thread_common.c 21.8 KB
Newer Older
Jingning Han's avatar
Jingning Han committed
1
/*
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
Jingning Han's avatar
Jingning Han committed
3
 *
4 5 6 7 8 9
 * 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.
Jingning Han's avatar
Jingning Han committed
10 11
 */

Adrian Grange's avatar
Adrian Grange committed
12 13 14
#include "./aom_config.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"
Yaowu Xu's avatar
Yaowu Xu committed
15 16 17 18
#include "av1/common/entropymode.h"
#include "av1/common/thread_common.h"
#include "av1/common/reconinter.h"
#include "av1/common/loopfilter.h"
Jingning Han's avatar
Jingning Han committed
19 20 21 22 23 24 25 26 27 28 29 30 31 32

#if CONFIG_MULTITHREAD
static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
  const int kMaxTryLocks = 4000;
  int locked = 0;
  int i;

  for (i = 0; i < kMaxTryLocks; ++i) {
    if (!pthread_mutex_trylock(mutex)) {
      locked = 1;
      break;
    }
  }

clang-format's avatar
clang-format committed
33
  if (!locked) pthread_mutex_lock(mutex);
Jingning Han's avatar
Jingning Han committed
34 35 36
}
#endif  // CONFIG_MULTITHREAD

37
static INLINE void sync_read(AV1LfSync *const lf_sync, int r, int c) {
Jingning Han's avatar
Jingning Han committed
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56
#if CONFIG_MULTITHREAD
  const int nsync = lf_sync->sync_range;

  if (r && !(c & (nsync - 1))) {
    pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
    mutex_lock(mutex);

    while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
      pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
    }
    pthread_mutex_unlock(mutex);
  }
#else
  (void)lf_sync;
  (void)r;
  (void)c;
#endif  // CONFIG_MULTITHREAD
}

57
static INLINE void sync_write(AV1LfSync *const lf_sync, int r, int c,
Jingning Han's avatar
Jingning Han committed
58 59 60 61 62 63 64 65 66
                              const int sb_cols) {
#if CONFIG_MULTITHREAD
  const int nsync = lf_sync->sync_range;
  int cur;
  // Only signal when there are enough filtered SB for next row to run.
  int sig = 1;

  if (c < sb_cols - 1) {
    cur = c;
clang-format's avatar
clang-format committed
67
    if (c % nsync) sig = 0;
Jingning Han's avatar
Jingning Han committed
68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
  } else {
    cur = sb_cols + nsync;
  }

  if (sig) {
    mutex_lock(&lf_sync->mutex_[r]);

    lf_sync->cur_sb_col[r] = cur;

    pthread_cond_signal(&lf_sync->cond_[r]);
    pthread_mutex_unlock(&lf_sync->mutex_[r]);
  }
#else
  (void)lf_sync;
  (void)r;
  (void)c;
  (void)sb_cols;
#endif  // CONFIG_MULTITHREAD
}

88 89
static INLINE enum lf_path get_loop_filter_path(
    int y_only, struct macroblockd_plane planes[MAX_MB_PLANE]) {
Jingning Han's avatar
Jingning Han committed
90
  if (y_only)
91
    return LF_PATH_444;
Jingning Han's avatar
Jingning Han committed
92
  else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
93
    return LF_PATH_420;
Jingning Han's avatar
Jingning Han committed
94
  else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
95
    return LF_PATH_444;
Jingning Han's avatar
Jingning Han committed
96
  else
97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150
    return LF_PATH_SLOW;
}

static INLINE void loop_filter_block_plane_ver(
    AV1_COMMON *cm, struct macroblockd_plane planes[MAX_MB_PLANE], int plane,
    MODE_INFO **mi, int mi_row, int mi_col, enum lf_path path,
    LOOP_FILTER_MASK *lfm) {
  if (plane == 0) {
    av1_filter_block_plane_ss00_ver(cm, &planes[0], mi_row, lfm);
  } else {
    switch (path) {
      case LF_PATH_420:
        av1_filter_block_plane_ss11_ver(cm, &planes[plane], mi_row, lfm);
        break;
      case LF_PATH_444:
        av1_filter_block_plane_ss00_ver(cm, &planes[plane], mi_row, lfm);
        break;
      case LF_PATH_SLOW:
        av1_filter_block_plane_non420_ver(cm, &planes[plane], mi + mi_col,
                                          mi_row, mi_col);
        break;
    }
  }
}

