libvpx: Pull from upstream

source/libvpx/vp9/decoder/vp9_decodeframe.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 258 matching lines @@
           vp9_idct32x32_add(dqcoeff, dst, stride, eob);
           break;
         default:
           assert(0 && "Invalid transform size");
           return;
       }
     }
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 
     if (eob == 1) {
-      vpx_memset(dqcoeff, 0, 2 * sizeof(dqcoeff[0]));
+      memset(dqcoeff, 0, 2 * sizeof(dqcoeff[0]));
     } else {
       if (tx_type == DCT_DCT && tx_size <= TX_16X16 && eob <= 10)
-        vpx_memset(dqcoeff, 0, 4 * (4 << tx_size) * sizeof(dqcoeff[0]));
+        memset(dqcoeff, 0, 4 * (4 << tx_size) * sizeof(dqcoeff[0]));
       else if (tx_size == TX_32X32 && eob <= 34)
-        vpx_memset(dqcoeff, 0, 256 * sizeof(dqcoeff[0]));
+        memset(dqcoeff, 0, 256 * sizeof(dqcoeff[0]));
       else
-        vpx_memset(dqcoeff, 0, (16 << (tx_size << 1)) * sizeof(dqcoeff[0]));
+        memset(dqcoeff, 0, (16 << (tx_size << 1)) * sizeof(dqcoeff[0]));
     }
   }
 }
 
 struct intra_args {
   VP9_COMMON *cm;
   MACROBLOCKD *xd;
   FRAME_COUNTS *counts;
   vp9_reader *r;
-  const int16_t *const y_dequant;
-  const int16_t *const uv_dequant;
+  int seg_id;
 };
 
 static void predict_and_reconstruct_intra_block(int plane, int block,
                                                 BLOCK_SIZE plane_bsize,
                                                 TX_SIZE tx_size, void *arg) {
   struct intra_args *const args = (struct intra_args *)arg;
   VP9_COMMON *const cm = args->cm;
   MACROBLOCKD *const xd = args->xd;
   struct macroblockd_plane *const pd = &xd->plane[plane];
-  MODE_INFO *const mi = xd->mi[0].src_mi;
+  MODE_INFO *const mi = xd->mi[0];
   const PREDICTION_MODE mode = (plane == 0) ? get_y_mode(mi, block)
                                             : mi->mbmi.uv_mode;
-  const int16_t *const dequant = (plane == 0) ? args->y_dequant
-                                              : args->uv_dequant;
   int x, y;
   uint8_t *dst;
   txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
   dst = &pd->dst.buf[4 * y * pd->dst.stride + 4 * x];
 
   vp9_predict_intra_block(xd, block >> (tx_size << 1),
                           b_width_log2_lookup[plane_bsize], tx_size, mode,
                           dst, pd->dst.stride, dst, pd->dst.stride,
                           x, y, plane);
 
   if (!mi->mbmi.skip) {
     const int eob = vp9_decode_block_tokens(cm, xd, args->counts, plane, block,
                                             plane_bsize, x, y, tx_size,
-                                            args->r, dequant);
+                                            args->r, args->seg_id);
     inverse_transform_block(xd, plane, block, tx_size, dst, pd->dst.stride,
                             eob);
   }
 }
 
 struct inter_args {
   VP9_COMMON *cm;
   MACROBLOCKD *xd;
   vp9_reader *r;
   FRAME_COUNTS *counts;
   int *eobtotal;
-  const int16_t *const y_dequant;
-  const int16_t *const uv_dequant;
+  int seg_id;
 };
 
 static void reconstruct_inter_block(int plane, int block,
                                     BLOCK_SIZE plane_bsize,
                                     TX_SIZE tx_size, void *arg) {
   struct inter_args *args = (struct inter_args *)arg;
   VP9_COMMON *const cm = args->cm;
   MACROBLOCKD *const xd = args->xd;
   struct macroblockd_plane *const pd = &xd->plane[plane];
-  const int16_t *const dequant = (plane == 0) ? args->y_dequant
-                                              : args->uv_dequant;
   int x, y, eob;
   txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
   eob = vp9_decode_block_tokens(cm, xd, args->counts, plane, block, plane_bsize,
-                                x, y, tx_size, args->r, dequant);
+                                x, y, tx_size, args->r, args->seg_id);
   inverse_transform_block(xd, plane, block, tx_size,
                           &pd->dst.buf[4 * y * pd->dst.stride + 4 * x],
                           pd->dst.stride, eob);
   *args->eobtotal += eob;
 }
 
