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.
  */
 
 #include <assert.h>
 #include <stdlib.h>  // qsort()
 
 #include "./vp9_rtcd.h"
 #include "./vpx_scale_rtcd.h"
 
 #include "vpx_mem/vpx_mem.h"
 #include "vpx_ports/mem_ops.h"
 #include "vpx_scale/vpx_scale.h"
 
 #include "vp9/common/vp9_alloccommon.h"
 #include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_entropy.h"
 #include "vp9/common/vp9_entropymode.h"
 #include "vp9/common/vp9_idct.h"
+#include "vp9/common/vp9_loopfilter_thread.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_thread.h"
 #include "vp9/common/vp9_tile_common.h"
 
 #include "vp9/decoder/vp9_decodeframe.h"
 #include "vp9/decoder/vp9_detokenize.h"
@@ skipping 255 matching lines @@
         vpx_memset(dqcoeff, 0, 256 * sizeof(dqcoeff[0]));
       else
         vpx_memset(dqcoeff, 0, (16 << (tx_size << 1)) * sizeof(dqcoeff[0]));
     }
   }
 }
 
 struct intra_args {
   VP9_COMMON *cm;
   MACROBLOCKD *xd;
+  FRAME_COUNTS *counts;
   vp9_reader *r;
 };
 
 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;
   const PREDICTION_MODE mode = (plane == 0) ? get_y_mode(mi, block)
                                             : mi->mbmi.uv_mode;
   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, plane, block,
+    const int eob = vp9_decode_block_tokens(cm, xd, args->counts, plane, block,
                                             plane_bsize, x, y, tx_size,
                                             args->r);
     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;
 };
 
 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];
   int x, y, eob;
   txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
-  eob = vp9_decode_block_tokens(cm, xd, plane, block, plane_bsize, x, y,
-                                tx_size, args->r);
+  eob = vp9_decode_block_tokens(cm, xd, args->counts, plane, block, plane_bsize,
+                                x, y, tx_size, args->r);
   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];
@@ skipping 15 matching lines @@
   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;
 }
 
-static void decode_block(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+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;
   MB_MODE_INFO *mbmi = set_offsets(cm, xd, tile, bsize, mi_row, mi_col);
-  vp9_read_mode_info(cm, xd, tile, mi_row, mi_col, r);
+  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)
       setup_plane_dequants(cm, xd, vp9_get_qindex(&cm->seg, mbmi->segment_id,
                                                   cm->base_qindex));
   }
 
   if (!is_inter_block(mbmi)) {
-    struct intra_args arg = { cm, xd, r };
+    struct intra_args arg = { cm, xd, counts, r };
     vp9_foreach_transformed_block(xd, bsize,
                                   predict_and_reconstruct_intra_block, &arg);
   } else {
     // Prediction
-    vp9_dec_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
+    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, &eobtotal };
+      struct inter_args arg = { cm, xd, r, counts, &eobtotal };
       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, int hbs,
+static PARTITION_TYPE read_partition(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                     FRAME_COUNTS *counts, int hbs,
                                      int mi_row, int mi_col, BLOCK_SIZE bsize,
                                      vp9_reader *r) {
   const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
   const vp9_prob *const probs = get_partition_probs(cm, ctx);
   const int has_rows = (mi_row + hbs) < cm->mi_rows;
   const int has_cols = (mi_col + hbs) < cm->mi_cols;
   PARTITION_TYPE p;
 
   if (has_rows && has_cols)
     p = (PARTITION_TYPE)vp9_read_tree(r, vp9_partition_tree, probs);
   else if (!has_rows && has_cols)
     p = vp9_read(r, probs[1]) ? PARTITION_SPLIT : PARTITION_HORZ;
   else if (has_rows && !has_cols)
     p = vp9_read(r, probs[2]) ? PARTITION_SPLIT : PARTITION_VERT;
   else
     p = PARTITION_SPLIT;
 
   if (!cm->frame_parallel_decoding_mode)
-    ++cm->counts.partition[ctx][p];
+    ++counts->partition[ctx][p];
 
   return p;
 }
 
-static void decode_partition(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+static void decode_partition(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 hbs = num_8x8_blocks_wide_lookup[bsize] / 2;
   PARTITION_TYPE partition;
   BLOCK_SIZE subsize, uv_subsize;
 
