vp9_thread_common.c 13.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11
/*
 *  Copyright (c) 2014 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 "./vpx_config.h"
12
#include "vpx_mem/vpx_mem.h"
13
#include "vp9/common/vp9_entropymode.h"
14
#include "vp9/common/vp9_thread_common.h"
15
#include "vp9/common/vp9_reconinter.h"
16
#include "vp9/common/vp9_loopfilter.h"
17

18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
#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;
    }
  }

  if (!locked)
    pthread_mutex_lock(mutex);
}
#endif  // CONFIG_MULTITHREAD

static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {
#if CONFIG_MULTITHREAD
  const int nsync = lf_sync->sync_range;

  if (r && !(c & (nsync - 1))) {
41 42
    pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
    mutex_lock(mutex);
43 44

    while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
45
      pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
46
    }
47
    pthread_mutex_unlock(mutex);
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88
  }
#else
  (void)lf_sync;
  (void)r;
  (void)c;
#endif  // CONFIG_MULTITHREAD
}

static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
                              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;
    if (c % nsync)
      sig = 0;
  } 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
}

// Implement row loopfiltering for each thread.
89 90 91 92 93 94
static INLINE
void thread_loop_filter_rows(const YV12_BUFFER_CONFIG *const frame_buffer,
                             VP9_COMMON *const cm,
                             struct macroblockd_plane planes[MAX_MB_PLANE],
                             int start, int stop, int y_only,
                             VP9LfSync *const lf_sync) {
95 96
  const int num_planes = y_only ? 1 : MAX_MB_PLANE;
  const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
97
  int mi_row, mi_col;
98 99 100 101 102 103 104 105 106
  enum lf_path path;
  if (y_only)
    path = LF_PATH_444;
  else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
    path = LF_PATH_420;
  else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
    path = LF_PATH_444;
  else
    path = LF_PATH_SLOW;
107

108 109
  for (mi_row = start; mi_row < stop;
       mi_row += lf_sync->num_workers * MI_BLOCK_SIZE) {
110
    MODE_INFO **const mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
111

112 113 114
    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
      const int r = mi_row >> MI_BLOCK_SIZE_LOG2;
      const int c = mi_col >> MI_BLOCK_SIZE_LOG2;
115
      LOOP_FILTER_MASK lfm;
116 117 118 119
      int plane;

      sync_read(lf_sync, r, c);

120
      vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
121 122

      // TODO(JBB): Make setup_mask work for non 420.
123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139
      vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
                     &lfm);

      vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, &lfm);
      for (plane = 1; plane < num_planes; ++plane) {
        switch (path) {
          case LF_PATH_420:
            vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, &lfm);
            break;
          case LF_PATH_444:
            vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, &lfm);
            break;
          case LF_PATH_SLOW:
            vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
                                          mi_row, mi_col);
            break;
        }
140 141 142 143 144 145 146 147
      }

      sync_write(lf_sync, r, c, sb_cols);
    }
  }
}

// Row-based multi-threaded loopfilter hook
148 149
static int loop_filter_row_worker(VP9LfSync *const lf_sync,
                                  LFWorkerData *const lf_data) {
150 151 152
  thread_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
                          lf_data->start, lf_data->stop, lf_data->y_only,
                          lf_sync);
153 154 155
  return 1;
}

156 157 158 159
static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame,
                                VP9_COMMON *cm,
                                struct macroblockd_plane planes[MAX_MB_PLANE],
                                int start, int stop, int y_only,
160
                                VPxWorker *workers, int nworkers,
161
                                VP9LfSync *lf_sync) {
162
  const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
163 164
  // Number of superblock rows and cols
  const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
165 166
  // Decoder may allocate more threads than number of tiles based on user's
  // input.
167
  const int tile_cols = 1 << cm->log2_tile_cols;
168
  const int num_workers = MIN(nworkers, tile_cols);
169 170
  int i;

171
  if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
172
      num_workers > lf_sync->num_workers) {
173
    vp9_loop_filter_dealloc(lf_sync);
174
    vp9_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
175 176 177
  }

  // Initialize cur_sb_col to -1 for all SB rows.
James Zern's avatar
James Zern committed
178
  memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
179 180

  // Set up loopfilter thread data.
181 182 183 184 185 186 187
  // 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.
188
  for (i = 0; i < num_workers; ++i) {
189
    VPxWorker *const worker = &workers[i];
190
    LFWorkerData *const lf_data = &lf_sync->lfdata[i];
191

192
    worker->hook = (VPxWorkerHook)loop_filter_row_worker;
193 194
    worker->data1 = lf_sync;
    worker->data2 = lf_data;
195 196