 static MB_MODE_INFO *set_offsets(VP9_COMMON *const cm, MACROBLOCKD *const xd,
                                  const TileInfo *const tile,
                                  BLOCK_SIZE bsize, int mi_row, int mi_col) {
   const int bw = num_8x8_blocks_wide_lookup[bsize];
   const int bh = num_8x8_blocks_high_lookup[bsize];
   const int x_mis = MIN(bw, cm->mi_cols - mi_col);
   const int y_mis = MIN(bh, cm->mi_rows - mi_row);
   const int offset = mi_row * cm->mi_stride + mi_col;
   int x, y;
 
-  xd->mi = cm->mi + offset;
-  xd->mi[0].src_mi = &xd->mi[0];  // Point to self.
-  xd->mi[0].mbmi.sb_type = bsize;
-
+  xd->mi = cm->mi_grid_visible + offset;
+  xd->mi[0] = &cm->mi[offset];
+  xd->mi[0]->mbmi.sb_type = bsize;
   for (y = 0; y < y_mis; ++y)
     for (x = !y; x < x_mis; ++x) {
-      xd->mi[y * cm->mi_stride + x].src_mi = &xd->mi[0];
+      xd->mi[y * cm->mi_stride + x] = xd->mi[0];
     }
 
   set_skip_context(xd, mi_row, mi_col);
 
   // Distance of Mb to the various image edges. These are specified to 8th pel
   // as they are always compared to values that are in 1/8th pel units
   set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
 
   vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
-  return &xd->mi[0].mbmi;
+  return &xd->mi[0]->mbmi;
 }
 
 static void decode_block(VP9Decoder *const pbi, MACROBLOCKD *const xd,
                          FRAME_COUNTS *counts,
                          const TileInfo *const tile,
                          int mi_row, int mi_col,
                          vp9_reader *r, BLOCK_SIZE bsize) {
   VP9_COMMON *const cm = &pbi->common;
   const int less8x8 = bsize < BLOCK_8X8;
-  int16_t y_dequant[2], uv_dequant[2];
-  int qindex = cm->base_qindex;
   MB_MODE_INFO *mbmi = set_offsets(cm, xd, tile, bsize, mi_row, mi_col);
   vp9_read_mode_info(pbi, xd, counts, tile, mi_row, mi_col, r);
 
   if (less8x8)
     bsize = BLOCK_8X8;
 
   if (mbmi->skip) {
     reset_skip_context(xd, bsize);
-  } else if (cm->seg.enabled) {
-    qindex = vp9_get_qindex(&cm->seg, mbmi->segment_id, cm->base_qindex);
   }
 
-  y_dequant[0] = vp9_dc_quant(qindex, cm->y_dc_delta_q, cm->bit_depth);
-  y_dequant[1] = vp9_ac_quant(qindex, 0, cm->bit_depth);
-  uv_dequant[0] = vp9_dc_quant(qindex, cm->uv_dc_delta_q, cm->bit_depth);
-  uv_dequant[1] = vp9_ac_quant(qindex, cm->uv_ac_delta_q, cm->bit_depth);
-
   if (!is_inter_block(mbmi)) {
-    struct intra_args arg = {cm, xd, counts, r , y_dequant, uv_dequant};
+    struct intra_args arg = {cm, xd, counts, r, mbmi->segment_id};
     vp9_foreach_transformed_block(xd, bsize,
                                   predict_and_reconstruct_intra_block, &arg);
   } else {
     // Prediction
     vp9_dec_build_inter_predictors_sb(pbi, xd, mi_row, mi_col, bsize);
 
     // Reconstruction
     if (!mbmi->skip) {
       int eobtotal = 0;
-      struct inter_args arg = {cm, xd, r, counts, &eobtotal, y_dequant,
-                               uv_dequant};
+      struct inter_args arg = {cm, xd, r, counts, &eobtotal, mbmi->segment_id};
       vp9_foreach_transformed_block(xd, bsize, reconstruct_inter_block, &arg);
       if (!less8x8 && eobtotal == 0)
         mbmi->skip = 1;  // skip loopfilter
     }
   }
 