   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
     return;
 
-  partition = read_partition(cm, xd, hbs, mi_row, mi_col, bsize, r);
+  partition = read_partition(cm, xd, counts, hbs, mi_row, mi_col, bsize, r);
   subsize = get_subsize(bsize, partition);
   uv_subsize = ss_size_lookup[subsize][cm->subsampling_x][cm->subsampling_y];
   if (subsize >= BLOCK_8X8 && uv_subsize == BLOCK_INVALID)
     vpx_internal_error(xd->error_info,
                        VPX_CODEC_CORRUPT_FRAME, "Invalid block size.");
   if (subsize < BLOCK_8X8) {
-    decode_block(cm, xd, tile, mi_row, mi_col, r, subsize);
+    decode_block(pbi, xd, counts, tile, mi_row, mi_col, r, subsize);
   } else {
     switch (partition) {
       case PARTITION_NONE:
-        decode_block(cm, xd, tile, mi_row, mi_col, r, subsize);
+        decode_block(pbi, xd, counts, tile, mi_row, mi_col, r, subsize);
         break;
       case PARTITION_HORZ:
-        decode_block(cm, xd, tile, mi_row, mi_col, r, subsize);
+        decode_block(pbi, xd, counts, tile, mi_row, mi_col, r, subsize);
         if (mi_row + hbs < cm->mi_rows)
-          decode_block(cm, xd, tile, mi_row + hbs, mi_col, r, subsize);
+          decode_block(pbi, xd, counts, tile, mi_row + hbs, mi_col, r, subsize);
         break;
       case PARTITION_VERT:
-        decode_block(cm, xd, tile, mi_row, mi_col, r, subsize);
+        decode_block(pbi, xd, counts, tile, mi_row, mi_col, r, subsize);
         if (mi_col + hbs < cm->mi_cols)
-          decode_block(cm, xd, tile, mi_row, mi_col + hbs, r, subsize);
+          decode_block(pbi, xd, counts, tile, mi_row, mi_col + hbs, r, subsize);
         break;
       case PARTITION_SPLIT:
-        decode_partition(cm, xd, tile, mi_row,       mi_col,       r, subsize);
-        decode_partition(cm, xd, tile, mi_row,       mi_col + hbs, r, subsize);
-        decode_partition(cm, xd, tile, mi_row + hbs, mi_col,       r, subsize);
-        decode_partition(cm, xd, tile, mi_row + hbs, mi_col + hbs, r, subsize);
+        decode_partition(pbi, xd, counts, tile, mi_row, mi_col, r, subsize);
+        decode_partition(pbi, xd, counts, tile, mi_row, mi_col + hbs, r,
+                         subsize);
+        decode_partition(pbi, xd, counts, tile, mi_row + hbs, mi_col, r,
+                         subsize);
+        decode_partition(pbi, xd, counts, tile, mi_row + hbs, mi_col + hbs, r,
+                         subsize);
         break;
       default:
         assert(0 && "Invalid partition type");
     }
   }
 
   // update partition context
   if (bsize >= BLOCK_8X8 &&
       (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
     update_partition_context(xd, mi_row, mi_col, subsize, bsize);
@@ skipping 201 matching lines @@
     cm->height = height;
   }
   if (cm->cur_frame->mvs == NULL || cm->mi_rows > cm->cur_frame->mi_rows ||
       cm->mi_cols > cm->cur_frame->mi_cols) {
     resize_mv_buffer(cm);
   }
 }
 
 static void setup_frame_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
   int width, height;
+  BufferPool *const pool = cm->buffer_pool;
   vp9_read_frame_size(rb, &width, &height);
   resize_context_buffers(cm, width, height);
   setup_display_size(cm, rb);
 
