon)
-    encode_sb(cpi, tile, tp, mi_row, mi_col, bsize == BLOCK_64X64, bsize);
+  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, chosen_rate);
+    }
+
+    encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize);
+  }
 
   *rate = chosen_rate;
   *dist = chosen_dist;
 }
 
 static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
   BLOCK_4X4, BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
   BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, BLOCK_8X8,
   BLOCK_8X8, BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, BLOCK_16X16
 };
@@ skipping 94 matching lines @@
   // Check border cases where max and min from neighbours may not be legal.
   *max_block_size = find_partition_size(*max_block_size,
                                         row8x8_remaining, col8x8_remaining,
                                         &bh, &bw);
   *min_block_size = MIN(*min_block_size, *max_block_size);
 }
 
 static void compute_fast_motion_search_level(VP9_COMP *cpi, BLOCK_SIZE bsize) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
-  MACROBLOCKD *const xd = &x->e_mbd;
 
   // Only use 8x8 result for non HD videos.
   // int use_8x8 = (MIN(cpi->common.width, cpi->common.height) < 720) ? 1 : 0;
   int use_8x8 = 1;
 
   if (cm->frame_type && !cpi->is_src_frame_alt_ref &&
       ((use_8x8 && bsize == BLOCK_16X16) ||
       bsize == BLOCK_32X32 || bsize == BLOCK_64X64)) {
     int ref0 = 0, ref1 = 0, ref2 = 0, ref3 = 0;
     PICK_MODE_CONTEXT *block_context = NULL;
 
     if (bsize == BLOCK_16X16) {
-      block_context = x->sb8x8_context[xd->sb_index][xd->mb_index];
+      block_context = x->sb8x8_context[x->sb_index][x->mb_index];
     } else if (bsize == BLOCK_32X32) {
-      block_context = x->mb_context[xd->sb_index];
+      block_context = x->mb_context[x->sb_index];
     } else if (bsize == BLOCK_64X64) {
       block_context = x->sb32_context;
     }
 
     if (block_context) {
       ref0 = block_context[0].mic.mbmi.ref_frame[0];
       ref1 = block_context[1].mic.mbmi.ref_frame[0];
       ref2 = block_context[2].mic.mbmi.ref_frame[0];
       ref3 = block_context[3].mic.mbmi.ref_frame[0];
     }
@@ skipping 67 matching lines @@
   vpx_memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
 }
 
 // TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
 // unlikely to be selected depending on previous rate-distortion optimization
 // results, for encoding speed-up.
 static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
                               TOKENEXTRA **tp, int mi_row,
                               int mi_col, BLOCK_SIZE bsize, int *rate,
                               int64_t *dist, int do_recon, int64_t best_rd) {
-  VP9_COMMON * const cm = &cpi->common;
-  MACROBLOCK * const x = &cpi->mb;
-  MACROBLOCKD * const xd = &x->e_mbd;
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->mb;
   const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
   ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
   PARTITION_CONTEXT sl[8], sa[8];
   TOKENEXTRA *tp_orig = *tp;
   int i, pl;
   BLOCK_SIZE subsize;
   int this_rate, sum_rate = 0, best_rate = INT_MAX;
   int64_t this_dist, sum_dist = 0, best_dist = INT64_MAX;
   int64_t sum_rd = 0;
   int do_split = bsize >= BLOCK_8X8;
   int do_rect = 1;
   // Override skipping rectangular partition operations for edge blocks
   const int force_horz_split = (mi_row + ms >= cm->mi_rows);
   const int force_vert_split = (mi_col + ms >= cm->mi_cols);
+  const int xss = x->e_mbd.plane[1].subsampling_x;
+  const int yss = x->e_mbd.plane[1].subsampling_y;
 
   int partition_none_allowed = !force_horz_split && !force_vert_split;
-  int partition_horz_allowed = !force_vert_split && bsize >= BLOCK_8X8;
-  int partition_vert_allowed = !force_horz_split && bsize >= BLOCK_8X8;
+  int partition_horz_allowed = !force_vert_split && yss <= xss &&
+                               bsize >= BLOCK_8X8;
+  int partition_vert_allowed = !force_horz_split && xss <= yss &&
+                               bsize >= BLOCK_8X8;
 
   int partition_split_done = 0;
   (void) *tp_orig;
 