   xd->corrupted |= vp9_reader_has_error(r);
 }
 
 static PARTITION_TYPE read_partition(VP9_COMMON *cm, MACROBLOCKD *xd,
@@ skipping 203 matching lines @@
                                struct vp9_read_bit_buffer *rb) {
   cm->base_qindex = vp9_rb_read_literal(rb, QINDEX_BITS);
   cm->y_dc_delta_q = read_delta_q(rb);
   cm->uv_dc_delta_q = read_delta_q(rb);
   cm->uv_ac_delta_q = read_delta_q(rb);
   cm->dequant_bit_depth = cm->bit_depth;
   xd->lossless = cm->base_qindex == 0 &&
                  cm->y_dc_delta_q == 0 &&
                  cm->uv_dc_delta_q == 0 &&
                  cm->uv_ac_delta_q == 0;
+
 #if CONFIG_VP9_HIGHBITDEPTH
   xd->bd = (int)cm->bit_depth;
 #endif
 }
 
+static void setup_segmentation_dequant(VP9_COMMON *const cm) {
+  // Build y/uv dequant values based on segmentation.
+  if (cm->seg.enabled) {
+    int i;
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      const int qindex = vp9_get_qindex(&cm->seg, i, cm->base_qindex);
+      cm->y_dequant[i][0] = vp9_dc_quant(qindex, cm->y_dc_delta_q,
+                                         cm->bit_depth);
+      cm->y_dequant[i][1] = vp9_ac_quant(qindex, 0, cm->bit_depth);
+      cm->uv_dequant[i][0] = vp9_dc_quant(qindex, cm->uv_dc_delta_q,
+                                          cm->bit_depth);
+      cm->uv_dequant[i][1] = vp9_ac_quant(qindex, cm->uv_ac_delta_q,
+                                          cm->bit_depth);
+    }
+  } else {
+    const int qindex = cm->base_qindex;
+    // When segmentation is disabled, only the first value is used.  The
+    // remaining are don't cares.
+    cm->y_dequant[0][0] = vp9_dc_quant(qindex, cm->y_dc_delta_q, cm->bit_depth);
+    cm->y_dequant[0][1] = vp9_ac_quant(qindex, 0, cm->bit_depth);
+    cm->uv_dequant[0][0] = vp9_dc_quant(qindex, cm->uv_dc_delta_q,
+                                        cm->bit_depth);
+    cm->uv_dequant[0][1] = vp9_ac_quant(qindex, cm->uv_ac_delta_q,
+                                        cm->bit_depth);
+  }
+}
+
 static INTERP_FILTER read_interp_filter(struct vp9_read_bit_buffer *rb) {
   const INTERP_FILTER literal_to_filter[] = { EIGHTTAP_SMOOTH,
                                               EIGHTTAP,
                                               EIGHTTAP_SHARP,
                                               BILINEAR };
   return vp9_rb_read_bit(rb) ? SWITCHABLE
                              : literal_to_filter[vp9_rb_read_literal(rb, 2)];
 }
 
 void vp9_read_frame_size(struct vp9_read_bit_buffer *rb,
@@ skipping 268 matching lines @@
     winterface->sync(&pbi->lf_worker);
     vp9_loop_filter_data_reset(lf_data, get_frame_new_buffer(cm), cm,
                                pbi->mb.plane);
   }
 
   assert(tile_rows <= 4);
   assert(tile_cols <= (1 << 6));
 
   // Note: this memset assumes above_context[0], [1] and [2]
   // are allocated as part of the same buffer.
-  vpx_memset(cm->above_context, 0,
-             sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_cols);
+  memset(cm->above_context, 0,
+         sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_cols);
 
-  vpx_memset(cm->above_seg_context, 0,
-             sizeof(*cm->above_seg_context) * aligned_cols);
+  memset(cm->above_seg_context, 0,
+         sizeof(*cm->above_seg_context) * aligned_cols);
 
   get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);
 
   if (pbi->tile_data == NULL ||
       (tile_cols * tile_rows) != pbi->total_tiles) {
     vpx_free(pbi->tile_data);
     CHECK_MEM_ERROR(
         cm,
         pbi->tile_data,
         vpx_memalign(32, tile_cols * tile_rows * (sizeof(*pbi->tile_data))));
@@ skipping 109 matching lines @@
                        BLOCK_64X64);
     }
   }
   return !tile_data->xd.corrupted;
 }
 
