1 /* 2 * Copyright (c) 2012 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 12 #include <limits.h> 13 #include "vpx_mem/vpx_mem.h" 14 #include "vp9/encoder/vp9_segmentation.h" 15 #include "vp9/common/vp9_pred_common.h" 16 #include "vp9/common/vp9_tile_common.h" 17 18 void vp9_enable_segmentation(VP9_PTR ptr) { 19 VP9_COMP *cpi = (VP9_COMP *)ptr; 20 struct segmentation *const seg = &cpi->common.seg; 21 22 seg->enabled = 1; 23 seg->update_map = 1; 24 seg->update_data = 1; 25 } 26 27 void vp9_disable_segmentation(VP9_PTR ptr) { 28 VP9_COMP *cpi = (VP9_COMP *)ptr; 29 struct segmentation *const seg = &cpi->common.seg; 30 seg->enabled = 0; 31 } 32 33 void vp9_set_segmentation_map(VP9_PTR ptr, 34 unsigned char *segmentation_map) { 35 VP9_COMP *cpi = (VP9_COMP *)ptr; 36 struct segmentation *const seg = &cpi->common.seg; 37 38 // Copy in the new segmentation map 39 vpx_memcpy(cpi->segmentation_map, segmentation_map, 40 (cpi->common.mi_rows * cpi->common.mi_cols)); 41 42 // Signal that the map should be updated. 43 seg->update_map = 1; 44 seg->update_data = 1; 45 } 46 47 void vp9_set_segment_data(VP9_PTR ptr, 48 signed char *feature_data, 49 unsigned char abs_delta) { 50 VP9_COMP *cpi = (VP9_COMP *)ptr; 51 struct segmentation *const seg = &cpi->common.seg; 52 53 seg->abs_delta = abs_delta; 54 55 vpx_memcpy(seg->feature_data, feature_data, sizeof(seg->feature_data)); 56 57 // TBD ?? Set the feature mask 58 // vpx_memcpy(cpi->mb.e_mbd.segment_feature_mask, 0, 59 // sizeof(cpi->mb.e_mbd.segment_feature_mask)); 60 } 61 62 // Based on set of segment counts calculate a probability tree 63 static void calc_segtree_probs(int *segcounts, vp9_prob *segment_tree_probs) { 64 // Work out probabilities of each segment 65 const int c01 = segcounts[0] + segcounts[1]; 66 const int c23 = segcounts[2] + segcounts[3]; 67 const int c45 = segcounts[4] + segcounts[5]; 68 const int c67 = segcounts[6] + segcounts[7]; 69 70 segment_tree_probs[0] = get_binary_prob(c01 + c23, c45 + c67); 71 segment_tree_probs[1] = get_binary_prob(c01, c23); 72 segment_tree_probs[2] = get_binary_prob(c45, c67); 73 segment_tree_probs[3] = get_binary_prob(segcounts[0], segcounts[1]); 74 segment_tree_probs[4] = get_binary_prob(segcounts[2], segcounts[3]); 75 segment_tree_probs[5] = get_binary_prob(segcounts[4], segcounts[5]); 76 segment_tree_probs[6] = get_binary_prob(segcounts[6], segcounts[7]); 77 } 78 79 // Based on set of segment counts and probabilities calculate a cost estimate 80 static int cost_segmap(int *segcounts, vp9_prob *probs) { 81 const int c01 = segcounts[0] + segcounts[1]; 82 const int c23 = segcounts[2] + segcounts[3]; 83 const int c45 = segcounts[4] + segcounts[5]; 84 const int c67 = segcounts[6] + segcounts[7]; 85 const int c0123 = c01 + c23; 86 const int c4567 = c45 + c67; 87 88 // Cost the top node of the tree 89 int cost = c0123 * vp9_cost_zero(probs[0]) + 90 c4567 * vp9_cost_one(probs[0]); 91 92 // Cost subsequent levels 93 if (c0123 > 0) { 94 cost += c01 * vp9_cost_zero(probs[1]) + 95 c23 * vp9_cost_one(probs[1]); 96 97 if (c01 > 0) 98 cost += segcounts[0] * vp9_cost_zero(probs[3]) + 99 segcounts[1] * vp9_cost_one(probs[3]); 100 if (c23 > 0) 101 cost += segcounts[2] * vp9_cost_zero(probs[4]) + 