+  lock_buffer_pool(pool);
   if (vp9_realloc_frame_buffer(
           get_frame_new_buffer(cm), cm->width, cm->height,
           cm->subsampling_x, cm->subsampling_y,
 #if CONFIG_VP9_HIGHBITDEPTH
           cm->use_highbitdepth,
 #endif
           VP9_DEC_BORDER_IN_PIXELS,
           cm->byte_alignment,
-          &cm->frame_bufs[cm->new_fb_idx].raw_frame_buffer, cm->get_fb_cb,
-          cm->cb_priv)) {
+          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
+          pool->cb_priv)) {
+    unlock_buffer_pool(pool);
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate frame buffer");
   }
-  cm->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
-  cm->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
-  cm->frame_bufs[cm->new_fb_idx].buf.color_space =
-      (vpx_color_space_t)cm->color_space;
-  cm->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
+  unlock_buffer_pool(pool);
+
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
+  pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
 }
 
 static INLINE int valid_ref_frame_img_fmt(vpx_bit_depth_t ref_bit_depth,
                                           int ref_xss, int ref_yss,
                                           vpx_bit_depth_t this_bit_depth,
                                           int this_xss, int this_yss) {
   return ref_bit_depth == this_bit_depth && ref_xss == this_xss &&
          ref_yss == this_yss;
 }
 
 static void setup_frame_size_with_refs(VP9_COMMON *cm,
                                        struct vp9_read_bit_buffer *rb) {
   int width, height;
   int found = 0, i;
   int has_valid_ref_frame = 0;
+  BufferPool *const pool = cm->buffer_pool;
   for (i = 0; i < REFS_PER_FRAME; ++i) {
     if (vp9_rb_read_bit(rb)) {
       YV12_BUFFER_CONFIG *const buf = cm->frame_refs[i].buf;
       width = buf->y_crop_width;
       height = buf->y_crop_height;
       found = 1;
       break;
     }
   }
 
@@ skipping 24 matching lines @@
             cm->bit_depth,
             cm->subsampling_x,
             cm->subsampling_y))
       vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                          "Referenced frame has incompatible color format");
   }
 
   resize_context_buffers(cm, width, height);
   setup_display_size(cm, rb);
 
+  lock_buffer_pool(pool);
   if (vp9_realloc_frame_buffer(
           get_frame_new_buffer(cm), cm->width, cm->height,
           cm->subsampling_x, cm->subsampling_y,
 #if CONFIG_VP9_HIGHBITDEPTH
           cm->use_highbitdepth,
 #endif
           VP9_DEC_BORDER_IN_PIXELS,
           cm->byte_alignment,
-          &cm->frame_bufs[cm->new_fb_idx].raw_frame_buffer, cm->get_fb_cb,
-          cm->cb_priv)) {
+          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
+          pool->cb_priv)) {
+    unlock_buffer_pool(pool);
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate frame buffer");
   }
-  cm->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
-  cm->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
-  cm->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
+  unlock_buffer_pool(pool);
+
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
+  pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
 }
 
 static void setup_tile_info(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
   int min_log2_tile_cols, max_log2_tile_cols, max_ones;
   vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
 
   // columns
   max_ones = max_log2_tile_cols - min_log2_tile_cols;
   cm->log2_tile_cols = min_log2_tile_cols;
   while (max_ones-- && vp9_rb_read_bit(rb))
@@ skipping 148 matching lines @@
          mi_row += MI_BLOCK_SIZE) {
       for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
         const int col = pbi->inv_tile_order ?
                         tile_cols - tile_col - 1 : tile_col;
         tile_data = pbi->tile_data + tile_cols * tile_row + col;
         vp9_tile_set_col(&tile, tile_data->cm, col);
         vp9_zero(tile_data->xd.left_context);
         vp9_zero(tile_data->xd.left_seg_context);
         for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
              mi_col += MI_BLOCK_SIZE) {
-          decode_partition(tile_data->cm, &tile_data->xd, &tile, mi_row, mi_col,
-                           &tile_data->bit_reader, BLOCK_64X64);
+          decode_partition(pbi, &tile_data->xd, &cm->counts, &tile, mi_row,
+                           mi_col, &tile_data->bit_reader, BLOCK_64X64);
         }
         pbi->mb.corrupted |= tile_data->xd.corrupted;
         if (pbi->mb.corrupted)
             vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                                "Failed to decode tile data");
       }
       // Loopfilter one row.
       if (cm->lf.filter_level) {
         const int lf_start = mi_row - MI_BLOCK_SIZE;
         LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
 
