vpx/av1/encoder/segmentation.c

/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include <limits.h>

#include "aom_mem/aom_mem.h"

#include "av1/common/pred_common.h"
#include "av1/common/tile_common.h"

#include "av1/encoder/cost.h"
#include "av1/encoder/segmentation.h"
#include "av1/encoder/subexp.h"

void av1_enable_segmentation(struct segmentation *seg) {
  seg->enabled = 1;
  seg->update_map = 1;
  seg->update_data = 1;
}

void av1_disable_segmentation(struct segmentation *seg) {
  seg->enabled = 0;
  seg->update_map = 0;
  seg->update_data = 0;
}

void av1_set_segment_data(struct segmentation *seg, signed char *feature_data,
                          unsigned char abs_delta) {
  seg->abs_delta = abs_delta;

  memcpy(seg->feature_data, feature_data, sizeof(seg->feature_data));
}
void av1_disable_segfeature(struct segmentation *seg, int segment_id,
                            SEG_LVL_FEATURES feature_id) {
  seg->feature_mask[segment_id] &= ~(1 << feature_id);
}

void av1_clear_segdata(struct segmentation *seg, int segment_id,
                       SEG_LVL_FEATURES feature_id) {
  seg->feature_data[segment_id][feature_id] = 0;
}

// Based on set of segment counts calculate a probability tree
static void calc_segtree_probs(unsigned *segcounts,
                               aom_prob *segment_tree_probs,
                               const aom_prob *cur_tree_probs,
                               const int probwt) {
  // Work out probabilities of each segment
  const unsigned cc[4] = { segcounts[0] + segcounts[1],
                           segcounts[2] + segcounts[3],
                           segcounts[4] + segcounts[5],
                           segcounts[6] + segcounts[7] };
  const unsigned ccc[2] = { cc[0] + cc[1], cc[2] + cc[3] };
  int i;

  segment_tree_probs[0] = get_binary_prob(ccc[0], ccc[1]);
  segment_tree_probs[1] = get_binary_prob(cc[0], cc[1]);
  segment_tree_probs[2] = get_binary_prob(cc[2], cc[3]);
  segment_tree_probs[3] = get_binary_prob(segcounts[0], segcounts[1]);
  segment_tree_probs[4] = get_binary_prob(segcounts[2], segcounts[3]);
  segment_tree_probs[5] = get_binary_prob(segcounts[4], segcounts[5]);
  segment_tree_probs[6] = get_binary_prob(segcounts[6], segcounts[7]);

  for (i = 0; i < 7; i++) {
    const unsigned *ct =
        i == 0 ? ccc : i < 3 ? cc + (i & 2) : segcounts + (i - 3) * 2;
    av1_prob_diff_update_savings_search(ct, cur_tree_probs[i],
                                        &segment_tree_probs[i],
                                        DIFF_UPDATE_PROB, probwt);
  }
}

// Based on set of segment counts and probabilities calculate a cost estimate
static int cost_segmap(unsigned *segcounts, aom_prob *probs) {
  const int c01 = segcounts[0] + segcounts[1];
  const int c23 = segcounts[2] + segcounts[3];
  const int c45 = segcounts[4] + segcounts[5];
  const int c67 = segcounts[6] + segcounts[7];
  const int c0123 = c01 + c23;
  const int c4567 = c45 + c67;

  // Cost the top node of the tree
  int cost = c0123 * av1_cost_zero(probs[0]) + c4567 * av1_cost_one(probs[0]);

  // Cost subsequent levels
  if (c0123 > 0) {
    cost += c01 * av1_cost_zero(probs[1]) + c23 * av1_cost_one(probs[1]);

    if (c01 > 0)
      cost += segcounts[0] * av1_cost_zero(probs[3]) +
              segcounts[1] * av1_cost_one(probs[3]);
    if (c23 > 0)
      cost += segcounts[2] * av1_cost_zero(probs[4]) +
              segcounts[3] * av1_cost_one(probs[4]);
  }

  if (c4567 > 0) {
    cost += c45 * av1_cost_zero(probs[2]) + c67 * av1_cost_one(probs[2]);

    if (c45 > 0)
      cost += segcounts[4] * av1_cost_zero(probs[5]) +
              segcounts[5] * av1_cost_one(probs[5]);
    if (c67 > 0)
      cost += segcounts[6] * av1_cost_zero(probs[6]) +
              segcounts[7] * av1_cost_one(probs[6]);
  }

  return cost;
}

static void count_segs(const AV1_COMMON *cm, MACROBLOCKD *xd,
                       const TileInfo *tile, MODE_INFO **mi,
                       unsigned *no_pred_segcounts,
                       unsigned (*temporal_predictor_count)[2],
                       unsigned *t_unpred_seg_counts, int bw, int bh,
                       int mi_row, int mi_col) {
  int segment_id;