   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 (xd->ab_index != 0) {
+    if (x->ab_index != 0) {
       *rate = 0;
       *dist = 0;
       return;
     }
   }
-  assert(mi_height_log2(bsize) == mi_width_log2(bsize));
+  assert(num_8x8_blocks_wide_lookup[bsize] ==
+             num_8x8_blocks_high_lookup[bsize]);
 
   if (bsize == BLOCK_16X16) {
     set_offsets(cpi, tile, mi_row, mi_col, bsize);
     x->mb_energy = vp9_block_energy(cpi, x, bsize);
   }
 
   // Determine partition types in search according to the speed features.
   // The threshold set here has to be of square block size.
   if (cpi->sf.auto_min_max_partition_size) {
     partition_none_allowed &= (bsize <= cpi->sf.max_partition_size &&
@@ skipping 30 matching lines @@
                   get_block_context(x, bsize), best_rd);
     if (this_rate != INT_MAX) {
       if (bsize >= BLOCK_8X8) {
         pl = partition_plane_context(cpi->above_seg_context,
                                      cpi->left_seg_context,
                                      mi_row, mi_col, bsize);
         this_rate += x->partition_cost[pl][PARTITION_NONE];
       }
       sum_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_dist);
       if (sum_rd < best_rd) {
-        int64_t stop_thresh = 2048;
+        int64_t stop_thresh = 4096;
+        int64_t stop_thresh_rd;
 
         best_rate = this_rate;
         best_dist = this_dist;
         best_rd = sum_rd;
         if (bsize >= BLOCK_8X8)
           *(get_sb_partitioning(x, bsize)) = bsize;
 
         // Adjust threshold according to partition size.
         stop_thresh >>= 8 - (b_width_log2_lookup[bsize] +
             b_height_log2_lookup[bsize]);
 
+        stop_thresh_rd = RDCOST(x->rdmult, x->rddiv, 0, stop_thresh);
         // If obtained distortion is very small, choose current partition
         // and stop splitting.
-        if (this_dist < stop_thresh) {
+        if (!x->e_mbd.lossless && best_rd < stop_thresh_rd) {
           do_split = 0;
           do_rect = 0;
         }
       }
     }
     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
   }
 
   // store estimated motion vector
   if (cpi->sf.adaptive_motion_search)
     store_pred_mv(x, get_block_context(x, bsize));
 
   // PARTITION_SPLIT
   sum_rd = 0;
   // TODO(jingning): use the motion vectors given by the above search as
   // the starting point of motion search in the following partition type check.
   if (do_split) {
     subsize = get_subsize(bsize, PARTITION_SPLIT);
     for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
       const int x_idx = (i & 1) * ms;
       const int y_idx = (i >> 1) * ms;
 
       if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
         continue;
 
-      *get_sb_index(xd, subsize) = i;
+      *get_sb_index(x, subsize) = i;
       if (cpi->sf.adaptive_motion_search)
         load_pred_mv(x, get_block_context(x, bsize));
+      if (cpi->sf.adaptive_pred_filter_type && bsize == BLOCK_8X8 &&
+          partition_none_allowed)
+        get_block_context(x, subsize)->pred_filter_type =
+            get_block_context(x, bsize)->mic.mbmi.interp_filter;
       rd_pick_partition(cpi, tile, tp, mi_row + y_idx, mi_col + x_idx, subsize,
                         &this_rate, &this_dist, i != 3, best_rd - sum_rd);
 
       if (this_rate == INT_MAX) {
         sum_rd = INT64_MAX;
       } else {
         sum_rate += this_rate;
         sum_dist += this_dist;
         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
       }
@@ skipping 24 matching lines @@
   x->subblock_ref = 0;
 
   if (partition_split_done &&
       cpi->sf.using_small_partition_info) {
     compute_fast_motion_search_level(cpi, bsize);
   }
 