 // sorts in descending order
 static int compare_tile_buffers(const void *a, const void *b) {
   const TileBuffer *const buf1 = (const TileBuffer*)a;
   const TileBuffer *const buf2 = (const TileBuffer*)b;
-  if (buf1->size < buf2->size) {
-    return 1;
-  } else if (buf1->size == buf2->size) {
-    return 0;
-  } else {
-    return -1;
-  }
+  return (int)(buf2->size - buf1->size);
 }
 
 static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
                                       const uint8_t *data,
                                       const uint8_t *data_end) {
   VP9_COMMON *const cm = &pbi->common;
   const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
   const uint8_t *bit_reader_end = NULL;
   const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
   const int tile_cols = 1 << cm->log2_tile_cols;
@@ skipping 40 matching lines @@
   for (n = 0; n < num_workers; ++n) {
     VP9Worker *const worker = &pbi->tile_workers[n];
     winterface->sync(worker);
     worker->hook = (VP9WorkerHook)tile_worker_hook;
     worker->data1 = &pbi->tile_worker_data[n];
     worker->data2 = &pbi->tile_worker_info[n];
   }
 
   // Note: this memset assumes above_context[0], [1] and [2]
   // are allocated as part of the same buffer.
-  vpx_memset(cm->above_context, 0,
-             sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_mi_cols);
-  vpx_memset(cm->above_seg_context, 0,
-             sizeof(*cm->above_seg_context) * aligned_mi_cols);
+  memset(cm->above_context, 0,
+         sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_mi_cols);
+  memset(cm->above_seg_context, 0,
+         sizeof(*cm->above_seg_context) * aligned_mi_cols);
 
   // Load tile data into tile_buffers
   get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);
 
   // Sort the buffers based on size in descending order.
   qsort(tile_buffers[0], tile_cols, sizeof(tile_buffers[0][0]),
         compare_tile_buffers);
 
   // Rearrange the tile buffers such that per-tile group the largest, and
   // presumably the most difficult, tile will be decoded in the main thread.
@@ skipping 201 matching lines @@
     read_bitdepth_colorspace_sampling(cm, rb);
     pbi->refresh_frame_flags = (1 << REF_FRAMES) - 1;
 
     for (i = 0; i < REFS_PER_FRAME; ++i) {
       cm->frame_refs[i].idx = INVALID_IDX;
       cm->frame_refs[i].buf = NULL;
     }
 
     setup_frame_size(cm, rb);
     if (pbi->need_resync) {
-      vpx_memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+      memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
       pbi->need_resync = 0;
     }
   } else {
     cm->intra_only = cm->show_frame ? 0 : vp9_rb_read_bit(rb);
 
     cm->reset_frame_context = cm->error_resilient_mode ?
         0 : vp9_rb_read_literal(rb, 2);
 
     if (cm->intra_only) {
       if (!vp9_read_sync_code(rb))
@@ skipping 10 matching lines @@
         cm->subsampling_y = cm->subsampling_x = 1;
         cm->bit_depth = VPX_BITS_8;
 #if CONFIG_VP9_HIGHBITDEPTH
         cm->use_highbitdepth = 0;
 #endif
       }
 
       pbi->refresh_frame_flags = vp9_rb_read_literal(rb, REF_FRAMES);
       setup_frame_size(cm, rb);
       if (pbi->need_resync) {
-        vpx_memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+        memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
         pbi->need_resync = 0;
       }
     } else if (pbi->need_resync != 1) {  /* Skip if need resync */
       pbi->refresh_frame_flags = vp9_rb_read_literal(rb, REF_FRAMES);
       for (i = 0; i < REFS_PER_FRAME; ++i) {
         const int ref = vp9_rb_read_literal(rb, REF_FRAMES_LOG2);
         const int idx = cm->ref_frame_map[ref];
         RefBuffer *const ref_frame = &cm->frame_refs[i];
         ref_frame->idx = idx;
         ref_frame->buf = &frame_bufs[idx].buf;
@@ skipping 68 matching lines @@
   }
   unlock_buffer_pool(pool);
   pbi->hold_ref_buf = 1;
 
   if (frame_is_intra_only(cm) || cm->error_resilient_mode)
     vp9_setup_past_independence(cm);
 
   setup_loopfilter(&cm->lf, rb);
   setup_quantization(cm, &pbi->mb, rb);
   setup_segmentation(&cm->seg, rb);
+  setup_segmentation_dequant(cm);
 