static INLINE void loop_filter_block_plane_hor(
    AV1_COMMON *cm, struct macroblockd_plane planes[MAX_MB_PLANE], int plane,
    int mi_row, enum lf_path path, LOOP_FILTER_MASK *lfm) {
  if (plane == 0) {
    av1_filter_block_plane_ss00_hor(cm, &planes[0], mi_row, lfm);
  } else {
    switch (path) {
      case LF_PATH_420:
        av1_filter_block_plane_ss11_hor(cm, &planes[plane], mi_row, lfm);
        break;
      case LF_PATH_444:
        av1_filter_block_plane_ss00_hor(cm, &planes[plane], mi_row, lfm);
        break;
      case LF_PATH_SLOW:
        av1_filter_block_plane_non420_hor(cm, &planes[plane], mi_row);
        break;
    }
  }
}

// Row-based multi-threaded loopfilter hook
#if CONFIG_PARALLEL_DEBLOCKING
static int loop_filter_ver_row_worker(AV1LfSync *const lf_sync,
                                      LFWorkerData *const lf_data) {
  const int num_planes = lf_data->y_only ? 1 : MAX_MB_PLANE;
  int mi_row, mi_col;
  enum lf_path path = get_loop_filter_path(lf_data->y_only, lf_data->planes);

  for (mi_row = lf_data->start; mi_row < lf_data->stop;
151
       mi_row += lf_sync->num_workers * MAX_MIB_SIZE) {
152 153 154
    MODE_INFO **const mi =
        lf_data->cm->mi_grid_visible + mi_row * lf_data->cm->mi_stride;

155
    for (mi_col = 0; mi_col < lf_data->cm->mi_cols; mi_col += MAX_MIB_SIZE) {
156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175
      LOOP_FILTER_MASK lfm;
      int plane;

      av1_setup_dst_planes(lf_data->planes, lf_data->frame_buffer, mi_row,
                           mi_col);
      av1_setup_mask(lf_data->cm, mi_row, mi_col, mi + mi_col,
                     lf_data->cm->mi_stride, &lfm);

      for (plane = 0; plane < num_planes; ++plane)
        loop_filter_block_plane_ver(lf_data->cm, lf_data->planes, plane, mi,
                                    mi_row, mi_col, path, &lfm);
    }
  }
  return 1;
}

static int loop_filter_hor_row_worker(AV1LfSync *const lf_sync,
                                      LFWorkerData *const lf_data) {
  const int num_planes = lf_data->y_only ? 1 : MAX_MB_PLANE;
  const int sb_cols =
176
      mi_cols_aligned_to_sb(lf_data->cm->mi_cols) >> MAX_MIB_SIZE_LOG2;
177 178
  int mi_row, mi_col;
  enum lf_path path = get_loop_filter_path(lf_data->y_only, lf_data->planes);
Jingning Han's avatar
Jingning Han committed
179

180
  for (mi_row = lf_data->start; mi_row < lf_data->stop;
181
       mi_row += lf_sync->num_workers * MAX_MIB_SIZE) {
182 183
    MODE_INFO **const mi =
        lf_data->cm->mi_grid_visible + mi_row * lf_data->cm->mi_stride;
Jingning Han's avatar
Jingning Han committed
184

185 186 187
    for (mi_col = 0; mi_col < lf_data->cm->mi_cols; mi_col += MAX_MIB_SIZE) {
      const int r = mi_row >> MAX_MIB_SIZE_LOG2;
      const int c = mi_col >> MAX_MIB_SIZE_LOG2;
Jingning Han's avatar
Jingning Han committed
188 189 190
      LOOP_FILTER_MASK lfm;
      int plane;

191 192
      // TODO(wenhao.zhang@intel.com): For better parallelization, reorder
      // the outer loop to column-based and remove the synchronizations here.
Jingning Han's avatar
Jingning Han committed
193 194
      sync_read(lf_sync, r, c);