   // PARTITION_HORZ
   if (partition_horz_allowed && do_rect) {
     subsize = get_subsize(bsize, PARTITION_HORZ);
-    *get_sb_index(xd, subsize) = 0;
+    *get_sb_index(x, subsize) = 0;
     if (cpi->sf.adaptive_motion_search)
       load_pred_mv(x, get_block_context(x, bsize));
+    if (cpi->sf.adaptive_pred_filter_type && bsize == BLOCK_8X8 &&
+        partition_none_allowed)
+      get_block_context(x, subsize)->pred_filter_type =
+          get_block_context(x, bsize)->mic.mbmi.interp_filter;
     pick_sb_modes(cpi, tile, mi_row, mi_col, &sum_rate, &sum_dist, subsize,
                   get_block_context(x, subsize), best_rd);
     sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
 
     if (sum_rd < best_rd && mi_row + ms < 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(xd, subsize) = 1;
+      *get_sb_index(x, subsize) = 1;
       if (cpi->sf.adaptive_motion_search)
         load_pred_mv(x, get_block_context(x, bsize));
+      if (cpi->sf.adaptive_pred_filter_type && bsize == BLOCK_8X8 &&
+          partition_none_allowed)
+        get_block_context(x, subsize)->pred_filter_type =
+            get_block_context(x, bsize)->mic.mbmi.interp_filter;
       pick_sb_modes(cpi, tile, mi_row + ms, mi_col, &this_rate,
                     &this_dist, subsize, get_block_context(x, subsize),
                     best_rd - sum_rd);
       if (this_rate == INT_MAX) {
         sum_rd = INT64_MAX;
       } else {
         sum_rate += this_rate;
         sum_dist += this_dist;
         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
       }
@@ skipping 11 matching lines @@
         *(get_sb_partitioning(x, bsize)) = subsize;
       }
     }
     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
   }
 
   // PARTITION_VERT
   if (partition_vert_allowed && do_rect) {
     subsize = get_subsize(bsize, PARTITION_VERT);
 
-    *get_sb_index(xd, subsize) = 0;
+    *get_sb_index(x, subsize) = 0;
     if (cpi->sf.adaptive_motion_search)
       load_pred_mv(x, get_block_context(x, bsize));
+    if (cpi->sf.adaptive_pred_filter_type && bsize == BLOCK_8X8 &&
+        partition_none_allowed)
+      get_block_context(x, subsize)->pred_filter_type =
+          get_block_context(x, bsize)->mic.mbmi.interp_filter;
     pick_sb_modes(cpi, tile, mi_row, mi_col, &sum_rate, &sum_dist, subsize,
                   get_block_context(x, subsize), best_rd);
     sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
     if (sum_rd < best_rd && mi_col + ms < 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(xd, subsize) = 1;
+      *get_sb_index(x, subsize) = 1;
       if (cpi->sf.adaptive_motion_search)
         load_pred_mv(x, get_block_context(x, bsize));
+      if (cpi->sf.adaptive_pred_filter_type && bsize == BLOCK_8X8 &&
+          partition_none_allowed)
+        get_block_context(x, subsize)->pred_filter_type =
+            get_block_context(x, bsize)->mic.mbmi.interp_filter;
       pick_sb_modes(cpi, tile, mi_row, mi_col + ms, &this_rate,
                     &this_dist, subsize, get_block_context(x, subsize),
                     best_rd - sum_rd);
       if (this_rate == INT_MAX) {
         sum_rd = INT64_MAX;
       } else {
         sum_rate += this_rate;
         sum_dist += this_dist;
         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
       }
@@ skipping 11 matching lines @@
         *(get_sb_partitioning(x, bsize)) = subsize;
       }
     }
     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
   }
 
 
   *rate = best_rate;
   *dist = best_dist;
 