   setup_tile_info(cm, rb);
   sz = vp9_rb_read_literal(rb, 16);
 
   if (sz == 0)
     vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                        "Invalid header size");
 
   return sz;
 }
@@ skipping 300 matching lines @@
     dst += dst_stride;
     ++y;
 
     if (y > 0 && y < h)
       ref_row += src_stride;
   } while (--b_h);
 }
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 
 void dec_build_inter_predictors(VP9Decoder *const pbi, MACROBLOCKD *xd,
-                                int plane, int block, int bw, int bh, int x,
-                                int y, int w, int h, int mi_x, int mi_y) {
+                                int plane, int bw, int bh, int x,
+                                int y, int w, int h, int mi_x, int mi_y,
+                                const InterpKernel *kernel,
+                                const struct scale_factors *sf,
+                                struct buf_2d *pre_buf, struct buf_2d *dst_buf,
+                                const MV* mv, RefCntBuffer *ref_frame_buf,
+                                int is_scaled, int ref) {
   struct macroblockd_plane *const pd = &xd->plane[plane];
-  const MODE_INFO *mi = xd->mi[0].src_mi;
-  const int is_compound = has_second_ref(&mi->mbmi);
-  const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
-  int ref;
+  uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
+  MV32 scaled_mv;
+  int xs, ys, x0, y0, x0_16, y0_16, frame_width, frame_height,
+      buf_stride, subpel_x, subpel_y;
+  uint8_t *ref_frame, *buf_ptr;
 
-  for (ref = 0; ref < 1 + is_compound; ++ref) {
-    const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
-    struct buf_2d *const pre_buf = &pd->pre[ref];
-    struct buf_2d *const dst_buf = &pd->dst;
-    uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
-    const MV mv = mi->mbmi.sb_type < BLOCK_8X8
-               ? average_split_mvs(pd, mi, ref, block)
-               : mi->mbmi.mv[ref].as_mv;
+  // Get reference frame pointer, width and height.
+  if (plane == 0) {
+    frame_width = ref_frame_buf->buf.y_crop_width;
+    frame_height = ref_frame_buf->buf.y_crop_height;
+    ref_frame = ref_frame_buf->buf.y_buffer;
+  } else {
+    frame_width = ref_frame_buf->buf.uv_crop_width;
+    frame_height = ref_frame_buf->buf.uv_crop_height;
+    ref_frame = plane == 1 ? ref_frame_buf->buf.u_buffer
+                         : ref_frame_buf->buf.v_buffer;
+  }
 
-    const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, &mv, bw, bh,
+  if (is_scaled) {
+    const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, mv, bw, bh,
                                                pd->subsampling_x,
                                                pd->subsampling_y);
+    // Co-ordinate of containing block to pixel precision.
+    int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
+    int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
 
-    MV32 scaled_mv;
-    int xs, ys, x0, y0, x0_16, y0_16, y1, frame_width, frame_height,
-        buf_stride, subpel_x, subpel_y;
-    uint8_t *ref_frame, *buf_ptr;
-    const int idx = xd->block_refs[ref]->idx;
-    BufferPool *const pool = pbi->common.buffer_pool;
-    RefCntBuffer *const ref_frame_buf = &pool->frame_bufs[idx];
-    const int is_scaled = vp9_is_scaled(sf);
+    // Co-ordinate of the block to 1/16th pixel precision.
+    x0_16 = (x_start + x) << SUBPEL_BITS;
+    y0_16 = (y_start + y) << SUBPEL_BITS;
 