195 196 197 198 199 200 201 202
      av1_setup_dst_planes(lf_data->planes, lf_data->frame_buffer, mi_row,
                           mi_col);
      av1_setup_mask(lf_data->cm, mi_row, mi_col, mi + mi_col,
                     lf_data->cm->mi_stride, &lfm);

      for (plane = 0; plane < num_planes; ++plane)
        loop_filter_block_plane_hor(lf_data->cm, lf_data->planes, plane, mi_row,
                                    path, &lfm);
Jingning Han's avatar
Jingning Han committed
203 204 205 206

      sync_write(lf_sync, r, c, sb_cols);
    }
  }
207
  return 1;
Jingning Han's avatar
Jingning Han committed
208
}
209
#else   //  CONFIG_PARALLEL_DEBLOCKING
210
static int loop_filter_row_worker(AV1LfSync *const lf_sync,
Jingning Han's avatar
Jingning Han committed
211
                                  LFWorkerData *const lf_data) {
212 213
  const int num_planes = lf_data->y_only ? 1 : MAX_MB_PLANE;
  const int sb_cols =
214
      mi_cols_aligned_to_sb(lf_data->cm->mi_cols) >> MAX_MIB_SIZE_LOG2;
215 216 217 218
  int mi_row, mi_col;
  enum lf_path path = get_loop_filter_path(lf_data->y_only, lf_data->planes);

  for (mi_row = lf_data->start; mi_row < lf_data->stop;
219
       mi_row += lf_sync->num_workers * MAX_MIB_SIZE) {
220 221 222
    MODE_INFO **const mi =
        lf_data->cm->mi_grid_visible + mi_row * lf_data->cm->mi_stride;

223 224 225
    for (mi_col = 0; mi_col < lf_data->cm->mi_cols; mi_col += MAX_MIB_SIZE) {
      const int r = mi_row >> MAX_MIB_SIZE_LOG2;
      const int c = mi_col >> MAX_MIB_SIZE_LOG2;
226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245
      LOOP_FILTER_MASK lfm;
      int plane;

      sync_read(lf_sync, r, c);

      av1_setup_dst_planes(lf_data->planes, lf_data->frame_buffer, mi_row,
                           mi_col);
      av1_setup_mask(lf_data->cm, mi_row, mi_col, mi + mi_col,
                     lf_data->cm->mi_stride, &lfm);

      for (plane = 0; plane < num_planes; ++plane) {
        loop_filter_block_plane_ver(lf_data->cm, lf_data->planes, plane, mi,
                                    mi_row, mi_col, path, &lfm);
        loop_filter_block_plane_hor(lf_data->cm, lf_data->planes, plane, mi_row,
                                    path, &lfm);
      }

      sync_write(lf_sync, r, c, sb_cols);
    }
  }
Jingning Han's avatar
Jingning Han committed
246 247
  return 1;
}
248
#endif  //  CONFIG_PARALLEL_DEBLOCKING
Jingning Han's avatar
Jingning Han committed
249

250
static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
Jingning Han's avatar
Jingning Han committed
251 252
                                struct macroblockd_plane planes[MAX_MB_PLANE],
                                int start, int stop, int y_only,
Adrian Grange's avatar
Adrian Grange committed
253
                                AVxWorker *workers, int nworkers,
254
                                AV1LfSync *lf_sync) {
Adrian Grange's avatar
Adrian Grange committed
255
  const AVxWorkerInterface *const winterface = aom_get_worker_interface();
Jingning Han's avatar
Jingning Han committed
256
  // Number of superblock rows and cols
257
  const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MAX_MIB_SIZE_LOG2;
Jingning Han's avatar
Jingning Han committed
258 259 260
  // Decoder may allocate more threads than number of tiles based on user's
  // input.
  const int tile_cols = 1 << cm->log2_tile_cols;
Adrian Grange's avatar
Adrian Grange committed
261
  const int num_workers = AOMMIN(nworkers, tile_cols);
Jingning Han's avatar
Jingning Han committed
262 263 264 265
  int i;

  if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
      num_workers > lf_sync->num_workers) {
266 267
    av1_loop_filter_dealloc(lf_sync);
    av1_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
Jingning Han's avatar
Jingning Han committed
268 269
  }