  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;

  xd->mi = mi;
  segment_id = xd->mi[0]->mbmi.segment_id;

  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);

  // Count the number of hits on each segment with no prediction
  no_pred_segcounts[segment_id]++;

  // Temporal prediction not allowed on key frames
  if (cm->frame_type != KEY_FRAME) {
    const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
    // Test to see if the segment id matches the predicted value.
    const int pred_segment_id =
        get_segment_id(cm, cm->last_frame_seg_map, bsize, mi_row, mi_col);
    const int pred_flag = pred_segment_id == segment_id;
    const int pred_context = av1_get_pred_context_seg_id(xd);

    // Store the prediction status for this mb and update counts
    // as appropriate
    xd->mi[0]->mbmi.seg_id_predicted = pred_flag;
    temporal_predictor_count[pred_context][pred_flag]++;

    // Update the "unpredicted" segment count
    if (!pred_flag) t_unpred_seg_counts[segment_id]++;
  }
}

static void count_segs_sb(const AV1_COMMON *cm, MACROBLOCKD *xd,
                          const TileInfo *tile, MODE_INFO **mi,
                          unsigned *no_pred_segcounts,
                          unsigned (*temporal_predictor_count)[2],
                          unsigned *t_unpred_seg_counts, int mi_row, int mi_col,
                          BLOCK_SIZE bsize) {
  const int mis = cm->mi_stride;
  const int bs = num_8x8_blocks_wide_lookup[bsize], hbs = bs / 2;
#if CONFIG_EXT_PARTITION_TYPES
  PARTITION_TYPE partition;
#else
  int bw, bh;
#endif  // CONFIG_EXT_PARTITION_TYPES

  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;

#if CONFIG_EXT_PARTITION_TYPES
  if (bsize == BLOCK_8X8)
    partition = PARTITION_NONE;
  else
    partition = get_partition(cm, mi_row, mi_col, bsize);
  switch (partition) {
    case PARTITION_NONE:
      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
                 t_unpred_seg_counts, bs, bs, mi_row, mi_col);
      break;
    case PARTITION_HORZ:
      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
                 t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
                 temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
                 mi_row + hbs, mi_col);
      break;
    case PARTITION_VERT:
      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
                 t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
                 temporal_predictor_count, t_unpred_seg_counts, hbs, bs, mi_row,
                 mi_col + hbs);
      break;
    case PARTITION_HORZ_A:
      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
                 t_unpred_seg_counts, hbs, hbs, mi_row, mi_col);
      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
                 mi_row, mi_col + hbs);
      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
                 temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
                 mi_row + hbs, mi_col);
      break;
    case PARTITION_HORZ_B:
      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
                 t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
                 mi_row + hbs, mi_col);
      count_segs(cm, xd, tile, mi + hbs + hbs * mis, no_pred_segcounts,
                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
                 mi_row + hbs, mi_col + hbs);
      break;
    case PARTITION_VERT_A:
      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
                 t_unpred_seg_counts, hbs, hbs, mi_row, mi_col);
      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
                 mi_row + hbs, mi_col);
      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
                 temporal_predictor_count, t_unpred_seg_counts, hbs, bs, mi_row,
                 mi_col + hbs);
      break;
    case PARTITION_VERT_B:
      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
                 t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
                 mi_row, mi_col + hbs);
      count_segs(cm, xd, tile, mi + hbs + hbs * mis, no_pred_segcounts,
                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
                 mi_row + hbs, mi_col + hbs);
      break;
    case PARTITION_SPLIT: {
      const BLOCK_SIZE subsize = subsize_lookup[PARTITION_SPLIT][bsize];
      int n;

      assert(num_8x8_blocks_wide_lookup[mi[0]->mbmi.sb_type] < bs &&
             num_8x8_blocks_high_lookup[mi[0]->mbmi.sb_type] < bs);

      for (n = 0; n < 4; n++) {
        const int mi_dc = hbs * (n & 1);
        const int mi_dr = hbs * (n >> 1);

        count_segs_sb(cm, xd, tile, &mi[mi_dr * mis + mi_dc], no_pred_segcounts,
                      temporal_predictor_count, t_unpred_seg_counts,
                      mi_row + mi_dr, mi_col + mi_dc, subsize);
      }
    } break;
    default: assert(0);
  }
#else
  bw = num_8x8_blocks_wide_lookup[mi[0]->mbmi.sb_type];
  bh = num_8x8_blocks_high_lookup[mi[0]->mbmi.sb_type];

  if (bw == bs && bh == bs) {
    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
               t_unpred_seg_counts, bs, bs, mi_row, mi_col);
  } else if (bw == bs && bh < bs) {
    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
               t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
    count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
               temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
               mi_row + hbs, mi_col);
  } else if (bw < bs && bh == bs) {
    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
               t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
    count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
               temporal_predictor_count, t_unpred_seg_counts, hbs, bs, mi_row,
               mi_col + hbs);
  } else {
    const BLOCK_SIZE subsize = subsize_lookup[PARTITION_SPLIT][bsize];
    int n;