102 segcounts[3] * vp9_cost_one(probs[4]); 103 } 104 105 if (c4567 > 0) { 106 cost += c45 * vp9_cost_zero(probs[2]) + 107 c67 * vp9_cost_one(probs[2]); 108 109 if (c45 > 0) 110 cost += segcounts[4] * vp9_cost_zero(probs[5]) + 111 segcounts[5] * vp9_cost_one(probs[5]); 112 if (c67 > 0) 113 cost += segcounts[6] * vp9_cost_zero(probs[6]) + 114 segcounts[7] * vp9_cost_one(probs[6]); 115 } 116 117 return cost; 118 } 119 120 static void count_segs(VP9_COMP *cpi, MODE_INFO **mi_8x8, 121 int *no_pred_segcounts, 122 int (*temporal_predictor_count)[2], 123 int *t_unpred_seg_counts, 124 int bw, int bh, int mi_row, int mi_col) { 125 VP9_COMMON *const cm = &cpi->common; 126 MACROBLOCKD *const xd = &cpi->mb.e_mbd; 127 int segment_id; 128 129 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) 130 return; 131 132 segment_id = mi_8x8[0]->mbmi.segment_id; 133 134 set_mi_row_col(cm, xd, mi_row, bh, mi_col, bw); 135 136 // Count the number of hits on each segment with no prediction 137 no_pred_segcounts[segment_id]++; 138 139 // Temporal prediction not allowed on key frames 140 if (cm->frame_type != KEY_FRAME) { 141 const BLOCK_SIZE bsize = mi_8x8[0]->mbmi.sb_type; 142 // Test to see if the segment id matches the predicted value. 143 const int pred_segment_id = vp9_get_segment_id(cm, cm->last_frame_seg_map, 144 bsize, mi_row, mi_col); 145 const int pred_flag = pred_segment_id == segment_id; 146 const int pred_context = vp9_get_pred_context_seg_id(xd); 147 148 // Store the prediction status for this mb and update counts 149 // as appropriate 150 vp9_set_pred_flag_seg_id(xd, pred_flag); 151 temporal_predictor_count[pred_context][pred_flag]++; 152 153 if (!pred_flag) 154 // Update the "unpredicted" segment count 155 t_unpred_seg_counts[segment_id]++; 156 } 157 } 158 159 static void count_segs_sb(VP9_COMP *cpi, MODE_INFO **mi_8x8, 160 int *no_pred_segcounts, 161 int (*temporal_predictor_count)[2], 162 int *t_unpred_seg_counts, 163 int mi_row, int mi_col, 164 BLOCK_SIZE bsize) { 165 const VP9_COMMON *const cm = &cpi->common; 166 const int mis = cm->mode_info_stride; 167 int bw, bh; 168 const int bs = num_8x8_blocks_wide_lookup[bsize], hbs = bs / 2; 169 170 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) 171 return; 172 173 bw = num_8x8_blocks_wide_lookup[mi_8x8[0]->mbmi.sb_type]; 174 bh = num_8x8_blocks_high_lookup[mi_8x8[0]->mbmi.sb_type]; 175 176 if (bw == bs && bh == bs) { 177 count_segs(cpi, mi_8x8, no_pred_segcounts, temporal_predictor_count, 178 t_unpred_seg_counts, bs, bs, mi_row, mi_col); 179 } else if (bw == bs && bh < bs) { 180 count_segs(cpi, mi_8x8, no_pred_segcounts, temporal_predictor_count, 181 t_unpred_seg_counts, bs, hbs, mi_row, mi_col); 182 count_segs(cpi, mi_8x8 + hbs * mis, no_pred_segcounts, 183 temporal_predictor_count, t_unpred_seg_counts, bs, hbs, 184 mi_row + hbs, mi_col); 185 } else if (bw < bs && bh == bs) { 186 count_segs(cpi, mi_8x8, no_pred_segcounts, temporal_predictor_count, 187 t_unpred_seg_counts, hbs, bs, mi_row, mi_col); 188 count_segs(cpi, mi_8x8 + hbs, no_pred_segcounts, temporal_predictor_count, 189 t_unpred_seg_counts, hbs, bs, mi_row, mi_col + hbs); 190 } else { 191 const BLOCK_SIZE subsize = subsize_lookup[PARTITION_SPLIT][bsize]; 192 