270 271 272 273 274 275 276 277
// Set up loopfilter thread data.
// The decoder is capping num_workers because it has been observed that using
// more threads on the loopfilter than there are cores will hurt performance
// on Android. This is because the system will only schedule the tile decode
// workers on cores equal to the number of tile columns. Then if the decoder
// tries to use more threads for the loopfilter, it will hurt performance
// because of contention. If the multithreading code changes in the future
// then the number of workers used by the loopfilter should be revisited.
278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293

#if CONFIG_PARALLEL_DEBLOCKING
  // Initialize cur_sb_col to -1 for all SB rows.
  memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);

  // Filter all the vertical edges in the whole frame
  for (i = 0; i < num_workers; ++i) {
    AVxWorker *const worker = &workers[i];
    LFWorkerData *const lf_data = &lf_sync->lfdata[i];

    worker->hook = (AVxWorkerHook)loop_filter_ver_row_worker;
    worker->data1 = lf_sync;
    worker->data2 = lf_data;

    // Loopfilter data
    av1_loop_filter_data_reset(lf_data, frame, cm, planes);
294
    lf_data->start = start + i * MAX_MIB_SIZE;
295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322
    lf_data->stop = stop;
    lf_data->y_only = y_only;

    // Start loopfiltering
    if (i == num_workers - 1) {
      winterface->execute(worker);
    } else {
      winterface->launch(worker);
    }
  }

  // Wait till all rows are finished
  for (i = 0; i < num_workers; ++i) {
    winterface->sync(&workers[i]);
  }

  memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
  // Filter all the horizontal edges in the whole frame
  for (i = 0; i < num_workers; ++i) {
    AVxWorker *const worker = &workers[i];
    LFWorkerData *const lf_data = &lf_sync->lfdata[i];

    worker->hook = (AVxWorkerHook)loop_filter_hor_row_worker;
    worker->data1 = lf_sync;
    worker->data2 = lf_data;

    // Loopfilter data
    av1_loop_filter_data_reset(lf_data, frame, cm, planes);
323
    lf_data->start = start + i * MAX_MIB_SIZE;
324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342
    lf_data->stop = stop;
    lf_data->y_only = y_only;

    // Start loopfiltering
    if (i == num_workers - 1) {
      winterface->execute(worker);
    } else {
      winterface->launch(worker);
    }
  }

  // Wait till all rows are finished
  for (i = 0; i < num_workers; ++i) {
    winterface->sync(&workers[i]);
  }
#else   // CONFIG_PARALLEL_DEBLOCKING
  // Initialize cur_sb_col to -1 for all SB rows.
  memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);

Jingning Han's avatar
Jingning Han committed
343
  for (i = 0; i < num_workers; ++i) {
Adrian Grange's avatar
Adrian Grange committed
344
    AVxWorker *const worker = &workers[i];
Jingning Han's avatar
Jingning Han committed
345 346
    LFWorkerData *const lf_data = &lf_sync->lfdata[i];

Adrian Grange's avatar
Adrian Grange committed
347
    worker->hook = (AVxWorkerHook)loop_filter_row_worker;
Jingning Han's avatar
Jingning Han committed
348 349 350 351
    worker->data1 = lf_sync;
    worker->data2 = lf_data;

    // Loopfilter data
352
    av1_loop_filter_data_reset(lf_data, frame, cm, planes);
353
    lf_data->start = start + i * MAX_MIB_SIZE;
Jingning Han's avatar
Jingning Han committed
354 355 356 357 358 359 360 361 362 363 364 365 366 367 368
    lf_data->stop = stop;
    lf_data->y_only = y_only;

    // Start loopfiltering
    if (i == num_workers - 1) {
      winterface->execute(worker);
    } else {
      winterface->launch(worker);
    }
  }

  // Wait till all rows are finished
  for (i = 0; i < num_workers; ++i) {
    winterface->sync(&workers[i]);
  }
369
#endif  // CONFIG_PARALLEL_DEBLOCKING
Jingning Han's avatar
Jingning Han committed
370 371
}