-    // Get reference frame pointer, width and height.
-    if (plane == 0) {
-      frame_width = ref_frame_buf->buf.y_crop_width;
-      frame_height = ref_frame_buf->buf.y_crop_height;
-      ref_frame = ref_frame_buf->buf.y_buffer;
-    } else {
-      frame_width = ref_frame_buf->buf.uv_crop_width;
-      frame_height = ref_frame_buf->buf.uv_crop_height;
-      ref_frame = plane == 1 ? ref_frame_buf->buf.u_buffer
-                           : ref_frame_buf->buf.v_buffer;
+    // Co-ordinate of current block in reference frame
+    // to 1/16th pixel precision.
+    x0_16 = sf->scale_value_x(x0_16, sf);
+    y0_16 = sf->scale_value_y(y0_16, sf);
+
+    // Map the top left corner of the block into the reference frame.
+    x0 = sf->scale_value_x(x_start + x, sf);
+    y0 = sf->scale_value_y(y_start + y, sf);
+
+    // Scale the MV and incorporate the sub-pixel offset of the block
+    // in the reference frame.
+    scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
+    xs = sf->x_step_q4;
+    ys = sf->y_step_q4;
+  } else {
+    // Co-ordinate of containing block to pixel precision.
+    x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
+    y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
+
+    // Co-ordinate of the block to 1/16th pixel precision.
+    x0_16 = x0 << SUBPEL_BITS;
+    y0_16 = y0 << SUBPEL_BITS;
+
+    scaled_mv.row = mv->row * (1 << (1 - pd->subsampling_y));
+    scaled_mv.col = mv->col * (1 << (1 - pd->subsampling_x));
+    xs = ys = 16;
+  }
+  subpel_x = scaled_mv.col & SUBPEL_MASK;
+  subpel_y = scaled_mv.row & SUBPEL_MASK;
+
+  // Calculate the top left corner of the best matching block in the
+  // reference frame.
+  x0 += scaled_mv.col >> SUBPEL_BITS;
+  y0 += scaled_mv.row >> SUBPEL_BITS;
+  x0_16 += scaled_mv.col;
+  y0_16 += scaled_mv.row;
+
+  // Get reference block pointer.
+  buf_ptr = ref_frame + y0 * pre_buf->stride + x0;
+  buf_stride = pre_buf->stride;
+
+  // Do border extension if there is motion or the
+  // width/height is not a multiple of 8 pixels.
+  if (is_scaled || scaled_mv.col || scaled_mv.row ||
+      (frame_width & 0x7) || (frame_height & 0x7)) {
+    int y1 = (y0_16 + (h - 1) * ys) >> SUBPEL_BITS;
+
+    // Get reference block bottom right horizontal coordinate.
+    int x1 = (x0_16 + (w - 1) * xs) >> SUBPEL_BITS;
+    int x_pad = 0, y_pad = 0;
+
+    if (subpel_x || (sf->x_step_q4 != SUBPEL_SHIFTS)) {
+      x0 -= VP9_INTERP_EXTEND - 1;
+      x1 += VP9_INTERP_EXTEND;
+      x_pad = 1;
     }
 
-    if (is_scaled) {
-      // Co-ordinate of containing block to pixel precision.
-      int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
-      int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
+    if (subpel_y || (sf->y_step_q4 != SUBPEL_SHIFTS)) {
+      y0 -= VP9_INTERP_EXTEND - 1;
+      y1 += VP9_INTERP_EXTEND;
+      y_pad = 1;
+    }
 
-      // Co-ordinate of the block to 1/16th pixel precision.
-      x0_16 = (x_start + x) << SUBPEL_BITS;
-      y0_16 = (y_start + y) << SUBPEL_BITS;
+    // Wait until reference block is ready. Pad 7 more pixels as last 7
+    // pixels of each superblock row can be changed by next superblock row.
+    if (pbi->frame_parallel_decode)
+      vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
+                           MAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
 