    assert(bw < bs && bh < bs);

    for (n = 0; n < 4; n++) {
      const int mi_dc = hbs * (n & 1);
      const int mi_dr = hbs * (n >> 1);

      count_segs_sb(cm, xd, tile, &mi[mi_dr * mis + mi_dc], no_pred_segcounts,
                    temporal_predictor_count, t_unpred_seg_counts,
                    mi_row + mi_dr, mi_col + mi_dc, subsize);
    }
  }
#endif  // CONFIG_EXT_PARTITION_TYPES
}

void av1_choose_segmap_coding_method(AV1_COMMON *cm, MACROBLOCKD *xd) {
  struct segmentation *seg = &cm->seg;
  struct segmentation_probs *segp = &cm->fc->seg;

  int no_pred_cost;
  int t_pred_cost = INT_MAX;

  int i, tile_col, tile_row, mi_row, mi_col;
#if CONFIG_TILE_GROUPS
  const int probwt = cm->num_tg;
#else
  const int probwt = 1;
#endif

  unsigned(*temporal_predictor_count)[2] = cm->counts.seg.pred;
  unsigned *no_pred_segcounts = cm->counts.seg.tree_total;
  unsigned *t_unpred_seg_counts = cm->counts.seg.tree_mispred;

  aom_prob no_pred_tree[SEG_TREE_PROBS];
  aom_prob t_pred_tree[SEG_TREE_PROBS];
  aom_prob t_nopred_prob[PREDICTION_PROBS];

  (void)xd;

  // We are about to recompute all the segment counts, so zero the accumulators.
  av1_zero(cm->counts.seg);

  // First of all generate stats regarding how well the last segment map
  // predicts this one
  for (tile_row = 0; tile_row < cm->tile_rows; tile_row++) {
    TileInfo tile_info;
    av1_tile_set_row(&tile_info, cm, tile_row);
    for (tile_col = 0; tile_col < cm->tile_cols; tile_col++) {
      MODE_INFO **mi_ptr;
      av1_tile_set_col(&tile_info, cm, tile_col);
      mi_ptr = cm->mi_grid_visible + tile_info.mi_row_start * cm->mi_stride +
               tile_info.mi_col_start;
      for (mi_row = tile_info.mi_row_start; mi_row < tile_info.mi_row_end;
           mi_row += cm->mib_size, mi_ptr += cm->mib_size * cm->mi_stride) {
        MODE_INFO **mi = mi_ptr;
        for (mi_col = tile_info.mi_col_start; mi_col < tile_info.mi_col_end;
             mi_col += cm->mib_size, mi += cm->mib_size) {
          count_segs_sb(cm, xd, &tile_info, mi, no_pred_segcounts,
                        temporal_predictor_count, t_unpred_seg_counts, mi_row,
                        mi_col, cm->sb_size);
        }
      }
    }
  }

  // Work out probability tree for coding segments without prediction
  // and the cost.
  calc_segtree_probs(no_pred_segcounts, no_pred_tree, segp->tree_probs, probwt);
  no_pred_cost = cost_segmap(no_pred_segcounts, no_pred_tree);

  // Key frames cannot use temporal prediction
  if (!frame_is_intra_only(cm) && !cm->error_resilient_mode) {
    // Work out probability tree for coding those segments not
    // predicted using the temporal method and the cost.
    calc_segtree_probs(t_unpred_seg_counts, t_pred_tree, segp->tree_probs,
                       probwt);
    t_pred_cost = cost_segmap(t_unpred_seg_counts, t_pred_tree);

    // Add in the cost of the signaling for each prediction context.
    for (i = 0; i < PREDICTION_PROBS; i++) {
      const int count0 = temporal_predictor_count[i][0];
      const int count1 = temporal_predictor_count[i][1];

      t_nopred_prob[i] = get_binary_prob(count0, count1);
      av1_prob_diff_update_savings_search(
          temporal_predictor_count[i], segp->pred_probs[i], &t_nopred_prob[i],
          DIFF_UPDATE_PROB, probwt);

      // Add in the predictor signaling cost
      t_pred_cost += count0 * av1_cost_zero(t_nopred_prob[i]) +
                     count1 * av1_cost_one(t_nopred_prob[i]);
    }
  }

  // Now choose which coding method to use.
  if (t_pred_cost < no_pred_cost) {
    assert(!cm->error_resilient_mode);
    seg->temporal_update = 1;
  } else {
    seg->temporal_update = 0;
  }
}

void av1_reset_segment_features(AV1_COMMON *cm) {
  struct segmentation *seg = &cm->seg;

  // Set up default state for MB feature flags
  seg->enabled = 0;
  seg->update_map = 0;
  seg->update_data = 0;
  av1_clearall_segfeatures(seg);
}