372
void av1_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
clang-format's avatar
clang-format committed
373 374 375 376
                              struct macroblockd_plane planes[MAX_MB_PLANE],
                              int frame_filter_level, int y_only,
                              int partial_frame, AVxWorker *workers,
                              int num_workers, AV1LfSync *lf_sync) {
Jingning Han's avatar
Jingning Han committed
377 378 379 380 381 382 383 384 385
  int start_mi_row, end_mi_row, mi_rows_to_filter;

  if (!frame_filter_level) return;

  start_mi_row = 0;
  mi_rows_to_filter = cm->mi_rows;
  if (partial_frame && cm->mi_rows > 8) {
    start_mi_row = cm->mi_rows >> 1;
    start_mi_row &= 0xfffffff8;
Adrian Grange's avatar
Adrian Grange committed
386
    mi_rows_to_filter = AOMMAX(cm->mi_rows / 8, 8);
Jingning Han's avatar
Jingning Han committed
387 388
  }
  end_mi_row = start_mi_row + mi_rows_to_filter;
389
  av1_loop_filter_frame_init(cm, frame_filter_level);
Jingning Han's avatar
Jingning Han committed
390

clang-format's avatar
clang-format committed
391 392
  loop_filter_rows_mt(frame, cm, planes, start_mi_row, end_mi_row, y_only,
                      workers, num_workers, lf_sync);
Jingning Han's avatar
Jingning Han committed
393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409
}

// Set up nsync by width.
static INLINE int get_sync_range(int width) {
  // nsync numbers are picked by testing. For example, for 4k
  // video, using 4 gives best performance.
  if (width < 640)
    return 1;
  else if (width <= 1280)
    return 2;
  else if (width <= 4096)
    return 4;
  else
    return 8;
}

// Allocate memory for lf row synchronization
410
void av1_loop_filter_alloc(AV1LfSync *lf_sync, AV1_COMMON *cm, int rows,
clang-format's avatar
clang-format committed
411
                           int width, int num_workers) {
Jingning Han's avatar
Jingning Han committed
412 413 414 415 416 417
  lf_sync->rows = rows;
#if CONFIG_MULTITHREAD
  {
    int i;

    CHECK_MEM_ERROR(cm, lf_sync->mutex_,
Adrian Grange's avatar
Adrian Grange committed
418
                    aom_malloc(sizeof(*lf_sync->mutex_) * rows));
Jingning Han's avatar
Jingning Han committed
419 420 421 422 423 424 425
    if (lf_sync->mutex_) {
      for (i = 0; i < rows; ++i) {
        pthread_mutex_init(&lf_sync->mutex_[i], NULL);
      }
    }

    CHECK_MEM_ERROR(cm, lf_sync->cond_,
Adrian Grange's avatar
Adrian Grange committed
426
                    aom_malloc(sizeof(*lf_sync->cond_) * rows));
Jingning Han's avatar
Jingning Han committed
427 428 429 430 431 432 433 434 435
    if (lf_sync->cond_) {
      for (i = 0; i < rows; ++i) {
        pthread_cond_init(&lf_sync->cond_[i], NULL);
      }
    }
  }
#endif  // CONFIG_MULTITHREAD

  CHECK_MEM_ERROR(cm, lf_sync->lfdata,
Adrian Grange's avatar
Adrian Grange committed
436
                  aom_malloc(num_workers * sizeof(*lf_sync->lfdata)));
Jingning Han's avatar
Jingning Han committed
437 438 439
  lf_sync->num_workers = num_workers;

  CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
Adrian Grange's avatar
Adrian Grange committed
440
                  aom_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
Jingning Han's avatar
Jingning Han committed
441 442 443 444 445 446