    // Loopfilter data
James Zern's avatar
James Zern committed
197
    vp9_loop_filter_data_reset(lf_data, frame, cm, planes);
198 199
    lf_data->start = start + i * MI_BLOCK_SIZE;
    lf_data->stop = stop;
200
    lf_data->y_only = y_only;
201 202

    // Start loopfiltering
203
    if (i == num_workers - 1) {
James Zern's avatar
James Zern committed
204
      winterface->execute(worker);
205
    } else {
James Zern's avatar
James Zern committed
206
      winterface->launch(worker);
207 208 209 210
    }
  }

  // Wait till all rows are finished
211
  for (i = 0; i < num_workers; ++i) {
212
    winterface->sync(&workers[i]);
213 214 215
  }
}

216 217 218 219 220
void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
                              VP9_COMMON *cm,
                              struct macroblockd_plane planes[MAX_MB_PLANE],
                              int frame_filter_level,
                              int y_only, int partial_frame,
221
                              VPxWorker *workers, int num_workers,
222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240
                              VP9LfSync *lf_sync) {
  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;
    mi_rows_to_filter = MAX(cm->mi_rows / 8, 8);
  }
  end_mi_row = start_mi_row + mi_rows_to_filter;
  vp9_loop_filter_frame_init(cm, frame_filter_level);

  loop_filter_rows_mt(frame, cm, planes, start_mi_row, end_mi_row,
                      y_only, workers, num_workers, lf_sync);
}

241
// Set up nsync by width.
242
static INLINE int get_sync_range(int width) {
243 244 245 246 247 248 249 250 251 252 253 254 255
  // 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
256
void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows,
257
                           int width, int num_workers) {
258
  lf_sync->rows = rows;
259
#if CONFIG_MULTITHREAD
260 261
  {
    int i;
262

263 264
    CHECK_MEM_ERROR(cm, lf_sync->mutex_,
                    vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
265 266 267 268
    if (lf_sync->mutex_) {
      for (i = 0; i < rows; ++i) {
        pthread_mutex_init(&lf_sync->mutex_[i], NULL);
      }
269
    }
270

271 272
    CHECK_MEM_ERROR(cm, lf_sync->cond_,
                    vpx_malloc(sizeof(*lf_sync->cond_) * rows));
273 274 275 276
    if (lf_sync->cond_) {
      for (i = 0; i < rows; ++i) {
        pthread_cond_init(&lf_sync->cond_[i], NULL);
      }
277
    }
278 279 280
  }
#endif  // CONFIG_MULTITHREAD

281 282 283 284
  CHECK_MEM_ERROR(cm, lf_sync->lfdata,
                  vpx_malloc(num_workers * sizeof(*lf_sync->lfdata)));
  lf_sync->num_workers = num_workers;

285 286 287 288 289 290 291 292
  CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
                  vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));

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

// Deallocate lf synchronization related mutex and data
293
void vp9_loop_filter_dealloc(VP9LfSync *lf_sync) {
294
  if (lf_sync != NULL) {
295
#if CONFIG_MULTITHREAD
296 297
    int i;

298
    if (lf_sync->mutex_ != NULL) {
299
      for (i = 0; i < lf_sync->rows; ++i) {
300 301 302 303 304
        pthread_mutex_destroy(&lf_sync->mutex_[i]);
      }
      vpx_free(lf_sync->mutex_);
    }
    if (lf_sync->cond_ != NULL) {
305
      for (i = 0; i < lf_sync->rows; ++i) {
306 307 308
        pthread_cond_destroy(&lf_sync->cond_[i]);
      }
      vpx_free(lf_sync->cond_);
309
    }
310
#endif  // CONFIG_MULTITHREAD
311
    vpx_free(lf_sync->lfdata);
312
    vpx_free(lf_sync->cur_sb_col);
313 314
    // 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.
315
    vp9_zero(*lf_sync);
316 317
  }
}
318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399

// Accumulate frame counts.
void vp9_accumulate_frame_counts(VP9_COMMON *cm, FRAME_COUNTS *counts,
                                 int is_dec) {
  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];
                // 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];
  }

  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];

  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];

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

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

  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];
  }

400 401 402
  for (i = 0; i < TX_SIZES; i++)
    cm->counts.tx.tx_totals[i] += counts->tx.tx_totals[i];

403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436
  for (i = 0; i < SKIP_CONTEXTS; i++)
    for (j = 0; j < 2; j++)
      cm->counts.skip[i][j] += counts->skip[i][j];

  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];
    }

    for (i = 0; i < MV_CLASSES; i++)
      comps->classes[i] += comps_t->classes[i];

    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++)
      for (j = 0; j < 2; j++)
        comps->bits[i][j] += comps_t->bits[i][j];

    for (i = 0; i < MV_FP_SIZE; i++)
      comps->fp[i] += comps_t->fp[i];
  }
}