         // delay the loopfilter by 1 macroblock row.
         if (lf_start < 0) continue;
 
         // decoding has completed: finish up the loop filter in this thread.
         if (mi_row + MI_BLOCK_SIZE >= cm->mi_rows) continue;
 
         winterface->sync(&pbi->lf_worker);
         lf_data->start = lf_start;
         lf_data->stop = mi_row;
         if (pbi->max_threads > 1) {
           winterface->launch(&pbi->lf_worker);
         } else {
           winterface->execute(&pbi->lf_worker);
         }
       }
+      // After loopfiltering, the last 7 row pixels in each superblock row may
+      // still be changed by the longest loopfilter of the next superblock
+      // row.
+      if (pbi->frame_parallel_decode)
+        vp9_frameworker_broadcast(pbi->cur_buf,
+                                  mi_row << MI_BLOCK_SIZE_LOG2);
     }
   }
 
   // Loopfilter remaining rows in the frame.
   if (cm->lf.filter_level) {
     LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
     winterface->sync(&pbi->lf_worker);
     lf_data->start = lf_data->stop;
     lf_data->stop = cm->mi_rows;
     winterface->execute(&pbi->lf_worker);
   }
 
   // Get last tile data.
   tile_data = pbi->tile_data + tile_cols * tile_rows - 1;
 
+  if (pbi->frame_parallel_decode)
+    vp9_frameworker_broadcast(pbi->cur_buf, INT_MAX);
   return vp9_reader_find_end(&tile_data->bit_reader);
 }
 
 static int tile_worker_hook(TileWorkerData *const tile_data,
                             const TileInfo *const tile) {
   int mi_row, mi_col;
 
   if (setjmp(tile_data->error_info.jmp)) {
     tile_data->error_info.setjmp = 0;
     tile_data->xd.corrupted = 1;
     return 0;
   }
 
   tile_data->error_info.setjmp = 1;
   tile_data->xd.error_info = &tile_data->error_info;
 
   for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
        mi_row += MI_BLOCK_SIZE) {
     vp9_zero(tile_data->xd.left_context);
     vp9_zero(tile_data->xd.left_seg_context);
     for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
          mi_col += MI_BLOCK_SIZE) {
-      decode_partition(tile_data->cm, &tile_data->xd, tile,
-                       mi_row, mi_col, &tile_data->bit_reader, BLOCK_64X64);
+      decode_partition(tile_data->pbi, &tile_data->xd,
+                       &tile_data->pbi->common.counts,
+                       tile, mi_row, mi_col, &tile_data->bit_reader,
+                       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) {
@@ skipping 93 matching lines @@
 
   n = 0;
   while (n < tile_cols) {
     int i;
     for (i = 0; i < num_workers && n < tile_cols; ++i) {
       VP9Worker *const worker = &pbi->tile_workers[i];
       TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
       TileInfo *const tile = (TileInfo*)worker->data2;
       TileBuffer *const buf = &tile_buffers[0][n];
 