  // Set up nsync.
  lf_sync->sync_range = get_sync_range(width);
}

// Deallocate lf synchronization related mutex and data
447
void av1_loop_filter_dealloc(AV1LfSync *lf_sync) {
Jingning Han's avatar
Jingning Han committed
448 449 450 451 452 453 454 455
  if (lf_sync != NULL) {
#if CONFIG_MULTITHREAD
    int i;

    if (lf_sync->mutex_ != NULL) {
      for (i = 0; i < lf_sync->rows; ++i) {
        pthread_mutex_destroy(&lf_sync->mutex_[i]);
      }
Adrian Grange's avatar
Adrian Grange committed
456
      aom_free(lf_sync->mutex_);
Jingning Han's avatar
Jingning Han committed
457 458 459 460 461
    }
    if (lf_sync->cond_ != NULL) {
      for (i = 0; i < lf_sync->rows; ++i) {
        pthread_cond_destroy(&lf_sync->cond_[i]);
      }
Adrian Grange's avatar
Adrian Grange committed
462
      aom_free(lf_sync->cond_);
Jingning Han's avatar
Jingning Han committed
463 464
    }
#endif  // CONFIG_MULTITHREAD
Adrian Grange's avatar
Adrian Grange committed
465 466
    aom_free(lf_sync->lfdata);
    aom_free(lf_sync->cur_sb_col);
Jingning Han's avatar
Jingning Han committed
467 468
    // clear the structure as the source of this call may be a resize in which
    // case this call will be followed by an _alloc() which may fail.
469
    av1_zero(*lf_sync);
Jingning Han's avatar
Jingning Han committed
470 471 472 473
  }
}

// Accumulate frame counts.
474
void av1_accumulate_frame_counts(AV1_COMMON *cm, FRAME_COUNTS *counts,
clang-format's avatar
clang-format committed
475
                                 int is_dec) {
Jingning Han's avatar
Jingning Han committed
476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510
  int i, j, k, l, m;

  for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
    for (j = 0; j < INTRA_MODES; j++)
      cm->counts.y_mode[i][j] += counts->y_mode[i][j];

  for (i = 0; i < INTRA_MODES; i++)
    for (j = 0; j < INTRA_MODES; j++)
      cm->counts.uv_mode[i][j] += counts->uv_mode[i][j];

  for (i = 0; i < PARTITION_CONTEXTS; i++)
    for (j = 0; j < PARTITION_TYPES; j++)
      cm->counts.partition[i][j] += counts->partition[i][j];

  if (is_dec) {
    int n;
    for (i = 0; i < TX_SIZES; i++)
      for (j = 0; j < PLANE_TYPES; j++)
        for (k = 0; k < REF_TYPES; k++)
          for (l = 0; l < COEF_BANDS; l++)
            for (m = 0; m < COEFF_CONTEXTS; m++) {
              cm->counts.eob_branch[i][j][k][l][m] +=
                  counts->eob_branch[i][j][k][l][m];
              for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
                cm->counts.coef[i][j][k][l][m][n] +=
                    counts->coef[i][j][k][l][m][n];
            }
  } else {
    for (i = 0; i < TX_SIZES; i++)
      for (j = 0; j < PLANE_TYPES; j++)
        for (k = 0; k < REF_TYPES; k++)
          for (l = 0; l < COEF_BANDS; l++)
            for (m = 0; m < COEFF_CONTEXTS; m++)
              cm->counts.eob_branch[i][j][k][l][m] +=
                  counts->eob_branch[i][j][k][l][m];
clang-format's avatar
clang-format committed
511 512 513 514 515
    // In the encoder, cm->counts.coef is only updated at frame
    // level, so not need to accumulate it here.
    // for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
    //   cm->counts.coef[i][j][k][l][m][n] +=
    //       counts->coef[i][j][k][l][m][n];
Jingning Han's avatar
Jingning Han committed
516 517 518 519 520 521
  }

  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
    for (j = 0; j < SWITCHABLE_FILTERS; j++)
      cm->counts.switchable_interp[i][j] += counts->switchable_interp[i][j];

522 523 524 525 526 527 528 529 530 531 532 533
#if CONFIG_REF_MV
  for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i)
    for (j = 0; j < 2; ++j)
      cm->counts.newmv_mode[i][j] += counts->newmv_mode[i][j];

  for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i)
    for (j = 0; j < 2; ++j)
      cm->counts.zeromv_mode[i][j] += counts->zeromv_mode[i][j];

  for (i = 0; i < REFMV_MODE_CONTEXTS; ++i)
    for (j = 0; j < 2; ++j)
      cm->counts.refmv_mode[i][j] += counts->refmv_mode[i][j];
534 535

  for (i = 0; i < DRL_MODE_CONTEXTS; ++i)
clang-format's avatar
clang-format committed
536
    for (j = 0; j < 2; ++j) cm->counts.drl_mode[i][j] += counts->drl_mode[i][j];
537 538
#endif