-  if (best_rate < INT_MAX && best_dist < INT64_MAX && do_recon)
-    encode_sb(cpi, tile, tp, mi_row, mi_col, bsize == BLOCK_64X64, bsize);
+  if (best_rate < INT_MAX && best_dist < INT64_MAX && 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, best_rate);
+    }
+    encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize);
+  }
   if (bsize == BLOCK_64X64) {
     assert(tp_orig < *tp);
     assert(best_rate < INT_MAX);
     assert(best_dist < INT_MAX);
   } else {
     assert(tp_orig == *tp);
   }
 }
 
 // Examines 64x64 block and chooses a best reference frame
@@ skipping 25 matching lines @@
     r += x->partition_cost[pl][PARTITION_NONE];
 
     *(get_sb_partitioning(x, BLOCK_64X64)) = BLOCK_64X64;
     cpi->set_ref_frame_mask = 0;
   }
 
   restore_context(cpi, mi_row, mi_col, a, l, sa, sl, BLOCK_64X64);
 }
 
 static void encode_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
-                          int mi_row, TOKENEXTRA **tp, int *totalrate) {
+                          int mi_row, TOKENEXTRA **tp) {
   VP9_COMMON * const cm = &cpi->common;
   int mi_col;
 
   // 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;
 
+    BLOCK_SIZE i;
+    MACROBLOCK *x = &cpi->mb;
+    for (i = BLOCK_4X4; i < BLOCK_8X8; ++i) {
+      const int num_4x4_w = num_4x4_blocks_wide_lookup[i];
+      const int num_4x4_h = num_4x4_blocks_high_lookup[i];
+      const int num_4x4_blk = MAX(4, num_4x4_w * num_4x4_h);
+      for (x->sb_index = 0; x->sb_index < 4; ++x->sb_index)
+        for (x->mb_index = 0; x->mb_index < 4; ++x->mb_index)
+          for (x->b_index = 0; x->b_index < 16 / num_4x4_blk; ++x->b_index)
+            get_block_context(x, i)->pred_filter_type = SWITCHABLE;
+    }
+
     vp9_zero(cpi->mb.pred_mv);
 
     if (cpi->sf.reference_masking)
       rd_pick_reference_frame(cpi, tile, mi_row, mi_col);
 
     if (cpi->sf.use_lastframe_partitioning ||
         cpi->sf.use_one_partition_size_always ) {
       const int idx_str = cm->mode_info_stride * mi_row + mi_col;
       MODE_INFO **mi_8x8 = cm->mi_grid_visible + idx_str;
       MODE_INFO **prev_mi_8x8 = cm->prev_mi_grid_visible + idx_str;
@@ skipping 47 matching lines @@
   MACROBLOCK *const x = &cpi->mb;
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCKD *const xd = &x->e_mbd;
   const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
 
   x->act_zbin_adj = 0;
   cpi->seg0_idx = 0;
 
   xd->mode_info_stride = cm->mode_info_stride;
 
-  // reset intra mode contexts
-  if (frame_is_intra_only(cm))
-    vp9_init_mbmode_probs(cm);
-
   // Copy data over into macro block data structures.
   vp9_setup_src_planes(x, cpi->Source, 0, 0);
 
   // TODO(jkoleszar): are these initializations required?
   setup_pre_planes(xd, 0, &cm->yv12_fb[cm->ref_frame_map[cpi->lst_fb_idx]],
                    0, 0, NULL);
   setup_dst_planes(xd, get_frame_new_buffer(cm), 0, 0);
 
   setup_block_dptrs(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
 
   xd->mi_8x8[0]->mbmi.mode = DC_PRED;
   xd->mi_8x8[0]->mbmi.uv_mode = DC_PRED;
 
   vp9_zero(cpi->y_mode_count);
   vp9_zero(cpi->y_uv_mode_count);
   vp9_zero(cm->counts.inter_mode);
-  vp9_zero(cpi->partition_count);
+  vp9_zero(cm->counts.partition);
   vp9_zero(cpi->intra_inter_count);
   vp9_zero(cpi->comp_inter_count);
   vp9_zero(cpi->single_ref_count);
   vp9_zero(cpi->comp_ref_count);
   vp9_zero(cm->counts.tx);
   vp9_zero(cm->counts.mbskip);
 