-      tile_data->cm = cm;
+      tile_data->pbi = pbi;
       tile_data->xd = pbi->mb;
       tile_data->xd.corrupted = 0;
-      vp9_tile_init(tile, tile_data->cm, 0, buf->col);
+      vp9_tile_init(tile, &pbi->common, 0, buf->col);
       setup_token_decoder(buf->data, data_end, buf->size, &cm->error,
                           &tile_data->bit_reader, pbi->decrypt_cb,
                           pbi->decrypt_state);
       init_macroblockd(cm, &tile_data->xd);
 
       worker->had_error = 0;
       if (i == num_workers - 1 || n == tile_cols - 1) {
         winterface->execute(worker);
       } else {
         winterface->launch(worker);
@@ skipping 83 matching lines @@
     } else {
       vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                          "4:4:4 color not supported in profile 0 or 2");
     }
   }
 }
 
 static size_t read_uncompressed_header(VP9Decoder *pbi,
                                        struct vp9_read_bit_buffer *rb) {
   VP9_COMMON *const cm = &pbi->common;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+  BufferPool *const pool = pbi->common.buffer_pool;
+  int i, mask, ref_index = 0;
   size_t sz;
-  int i;
 
   cm->last_frame_type = cm->frame_type;
 
   if (vp9_rb_read_literal(rb, 2) != VP9_FRAME_MARKER)
       vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                          "Invalid frame marker");
 
   cm->profile = vp9_read_profile(rb);
 
   if (cm->profile >= MAX_PROFILES)
     vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                        "Unsupported bitstream profile");
 
   cm->show_existing_frame = vp9_rb_read_bit(rb);
   if (cm->show_existing_frame) {
     // Show an existing frame directly.
     const int frame_to_show = cm->ref_frame_map[vp9_rb_read_literal(rb, 3)];
-
-    if (frame_to_show < 0 || cm->frame_bufs[frame_to_show].ref_count < 1)
+    lock_buffer_pool(pool);
+    if (frame_to_show < 0 || frame_bufs[frame_to_show].ref_count < 1) {
+      unlock_buffer_pool(pool);
       vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                          "Buffer %d does not contain a decoded frame",
                          frame_to_show);
+    }
 
-    ref_cnt_fb(cm->frame_bufs, &cm->new_fb_idx, frame_to_show);
+    ref_cnt_fb(frame_bufs, &cm->new_fb_idx, frame_to_show);
+    unlock_buffer_pool(pool);
     pbi->refresh_frame_flags = 0;
     cm->lf.filter_level = 0;
     cm->show_frame = 1;
+
+    if (pbi->frame_parallel_decode) {
+      for (i = 0; i < REF_FRAMES; ++i)
+        cm->next_ref_frame_map[i] = cm->ref_frame_map[i];
+    }
     return 0;
   }
 
   cm->frame_type = (FRAME_TYPE) vp9_rb_read_bit(rb);
   cm->show_frame = vp9_rb_read_bit(rb);
   cm->error_resilient_mode = vp9_rb_read_bit(rb);
 
   if (cm->frame_type == KEY_FRAME) {
     if (!vp9_read_sync_code(rb))
       vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                          "Invalid frame sync code");
 
     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 = -1;
       cm->frame_refs[i].buf = NULL;
     }
 
     setup_frame_size(cm, rb);
-    pbi->need_resync = 0;
+    if (pbi->need_resync) {
+      vpx_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))
         vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                            "Invalid frame sync code");
       if (cm->profile > PROFILE_0) {
         read_bitdepth_colorspace_sampling(cm, rb);
       } else {
         // NOTE: The intra-only frame header does not include the specification
         // of either the color format or color sub-sampling in profile 0. VP9
         // specifies that the default color format should be YUV 4:2:0 in this
         // case (normative).
         cm->color_space = VPX_CS_BT_601;
         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);
-      pbi->need_resync = 0;
-    } else {
+      if (pbi->need_resync) {
+        vpx_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 = &cm->frame_bufs[idx].buf;
+        ref_frame->buf = &frame_bufs[idx].buf;
         cm->ref_frame_sign_bias[LAST_FRAME + i] = vp9_rb_read_bit(rb);
       }
 
       setup_frame_size_with_refs(cm, rb);
 
       cm->allow_high_precision_mv = vp9_rb_read_bit(rb);
       cm->interp_filter = read_interp_filter(rb);
 
       for (i = 0; i < REFS_PER_FRAME; ++i) {
         RefBuffer *const ref_buf = &cm->frame_refs[i];
@@ skipping 29 matching lines @@
     cm->frame_parallel_decoding_mode = vp9_rb_read_bit(rb);
   } else {
     cm->refresh_frame_context = 0;
     cm->frame_parallel_decoding_mode = 1;
   }
 