Jingning Han's avatar
Jingning Han committed
539 540 541 542
  for (i = 0; i < INTER_MODE_CONTEXTS; i++)
    for (j = 0; j < INTER_MODES; j++)
      cm->counts.inter_mode[i][j] += counts->inter_mode[i][j];

543 544 545 546 547 548
#if CONFIG_MOTION_VAR
  for (i = 0; i < BLOCK_SIZES; ++i)
    for (j = 0; j < MOTION_MODES; ++j)
      cm->counts.motion_mode[i][j] += counts->motion_mode[i][j];
#endif  // CONFIG_MOTION_VAR

Jingning Han's avatar
Jingning Han committed
549 550 551 552 553 554 555 556 557
  for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
    for (j = 0; j < 2; j++)
      cm->counts.intra_inter[i][j] += counts->intra_inter[i][j];

  for (i = 0; i < COMP_INTER_CONTEXTS; i++)
    for (j = 0; j < 2; j++)
      cm->counts.comp_inter[i][j] += counts->comp_inter[i][j];

  for (i = 0; i < REF_CONTEXTS; i++)
558
    for (j = 0; j < (SINGLE_REFS - 1); j++)
Jingning Han's avatar
Jingning Han committed
559
      for (k = 0; k < 2; k++)
clang-format's avatar
clang-format committed
560
        cm->counts.single_ref[i][j][k] += counts->single_ref[i][j][k];
Jingning Han's avatar
Jingning Han committed
561

562 563 564 565 566 567 568 569 570 571
#if CONFIG_EXT_REFS
  for (i = 0; i < REF_CONTEXTS; i++)
    for (j = 0; j < (FWD_REFS - 1); j++)
      for (k = 0; k < 2; k++)
        cm->counts.comp_fwdref[i][j][k] += counts->comp_fwdref[i][j][k];
  for (i = 0; i < REF_CONTEXTS; i++)
    for (j = 0; j < (BWD_REFS - 1); j++)
      for (k = 0; k < 2; k++)
        cm->counts.comp_bwdref[i][j][k] += counts->comp_bwdref[i][j][k];
#else
Jingning Han's avatar
Jingning Han committed
572
  for (i = 0; i < REF_CONTEXTS; i++)
clang-format's avatar
clang-format committed
573
    for (j = 0; j < 2; j++) cm->counts.comp_ref[i][j] += counts->comp_ref[i][j];
574
#endif  // CONFIG_EXT_REFS
Jingning Han's avatar
Jingning Han committed
575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590

  for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
    for (j = 0; j < TX_SIZES; j++)
      cm->counts.tx.p32x32[i][j] += counts->tx.p32x32[i][j];

    for (j = 0; j < TX_SIZES - 1; j++)
      cm->counts.tx.p16x16[i][j] += counts->tx.p16x16[i][j];

    for (j = 0; j < TX_SIZES - 2; j++)
      cm->counts.tx.p8x8[i][j] += counts->tx.p8x8[i][j];
  }

  for (i = 0; i < TX_SIZES; i++)
    cm->counts.tx.tx_totals[i] += counts->tx.tx_totals[i];

  for (i = 0; i < SKIP_CONTEXTS; i++)
clang-format's avatar
clang-format committed
591
    for (j = 0; j < 2; j++) cm->counts.skip[i][j] += counts->skip[i][j];
Jingning Han's avatar
Jingning Han committed
592