-      // Co-ordinate of current block in reference frame
-      // to 1/16th pixel precision.
-      x0_16 = sf->scale_value_x(x0_16, sf);
-      y0_16 = sf->scale_value_y(y0_16, sf);
-
-      // Map the top left corner of the block into the reference frame.
-      x0 = sf->scale_value_x(x_start + x, sf);
-      y0 = sf->scale_value_y(y_start + y, sf);
-
-      // Scale the MV and incorporate the sub-pixel offset of the block
-      // in the reference frame.
-      scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
-      xs = sf->x_step_q4;
-      ys = sf->y_step_q4;
-    } else {
-      // Co-ordinate of containing block to pixel precision.
-      x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
-      y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
-
-      // Co-ordinate of the block to 1/16th pixel precision.
-      x0_16 = x0 << SUBPEL_BITS;
-      y0_16 = y0 << SUBPEL_BITS;
-
-      scaled_mv.row = mv_q4.row;
-      scaled_mv.col = mv_q4.col;
-      xs = ys = 16;
-    }
-    subpel_x = scaled_mv.col & SUBPEL_MASK;
-    subpel_y = scaled_mv.row & SUBPEL_MASK;
-
-    // Calculate the top left corner of the best matching block in the
-    // reference frame.
-    x0 += scaled_mv.col >> SUBPEL_BITS;
-    y0 += scaled_mv.row >> SUBPEL_BITS;
-    x0_16 += scaled_mv.col;
-    y0_16 += scaled_mv.row;
-
-    // Get reference block pointer.
-    buf_ptr = ref_frame + y0 * pre_buf->stride + x0;
-    buf_stride = pre_buf->stride;
-
-    // Get reference block bottom right vertical coordinate.
-    y1 = ((y0_16 + (h - 1) * ys) >> SUBPEL_BITS) + 1;
-
-    // Do border extension if there is motion or the
-    // width/height is not a multiple of 8 pixels.
-    if (is_scaled || scaled_mv.col || scaled_mv.row ||
-        (frame_width & 0x7) || (frame_height & 0x7)) {
-      // Get reference block bottom right horizontal coordinate.
-      int x1 = ((x0_16 + (w - 1) * xs) >> SUBPEL_BITS) + 1;
-      int x_pad = 0, y_pad = 0;
-
-      if (subpel_x || (sf->x_step_q4 != SUBPEL_SHIFTS)) {
-        x0 -= VP9_INTERP_EXTEND - 1;
-        x1 += VP9_INTERP_EXTEND;
-        x_pad = 1;
-      }
-
-      if (subpel_y || (sf->y_step_q4 != SUBPEL_SHIFTS)) {
-        y0 -= VP9_INTERP_EXTEND - 1;
-        y1 += VP9_INTERP_EXTEND;
-        y_pad = 1;
-      }
-
-      // Wait until reference block is ready. Pad 7 more pixels as last 7
-      // pixels of each superblock row can be changed by next superblock row.
-       if (pbi->frame_parallel_decode)
-         vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
-                              MAX(0, (y1 + 7) << (plane == 0 ? 0 : 1)));
-
-      // Skip border extension if block is inside the frame.
-      if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width - 1 ||
-          y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
-        uint8_t *buf_ptr1 = ref_frame + y0 * pre_buf->stride + x0;
-        // Extend the border.
+    // Skip border extension if block is inside the frame.
+    if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width - 1 ||
+        y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
+      uint8_t *buf_ptr1 = ref_frame + y0 * pre_buf->stride + x0;
+      // Extend the border.
 #if CONFIG_VP9_HIGHBITDEPTH
-        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
-          high_build_mc_border(buf_ptr1,
-                               pre_buf->stride,
-                               xd->mc_buf_high,
-                               x1 - x0 + 1,
-                               x0,
-                               y0,
-                               x1 - x0 + 1,
-                               y1 - y0 + 1,
-                               frame_width,
-                               frame_height);
-          buf_stride = x1 - x0 + 1;
-          buf_ptr = CONVERT_TO_BYTEPTR(xd->mc_buf_high) +
-              y_pad * 3 * buf_stride + x_pad * 3;
-        } else {
-          build_mc_border(buf_ptr1,
-                          pre_buf->stride,
-                          xd->mc_buf,
-                          x1 - x0 + 1,
-                          x0,
-                          y0,
-                          x1 - x0 + 1,
-                          y1 - y0 + 1,
-                          frame_width,
-                          frame_height);
-          buf_stride = x1 - x0 + 1;
-          buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
-        }
-#else
+      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        high_build_mc_border(buf_ptr1,
+                             pre_buf->stride,
+                             xd->mc_buf_high,
+                             x1 - x0 + 1,
+                             x0,
+                             y0,
+                             x1 - x0 + 1,
+                             y1 - y0 + 1,
+                             frame_width,
+                             frame_height);
+        buf_stride = x1 - x0 + 1;
+        buf_ptr = CONVERT_TO_BYTEPTR(xd->mc_buf_high) +
+            y_pad * 3 * buf_stride + x_pad * 3;
+      } else {
         build_mc_border(buf_ptr1,
                         pre_buf->stride,
                         xd->mc_buf,
                         x1 - x0 + 1,
                         x0,
                         y0,
                         x1 - x0 + 1,
                         y1 - y0 + 1,
                         frame_width,
                         frame_height);
         buf_stride = x1 - x0 + 1;
         buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
+      }
+#else
+      build_mc_border(buf_ptr1,
+                      pre_buf->stride,
+                      xd->mc_buf,
+                      x1 - x0 + 1,
+                      x0,
+                      y0,
+                      x1 - x0 + 1,
+                      y1 - y0 + 1,
+                      frame_width,
+                      frame_height);
+      buf_stride = x1 - x0 + 1;
+      buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
 #endif  // CONFIG_VP9_HIGHBITDEPTH
-      }
-    } else {
-      // Wait until reference block is ready. Pad 7 more pixels as last 7
-      // pixels of each superblock row can be changed by next superblock row.
-       if (pbi->frame_parallel_decode)
-         vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
-                              MAX(0, (y1 + 7) << (plane == 0 ? 0 : 1)));
     }
+  } else {
+    // Wait until reference block is ready. Pad 7 more pixels as last 7
+    // pixels of each superblock row can be changed by next superblock row.
+     if (pbi->frame_parallel_decode) {
+       const int y1 = ((y0_16 + (h - 1) * ys) >> SUBPEL_BITS) + 1;
+       vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
+                            MAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
+     }
+  }
 #if CONFIG_VP9_HIGHBITDEPTH
-    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
-      high_inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
-                           subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
-    } else {
-      inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
-                    subpel_y, sf, w, h, ref, kernel, xs, ys);
-    }
-#else
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    high_inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+                         subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
+  } else {
     inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
                     subpel_y, sf, w, h, ref, kernel, xs, ys);
+  }
+#else
+  inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+                  subpel_y, sf, w, h, ref, kernel, xs, ys);
 #endif  // CONFIG_VP9_HIGHBITDEPTH
-  }
 }
 