int n; 193 194 assert(bw < bs && bh < bs); 195 196 for (n = 0; n < 4; n++) { 197 const int mi_dc = hbs * (n & 1); 198 const int mi_dr = hbs * (n >> 1); 199 200 count_segs_sb(cpi, &mi_8x8[mi_dr * mis + mi_dc], 201 no_pred_segcounts, temporal_predictor_count, 202 t_unpred_seg_counts, 203 mi_row + mi_dr, mi_col + mi_dc, subsize); 204 } 205 } 206 } 207 208 void vp9_choose_segmap_coding_method(VP9_COMP *cpi) { 209 VP9_COMMON *const cm = &cpi->common; 210 struct segmentation *seg = &cm->seg; 211 212 int no_pred_cost; 213 int t_pred_cost = INT_MAX; 214 215 int i, tile_col, mi_row, mi_col; 216 217 int temporal_predictor_count[PREDICTION_PROBS][2] = { { 0 } }; 218 int no_pred_segcounts[MAX_SEGMENTS] = { 0 }; 219 int t_unpred_seg_counts[MAX_SEGMENTS] = { 0 }; 220 221 vp9_prob no_pred_tree[SEG_TREE_PROBS]; 222 vp9_prob t_pred_tree[SEG_TREE_PROBS]; 223 vp9_prob t_nopred_prob[PREDICTION_PROBS]; 224 225 const int mis = cm->mode_info_stride; 226 MODE_INFO **mi_ptr, **mi; 227 228 // Set default state for the segment tree probabilities and the 229 // temporal coding probabilities 230 vpx_memset(seg->tree_probs, 255, sizeof(seg->tree_probs)); 231 vpx_memset(seg->pred_probs, 255, sizeof(seg->pred_probs)); 232 233 // First of all generate stats regarding how well the last segment map 234 // predicts this one 235 for (tile_col = 0; tile_col < 1 << cm->log2_tile_cols; tile_col++) { 236 vp9_get_tile_col_offsets(cm, tile_col); 237 mi_ptr = cm->mi_grid_visible + cm->cur_tile_mi_col_start; 238 for (mi_row = 0; mi_row < cm->mi_rows; 239 mi_row += 8, mi_ptr += 8 * mis) { 240 mi = mi_ptr; 241 for (mi_col = cm->cur_tile_mi_col_start; mi_col < cm->cur_tile_mi_col_end; 242 mi_col += 8, mi += 8) 243 count_segs_sb(cpi, mi, no_pred_segcounts, temporal_predictor_count, 244 t_unpred_seg_counts, mi_row, mi_col, BLOCK_64X64); 245 } 246 } 247 248 // Work out probability tree for coding segments without prediction 249 // and the cost. 250 calc_segtree_probs(no_pred_segcounts, no_pred_tree); 251 no_pred_cost = cost_segmap(no_pred_segcounts, no_pred_tree); 252 253 // Key frames cannot use temporal prediction 254 if (cm->frame_type != KEY_FRAME) { 255 // Work out probability tree for coding those segments not 256 // predicted using the temporal method and the cost. 257 calc_segtree_probs(t_unpred_seg_counts, t_pred_tree); 258 t_pred_cost = cost_segmap(t_unpred_seg_counts, t_pred_tree); 259 260 // Add in the cost of the signalling for each prediction context 261 for (i = 0; i < PREDICTION_PROBS; i++) { 262 const int count0 = temporal_predictor_count[i][0]; 263 const int count1 = temporal_predictor_count[i][1]; 264 265 t_nopred_prob[i] = get_binary_prob(count0, count1); 266 267 // Add in the predictor signaling cost 268 t_pred_cost += count0 * vp9_cost_zero(t_nopred_prob[i]) + 269 count1 * vp9_cost_one(t_nopred_prob[i]); 270 } 271 } 272 273 // Now choose which coding method to use. 274 if (t_pred_cost < no_pred_cost) { 275 seg->temporal_update = 1; 276 vpx_memcpy(seg->tree_probs, t_pred_tree, sizeof(t_pred_tree)); 277 vpx_memcpy(seg->pred_probs, t_nopred_prob, sizeof(t_nopred_prob)); 278 } else { 279 seg->temporal_update = 0; 280 vpx_memcpy(seg->tree_probs, no_pred_tree, sizeof(no_pred_tree)); 281 } 282 } 283