593 594 595 596 597 598 599 600 601 602 603 604 605 606 607
#if CONFIG_REF_MV
  for (m = 0; m < NMV_CONTEXTS; ++m) {
    for (i = 0; i < MV_JOINTS; i++)
      cm->counts.mv[m].joints[i] += counts->mv[m].joints[i];

    for (k = 0; k < 2; k++) {
      nmv_component_counts *comps = &cm->counts.mv[m].comps[k];
      nmv_component_counts *comps_t = &counts->mv[m].comps[k];

      for (i = 0; i < 2; i++) {
        comps->sign[i] += comps_t->sign[i];
        comps->class0_hp[i] += comps_t->class0_hp[i];
        comps->hp[i] += comps_t->hp[i];
      }

clang-format's avatar
clang-format committed
608
      for (i = 0; i < MV_CLASSES; i++) comps->classes[i] += comps_t->classes[i];
609 610 611 612 613 614 615 616

      for (i = 0; i < CLASS0_SIZE; i++) {
        comps->class0[i] += comps_t->class0[i];
        for (j = 0; j < MV_FP_SIZE; j++)
          comps->class0_fp[i][j] += comps_t->class0_fp[i][j];
      }

      for (i = 0; i < MV_OFFSET_BITS; i++)
clang-format's avatar
clang-format committed
617
        for (j = 0; j < 2; j++) comps->bits[i][j] += comps_t->bits[i][j];
618

clang-format's avatar
clang-format committed
619
      for (i = 0; i < MV_FP_SIZE; i++) comps->fp[i] += comps_t->fp[i];
620 621 622
    }
  }
#else
Jingning Han's avatar
Jingning Han committed
623 624 625 626 627 628 629 630 631 632 633 634 635
  for (i = 0; i < MV_JOINTS; i++)
    cm->counts.mv.joints[i] += counts->mv.joints[i];

  for (k = 0; k < 2; k++) {
    nmv_component_counts *comps = &cm->counts.mv.comps[k];
    nmv_component_counts *comps_t = &counts->mv.comps[k];

    for (i = 0; i < 2; i++) {
      comps->sign[i] += comps_t->sign[i];
      comps->class0_hp[i] += comps_t->class0_hp[i];
      comps->hp[i] += comps_t->hp[i];
    }

clang-format's avatar
clang-format committed
636
    for (i = 0; i < MV_CLASSES; i++) comps->classes[i] += comps_t->classes[i];
Jingning Han's avatar
Jingning Han committed
637 638 639 640 641 642 643 644

    for (i = 0; i < CLASS0_SIZE; i++) {
      comps->class0[i] += comps_t->class0[i];
      for (j = 0; j < MV_FP_SIZE; j++)
        comps->class0_fp[i][j] += comps_t->class0_fp[i][j];
    }

    for (i = 0; i < MV_OFFSET_BITS; i++)
clang-format's avatar
clang-format committed
645
      for (j = 0; j < 2; j++) comps->bits[i][j] += comps_t->bits[i][j];
Jingning Han's avatar
Jingning Han committed
646

clang-format's avatar
clang-format committed
647
    for (i = 0; i < MV_FP_SIZE; i++) comps->fp[i] += comps_t->fp[i];
Jingning Han's avatar
Jingning Han committed
648
  }
649
#endif
650

651 652 653 654 655 656 657 658 659 660
  for (i = 0; i < EXT_TX_SIZES; i++) {
    for (j = 0; j < TX_TYPES; ++j)
      for (k = 0; k < TX_TYPES; k++)
        cm->counts.intra_ext_tx[i][j][k] += counts->intra_ext_tx[i][j][k];
  }
  for (i = 0; i < EXT_TX_SIZES; i++) {
    for (k = 0; k < TX_TYPES; k++)
      cm->counts.inter_ext_tx[i][k] += counts->inter_ext_tx[i][k];
  }

661 662
#if CONFIG_MISC_FIXES
  for (i = 0; i < PREDICTION_PROBS; i++)
clang-format's avatar
clang-format committed
663
    for (j = 0; j < 2; j++) cm->counts.seg.pred[i][j] += counts->seg.pred[i][j];
664 665 666 667 668 669

  for (i = 0; i < MAX_SEGMENTS; i++) {
    cm->counts.seg.tree_total[i] += counts->seg.tree_total[i];
    cm->counts.seg.tree_mispred[i] += counts->seg.tree_mispred[i];
  }
#endif
Jingning Han's avatar
Jingning Han committed
670
}