   // This flag will be overridden by the call to vp9_setup_past_independence
   // below, forcing the use of context 0 for those frame types.
   cm->frame_context_idx = vp9_rb_read_literal(rb, FRAME_CONTEXTS_LOG2);
 
+  // Generate next_ref_frame_map.
+  lock_buffer_pool(pool);
+  for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+    if (mask & 1) {
+      cm->next_ref_frame_map[ref_index] = cm->new_fb_idx;
+      ++frame_bufs[cm->new_fb_idx].ref_count;
+    } else {
+      cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+    }
+    // Current thread holds the reference frame.
+    if (cm->ref_frame_map[ref_index] >= 0)
+      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
+    ++ref_index;
+  }
+
+  for (; ref_index < REF_FRAMES; ++ref_index) {
+    cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+    // Current thread holds the reference frame.
+    if (cm->ref_frame_map[ref_index] >= 0)
+      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
+  }
+  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_tile_info(cm, rb);
   sz = vp9_rb_read_literal(rb, 16);
 
@@ skipping 125 matching lines @@
   }
   return rb;
 }
 
 void vp9_decode_frame(VP9Decoder *pbi,
                       const uint8_t *data, const uint8_t *data_end,
                       const uint8_t **p_data_end) {
   VP9_COMMON *const cm = &pbi->common;
   MACROBLOCKD *const xd = &pbi->mb;
   struct vp9_read_bit_buffer rb = { NULL, NULL, 0, NULL, 0};
-
+  int context_updated = 0;
   uint8_t clear_data[MAX_VP9_HEADER_SIZE];
   const size_t first_partition_size = read_uncompressed_header(pbi,
       init_read_bit_buffer(pbi, &rb, data, data_end, clear_data));
   const int tile_rows = 1 << cm->log2_tile_rows;
   const int tile_cols = 1 << cm->log2_tile_cols;
   YV12_BUFFER_CONFIG *const new_fb = get_frame_new_buffer(cm);
   xd->cur_buf = new_fb;
 
   if (!first_partition_size) {
     // showing a frame directly
@@ skipping 21 matching lines @@
                        "Uninitialized entropy context.");
 
   vp9_zero(cm->counts);
 
   xd->corrupted = 0;
   new_fb->corrupted = read_compressed_header(pbi, data, first_partition_size);
   if (new_fb->corrupted)
     vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                        "Decode failed. Frame data header is corrupted.");
 
+  if (cm->lf.filter_level) {
+    vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
+  }
+
+  // If encoded in frame parallel mode, frame context is ready after decoding
+  // the frame header.
+  if (pbi->frame_parallel_decode && cm->frame_parallel_decoding_mode) {
+    VP9Worker *const worker = pbi->frame_worker_owner;
+    FrameWorkerData *const frame_worker_data = worker->data1;
+    if (cm->refresh_frame_context) {
+      context_updated = 1;
+      cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+    }
+    vp9_frameworker_lock_stats(worker);
+    pbi->cur_buf->row = -1;
+    pbi->cur_buf->col = -1;
+    frame_worker_data->frame_context_ready = 1;
+    // Signal the main thread that context is ready.
+    vp9_frameworker_signal_stats(worker);
+    vp9_frameworker_unlock_stats(worker);
+  }
+
   // TODO(jzern): remove frame_parallel_decoding_mode restriction for
   // single-frame tile decoding.
   if (pbi->max_threads > 1 && tile_rows == 1 && tile_cols > 1 &&
       cm->frame_parallel_decoding_mode) {
     *p_data_end = decode_tiles_mt(pbi, data + first_partition_size, data_end);
     if (!xd->corrupted) {
       // If multiple threads are used to decode tiles, then we use those threads
       // to do parallel loopfiltering.
       vp9_loop_filter_frame_mt(new_fb, cm, pbi->mb.plane, cm->lf.filter_level,
                                0, 0, pbi->tile_workers, pbi->num_tile_workers,
                                &pbi->lf_row_sync);
     } else {
       vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                          "Decode failed. Frame data is corrupted.");
 