 void vp9_dec_build_inter_predictors_sb(VP9Decoder *const pbi, MACROBLOCKD *xd,
                                        int mi_row, int mi_col,
                                        BLOCK_SIZE bsize) {
   int plane;
   const int mi_x = mi_col * MI_SIZE;
   const int mi_y = mi_row * MI_SIZE;
+  const MODE_INFO *mi = xd->mi[0];
+  const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
+  const BLOCK_SIZE sb_type = mi->mbmi.sb_type;
+  const int is_compound = has_second_ref(&mi->mbmi);
+
   for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
     const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize,
                                                         &xd->plane[plane]);
+    struct macroblockd_plane *const pd = &xd->plane[plane];
+    struct buf_2d *const dst_buf = &pd->dst;
     const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
     const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+
     const int bw = 4 * num_4x4_w;
     const int bh = 4 * num_4x4_h;
+    int ref;
 
-    if (xd->mi[0].src_mi->mbmi.sb_type < BLOCK_8X8) {
-      int i = 0, x, y;
-      assert(bsize == BLOCK_8X8);
-      for (y = 0; y < num_4x4_h; ++y)
-        for (x = 0; x < num_4x4_w; ++x)
-          dec_build_inter_predictors(pbi, xd, plane, i++, bw, bh,
-                                     4 * x, 4 * y, 4, 4, mi_x, mi_y);
-    } else {
-      dec_build_inter_predictors(pbi, xd, plane, 0, bw, bh,
-                                 0, 0, bw, bh, mi_x, mi_y);
+    for (ref = 0; ref < 1 + is_compound; ++ref) {
+      const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
+      struct buf_2d *const pre_buf = &pd->pre[ref];
+      const int idx = xd->block_refs[ref]->idx;
+      BufferPool *const pool = pbi->common.buffer_pool;
+      RefCntBuffer *const ref_frame_buf = &pool->frame_bufs[idx];
+      const int is_scaled = vp9_is_scaled(sf);
+
+      if (sb_type < BLOCK_8X8) {
+        int i = 0, x, y;
+        assert(bsize == BLOCK_8X8);
+        for (y = 0; y < num_4x4_h; ++y) {
+          for (x = 0; x < num_4x4_w; ++x) {
+            const MV mv = average_split_mvs(pd, mi, ref, i++);
+            dec_build_inter_predictors(pbi, xd, plane, bw, bh,
+                                       4 * x, 4 * y, 4, 4, mi_x, mi_y, kernel,
+                                       sf, pre_buf, dst_buf, &mv,
+                                       ref_frame_buf, is_scaled, ref);
+          }
+        }
+      } else {
+        const MV mv = mi->mbmi.mv[ref].as_mv;
+        dec_build_inter_predictors(pbi, xd, plane, bw, bh,
+                                   0, 0, bw, bh, mi_x, mi_y, kernel,
+                                   sf, pre_buf, dst_buf, &mv, ref_frame_buf,
+                                   is_scaled, ref);
+      }
     }
   }
 }