   // Note: this memset assumes above_context[0], [1] and [2]
   // are allocated as part of the same buffer.
   vpx_memset(cpi->above_context[0], 0,
@@ skipping 23 matching lines @@
   if (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
       cpi->common.tx_mode >= ALLOW_32X32)
     cpi->common.tx_mode = ALLOW_32X32;
 }
 
 static void encode_frame_internal(VP9_COMP *cpi) {
   int mi_row;
   MACROBLOCK * const x = &cpi->mb;
   VP9_COMMON * const cm = &cpi->common;
   MACROBLOCKD * const xd = &x->e_mbd;
-  int totalrate;
 
 //  fprintf(stderr, "encode_frame_internal frame %d (%d) type %d\n",
 //           cpi->common.current_video_frame, cpi->common.show_frame,
 //           cm->frame_type);
 
 // debug output
 #if DBG_PRNT_SEGMAP
   {
     FILE *statsfile;
     statsfile = fopen("segmap2.stt", "a");
     fprintf(statsfile, "\n");
     fclose(statsfile);
   }
 #endif
 
-  totalrate = 0;
-
-  // Reset frame count of inter 0,0 motion vector usage.
-  cpi->inter_zz_count = 0;
-
   vp9_zero(cm->counts.switchable_interp);
   vp9_zero(cpi->tx_stepdown_count);
 
   xd->mi_8x8 = cm->mi_grid_visible;
   // required for vp9_frame_init_quantizer
   xd->mi_8x8[0] = cm->mi;
 
   xd->last_mi = cm->prev_mi;
 
   vp9_zero(cpi->NMVcount);
@@ skipping 41 matching lines @@
 
       for (tile_row = 0; tile_row < tile_rows; tile_row++) {
         for (tile_col = 0; tile_col < tile_cols; tile_col++) {
           TileInfo tile;
           TOKENEXTRA *tp_old = tp;
 
           // For each row of SBs in the frame
           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)
-            encode_sb_row(cpi, &tile, mi_row, &tp, &totalrate);
+            encode_sb_row(cpi, &tile, mi_row, &tp);
 
           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));
         }
       }
     }
 
     vpx_usec_timer_mark(&emr_timer);
     cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
   }
 
   if (cpi->sf.skip_encode_sb) {
     int j;
     unsigned int intra_count = 0, inter_count = 0;
     for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
       intra_count += cpi->intra_inter_count[j][0];
       inter_count += cpi->intra_inter_count[j][1];
     }
     cpi->sf.skip_encode_frame = ((intra_count << 2) < inter_count);
     cpi->sf.skip_encode_frame &= (cm->frame_type != KEY_FRAME);
     cpi->sf.skip_encode_frame &= cm->show_frame;
   } else {
     cpi->sf.skip_encode_frame = 0;
   }
 
-  // 256 rate units to the bit,
-  // projected_frame_size in units of BYTES
-  cpi->projected_frame_size = totalrate >> 8;
-
 #if 0
   // Keep record of the total distortion this time around for future use
   cpi->last_frame_distortion = cpi->frame_distortion;
 #endif
 }
 
 static int check_dual_ref_flags(VP9_COMP *cpi) {
   const int ref_flags = cpi->ref_frame_flags;
 
   if (vp9_segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
@@ skipping 180 matching lines @@
      * worked best for that type of frame in the past.
      * It also predicts whether another coding mode would have worked
      * better that this coding mode. If that is the case, it remembers
      * that for subsequent frames.
      * It does the same analysis for transform size selection also.
      */
     int frame_type = get_frame_type(cpi);
 