     }
   } else {
     *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
   }
 
-  new_fb->corrupted |= xd->corrupted;
-
-  if (!new_fb->corrupted) {
+  if (!xd->corrupted) {
     if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
       vp9_adapt_coef_probs(cm);
 
       if (!frame_is_intra_only(cm)) {
         vp9_adapt_mode_probs(cm);
         vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
       }
     } else {
       debug_check_frame_counts(cm);
     }
   } else {
     vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                        "Decode failed. Frame data is corrupted.");
   }
 
-  if (cm->refresh_frame_context)
+  // Non frame parallel update frame context here.
+  if (cm->refresh_frame_context && !context_updated)
     cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
 }
+
+static void build_mc_border(const uint8_t *src, int src_stride,
+                            uint8_t *dst, int dst_stride,
+                            int x, int y, int b_w, int b_h, int w, int h) {
+  // Get a pointer to the start of the real data for this row.
+  const uint8_t *ref_row = src - x - y * src_stride;
+
+  if (y >= h)
+    ref_row += (h - 1) * src_stride;
+  else if (y > 0)
+    ref_row += y * src_stride;
+
+  do {
+    int right = 0, copy;
+    int left = x < 0 ? -x : 0;
+
+    if (left > b_w)
+      left = b_w;
+
+    if (x + b_w > w)
+      right = x + b_w - w;
+
+    if (right > b_w)
+      right = b_w;
+
+    copy = b_w - left - right;
+
+    if (left)
+      memset(dst, ref_row[0], left);
+
+    if (copy)
+      memcpy(dst + left, ref_row + x + left, copy);
+
+    if (right)
+      memset(dst + left + copy, ref_row[w - 1], right);
+
+    dst += dst_stride;
+    ++y;
+
+    if (y > 0 && y < h)
+      ref_row += src_stride;
+  } while (--b_h);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void high_build_mc_border(const uint8_t *src8, int src_stride,
+                                 uint16_t *dst, int dst_stride,
+                                 int x, int y, int b_w, int b_h,
+                                 int w, int h) {
+  // Get a pointer to the start of the real data for this row.
+  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  const uint16_t *ref_row = src - x - y * src_stride;
+
+  if (y >= h)
+    ref_row += (h - 1) * src_stride;
+  else if (y > 0)
+    ref_row += y * src_stride;
+
+  do {
+    int right = 0, copy;
+    int left = x < 0 ? -x : 0;
+
+    if (left > b_w)
+      left = b_w;
+
+    if (x + b_w > w)
+      right = x + b_w - w;
+
+    if (right > b_w)
+      right = b_w;
+
+    copy = b_w - left - right;
+
+    if (left)
+      vpx_memset16(dst, ref_row[0], left);
+
+    if (copy)
+      memcpy(dst + left, ref_row + x + left, copy * sizeof(uint16_t));
+
+    if (right)
+      vpx_memset16(dst + left + copy, ref_row[w - 1], right);
+
+    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) {
+  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;
+
+  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;
+
+    const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, &mv, bw, bh,
+                                               pd->subsampling_x,
+                                               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);
+
+    // 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;
+    }
+
+    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));
+
+      // Co-ordinate of the block to 1/16th pixel precision.
+      x0_16 = (x_start + x) << SUBPEL_BITS;
+      y0_16 = (y_start + y) << SUBPEL_BITS;
+
+      // 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,
+                              (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.
+#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
+        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,
+                              (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
+    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;
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize,
+                                                        &xd->plane[plane]);
+    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;
+
+    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);
+    }
+  }
+}