     /* prediction (compound, single or hybrid) mode selection */
     if (frame_type == 3 || !cm->allow_comp_inter_inter)
-      pred_type = SINGLE_PREDICTION_ONLY;
+      pred_type = SINGLE_REFERENCE;
     else if (cpi->rd_prediction_type_threshes[frame_type][1]
              > cpi->rd_prediction_type_threshes[frame_type][0]
              && cpi->rd_prediction_type_threshes[frame_type][1]
              > cpi->rd_prediction_type_threshes[frame_type][2]
              && check_dual_ref_flags(cpi) && cpi->static_mb_pct == 100)
-      pred_type = COMP_PREDICTION_ONLY;
+      pred_type = COMPOUND_REFERENCE;
     else if (cpi->rd_prediction_type_threshes[frame_type][0]
              > cpi->rd_prediction_type_threshes[frame_type][2])
-      pred_type = SINGLE_PREDICTION_ONLY;
+      pred_type = SINGLE_REFERENCE;
     else
-      pred_type = HYBRID_PREDICTION;
+      pred_type = REFERENCE_MODE_SELECT;
 
     /* filter type selection */
     // FIXME(rbultje) for some odd reason, we often select smooth_filter
     // as default filter for ARF overlay frames. This is a REALLY BAD
     // IDEA so we explicitly disable it here.
     if (frame_type != 3 &&
         cpi->rd_filter_threshes[frame_type][1] >
             cpi->rd_filter_threshes[frame_type][0] &&
         cpi->rd_filter_threshes[frame_type][1] >
             cpi->rd_filter_threshes[frame_type][2] &&
@@ skipping 12 matching lines @@
       filter_type = SWITCHABLE;
     }
 
     cpi->mb.e_mbd.lossless = 0;
     if (cpi->oxcf.lossless) {
       cpi->mb.e_mbd.lossless = 1;
     }
 
     /* transform size selection (4x4, 8x8, 16x16 or select-per-mb) */
     select_tx_mode(cpi);
-    cpi->common.comp_pred_mode = pred_type;
+    cpi->common.reference_mode = pred_type;
     cpi->common.mcomp_filter_type = filter_type;
     encode_frame_internal(cpi);
 
-    for (i = 0; i < NB_PREDICTION_TYPES; ++i) {
+    for (i = 0; i < REFERENCE_MODES; ++i) {
       const int diff = (int) (cpi->rd_comp_pred_diff[i] / cpi->common.MBs);
       cpi->rd_prediction_type_threshes[frame_type][i] += diff;
       cpi->rd_prediction_type_threshes[frame_type][i] >>= 1;
     }
 
     for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
       const int64_t diff = cpi->rd_filter_diff[i] / cpi->common.MBs;
       cpi->rd_filter_threshes[frame_type][i] =
           (cpi->rd_filter_threshes[frame_type][i] + diff) / 2;
     }
 
     for (i = 0; i < TX_MODES; ++i) {
       int64_t pd = cpi->rd_tx_select_diff[i];
       int diff;
       if (i == TX_MODE_SELECT)
         pd -= RDCOST(cpi->mb.rdmult, cpi->mb.rddiv,
                      2048 * (TX_SIZES - 1), 0);
       diff = (int) (pd / cpi->common.MBs);
       cpi->rd_tx_select_threshes[frame_type][i] += diff;
       cpi->rd_tx_select_threshes[frame_type][i] /= 2;
     }
 
-    if (cpi->common.comp_pred_mode == HYBRID_PREDICTION) {
+    if (cpi->common.reference_mode == REFERENCE_MODE_SELECT) {
       int single_count_zero = 0;
       int comp_count_zero = 0;
 
       for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
         single_count_zero += cpi->comp_inter_count[i][0];
         comp_count_zero += cpi->comp_inter_count[i][1];
       }
 
       if (comp_count_zero == 0) {
-        cpi->common.comp_pred_mode = SINGLE_PREDICTION_ONLY;
+        cpi->common.reference_mode = SINGLE_REFERENCE;
         vp9_zero(cpi->comp_inter_count);
       } else if (single_count_zero == 0) {
-        cpi->common.comp_pred_mode = COMP_PREDICTION_ONLY;
+        cpi->common.reference_mode = COMPOUND_REFERENCE;
         vp9_zero(cpi->comp_inter_count);
       }
     }
 
     if (cpi->common.tx_mode == TX_MODE_SELECT) {
       int count4x4 = 0;
       int count8x8_lp = 0, count8x8_8x8p = 0;
       int count16x16_16x16p = 0, count16x16_lp = 0;
       int count32x32 = 0;
 
@@ skipping 71 matching lines @@
 #endif
 }
 static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
                               int mi_row, int mi_col, BLOCK_SIZE bsize) {
   VP9_COMMON * const cm = &cpi->common;
   MACROBLOCK * const x = &cpi->mb;
   MACROBLOCKD * const xd = &x->e_mbd;
   MODE_INFO **mi_8x8 = xd->mi_8x8;
   MODE_INFO *mi = mi_8x8[0];
   MB_MODE_INFO *mbmi = &mi->mbmi;
+  PICK_MODE_CONTEXT *ctx = get_block_context(x, bsize);
   unsigned int segment_id = mbmi->segment_id;
   const int mis = cm->mode_info_stride;
   const int mi_width = num_8x8_blocks_wide_lookup[bsize];
   const int mi_height = num_8x8_blocks_high_lookup[bsize];
+  x->skip_recode = !x->select_txfm_size && mbmi->sb_type >= BLOCK_8X8 &&
+                   (cpi->oxcf.aq_mode != COMPLEXITY_AQ);
+  x->skip_optimize = ctx->is_coded;
+  ctx->is_coded = 1;
   x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
   x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
-                    xd->q_index < QIDX_SKIP_THRESH);
+                    x->q_index < QIDX_SKIP_THRESH);
   if (x->skip_encode)
     return;
 
   if (cm->frame_type == KEY_FRAME) {
     if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
       adjust_act_zbin(cpi, x);
       vp9_update_zbin_extra(cpi, x);
     }
   } else {
     vp9_setup_interp_filters(xd, mbmi->interp_filter, cm);
@@ skipping 64 matching lines @@
       cm->counts.mbskip[mb_skip_context][1]++;
     reset_skip_context(xd, MAX(bsize, BLOCK_8X8));
   }
 
   if (output_enabled) {
     if (cm->tx_mode == TX_MODE_SELECT &&
         mbmi->sb_type >= BLOCK_8X8  &&
         !(is_inter_block(mbmi) &&
             (mbmi->skip_coeff ||
              vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)))) {
-      const uint8_t context = vp9_get_pred_context_tx_size(xd);
-      ++get_tx_counts(bsize, context, &cm->counts.tx)[mbmi->tx_size];
+      ++get_tx_counts(max_txsize_lookup[bsize], vp9_get_tx_size_context(xd),
+                      &cm->counts.tx)[mbmi->tx_size];
     } else {
       int x, y;
-      TX_SIZE sz = tx_mode_to_biggest_tx_size[cm->tx_mode];
-      assert(sizeof(tx_mode_to_biggest_tx_size) /
-             sizeof(tx_mode_to_biggest_tx_size[0]) == TX_MODES);
+      TX_SIZE tx_size;
       // The new intra coding scheme requires no change of transform size
       if (is_inter_block(&mi->mbmi)) {
-        if (sz == TX_32X32 && bsize < BLOCK_32X32)
-          sz = TX_16X16;
-        if (sz == TX_16X16 && bsize < BLOCK_16X16)
-          sz = TX_8X8;
-        if (sz == TX_8X8 && bsize < BLOCK_8X8)
-          sz = TX_4X4;
-      } else if (bsize >= BLOCK_8X8) {
-        sz = mbmi->tx_size;
+        tx_size = MIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
+                      max_txsize_lookup[bsize]);
       } else {
-        sz = TX_4X4;
+        tx_size = (bsize >= BLOCK_8X8) ? mbmi->tx_size : TX_4X4;
       }
 
       for (y = 0; y < mi_height; y++)
         for (x = 0; x < mi_width; x++)
           if (mi_col + x < cm->mi_cols && mi_row + y < cm->mi_rows)
-            mi_8x8[mis * y + x]->mbmi.tx_size = sz;
+            mi_8x8[mis * y + x]->mbmi.tx_size = tx_size;
     }
   }
 }